From 447ae8d4d577d66dda8c6e84ff89b997e988948c Mon Sep 17 00:00:00 2001 From: hootandy321 Date: Sun, 24 Aug 2025 21:30:52 +0800 Subject: [PATCH 1/5] feat: sync local changes --- .DS_Store | Bin 0 -> 8196 bytes ...07\346\241\243\345\257\274\350\247\210.md" | 85 + infini-qwen3-moe/.DS_Store | Bin 0 -> 8196 bytes .../.xmake/linux/x86_64/cache/config | 7 + .../.xmake/linux/x86_64/cache/detect | 294 + .../.xmake/linux/x86_64/cache/history | 69 + .../.xmake/linux/x86_64/cache/toolchain | 22 + .../.xmake/linux/x86_64/project.lock | 0 .../.xmake/linux/x86_64/xmake.conf | 23 + .../.xmake/windows/x64/cache/config | 0 .../.xmake/windows/x64/cache/detect | 0 .../.xmake/windows/x64/cache/history | 0 .../.xmake/windows/x64/cache/option | 0 .../.xmake/windows/x64/cache/package | 0 .../.xmake/windows/x64/cache/project | 0 .../.xmake/windows/x64/cache/toolchain | 0 .../.xmake/windows/x64/project.lock | 0 .../.xmake/windows/x64/xmake.conf | 22 + infini-qwen3-moe/README.md | 8 + infini-qwen3-moe/build/.DS_Store | Bin 0 -> 6148 bytes .../x86_64/debug/libinfinicore_infer.so.d | 24 + .../src/allocator/memory_allocator.cpp.o.d | 20 + .../debug/src/models/jiuge/jiuge.cpp.o.d | 20 + .../src/models/jiuge/jiuge_kv_cache.cpp.o.d | 20 + .../x86_64/debug/src/models/qw/qwen3.cpp.o.d | 20 + .../src/models/qw/qwen3_kv_cache.cpp.o.d | 20 + .../x86_64/debug/src/tensor/strorage.cpp.o.d | 20 + .../x86_64/debug/src/tensor/tensor.cpp.o.d | 20 + .../x86_64/debug/src/tensor/transform.cpp.o.d | 20 + .../src/allocator/memory_allocator.cpp.o | Bin 0 -> 269784 bytes .../x86_64/debug/src/models/jiuge/jiuge.cpp.o | Bin 0 -> 1080 bytes .../src/models/jiuge/jiuge_kv_cache.cpp.o | Bin 0 -> 1088 bytes .../x86_64/debug/src/models/qw/qwen3.cpp.o | Bin 0 -> 1003384 bytes .../debug/src/models/qw/qwen3_kv_cache.cpp.o | Bin 0 -> 144584 bytes .../x86_64/debug/src/tensor/strorage.cpp.o | Bin 0 -> 49936 bytes .../x86_64/debug/src/tensor/tensor.cpp.o | Bin 0 -> 227392 bytes .../x86_64/debug/src/tensor/transform.cpp.o | Bin 0 -> 264688 bytes .../linux/x86_64/debug/libinfinicore_infer.so | Bin 0 -> 1050896 bytes infini-qwen3-moe/debug_output.json | 32782 ++++++++++++++++ infini-qwen3-moe/include/infinicore_infer.h | 8 + .../include/infinicore_infer/models/jiuge.h | 98 + .../include/infinicore_infer/models/qwen3.h | 120 + .../infinicore_infer/models/qwen3_moe.h | 158 + .../output/cpp_input_embeddings.txt | 108 + infini-qwen3-moe/output/cpp_input_ids.txt | 14 + .../output/cpp_layer_0_attn_k_normed.txt | 115 + .../output/cpp_layer_0_attn_k_proj_raw.txt | 115 + .../output/cpp_layer_0_attn_norm_output.txt | 116 + .../output/cpp_layer_0_attn_q_normed.txt | 115 + .../output/cpp_layer_0_attn_q_proj_raw.txt | 115 + .../cpp_layer_0_attn_residual_output.txt | 115 + .../output/cpp_layer_0_attn_v_proj_raw.txt | 115 + .../cpp_layer_0_input_hidden_states.txt | 108 + .../output/cpp_layer_0_layer_output.txt | 115 + .../output/cpp_layer_0_mlp_gate_proj.txt | 115 + .../output/cpp_layer_0_mlp_intermediate.txt | 115 + .../output/cpp_layer_0_mlp_norm_output.txt | 115 + .../output/cpp_layer_0_mlp_up_proj.txt | 115 + infini-qwen3-moe/reference/.DS_Store | Bin 0 -> 6148 bytes .../reference/qwen3_moe/__init__.py | 27 + .../qwen3_moe/configuration_qwen3_moe.py | 240 + .../reference/qwen3_moe/modeling_qwen3_moe.py | 1052 + .../reference/qwen3_moe/modular_qwen3_moe.py | 349 + infini-qwen3-moe/scripts/__init__.py | 0 .../__pycache__/infer_task.cpython-310.pyc | Bin 0 -> 2770 bytes .../libinfinicore_infer.cpython-310.pyc | Bin 0 -> 3917 bytes .../scripts/__pycache__/qwen3.cpython-310.pyc | Bin 0 -> 28544 bytes infini-qwen3-moe/scripts/infer_task.py | 230 + infini-qwen3-moe/scripts/jiuge.py | 625 + infini-qwen3-moe/scripts/kvcache_pool.py | 90 + infini-qwen3-moe/scripts/launch_server.py | 297 + .../scripts/libinfinicore_infer.py | 157 + infini-qwen3-moe/scripts/qwen3-moe_infer.md | 376 + infini-qwen3-moe/scripts/qwen3.py | 1019 + infini-qwen3-moe/scripts/qwen3_moe.py | 856 + infini-qwen3-moe/scripts/test_program.py | 28 + infini-qwen3-moe/scripts/test_server.py | 130 + infini-qwen3-moe/src/.DS_Store | Bin 0 -> 6148 bytes infini-qwen3-moe/src/allocator.hpp | 54 + .../src/allocator/memory_allocator.cpp | 136 + infini-qwen3-moe/src/models/.DS_Store | Bin 0 -> 6148 bytes infini-qwen3-moe/src/models/jiuge/jiuge.cpp | 0 .../src/models/jiuge/jiuge_impl.hpp | 0 .../src/models/jiuge/jiuge_kv_cache.cpp | 0 .../src/models/jiuge/jiuge_weight.hpp | 0 infini-qwen3-moe/src/models/qw/qwen3.cpp | 2393 ++ infini-qwen3-moe/src/models/qw/qwen3_impl.hpp | 272 + .../src/models/qw/qwen3_kv_cache.cpp | 232 + infini-qwen3-moe/src/models/qw/qwen3_moe.cpp | 718 + .../src/models/qw/qwen3_moe_impl.hpp | 272 + .../src/models/qw/qwen3_moe_kv_cache.cpp | 342 + .../src/models/qw/qwen3_moe_weight.hpp | 413 + .../src/models/qw/qwen3_weight.hpp | 759 + infini-qwen3-moe/src/tensor.hpp | 181 + infini-qwen3-moe/src/tensor/strorage.cpp | 53 + infini-qwen3-moe/src/tensor/tensor.cpp | 322 + infini-qwen3-moe/src/tensor/transform.cpp | 158 + infini-qwen3-moe/src/utils.hpp | 122 + .../third_party/nlohmann/json.hpp | 13908 +++++++ infini-qwen3-moe/trace.log | 6746 ++++ infini-qwen3-moe/xmake.lua | 23 + infini-qwen3/.DS_Store | Bin 0 -> 8196 bytes infini-qwen3/.xmake/.DS_Store | Bin 0 -> 6148 bytes infini-qwen3/.xmake/linux/x86_64/cache/config | 7 + infini-qwen3/.xmake/linux/x86_64/cache/detect | 294 + .../.xmake/linux/x86_64/cache/history | 69 + .../.xmake/linux/x86_64/cache/toolchain | 22 + infini-qwen3/.xmake/linux/x86_64/project.lock | 0 infini-qwen3/.xmake/linux/x86_64/xmake.conf | 23 + infini-qwen3/.xmake/windows/x64/cache/config | 0 infini-qwen3/.xmake/windows/x64/cache/detect | 0 infini-qwen3/.xmake/windows/x64/cache/history | 0 infini-qwen3/.xmake/windows/x64/cache/option | 0 infini-qwen3/.xmake/windows/x64/cache/package | 0 infini-qwen3/.xmake/windows/x64/cache/project | 0 .../.xmake/windows/x64/cache/toolchain | 0 infini-qwen3/.xmake/windows/x64/project.lock | 0 infini-qwen3/.xmake/windows/x64/xmake.conf | 22 + infini-qwen3/README.md | 27 + infini-qwen3/build/.DS_Store | Bin 0 -> 6148 bytes .../x86_64/debug/libinfinicore_infer.so.d | 24 + .../src/allocator/memory_allocator.cpp.o.d | 20 + .../debug/src/models/jiuge/jiuge.cpp.o.d | 20 + .../src/models/jiuge/jiuge_kv_cache.cpp.o.d | 20 + .../x86_64/debug/src/models/qw/qwen3.cpp.o.d | 20 + .../src/models/qw/qwen3_kv_cache.cpp.o.d | 20 + .../x86_64/debug/src/tensor/strorage.cpp.o.d | 20 + .../x86_64/debug/src/tensor/tensor.cpp.o.d | 20 + .../x86_64/debug/src/tensor/transform.cpp.o.d | 20 + .../src/allocator/memory_allocator.cpp.o | Bin 0 -> 269784 bytes .../x86_64/debug/src/models/jiuge/jiuge.cpp.o | Bin 0 -> 1080 bytes .../src/models/jiuge/jiuge_kv_cache.cpp.o | Bin 0 -> 1088 bytes .../x86_64/debug/src/models/qw/qwen3.cpp.o | Bin 0 -> 1015248 bytes .../debug/src/models/qw/qwen3_kv_cache.cpp.o | Bin 0 -> 144584 bytes .../x86_64/debug/src/tensor/strorage.cpp.o | Bin 0 -> 49936 bytes .../x86_64/debug/src/tensor/tensor.cpp.o | Bin 0 -> 227392 bytes .../x86_64/debug/src/tensor/transform.cpp.o | Bin 0 -> 264688 bytes .../linux/x86_64/debug/libinfinicore_infer.so | Bin 0 -> 1055920 bytes infini-qwen3/debug_output.json | 32782 ++++++++++++++++ infini-qwen3/include/infinicore_infer.h | 7 + .../include/infinicore_infer/models/jiuge.h | 98 + .../include/infinicore_infer/models/qwen3.h | 120 + infini-qwen3/output/cpp_input_embeddings.txt | 108 + infini-qwen3/output/cpp_input_ids.txt | 14 + .../output/cpp_layer_0_attn_k_normed.txt | 115 + .../output/cpp_layer_0_attn_k_proj_raw.txt | 115 + .../output/cpp_layer_0_attn_norm_output.txt | 116 + .../output/cpp_layer_0_attn_q_normed.txt | 115 + .../output/cpp_layer_0_attn_q_proj_raw.txt | 115 + .../cpp_layer_0_attn_residual_output.txt | 115 + .../output/cpp_layer_0_attn_v_proj_raw.txt | 115 + .../cpp_layer_0_input_hidden_states.txt | 108 + .../output/cpp_layer_0_layer_output.txt | 115 + .../output/cpp_layer_0_mlp_gate_proj.txt | 115 + .../output/cpp_layer_0_mlp_intermediate.txt | 115 + .../output/cpp_layer_0_mlp_norm_output.txt | 115 + .../output/cpp_layer_0_mlp_up_proj.txt | 115 + infini-qwen3/scripts/__init__.py | 0 .../__pycache__/infer_task.cpython-310.pyc | Bin 0 -> 2770 bytes .../libinfinicore_infer.cpython-310.pyc | Bin 0 -> 3917 bytes .../scripts/__pycache__/qwen3.cpython-310.pyc | Bin 0 -> 28544 bytes infini-qwen3/scripts/infer_task.py | 81 + infini-qwen3/scripts/jiuge.py | 625 + infini-qwen3/scripts/kvcache_pool.py | 90 + infini-qwen3/scripts/launch_server.py | 297 + infini-qwen3/scripts/libinfinicore_infer.py | 157 + infini-qwen3/scripts/qwen3.py | 1424 + infini-qwen3/scripts/test_program.py | 28 + infini-qwen3/scripts/test_server.py | 130 + infini-qwen3/src/.DS_Store | Bin 0 -> 6148 bytes infini-qwen3/src/allocator.hpp | 54 + .../src/allocator/memory_allocator.cpp | 136 + infini-qwen3/src/models/.DS_Store | Bin 0 -> 6148 bytes infini-qwen3/src/models/jiuge/jiuge.cpp | 2 + infini-qwen3/src/models/jiuge/jiuge_impl.hpp | 0 .../src/models/jiuge/jiuge_kv_cache.cpp | 0 .../src/models/jiuge/jiuge_weight.hpp | 0 infini-qwen3/src/models/qw/qwen3.cpp | 2203 ++ infini-qwen3/src/models/qw/qwen3_impl.hpp | 272 + infini-qwen3/src/models/qw/qwen3_kv_cache.cpp | 231 + infini-qwen3/src/models/qw/qwen3_weight.hpp | 759 + infini-qwen3/src/tensor.hpp | 181 + infini-qwen3/src/tensor/strorage.cpp | 53 + infini-qwen3/src/tensor/tensor.cpp | 322 + infini-qwen3/src/tensor/transform.cpp | 158 + infini-qwen3/src/utils.hpp | 122 + infini-qwen3/third_party/nlohmann/json.hpp | 13908 +++++++ infini-qwen3/trace.log | 6746 ++++ infini-qwen3/xmake.lua | 23 + infinicore/.DS_Store | Bin 0 -> 6148 bytes infinicore/.xmake/.DS_Store | Bin 0 -> 6148 bytes infinicore/.xmake/linux/x86_64/cache/config | 8 + infinicore/.xmake/linux/x86_64/cache/detect | 340 + infinicore/.xmake/linux/x86_64/cache/history | 6 + infinicore/.xmake/linux/x86_64/cache/option | 44 + infinicore/.xmake/linux/x86_64/cache/package | 14 + infinicore/.xmake/linux/x86_64/cache/project | 1 + .../.xmake/linux/x86_64/cache/references | 5 + .../.xmake/linux/x86_64/cache/repository | 12 + .../.xmake/linux/x86_64/cache/toolchain | 273 + infinicore/.xmake/linux/x86_64/project.lock | 0 infinicore/.xmake/linux/x86_64/xmake.conf | 36 + infinicore/.xmake/windows/x64/cache/config | 10 + infinicore/.xmake/windows/x64/cache/detect | 55 + infinicore/.xmake/windows/x64/cache/history | 6 + infinicore/.xmake/windows/x64/cache/option | 44 + infinicore/.xmake/windows/x64/cache/package | 9 + infinicore/.xmake/windows/x64/cache/project | 1 + .../.xmake/windows/x64/cache/references | 5 + .../.xmake/windows/x64/cache/repository | 12 + infinicore/.xmake/windows/x64/cache/toolchain | 289 + infinicore/.xmake/windows/x64/project.lock | 0 infinicore/.xmake/windows/x64/xmake.conf | 36 + infinicore/DEV.md | 204 + infinicore/LICENSE | 8 + infinicore/README.md | 167 + infinicore/build/.DS_Store | Bin 0 -> 6148 bytes infinicore/build/.build_cache/.DS_Store | Bin 0 -> 10244 bytes .../07/079094d91096cdd9f330f57b6a792f7b | Bin 0 -> 2840 bytes .../0a/0a73e44ba23fd795f3d4e09bb324fb87 | Bin 0 -> 2000 bytes .../0a/0ae45ce2e66eec59c15425663e400eab | Bin 0 -> 536 bytes .../0c/0ce03dc3ab55c9ab50da09511c741b58 | Bin 0 -> 167792 bytes .../0d/0db018060c11e1dde1c84662a77b70a0 | Bin 0 -> 5816 bytes .../16/16813d01d7491e13bc68a1fbc2d8485c | Bin 0 -> 22824 bytes .../19/1930a2362e3e9ce56aa85bf8a34c2b84 | Bin 0 -> 24712 bytes .../1a/1a590afb820efc0028c6ce22d2f45895 | Bin 0 -> 5816 bytes .../1b/1b7287ad34df1a6e007e32a18b9e4e33 | Bin 0 -> 4288 bytes .../1b/1b94fbad3a7729ed46b7ea4d26793b85 | Bin 0 -> 1080 bytes .../23/23bc33f6571fcc76ac0b30481c73592e | Bin 0 -> 18904 bytes .../23/23c1737cf6b3e7dea1b42e83fed588fc | Bin 0 -> 3240 bytes .../24/242bb4e0e02aac1f8f82db155d128974 | Bin 0 -> 21696 bytes .../28/28c56532e9af3e996ae0a1aab0ddb43a | Bin 0 -> 536 bytes .../2a/2a21c4abd4dbe8a5372e53ece00edecb | Bin 0 -> 16816 bytes .../2c/2ce267e5c00137d34b82c828463350c0 | Bin 0 -> 3224 bytes .../2e/2effc204d106ca3ade254203795dbe48 | Bin 0 -> 1280 bytes .../35/35799ef56d08df52008278bd80e12134 | Bin 0 -> 1192 bytes .../35/35799ef56d08df52008278bd80e12134.txt | 44 + .../3a/3a9735854fa4306a2103db5d249eff67 | Bin 0 -> 18008 bytes .../3b/3b6972275901f725ac6af975fd524301 | Bin 0 -> 35904 bytes .../3e/3e1e5f5a7f4018ca799116595dc2a371 | Bin 0 -> 1200 bytes .../3e/3e9968804b582619978d551bf7c2fa71 | Bin 0 -> 1280 bytes .../43/43fccb1f0e3dd562b18da2496db32ad3 | Bin 0 -> 28176 bytes .../48/48715c83102b0bd0122a823a18462ed5 | Bin 0 -> 3112 bytes .../4c/4ca9e7295d0b02bc979602a9798c8636 | Bin 0 -> 2008 bytes .../4f/4fdd279746cbccf1730367960f8571d3 | Bin 0 -> 5672 bytes .../51/510b11b3f3e77f76008cf051c8f20a9e | Bin 0 -> 5200 bytes .../58/5877c878b49eaaf94554ff6f97170d72 | Bin 0 -> 47224 bytes .../59/5967ec2dba3dca8078c4305b8c7c31fa | Bin 0 -> 5672 bytes .../5c/5c547ff9d8aa359f3d9451d41da67e35 | Bin 0 -> 1192 bytes .../5c/5c547ff9d8aa359f3d9451d41da67e35.txt | 44 + .../5e/5ef16d3b59b6ef16a6ee10e14c4abf32 | Bin 0 -> 69928 bytes .../5f/5f12bba0194596c745b7f8e0f7e7a933 | Bin 0 -> 3448 bytes .../61/6164ba502ffcbd8246dc051f0a1ccb4d | Bin 0 -> 1504 bytes .../65/65f33bf7911e944c233b098ba0d0cc6c | Bin 0 -> 4208 bytes .../6a/6ae6d73efbe28109fcdb1141db5f0daf | Bin 0 -> 3232 bytes .../6c/6cab78194d91d5ea32648c1790ee34a0 | Bin 0 -> 2032 bytes .../6c/6cab78194d91d5ea32648c1790ee34a0.txt | 14 + .../6d/6d5a8ac3b00042fce3bf4280f36ea468 | Bin 0 -> 2000 bytes .../77/77375e6c2de02c7da0e7f38d9b1e356f | Bin 0 -> 1344 bytes .../7b/7b0f3e8e44aaa0db4e3a755e069c5090 | Bin 0 -> 2008 bytes .../82/82253da2a9b6dbe64995385bf3430976 | Bin 0 -> 5728 bytes .../82/828e126938dedb7e5e4375af7de6483f | Bin 0 -> 1200 bytes .../89/8997150f6a1a0f99f60254427275cc3a | Bin 0 -> 2032 bytes .../89/8997150f6a1a0f99f60254427275cc3a.txt | 10 + .../8c/8c7dde4af721c0f837f053165ef4e436 | Bin 0 -> 3280 bytes .../91/91165646f8d377c87978019a0be2868b | Bin 0 -> 27776 bytes .../97/9748c47c72b101ed65cc7ceb4f172447 | Bin 0 -> 3816 bytes .../9e/9e189a5ef986c457cef19e935e2d747d | Bin 0 -> 23856 bytes .../9f/9f9df8837828a6b2269d6b824b68f479 | Bin 0 -> 22664 bytes .../9f/9fe4409b33f035700e8fc666a8d25af2 | Bin 0 -> 53304 bytes .../a0/a06cbe05cc4571da34aa734953a6af89 | Bin 0 -> 2008 bytes .../a2/a24cc3806ad790f5f2b97faad2072f1f | Bin 0 -> 13792 bytes .../a5/a5b3e124e86ade460f75a7adaa0b2871 | Bin 0 -> 1200 bytes .../a5/a5b3e124e86ade460f75a7adaa0b2871.txt | 44 + .../ad/ad2e31d31efdbda13c270e2f178d2bc5 | Bin 0 -> 26512 bytes .../af/afa0c9b9f5721f7730071f22f8c9c3a2 | Bin 0 -> 1240 bytes .../af/afa0c9b9f5721f7730071f22f8c9c3a2.txt | 44 + .../b0/b09e272a8ff712039da5f3f3b05f0154 | Bin 0 -> 2008 bytes .../b6/b6029d0ce31f67c9fa0cd7f280f8e550 | Bin 0 -> 3448 bytes .../b8/b854f6eb047932c12a12c4e6184c81cc | Bin 0 -> 26512 bytes .../b9/b9355f93dd12452230454ec21480a787 | Bin 0 -> 1192 bytes .../b9/b938c1086c7d7ff86ec34645aa82b8dc | Bin 0 -> 1192 bytes .../bf/bf9b14bd9d6fe3479195fb29eb2549be | Bin 0 -> 5664 bytes .../c0/c00777ccbe785ec6ac87fe1b6c78d549 | Bin 0 -> 3312 bytes .../c2/c25ff230fd6c08c038e68655b7f2acd2 | Bin 0 -> 1080 bytes .../c2/c27134bd84820557efa438dca1573592 | Bin 0 -> 14104 bytes .../c5/c531785beef19fdc65a1d1ce4cbaf75b | Bin 0 -> 1200 bytes .../c5/c531785beef19fdc65a1d1ce4cbaf75b.txt | 44 + .../c8/c89f5db66fe2c43876a3898c335eccdb | Bin 0 -> 2008 bytes .../c9/c902b460be650828fed61b8fb134426f | Bin 0 -> 1192 bytes .../cf/cfef3dabe2383e3a7ae9d0a0fc659f06 | Bin 0 -> 1240 bytes .../cf/cfef3dabe2383e3a7ae9d0a0fc659f06.txt | 44 + .../d3/d369e61bde40654d96cda75ca8d26421 | Bin 0 -> 1760 bytes .../d4/d4964a0b57d348c3160657520a75ac7d | Bin 0 -> 648 bytes .../d5/d5831114e2166fe232eba81718436b20 | Bin 0 -> 26512 bytes .../d6/d6098f785e367a2d837efdf44d097bdb | Bin 0 -> 2344 bytes .../d7/d715d807a8a3e47239ab84ac303ffe99 | Bin 0 -> 2032 bytes .../d7/d715d807a8a3e47239ab84ac303ffe99.txt | 23 + .../df/dfa56d0f177c8224b99a7e7f2bb46d98 | Bin 0 -> 2008 bytes .../df/dfd463214c3b5645cd204b724feed178 | Bin 0 -> 536 bytes .../e0/e09a324f71a18401d5d306e49772a72e | Bin 0 -> 18064 bytes .../e2/e28f98de25d79bf4b5cb6dc4a9ed0f78 | Bin 0 -> 32280 bytes .../e5/e54c931a6b89d000abfe3009a63e9b70 | Bin 0 -> 608 bytes .../e7/e71e476f674a806a21e061e52c3ed8d8 | Bin 0 -> 3232 bytes .../eb/eb4f29fb20cb1396e66ea71e61b4575b | Bin 0 -> 2032 bytes .../eb/eb4f29fb20cb1396e66ea71e61b4575b.txt | 10 + .../eb/ebae43a3754c4b5dcd2fc487fa865ceb | Bin 0 -> 3232 bytes .../ef/ef0c63f4bc5d819fd4c13bd7e7355359 | Bin 0 -> 36136 bytes .../f4/f4b7bb59dbfda1c1b216a74713ba8848 | Bin 0 -> 5696 bytes .../f4/f4c57700f62e99b9d9e7b6be4d6732bf | Bin 0 -> 536 bytes .../f7/f7978159b8ebb0680adf4a687e7a1b98 | Bin 0 -> 2000 bytes .../f9/f9a10f14dde71fd2f042191ac85ae61f | Bin 0 -> 5672 bytes infinicore/build/.deps/.DS_Store | Bin 0 -> 6148 bytes .../linux/x86_64/release/libinfini-utils.a.d | 12 + .../release/src/utils/custom_types.cc.o.d | 30 + .../x86_64/release/src/utils/rearrange.cc.o.d | 30 + .../release/src/utils/custom_types.cc.obj.d | 58 + .../x64/release/src/utils/rearrange.cc.obj.d | 169 + .../linux/x86_64/release/libinfiniccl.so.d | 21 + .../release/src/infiniccl/infiniccl.cc.o.d | 24 + .../release/src/infiniccl/infiniccl.cc.obj.d | 50 + .../linux/x86_64/release/libinfiniop-cpu.a.d | 29 + .../infiniop/devices/cpu/common_cpu.cc.o.d | 30 + .../infiniop/devices/cpu/cpu_handle.cc.o.d | 30 + .../src/infiniop/ops/add/cpu/add_cpu.cc.o.d | 30 + .../cpu/causal_softmax_cpu.cc.o.d | 30 + .../src/infiniop/ops/clip/cpu/clip_cpu.cc.o.d | 30 + .../src/infiniop/ops/conv/cpu/conv_cpu.cc.o.d | 30 + .../src/infiniop/ops/gemm/cpu/gemm_cpu.cc.o.d | 30 + .../infiniop/ops/linear/cpu/linear_cpu.cc.o.d | 30 + .../cpu/linear_backwards_cpu.cc.o.d | 30 + .../src/infiniop/ops/mul/cpu/mul_cpu.cc.o.d | 30 + .../cpu/random_sample_cpu.cc.o.d | 30 + .../ops/rearrange/cpu/rearrange_cpu.cc.o.d | 30 + .../src/infiniop/ops/relu/cpu/relu_cpu.cc.o.d | 30 + .../ops/rms_norm/cpu/rms_norm_cpu.cc.o.d | 30 + .../src/infiniop/ops/rope/cpu/rope_cpu.cc.o.d | 30 + .../src/infiniop/ops/sub/cpu/sub_cpu.cc.o.d | 30 + .../infiniop/ops/swiglu/cpu/swiglu_cpu.cc.o.d | 30 + .../src/infiniop/reduce/cpu/reduce.cc.o.d | 30 + .../infiniop/devices/cpu/common_cpu.cc.obj.d | 172 + .../infiniop/devices/cpu/cpu_handle.cc.obj.d | 51 + .../ops/rearrange/cpu/rearrange_cpu.cc.obj.d | 179 + .../infiniop/ops/relu/cpu/relu_cpu.cc.obj.d | 185 + .../ops/rms_norm/cpu/rms_norm_cpu.cc.obj.d | 181 + .../infiniop/ops/rope/cpu/rope_cpu.cc.obj.d | 179 + .../src/infiniop/ops/sub/cpu/sub_cpu.cc.obj.d | 185 + .../ops/swiglu/cpu/swiglu_cpu.cc.obj.d | 185 + .../linux/x86_64/release/libinfiniop.so.d | 41 + .../src/infiniop/devices/handle.cc.o.d | 24 + .../src/infiniop/operator_descriptor.cc.o.d | 24 + .../src/infiniop/ops/add/operator.cc.o.d | 24 + .../infiniop/ops/attention/operator.cc.o.d | 24 + .../ops/causal_softmax/operator.cc.o.d | 24 + .../src/infiniop/ops/clip/operator.cc.o.d | 24 + .../src/infiniop/ops/conv/operator.cc.o.d | 24 + .../src/infiniop/ops/gemm/operator.cc.o.d | 24 + .../src/infiniop/ops/linear/operator.cc.o.d | 24 + .../ops/linear_backwards/operator.cc.o.d | 24 + .../src/infiniop/ops/mul/operator.cc.o.d | 24 + .../ops/random_sample/operator.cc.o.d | 24 + .../infiniop/ops/rearrange/operator.cc.o.d | 24 + .../src/infiniop/ops/relu/operator.cc.o.d | 24 + .../src/infiniop/ops/rms_norm/operator.cc.o.d | 24 + .../src/infiniop/ops/rope/operator.cc.o.d | 24 + .../src/infiniop/ops/sub/operator.cc.o.d | 24 + .../src/infiniop/ops/swiglu/operator.cc.o.d | 24 + .../src/infiniop/tensor_descriptor.cc.o.d | 24 + .../src/infiniop/devices/handle.cc.o.d | 28 + .../ops/causal_softmax/operator.cc.o.d | 28 + .../src/infiniop/ops/conv/operator.cc.o.d | 28 + .../src/infiniop/ops/linear/operator.cc.o.d | 28 + .../ops/linear_backwards/operator.cc.o.d | 28 + .../infiniop/ops/rearrange/operator.cc.o.d | 28 + .../src/infiniop/ops/rms_norm/operator.cc.o.d | 28 + .../src/infiniop/ops/rope/operator.cc.o.d | 28 + .../src/infiniop/devices/handle.cc.obj.d | 167 + .../src/infiniop/operator_descriptor.cc.obj.d | 48 + .../src/infiniop/ops/add/operator.cc.obj.d | 182 + .../infiniop/ops/attention/operator.cc.obj.d | 175 + .../ops/causal_softmax/operator.cc.obj.d | 174 + .../src/infiniop/ops/clip/operator.cc.obj.d | 182 + .../src/infiniop/ops/conv/operator.cc.obj.d | 174 + .../src/infiniop/ops/gemm/operator.cc.obj.d | 177 + .../src/infiniop/ops/linear/operator.cc.obj.d | 177 + .../ops/linear_backwards/operator.cc.obj.d | 177 + .../src/infiniop/ops/mul/operator.cc.obj.d | 182 + .../ops/random_sample/operator.cc.obj.d | 174 + .../infiniop/ops/rearrange/operator.cc.obj.d | 171 + .../src/infiniop/ops/relu/operator.cc.obj.d | 182 + .../infiniop/ops/rms_norm/operator.cc.obj.d | 174 + .../src/infiniop/ops/rope/operator.cc.obj.d | 173 + .../src/infiniop/ops/sub/operator.cc.obj.d | 182 + .../src/infiniop/ops/swiglu/operator.cc.obj.d | 182 + .../src/infiniop/tensor_descriptor.cc.obj.d | 176 + .../linux/x86_64/release/libinfinirt-cpu.a.d | 12 + .../src/infinirt/cpu/infinirt_cpu.cc.o.d | 29 + .../src/infinirt/cpu/infinirt_cpu.cc.obj.d | 65 + .../linux/x86_64/release/libinfinirt.so.d | 20 + .../release/src/infinirt/infinirt.cc.o.d | 24 + .../release/src/infinirt/infinirt.cc.obj.d | 172 + .../linux/x86_64/release/infiniutils-test.d | 17 + .../x86_64/release/src/utils-test/main.cc.o.d | 27 + .../src/utils-test/test_rearrange.cc.o.d | 27 + infinicore/build/.objs/.DS_Store | Bin 0 -> 6148 bytes .../release/src/utils/custom_types.cc.o | Bin 0 -> 2000 bytes .../x86_64/release/src/utils/rearrange.cc.o | Bin 0 -> 21696 bytes .../x64/release/src/utils/custom_types.cc.obj | Bin 0 -> 1651 bytes .../x64/release/src/utils/rearrange.cc.obj | Bin 0 -> 186678 bytes .../release/src/infiniccl/infiniccl.cc.o | Bin 0 -> 3448 bytes .../release/src/infiniccl/infiniccl.cc.obj | Bin 0 -> 2976 bytes .../src/infiniop/devices/cpu/common_cpu.cc.o | Bin 0 -> 3816 bytes .../src/infiniop/devices/cpu/cpu_handle.cc.o | Bin 0 -> 1760 bytes .../src/infiniop/ops/add/cpu/add_cpu.cc.o | Bin 0 -> 26512 bytes .../cpu/causal_softmax_cpu.cc.o | Bin 0 -> 18008 bytes .../src/infiniop/ops/clip/cpu/clip_cpu.cc.o | Bin 0 -> 28176 bytes .../src/infiniop/ops/conv/cpu/conv_cpu.cc.o | Bin 0 -> 47224 bytes .../src/infiniop/ops/gemm/cpu/gemm_cpu.cc.o | Bin 0 -> 18064 bytes .../infiniop/ops/linear/cpu/linear_cpu.cc.o | Bin 0 -> 13792 bytes .../cpu/linear_backwards_cpu.cc.o | Bin 0 -> 22824 bytes .../src/infiniop/ops/mul/cpu/mul_cpu.cc.o | Bin 0 -> 26512 bytes .../random_sample/cpu/random_sample_cpu.cc.o | Bin 0 -> 167792 bytes .../ops/rearrange/cpu/rearrange_cpu.cc.o | Bin 0 -> 16816 bytes .../src/infiniop/ops/relu/cpu/relu_cpu.cc.o | Bin 0 -> 24712 bytes .../ops/rms_norm/cpu/rms_norm_cpu.cc.o | Bin 0 -> 23856 bytes .../src/infiniop/ops/rope/cpu/rope_cpu.cc.o | Bin 0 -> 69928 bytes .../src/infiniop/ops/sub/cpu/sub_cpu.cc.o | Bin 0 -> 26512 bytes .../infiniop/ops/swiglu/cpu/swiglu_cpu.cc.o | Bin 0 -> 27776 bytes .../src/infiniop/reduce/cpu/reduce.cc.o | Bin 0 -> 4208 bytes .../infiniop/devices/cpu/common_cpu.cc.obj | Bin 0 -> 38890 bytes .../infiniop/devices/cpu/cpu_handle.cc.obj | Bin 0 -> 1574 bytes .../ops/rearrange/cpu/rearrange_cpu.cc.obj | Bin 0 -> 97023 bytes .../src/infiniop/ops/relu/cpu/relu_cpu.cc.obj | Bin 0 -> 182576 bytes .../ops/rms_norm/cpu/rms_norm_cpu.cc.obj | Bin 0 -> 150686 bytes .../src/infiniop/ops/rope/cpu/rope_cpu.cc.obj | Bin 0 -> 181092 bytes .../src/infiniop/ops/sub/cpu/sub_cpu.cc.obj | Bin 0 -> 184708 bytes .../infiniop/ops/swiglu/cpu/swiglu_cpu.cc.obj | Bin 0 -> 188740 bytes .../release/src/infiniop/devices/handle.cc.o | Bin 0 -> 5200 bytes .../src/infiniop/operator_descriptor.cc.o | Bin 0 -> 1504 bytes .../src/infiniop/ops/add/operator.cc.o | Bin 0 -> 5672 bytes .../src/infiniop/ops/attention/operator.cc.o | Bin 0 -> 53304 bytes .../infiniop/ops/causal_softmax/operator.cc.o | Bin 0 -> 3280 bytes .../src/infiniop/ops/clip/operator.cc.o | Bin 0 -> 5728 bytes .../src/infiniop/ops/conv/operator.cc.o | Bin 0 -> 3232 bytes .../src/infiniop/ops/gemm/operator.cc.o | Bin 0 -> 3232 bytes .../src/infiniop/ops/linear/operator.cc.o | Bin 0 -> 3232 bytes .../ops/linear_backwards/operator.cc.o | Bin 0 -> 3312 bytes .../src/infiniop/ops/mul/operator.cc.o | Bin 0 -> 5672 bytes .../infiniop/ops/random_sample/operator.cc.o | Bin 0 -> 3448 bytes .../src/infiniop/ops/rearrange/operator.cc.o | Bin 0 -> 3112 bytes .../src/infiniop/ops/relu/operator.cc.o | Bin 0 -> 5664 bytes .../src/infiniop/ops/rms_norm/operator.cc.o | Bin 0 -> 3240 bytes .../src/infiniop/ops/rope/operator.cc.o | Bin 0 -> 3224 bytes .../src/infiniop/ops/sub/operator.cc.o | Bin 0 -> 5672 bytes .../src/infiniop/ops/swiglu/operator.cc.o | Bin 0 -> 5696 bytes .../src/infiniop/tensor_descriptor.cc.o | Bin 0 -> 32280 bytes .../release/src/infiniop/devices/handle.cc.o | Bin 0 -> 5816 bytes .../src/infiniop/ops/attention/operator.cc.o | Bin 0 -> 35904 bytes .../infiniop/ops/causal_softmax/operator.cc.o | Bin 0 -> 1240 bytes .../src/infiniop/ops/conv/operator.cc.o | Bin 0 -> 1200 bytes .../src/infiniop/ops/linear/operator.cc.o | Bin 0 -> 1192 bytes .../ops/linear_backwards/operator.cc.o | Bin 0 -> 1280 bytes .../infiniop/ops/random_sample/operator.cc.o | Bin 0 -> 1344 bytes .../src/infiniop/ops/rearrange/operator.cc.o | Bin 0 -> 1080 bytes .../src/infiniop/ops/rms_norm/operator.cc.o | Bin 0 -> 1200 bytes .../src/infiniop/ops/rope/operator.cc.o | Bin 0 -> 1192 bytes .../src/infiniop/tensor_descriptor.cc.o | Bin 0 -> 22656 bytes .../src/infiniop/devices/handle.cc.obj | Bin 0 -> 17387 bytes .../src/infiniop/operator_descriptor.cc.obj | Bin 0 -> 1052 bytes .../src/infiniop/ops/add/operator.cc.obj | Bin 0 -> 62068 bytes .../infiniop/ops/attention/operator.cc.obj | Bin 0 -> 94424 bytes .../ops/causal_softmax/operator.cc.obj | Bin 0 -> 3827 bytes .../src/infiniop/ops/clip/operator.cc.obj | Bin 0 -> 63058 bytes .../src/infiniop/ops/conv/operator.cc.obj | Bin 0 -> 3667 bytes .../src/infiniop/ops/gemm/operator.cc.obj | Bin 0 -> 4078 bytes .../src/infiniop/ops/linear/operator.cc.obj | Bin 0 -> 3697 bytes .../ops/linear_backwards/operator.cc.obj | Bin 0 -> 3875 bytes .../src/infiniop/ops/mul/operator.cc.obj | Bin 0 -> 62068 bytes .../ops/random_sample/operator.cc.obj | Bin 0 -> 4133 bytes .../infiniop/ops/rearrange/operator.cc.obj | Bin 0 -> 3062 bytes .../src/infiniop/ops/relu/operator.cc.obj | Bin 0 -> 61285 bytes .../src/infiniop/ops/rms_norm/operator.cc.obj | Bin 0 -> 3738 bytes .../src/infiniop/ops/rope/operator.cc.obj | Bin 0 -> 3661 bytes .../src/infiniop/ops/sub/operator.cc.obj | Bin 0 -> 62068 bytes .../src/infiniop/ops/swiglu/operator.cc.obj | Bin 0 -> 62147 bytes .../src/infiniop/tensor_descriptor.cc.obj | Bin 0 -> 233385 bytes .../src/infinirt/cpu/infinirt_cpu.cc.o | Bin 0 -> 4288 bytes .../src/infinirt/cpu/infinirt_cpu.cc.obj | Bin 0 -> 7724 bytes .../x86_64/release/src/infinirt/infinirt.cc.o | Bin 0 -> 36136 bytes .../x64/release/src/infinirt/infinirt.cc.obj | Bin 0 -> 35447 bytes .../x86_64/release/src/utils-test/main.cc.o | Bin 0 -> 2344 bytes .../src/utils-test/test_rearrange.cc.o | Bin 0 -> 14104 bytes .../linux/x86_64/release/infiniutils-test | Bin 0 -> 26928 bytes .../linux/x86_64/release/libinfini-utils.a | Bin 0 -> 25974 bytes .../linux/x86_64/release/libinfiniccl.so | Bin 0 -> 14256 bytes .../linux/x86_64/release/libinfiniop-cpu.a | Bin 0 -> 602830 bytes .../build/linux/x86_64/release/libinfiniop.so | Bin 0 -> 301880 bytes .../linux/x86_64/release/libinfinirt-cpu.a | Bin 0 -> 5348 bytes .../build/linux/x86_64/release/libinfinirt.so | Bin 0 -> 40072 bytes infinicore/comprehensive_demo.py | 263 + infinicore/include/.DS_Store | Bin 0 -> 6148 bytes infinicore/include/infiniccl.h | 35 + infinicore/include/infinicore.h | 75 + infinicore/include/infiniop.h | 23 + infinicore/include/infiniop/handle.h | 14 + .../include/infiniop/operator_descriptor.h | 13 + infinicore/include/infiniop/ops/add.h | 26 + infinicore/include/infiniop/ops/attention.h | 34 + .../include/infiniop/ops/causal_softmax.h | 26 + infinicore/include/infiniop/ops/clip.h | 28 + infinicore/include/infiniop/ops/conv.h | 25 + infinicore/include/infiniop/ops/gemm.h | 28 + infinicore/include/infiniop/ops/linear.h | 28 + .../include/infiniop/ops/linear_backwards.h | 32 + infinicore/include/infiniop/ops/mul.h | 26 + .../include/infiniop/ops/random_sample.h | 33 + infinicore/include/infiniop/ops/rearrange.h | 23 + infinicore/include/infiniop/ops/relu.h | 22 + infinicore/include/infiniop/ops/rms_norm.h | 23 + infinicore/include/infiniop/ops/rope.h | 32 + infinicore/include/infiniop/ops/sub.h | 26 + infinicore/include/infiniop/ops/swiglu.h | 26 + .../include/infiniop/tensor_descriptor.h | 14 + infinicore/include/infinirt.h | 56 + infinicore/infinicore_comparison.py | 341 + infinicore/performance_comparison.py | 259 + infinicore/quick_demo.py | 133 + infinicore/scripts/.DS_Store | Bin 0 -> 6148 bytes .../__pycache__/set_env.cpython-312.pyc | Bin 0 -> 3003 bytes .../__pycache__/set_env.cpython-313.pyc | Bin 0 -> 3155 bytes infinicore/scripts/build_ntops.py | 29 + infinicore/scripts/format.py | 204 + infinicore/scripts/install.py | 24 + infinicore/scripts/python_test.py | 46 + infinicore/scripts/set_env.py | 36 + infinicore/src/.DS_Store | Bin 0 -> 8196 bytes .../src/infiniccl-test/infiniccl_test.cpp | 185 + .../src/infiniccl-test/infiniccl_test.hpp | 10 + infinicore/src/infiniccl-test/main.cpp | 78 + infinicore/src/infiniccl/.DS_Store | Bin 0 -> 6148 bytes .../src/infiniccl/ascend/infiniccl_ascend.cc | 99 + .../src/infiniccl/ascend/infiniccl_ascend.h | 12 + .../src/infiniccl/cuda/infiniccl_cuda.cu | 94 + .../src/infiniccl/cuda/infiniccl_cuda.h | 13 + infinicore/src/infiniccl/infiniccl.cc | 75 + infinicore/src/infiniccl/infiniccl_impl.h | 37 + .../src/infiniccl/metax/infiniccl_metax.cc | 95 + .../src/infiniccl/metax/infiniccl_metax.h | 12 + infinicore/src/infiniop-test/.DS_Store | Bin 0 -> 6148 bytes .../infiniop-test/include/file_mapping.hpp | 31 + infinicore/src/infiniop-test/include/gguf.hpp | 248 + infinicore/src/infiniop-test/include/ops.hpp | 67 + .../src/infiniop-test/include/tensor.hpp | 80 + infinicore/src/infiniop-test/include/test.hpp | 132 + .../src/infiniop-test/include/utils.hpp | 52 + .../src/infiniop-test/src/file_mapping.cpp | 65 + infinicore/src/infiniop-test/src/gguf.cpp | 167 + infinicore/src/infiniop-test/src/main.cpp | 141 + infinicore/src/infiniop-test/src/ops/add.cpp | 109 + .../infiniop-test/src/ops/causal_softmax.cpp | 103 + infinicore/src/infiniop-test/src/ops/clip.cpp | 120 + infinicore/src/infiniop-test/src/ops/gemm.cpp | 131 + infinicore/src/infiniop-test/src/ops/mul.cpp | 109 + .../infiniop-test/src/ops/random_sample.cpp | 128 + .../src/infiniop-test/src/ops/rearrange.cpp | 97 + .../src/infiniop-test/src/ops/rms_norm.cpp | 128 + infinicore/src/infiniop-test/src/ops/rope.cpp | 132 + infinicore/src/infiniop-test/src/ops/sub.cpp | 109 + .../src/infiniop-test/src/ops/swiglu.cpp | 104 + infinicore/src/infiniop-test/src/tensor.cpp | 280 + infinicore/src/infiniop-test/src/test.cpp | 233 + infinicore/src/infiniop/.DS_Store | Bin 0 -> 6148 bytes infinicore/src/infiniop/binary/binary.h | 98 + .../src/infiniop/binary/cpu/binary_cpu.h | 57 + infinicore/src/infiniop/devices/.DS_Store | Bin 0 -> 6148 bytes .../infiniop/devices/ascend/CMakeLists.txt | 32 + .../src/infiniop/devices/ascend/Makefile | 10 + .../infiniop/devices/ascend/ascend_handle.cc | 14 + .../infiniop/devices/ascend/ascend_handle.h | 17 + .../devices/ascend/ascend_kernel_common.h | 21 + .../infiniop/devices/ascend/common_ascend.cc | 238 + .../infiniop/devices/ascend/common_ascend.h | 56 + .../src/infiniop/devices/bang/bang_handle.cc | 88 + .../src/infiniop/devices/bang/bang_handle.h | 33 + .../src/infiniop/devices/bang/common_bang.h | 37 + .../src/infiniop/devices/cpu/common_cpu.cc | 54 + .../src/infiniop/devices/cpu/common_cpu.h | 35 + .../src/infiniop/devices/cpu/cpu_handle.cc | 12 + .../src/infiniop/devices/cpu/cpu_handle.h | 17 + infinicore/src/infiniop/devices/handle.cc | 110 + .../infiniop/devices/kunlun/kunlun_handle.cc | 29 + .../infiniop/devices/kunlun/kunlun_handle.h | 47 + .../devices/kunlun/kunlun_kernel_common.h | 64 + .../devices/kunlun/kunlun_kernel_dtype.h | 22 + .../src/infiniop/devices/metax/metax_common.h | 42 + .../infiniop/devices/metax/metax_handle.cc | 86 + .../src/infiniop/devices/metax/metax_handle.h | 25 + .../devices/metax/metax_kernel_common.h | 69 + .../src/infiniop/devices/musa/common_musa.h | 42 + .../src/infiniop/devices/musa/musa_handle.cc | 70 + .../src/infiniop/devices/musa/musa_handle.h | 25 + .../infiniop/devices/nvidia/nvidia_common.cu | 107 + .../infiniop/devices/nvidia/nvidia_common.cuh | 15 + .../infiniop/devices/nvidia/nvidia_handle.cuh | 53 + .../infiniop/devices/nvidia/nvidia_handle.h | 40 + .../devices/nvidia/nvidia_kernel_common.cuh | 78 + infinicore/src/infiniop/devices/pool.h | 50 + infinicore/src/infiniop/elementwise/.DS_Store | Bin 0 -> 6148 bytes .../elementwise/cpu/elementwise_cpu.h | 201 + .../src/infiniop/elementwise/elementwise.h | 206 + .../elementwise/kunlun/elementwise_kunlun.h | 137 + .../kunlun/elementwise_kunlun_api.h | 50 + .../kunlun/elementwise_kunlun_kernel.h | 192 + .../elementwise/metax/elementwise_metax.h | 264 + .../elementwise/metax/elementwise_metax_api.h | 59 + .../elementwise/nvidia/elementwise_nvidia.cuh | 419 + .../nvidia/elementwise_nvidia_api.cuh | 109 + infinicore/src/infiniop/handle.h | 11 + infinicore/src/infiniop/ninetoothed/build.py | 112 + infinicore/src/infiniop/operator.h | 11 + .../src/infiniop/operator_descriptor.cc | 15 + infinicore/src/infiniop/ops/.DS_Store | Bin 0 -> 10244 bytes infinicore/src/infiniop/ops/add/.DS_Store | Bin 0 -> 6148 bytes .../src/infiniop/ops/add/cpu/add_cpu.cc | 54 + infinicore/src/infiniop/ops/add/cpu/add_cpu.h | 19 + .../src/infiniop/ops/add/cuda/kernel.cuh | 23 + .../src/infiniop/ops/add/metax/add_metax.h | 8 + .../src/infiniop/ops/add/metax/add_metax.maca | 62 + .../src/infiniop/ops/add/nvidia/add_nvidia.cu | 61 + .../infiniop/ops/add/nvidia/add_nvidia.cuh | 8 + infinicore/src/infiniop/ops/add/operator.cc | 145 + .../src/infiniop/ops/attention/attention.h | 37 + .../src/infiniop/ops/attention/operator.cc | 291 + .../src/infiniop/ops/causal_softmax/.DS_Store | Bin 0 -> 6148 bytes .../ascend/causal_softmax_ascend.cc | 140 + .../ascend/causal_softmax_ascend.h | 7 + .../ops/causal_softmax/causal_softmax.h | 46 + .../causal_softmax/cpu/causal_softmax_cpu.cc | 77 + .../causal_softmax/cpu/causal_softmax_cpu.h | 7 + .../ops/causal_softmax/cuda/kernel.cuh | 57 + .../src/infiniop/ops/causal_softmax/info.h | 78 + .../metax/causal_softmax_metax.h | 8 + .../metax/causal_softmax_metax.maca | 93 + .../nvidia/causal_softmax_nvidia.cu | 97 + .../nvidia/causal_softmax_nvidia.cuh | 8 + .../infiniop/ops/causal_softmax/operator.cc | 180 + infinicore/src/infiniop/ops/clip/.DS_Store | Bin 0 -> 6148 bytes .../src/infiniop/ops/clip/cpu/clip_cpu.cc | 56 + .../src/infiniop/ops/clip/cpu/clip_cpu.h | 23 + .../src/infiniop/ops/clip/cuda/kernel.cuh | 25 + .../src/infiniop/ops/clip/metax/clip_metax.h | 8 + .../infiniop/ops/clip/metax/clip_metax.maca | 63 + .../infiniop/ops/clip/nvidia/clip_nvidia.cu | 63 + .../infiniop/ops/clip/nvidia/clip_nvidia.cuh | 8 + infinicore/src/infiniop/ops/clip/operator.cc | 145 + infinicore/src/infiniop/ops/conv/.DS_Store | Bin 0 -> 6148 bytes infinicore/src/infiniop/ops/conv/conv.h | 56 + .../src/infiniop/ops/conv/cpu/conv_cpu.cc | 363 + .../src/infiniop/ops/conv/cpu/conv_cpu.h | 8 + infinicore/src/infiniop/ops/conv/info.h | 257 + .../infiniop/ops/conv/nvidia/conv_nvidia.cu | 456 + .../infiniop/ops/conv/nvidia/conv_nvidia.cuh | 8 + infinicore/src/infiniop/ops/conv/operator.cc | 138 + infinicore/src/infiniop/ops/gemm/.DS_Store | Bin 0 -> 6148 bytes .../infiniop/ops/gemm/ascend/gemm_ascend.cc | 150 + .../infiniop/ops/gemm/ascend/gemm_ascend.h | 8 + .../src/infiniop/ops/gemm/bang/gemm_bang.cc | 158 + .../src/infiniop/ops/gemm/bang/gemm_bang.h | 8 + .../src/infiniop/ops/gemm/cpu/gemm_cpu.cc | 100 + .../src/infiniop/ops/gemm/cpu/gemm_cpu.h | 8 + infinicore/src/infiniop/ops/gemm/gemm.h | 93 + infinicore/src/infiniop/ops/gemm/info.h | 132 + .../infiniop/ops/gemm/kunlun/gemm_kunlun.cc | 111 + .../infiniop/ops/gemm/kunlun/gemm_kunlun.h | 8 + .../src/infiniop/ops/gemm/metax/gemm_metax.cc | 107 + .../src/infiniop/ops/gemm/metax/gemm_metax.h | 8 + .../src/infiniop/ops/gemm/musa/gemm_musa.h | 8 + .../src/infiniop/ops/gemm/musa/gemm_musa.mu | 121 + .../infiniop/ops/gemm/nvidia/gemm_nvidia.cu | 124 + .../infiniop/ops/gemm/nvidia/gemm_nvidia.cuh | 8 + infinicore/src/infiniop/ops/gemm/operator.cc | 214 + infinicore/src/infiniop/ops/linear/.DS_Store | Bin 0 -> 6148 bytes .../src/infiniop/ops/linear/cpu/linear_cpu.cc | 126 + .../src/infiniop/ops/linear/cpu/linear_cpu.h | 8 + infinicore/src/infiniop/ops/linear/info.h | 102 + infinicore/src/infiniop/ops/linear/linear.h | 64 + .../ops/linear/nvidia/linear_nvidia.cu | 165 + .../ops/linear/nvidia/linear_nvidia.cuh | 8 + .../src/infiniop/ops/linear/operator.cc | 134 + .../infiniop/ops/linear_backwards/.DS_Store | Bin 0 -> 6148 bytes .../cpu/linear_backwards_cpu.cc | 207 + .../cpu/linear_backwards_cpu.h | 8 + .../src/infiniop/ops/linear_backwards/info.h | 149 + .../ops/linear_backwards/linear_backwards.h | 66 + .../nvidia/linear_backwards_nvidia.cu | 287 + .../nvidia/linear_backwards_nvidia.cuh | 8 + .../infiniop/ops/linear_backwards/operator.cc | 141 + infinicore/src/infiniop/ops/mul/.DS_Store | Bin 0 -> 6148 bytes .../src/infiniop/ops/mul/cpu/mul_cpu.cc | 53 + infinicore/src/infiniop/ops/mul/cpu/mul_cpu.h | 19 + .../src/infiniop/ops/mul/cuda/kernel.cuh | 23 + .../src/infiniop/ops/mul/metax/mul_metax.h | 8 + .../src/infiniop/ops/mul/metax/mul_metax.maca | 61 + .../src/infiniop/ops/mul/nvidia/mul_nvidia.cu | 61 + .../infiniop/ops/mul/nvidia/mul_nvidia.cuh | 8 + infinicore/src/infiniop/ops/mul/operator.cc | 146 + .../src/infiniop/ops/random_sample/.DS_Store | Bin 0 -> 6148 bytes .../ascend/random_sample_aclnn.h | 8 + .../ascend/random_sample_kernel.cpp | 268 + .../ascend/randomsample_aclnn.cc | 110 + .../random_sample/cpu/random_sample_cpu.cc | 139 + .../ops/random_sample/cpu/random_sample_cpu.h | 8 + .../src/infiniop/ops/random_sample/info.h | 32 + .../metax/random_sample_kernel.h | 259 + .../random_sample/metax/random_sample_metax.h | 8 + .../metax/random_sample_metax.maca | 103 + .../nvidia/random_sample_kernel.cuh | 259 + .../nvidia/random_sample_nvidia.cu | 102 + .../nvidia/random_sample_nvidia.cuh | 8 + .../infiniop/ops/random_sample/operator.cc | 171 + .../ops/random_sample/random_sample.h | 147 + .../src/infiniop/ops/rearrange/.DS_Store | Bin 0 -> 6148 bytes .../ops/rearrange/ascend/rearrange_ascend.cc | 94 + .../ops/rearrange/ascend/rearrange_ascend.h | 8 + .../ops/rearrange/cpu/rearrange_cpu.cc | 48 + .../ops/rearrange/cpu/rearrange_cpu.h | 8 + .../ops/rearrange/metax/rearrange_kernel.h | 331 + .../ops/rearrange/metax/rearrange_metax.h | 8 + .../ops/rearrange/metax/rearrange_metax.maca | 483 + .../ops/rearrange/nvidia/rearrange_kernel.cuh | 321 + .../ops/rearrange/nvidia/rearrange_nvidia.cu | 485 + .../ops/rearrange/nvidia/rearrange_nvidia.cuh | 8 + .../src/infiniop/ops/rearrange/operator.cc | 125 + .../src/infiniop/ops/rearrange/rearrange.h | 40 + infinicore/src/infiniop/ops/relu/.DS_Store | Bin 0 -> 6148 bytes .../src/infiniop/ops/relu/cpu/relu_cpu.cc | 52 + .../src/infiniop/ops/relu/cpu/relu_cpu.h | 22 + .../src/infiniop/ops/relu/metax/relu_metax.h | 12 + .../infiniop/ops/relu/metax/relu_metax.maca | 80 + .../infiniop/ops/relu/ninetoothed/build.py | 30 + .../infiniop/ops/relu/nvidia/relu_nvidia.cu | 80 + .../infiniop/ops/relu/nvidia/relu_nvidia.cuh | 12 + infinicore/src/infiniop/ops/relu/operator.cc | 170 + .../src/infiniop/ops/rms_norm/.DS_Store | Bin 0 -> 6148 bytes .../ops/rms_norm/ascend/rms_norm_aclnn.cc | 107 + .../ops/rms_norm/ascend/rms_norm_aclnn.h | 8 + .../infiniop/ops/rms_norm/cpu/rms_norm_cpu.cc | 105 + .../infiniop/ops/rms_norm/cpu/rms_norm_cpu.h | 7 + .../src/infiniop/ops/rms_norm/cuda/kernel.cuh | 34 + infinicore/src/infiniop/ops/rms_norm/info.h | 80 + .../ops/rms_norm/kunlun/rms_norm_kernel.xpu | 125 + .../ops/rms_norm/kunlun/rms_norm_kunlun.cc | 79 + .../ops/rms_norm/kunlun/rms_norm_kunlun.h | 8 + .../ops/rms_norm/metax/rms_norm_metax.cuh | 8 + .../ops/rms_norm/metax/rms_norm_metax.maca | 119 + .../ops/rms_norm/musa/rms_norm_musa.cuh | 8 + .../ops/rms_norm/musa/rms_norm_musa.mu | 97 + .../ops/rms_norm/nvidia/rms_norm_nvidia.cu | 123 + .../ops/rms_norm/nvidia/rms_norm_nvidia.cuh | 8 + .../src/infiniop/ops/rms_norm/operator.cc | 198 + .../src/infiniop/ops/rms_norm/rms_norm.h | 49 + infinicore/src/infiniop/ops/rope/.DS_Store | Bin 0 -> 6148 bytes .../infiniop/ops/rope/ascend/rope_ascend.cc | 50 + .../infiniop/ops/rope/ascend/rope_ascend.h | 25 + .../ops/rope/ascend/rope_ascend_kernel.cpp | 280 + .../src/infiniop/ops/rope/cpu/rope_cpu.cc | 128 + .../src/infiniop/ops/rope/cpu/rope_cpu.h | 8 + .../src/infiniop/ops/rope/cuda/kernel.cuh | 51 + .../src/infiniop/ops/rope/metax/rope_metax.h | 8 + .../infiniop/ops/rope/metax/rope_metax.maca | 144 + .../infiniop/ops/rope/nvidia/rope_nvidia.cu | 144 + .../infiniop/ops/rope/nvidia/rope_nvidia.cuh | 8 + infinicore/src/infiniop/ops/rope/operator.cc | 221 + infinicore/src/infiniop/ops/rope/rope.h | 125 + infinicore/src/infiniop/ops/sub/.DS_Store | Bin 0 -> 6148 bytes .../src/infiniop/ops/sub/cpu/sub_cpu.cc | 54 + infinicore/src/infiniop/ops/sub/cpu/sub_cpu.h | 19 + .../src/infiniop/ops/sub/cuda/kernel.cuh | 23 + .../src/infiniop/ops/sub/metax/sub_metax.h | 8 + .../src/infiniop/ops/sub/metax/sub_metax.maca | 61 + .../src/infiniop/ops/sub/nvidia/sub_nvidia.cu | 61 + .../infiniop/ops/sub/nvidia/sub_nvidia.cuh | 8 + infinicore/src/infiniop/ops/sub/operator.cc | 145 + infinicore/src/infiniop/ops/swiglu/.DS_Store | Bin 0 -> 6148 bytes .../ops/swiglu/ascend/swiglu_ascend.cc | 47 + .../ops/swiglu/ascend/swiglu_ascend.h | 79 + .../swiglu/ascend/swiglu_ascend_kernel.cpp | 180 + .../src/infiniop/ops/swiglu/cpu/swiglu_cpu.cc | 54 + .../src/infiniop/ops/swiglu/cpu/swiglu_cpu.h | 25 + .../src/infiniop/ops/swiglu/cuda/kernel.cuh | 63 + .../ops/swiglu/kunlun/swiglu_kunlun.cc | 63 + .../ops/swiglu/kunlun/swiglu_kunlun.h | 8 + .../swiglu/kunlun/swiglu_kunlun_internal.xpu | 33 + .../infiniop/ops/swiglu/metax/swiglu_metax.h | 8 + .../ops/swiglu/metax/swiglu_metax.maca | 61 + .../ops/swiglu/nvidia/swiglu_nvidia.cu | 61 + .../ops/swiglu/nvidia/swiglu_nvidia.cuh | 8 + .../src/infiniop/ops/swiglu/operator.cc | 229 + infinicore/src/infiniop/reduce/.DS_Store | Bin 0 -> 6148 bytes infinicore/src/infiniop/reduce/cpu/reduce.cc | 59 + infinicore/src/infiniop/reduce/cpu/reduce.h | 73 + .../src/infiniop/reduce/cuda/reduce.cuh | 66 + .../infiniop/reduce/kunlun/reduce_kunlun.h | 40 + infinicore/src/infiniop/tensor.h | 58 + infinicore/src/infiniop/tensor_descriptor.cc | 240 + infinicore/src/infinirt/.DS_Store | Bin 0 -> 6148 bytes .../src/infinirt/ascend/infinirt_ascend.cc | 146 + .../src/infinirt/ascend/infinirt_ascend.h | 14 + infinicore/src/infinirt/bang/infinirt_bang.cc | 136 + infinicore/src/infinirt/bang/infinirt_bang.h | 13 + infinicore/src/infinirt/cpu/infinirt_cpu.cc | 91 + infinicore/src/infinirt/cpu/infinirt_cpu.h | 13 + infinicore/src/infinirt/cuda/infinirt_cuda.cu | 137 + .../src/infinirt/cuda/infinirt_cuda.cuh | 13 + infinicore/src/infinirt/infinirt.cc | 168 + infinicore/src/infinirt/infinirt_impl.h | 36 + .../src/infinirt/kunlun/infinirt_kunlun.cc | 126 + .../src/infinirt/kunlun/infinirt_kunlun.h | 13 + .../src/infinirt/metax/infinirt_metax.cc | 127 + .../src/infinirt/metax/infinirt_metax.h | 13 + infinicore/src/infinirt/musa/infinirt_musa.cc | 132 + infinicore/src/infinirt/musa/infinirt_musa.h | 13 + infinicore/src/utils-test/main.cc | 8 + infinicore/src/utils-test/test_rearrange.cc | 72 + infinicore/src/utils-test/utils_test.h | 7 + infinicore/src/utils.h | 111 + infinicore/src/utils/check.h | 62 + infinicore/src/utils/custom_types.cc | 85 + infinicore/src/utils/custom_types.h | 51 + infinicore/src/utils/rearrange.cc | 213 + infinicore/src/utils/rearrange.h | 47 + infinicore/src/utils/result.hpp | 59 + infinicore/test/.DS_Store | Bin 0 -> 6148 bytes infinicore/test/infiniop-test/README.md | 75 + .../infiniop-test/test_generate/__init__.py | 1 + .../test_generate/infiniop_test.py | 83 + .../test_generate/testcases/__init__.py | 0 .../test_generate/testcases/add.py | 120 + .../test_generate/testcases/causal_softmax.py | 107 + .../test_generate/testcases/clip.py | 252 + .../test_generate/testcases/gemm.py | 196 + .../test_generate/testcases/mul.py | 124 + .../test_generate/testcases/random_sample.py | 240 + .../test_generate/testcases/rearrange.py | 146 + .../test_generate/testcases/rms_norm.py | 120 + .../test_generate/testcases/rope.py | 147 + .../test_generate/testcases/sub.py | 134 + .../test_generate/testcases/swiglu.py | 121 + infinicore/test/infiniop/.DS_Store | Bin 0 -> 6148 bytes infinicore/test/infiniop/__init__.py | 1 + infinicore/test/infiniop/add.py | 182 + infinicore/test/infiniop/attention.py | 275 + infinicore/test/infiniop/causal_softmax.py | 165 + infinicore/test/infiniop/clip.py | 198 + infinicore/test/infiniop/conv.py | 269 + infinicore/test/infiniop/gemm.py | 183 + .../test/infiniop/libinfiniop/__init__.py | 14 + .../__pycache__/__init__.cpython-311.pyc | Bin 0 -> 855 bytes .../__pycache__/datatypes.cpython-311.pyc | Bin 0 -> 1599 bytes .../__pycache__/devices.cpython-311.pyc | Bin 0 -> 1198 bytes .../__pycache__/liboperators.cpython-311.pyc | Bin 0 -> 4412 bytes .../__pycache__/op_register.cpython-311.pyc | Bin 0 -> 16066 bytes .../__pycache__/structs.cpython-311.pyc | Bin 0 -> 1198 bytes .../test/infiniop/libinfiniop/datatypes.py | 45 + .../test/infiniop/libinfiniop/devices.py | 36 + .../test/infiniop/libinfiniop/liboperators.py | 80 + .../test/infiniop/libinfiniop/op_register.py | 567 + .../test/infiniop/libinfiniop/structs.py | 22 + infinicore/test/infiniop/libinfiniop/utils.py | 615 + infinicore/test/infiniop/linear.py | 179 + infinicore/test/infiniop/linear_backwards.py | 238 + infinicore/test/infiniop/mul.py | 181 + infinicore/test/infiniop/random_sample.py | 191 + infinicore/test/infiniop/rearrange.py | 159 + infinicore/test/infiniop/relu.py | 162 + infinicore/test/infiniop/rms_norm.py | 162 + infinicore/test/infiniop/rope.py | 227 + infinicore/test/infiniop/sub.py | 180 + infinicore/test/infiniop/swiglu.py | 182 + infinicore/xmake.lua | 320 + infinicore/xmake/ascend.lua | 80 + infinicore/xmake/bang.lua | 62 + infinicore/xmake/cpu.lua | 50 + infinicore/xmake/cpu.lua.backup | 44 + infinicore/xmake/cpu.lua.bak | 44 + infinicore/xmake/iluvatar.lua | 106 + infinicore/xmake/kunlun.lua | 86 + infinicore/xmake/metax.lua | 74 + infinicore/xmake/musa.lua | 57 + infinicore/xmake/nvidia.lua | 114 + infinicore/xmake/test.lua | 53 + ...63\345\217\260\351\200\202\351\205\215.md" | 888 + ...47\350\203\275\345\210\206\346\236\220.md" | 162 + 893 files changed, 173980 insertions(+) create mode 100644 .DS_Store create mode 100644 "README-\346\212\200\346\234\257\346\226\207\346\241\243\345\257\274\350\247\210.md" create mode 100755 infini-qwen3-moe/.DS_Store create mode 100755 infini-qwen3-moe/.xmake/linux/x86_64/cache/config create mode 100755 infini-qwen3-moe/.xmake/linux/x86_64/cache/detect create mode 100755 infini-qwen3-moe/.xmake/linux/x86_64/cache/history create mode 100755 infini-qwen3-moe/.xmake/linux/x86_64/cache/toolchain create mode 100755 infini-qwen3-moe/.xmake/linux/x86_64/project.lock create mode 100755 infini-qwen3-moe/.xmake/linux/x86_64/xmake.conf create mode 100755 infini-qwen3-moe/.xmake/windows/x64/cache/config create mode 100755 infini-qwen3-moe/.xmake/windows/x64/cache/detect create mode 100755 infini-qwen3-moe/.xmake/windows/x64/cache/history create mode 100755 infini-qwen3-moe/.xmake/windows/x64/cache/option create mode 100755 infini-qwen3-moe/.xmake/windows/x64/cache/package create mode 100755 infini-qwen3-moe/.xmake/windows/x64/cache/project create mode 100755 infini-qwen3-moe/.xmake/windows/x64/cache/toolchain create mode 100755 infini-qwen3-moe/.xmake/windows/x64/project.lock create mode 100755 infini-qwen3-moe/.xmake/windows/x64/xmake.conf create mode 100755 infini-qwen3-moe/README.md create mode 100644 infini-qwen3-moe/build/.DS_Store create mode 100755 infini-qwen3-moe/build/.deps/infinicore_infer/linux/x86_64/debug/libinfinicore_infer.so.d create mode 100755 infini-qwen3-moe/build/.deps/infinicore_infer/linux/x86_64/debug/src/allocator/memory_allocator.cpp.o.d create mode 100755 infini-qwen3-moe/build/.deps/infinicore_infer/linux/x86_64/debug/src/models/jiuge/jiuge.cpp.o.d create mode 100755 infini-qwen3-moe/build/.deps/infinicore_infer/linux/x86_64/debug/src/models/jiuge/jiuge_kv_cache.cpp.o.d create mode 100755 infini-qwen3-moe/build/.deps/infinicore_infer/linux/x86_64/debug/src/models/qw/qwen3.cpp.o.d create mode 100755 infini-qwen3-moe/build/.deps/infinicore_infer/linux/x86_64/debug/src/models/qw/qwen3_kv_cache.cpp.o.d create mode 100755 infini-qwen3-moe/build/.deps/infinicore_infer/linux/x86_64/debug/src/tensor/strorage.cpp.o.d create mode 100755 infini-qwen3-moe/build/.deps/infinicore_infer/linux/x86_64/debug/src/tensor/tensor.cpp.o.d create mode 100755 infini-qwen3-moe/build/.deps/infinicore_infer/linux/x86_64/debug/src/tensor/transform.cpp.o.d create mode 100755 infini-qwen3-moe/build/.objs/infinicore_infer/linux/x86_64/debug/src/allocator/memory_allocator.cpp.o create mode 100755 infini-qwen3-moe/build/.objs/infinicore_infer/linux/x86_64/debug/src/models/jiuge/jiuge.cpp.o create mode 100755 infini-qwen3-moe/build/.objs/infinicore_infer/linux/x86_64/debug/src/models/jiuge/jiuge_kv_cache.cpp.o create mode 100755 infini-qwen3-moe/build/.objs/infinicore_infer/linux/x86_64/debug/src/models/qw/qwen3.cpp.o create mode 100755 infini-qwen3-moe/build/.objs/infinicore_infer/linux/x86_64/debug/src/models/qw/qwen3_kv_cache.cpp.o create mode 100755 infini-qwen3-moe/build/.objs/infinicore_infer/linux/x86_64/debug/src/tensor/strorage.cpp.o create mode 100755 infini-qwen3-moe/build/.objs/infinicore_infer/linux/x86_64/debug/src/tensor/tensor.cpp.o create mode 100755 infini-qwen3-moe/build/.objs/infinicore_infer/linux/x86_64/debug/src/tensor/transform.cpp.o create mode 100755 infini-qwen3-moe/build/linux/x86_64/debug/libinfinicore_infer.so create mode 100755 infini-qwen3-moe/debug_output.json create mode 100755 infini-qwen3-moe/include/infinicore_infer.h create mode 100755 infini-qwen3-moe/include/infinicore_infer/models/jiuge.h create mode 100755 infini-qwen3-moe/include/infinicore_infer/models/qwen3.h create mode 100755 infini-qwen3-moe/include/infinicore_infer/models/qwen3_moe.h create mode 100755 infini-qwen3-moe/output/cpp_input_embeddings.txt create mode 100755 infini-qwen3-moe/output/cpp_input_ids.txt create mode 100755 infini-qwen3-moe/output/cpp_layer_0_attn_k_normed.txt create mode 100755 infini-qwen3-moe/output/cpp_layer_0_attn_k_proj_raw.txt create mode 100755 infini-qwen3-moe/output/cpp_layer_0_attn_norm_output.txt create mode 100755 infini-qwen3-moe/output/cpp_layer_0_attn_q_normed.txt create mode 100755 infini-qwen3-moe/output/cpp_layer_0_attn_q_proj_raw.txt create mode 100755 infini-qwen3-moe/output/cpp_layer_0_attn_residual_output.txt create mode 100755 infini-qwen3-moe/output/cpp_layer_0_attn_v_proj_raw.txt create mode 100755 infini-qwen3-moe/output/cpp_layer_0_input_hidden_states.txt create mode 100755 infini-qwen3-moe/output/cpp_layer_0_layer_output.txt create mode 100755 infini-qwen3-moe/output/cpp_layer_0_mlp_gate_proj.txt create mode 100755 infini-qwen3-moe/output/cpp_layer_0_mlp_intermediate.txt create mode 100755 infini-qwen3-moe/output/cpp_layer_0_mlp_norm_output.txt create mode 100755 infini-qwen3-moe/output/cpp_layer_0_mlp_up_proj.txt create mode 100644 infini-qwen3-moe/reference/.DS_Store create mode 100755 infini-qwen3-moe/reference/qwen3_moe/__init__.py create mode 100755 infini-qwen3-moe/reference/qwen3_moe/configuration_qwen3_moe.py create mode 100755 infini-qwen3-moe/reference/qwen3_moe/modeling_qwen3_moe.py create mode 100755 infini-qwen3-moe/reference/qwen3_moe/modular_qwen3_moe.py create mode 100755 infini-qwen3-moe/scripts/__init__.py create mode 100755 infini-qwen3-moe/scripts/__pycache__/infer_task.cpython-310.pyc create mode 100755 infini-qwen3-moe/scripts/__pycache__/libinfinicore_infer.cpython-310.pyc create mode 100755 infini-qwen3-moe/scripts/__pycache__/qwen3.cpython-310.pyc create mode 100755 infini-qwen3-moe/scripts/infer_task.py create mode 100755 infini-qwen3-moe/scripts/jiuge.py create mode 100755 infini-qwen3-moe/scripts/kvcache_pool.py create mode 100755 infini-qwen3-moe/scripts/launch_server.py create mode 100755 infini-qwen3-moe/scripts/libinfinicore_infer.py create mode 100755 infini-qwen3-moe/scripts/qwen3-moe_infer.md create mode 100755 infini-qwen3-moe/scripts/qwen3.py create mode 100755 infini-qwen3-moe/scripts/qwen3_moe.py create mode 100755 infini-qwen3-moe/scripts/test_program.py create mode 100755 infini-qwen3-moe/scripts/test_server.py create mode 100755 infini-qwen3-moe/src/.DS_Store create mode 100755 infini-qwen3-moe/src/allocator.hpp create mode 100755 infini-qwen3-moe/src/allocator/memory_allocator.cpp create mode 100755 infini-qwen3-moe/src/models/.DS_Store create mode 100755 infini-qwen3-moe/src/models/jiuge/jiuge.cpp create mode 100755 infini-qwen3-moe/src/models/jiuge/jiuge_impl.hpp create mode 100755 infini-qwen3-moe/src/models/jiuge/jiuge_kv_cache.cpp create mode 100755 infini-qwen3-moe/src/models/jiuge/jiuge_weight.hpp create mode 100644 infini-qwen3-moe/src/models/qw/qwen3.cpp create mode 100755 infini-qwen3-moe/src/models/qw/qwen3_impl.hpp create mode 100755 infini-qwen3-moe/src/models/qw/qwen3_kv_cache.cpp create mode 100755 infini-qwen3-moe/src/models/qw/qwen3_moe.cpp create mode 100755 infini-qwen3-moe/src/models/qw/qwen3_moe_impl.hpp create mode 100755 infini-qwen3-moe/src/models/qw/qwen3_moe_kv_cache.cpp create mode 100755 infini-qwen3-moe/src/models/qw/qwen3_moe_weight.hpp create mode 100755 infini-qwen3-moe/src/models/qw/qwen3_weight.hpp create mode 100755 infini-qwen3-moe/src/tensor.hpp create mode 100755 infini-qwen3-moe/src/tensor/strorage.cpp create mode 100755 infini-qwen3-moe/src/tensor/tensor.cpp create mode 100755 infini-qwen3-moe/src/tensor/transform.cpp create mode 100755 infini-qwen3-moe/src/utils.hpp create mode 100755 infini-qwen3-moe/third_party/nlohmann/json.hpp create mode 100755 infini-qwen3-moe/trace.log create mode 100755 infini-qwen3-moe/xmake.lua create mode 100644 infini-qwen3/.DS_Store create mode 100644 infini-qwen3/.xmake/.DS_Store create mode 100644 infini-qwen3/.xmake/linux/x86_64/cache/config create mode 100644 infini-qwen3/.xmake/linux/x86_64/cache/detect create mode 100644 infini-qwen3/.xmake/linux/x86_64/cache/history create mode 100644 infini-qwen3/.xmake/linux/x86_64/cache/toolchain create mode 100644 infini-qwen3/.xmake/linux/x86_64/project.lock create mode 100644 infini-qwen3/.xmake/linux/x86_64/xmake.conf create mode 100644 infini-qwen3/.xmake/windows/x64/cache/config create mode 100644 infini-qwen3/.xmake/windows/x64/cache/detect create mode 100644 infini-qwen3/.xmake/windows/x64/cache/history create mode 100644 infini-qwen3/.xmake/windows/x64/cache/option create mode 100644 infini-qwen3/.xmake/windows/x64/cache/package create mode 100644 infini-qwen3/.xmake/windows/x64/cache/project create mode 100644 infini-qwen3/.xmake/windows/x64/cache/toolchain create mode 100644 infini-qwen3/.xmake/windows/x64/project.lock create mode 100644 infini-qwen3/.xmake/windows/x64/xmake.conf create mode 100644 infini-qwen3/README.md create mode 100644 infini-qwen3/build/.DS_Store create mode 100644 infini-qwen3/build/.deps/infinicore_infer/linux/x86_64/debug/libinfinicore_infer.so.d create mode 100644 infini-qwen3/build/.deps/infinicore_infer/linux/x86_64/debug/src/allocator/memory_allocator.cpp.o.d create mode 100644 infini-qwen3/build/.deps/infinicore_infer/linux/x86_64/debug/src/models/jiuge/jiuge.cpp.o.d create mode 100644 infini-qwen3/build/.deps/infinicore_infer/linux/x86_64/debug/src/models/jiuge/jiuge_kv_cache.cpp.o.d create mode 100644 infini-qwen3/build/.deps/infinicore_infer/linux/x86_64/debug/src/models/qw/qwen3.cpp.o.d create mode 100644 infini-qwen3/build/.deps/infinicore_infer/linux/x86_64/debug/src/models/qw/qwen3_kv_cache.cpp.o.d create mode 100644 infini-qwen3/build/.deps/infinicore_infer/linux/x86_64/debug/src/tensor/strorage.cpp.o.d create mode 100644 infini-qwen3/build/.deps/infinicore_infer/linux/x86_64/debug/src/tensor/tensor.cpp.o.d create mode 100644 infini-qwen3/build/.deps/infinicore_infer/linux/x86_64/debug/src/tensor/transform.cpp.o.d create mode 100644 infini-qwen3/build/.objs/infinicore_infer/linux/x86_64/debug/src/allocator/memory_allocator.cpp.o create mode 100644 infini-qwen3/build/.objs/infinicore_infer/linux/x86_64/debug/src/models/jiuge/jiuge.cpp.o create mode 100644 infini-qwen3/build/.objs/infinicore_infer/linux/x86_64/debug/src/models/jiuge/jiuge_kv_cache.cpp.o create mode 100644 infini-qwen3/build/.objs/infinicore_infer/linux/x86_64/debug/src/models/qw/qwen3.cpp.o create mode 100644 infini-qwen3/build/.objs/infinicore_infer/linux/x86_64/debug/src/models/qw/qwen3_kv_cache.cpp.o create mode 100644 infini-qwen3/build/.objs/infinicore_infer/linux/x86_64/debug/src/tensor/strorage.cpp.o create mode 100644 infini-qwen3/build/.objs/infinicore_infer/linux/x86_64/debug/src/tensor/tensor.cpp.o create mode 100644 infini-qwen3/build/.objs/infinicore_infer/linux/x86_64/debug/src/tensor/transform.cpp.o create mode 100644 infini-qwen3/build/linux/x86_64/debug/libinfinicore_infer.so create mode 100644 infini-qwen3/debug_output.json create mode 100644 infini-qwen3/include/infinicore_infer.h create mode 100644 infini-qwen3/include/infinicore_infer/models/jiuge.h create mode 100644 infini-qwen3/include/infinicore_infer/models/qwen3.h create mode 100644 infini-qwen3/output/cpp_input_embeddings.txt create mode 100644 infini-qwen3/output/cpp_input_ids.txt create mode 100644 infini-qwen3/output/cpp_layer_0_attn_k_normed.txt create mode 100644 infini-qwen3/output/cpp_layer_0_attn_k_proj_raw.txt create mode 100644 infini-qwen3/output/cpp_layer_0_attn_norm_output.txt create mode 100644 infini-qwen3/output/cpp_layer_0_attn_q_normed.txt create mode 100644 infini-qwen3/output/cpp_layer_0_attn_q_proj_raw.txt create mode 100644 infini-qwen3/output/cpp_layer_0_attn_residual_output.txt create mode 100644 infini-qwen3/output/cpp_layer_0_attn_v_proj_raw.txt create mode 100644 infini-qwen3/output/cpp_layer_0_input_hidden_states.txt create mode 100644 infini-qwen3/output/cpp_layer_0_layer_output.txt create mode 100644 infini-qwen3/output/cpp_layer_0_mlp_gate_proj.txt create mode 100644 infini-qwen3/output/cpp_layer_0_mlp_intermediate.txt create mode 100644 infini-qwen3/output/cpp_layer_0_mlp_norm_output.txt create mode 100644 infini-qwen3/output/cpp_layer_0_mlp_up_proj.txt create mode 100644 infini-qwen3/scripts/__init__.py create mode 100644 infini-qwen3/scripts/__pycache__/infer_task.cpython-310.pyc create mode 100644 infini-qwen3/scripts/__pycache__/libinfinicore_infer.cpython-310.pyc create mode 100644 infini-qwen3/scripts/__pycache__/qwen3.cpython-310.pyc create mode 100644 infini-qwen3/scripts/infer_task.py create mode 100644 infini-qwen3/scripts/jiuge.py create mode 100644 infini-qwen3/scripts/kvcache_pool.py create mode 100644 infini-qwen3/scripts/launch_server.py create mode 100644 infini-qwen3/scripts/libinfinicore_infer.py create mode 100644 infini-qwen3/scripts/qwen3.py create mode 100644 infini-qwen3/scripts/test_program.py create mode 100644 infini-qwen3/scripts/test_server.py create mode 100644 infini-qwen3/src/.DS_Store create mode 100644 infini-qwen3/src/allocator.hpp create mode 100644 infini-qwen3/src/allocator/memory_allocator.cpp create mode 100644 infini-qwen3/src/models/.DS_Store create mode 100644 infini-qwen3/src/models/jiuge/jiuge.cpp create mode 100644 infini-qwen3/src/models/jiuge/jiuge_impl.hpp create mode 100644 infini-qwen3/src/models/jiuge/jiuge_kv_cache.cpp create mode 100644 infini-qwen3/src/models/jiuge/jiuge_weight.hpp create mode 100644 infini-qwen3/src/models/qw/qwen3.cpp create mode 100644 infini-qwen3/src/models/qw/qwen3_impl.hpp create mode 100644 infini-qwen3/src/models/qw/qwen3_kv_cache.cpp create mode 100644 infini-qwen3/src/models/qw/qwen3_weight.hpp create mode 100644 infini-qwen3/src/tensor.hpp create mode 100644 infini-qwen3/src/tensor/strorage.cpp create mode 100644 infini-qwen3/src/tensor/tensor.cpp create mode 100644 infini-qwen3/src/tensor/transform.cpp create mode 100644 infini-qwen3/src/utils.hpp create mode 100644 infini-qwen3/third_party/nlohmann/json.hpp create mode 100644 infini-qwen3/trace.log create mode 100644 infini-qwen3/xmake.lua create mode 100644 infinicore/.DS_Store create mode 100644 infinicore/.xmake/.DS_Store create mode 100644 infinicore/.xmake/linux/x86_64/cache/config create mode 100644 infinicore/.xmake/linux/x86_64/cache/detect create mode 100644 infinicore/.xmake/linux/x86_64/cache/history create mode 100644 infinicore/.xmake/linux/x86_64/cache/option create mode 100644 infinicore/.xmake/linux/x86_64/cache/package create mode 100644 infinicore/.xmake/linux/x86_64/cache/project create mode 100644 infinicore/.xmake/linux/x86_64/cache/references create mode 100644 infinicore/.xmake/linux/x86_64/cache/repository create mode 100644 infinicore/.xmake/linux/x86_64/cache/toolchain create mode 100644 infinicore/.xmake/linux/x86_64/project.lock create mode 100644 infinicore/.xmake/linux/x86_64/xmake.conf create mode 100644 infinicore/.xmake/windows/x64/cache/config create mode 100644 infinicore/.xmake/windows/x64/cache/detect create mode 100644 infinicore/.xmake/windows/x64/cache/history create mode 100644 infinicore/.xmake/windows/x64/cache/option create mode 100644 infinicore/.xmake/windows/x64/cache/package create mode 100644 infinicore/.xmake/windows/x64/cache/project create mode 100644 infinicore/.xmake/windows/x64/cache/references create mode 100644 infinicore/.xmake/windows/x64/cache/repository create mode 100644 infinicore/.xmake/windows/x64/cache/toolchain create mode 100644 infinicore/.xmake/windows/x64/project.lock create mode 100644 infinicore/.xmake/windows/x64/xmake.conf create mode 100644 infinicore/DEV.md create mode 100644 infinicore/LICENSE create mode 100644 infinicore/README.md create mode 100644 infinicore/build/.DS_Store create mode 100644 infinicore/build/.build_cache/.DS_Store create mode 100644 infinicore/build/.build_cache/07/079094d91096cdd9f330f57b6a792f7b create mode 100644 infinicore/build/.build_cache/0a/0a73e44ba23fd795f3d4e09bb324fb87 create mode 100644 infinicore/build/.build_cache/0a/0ae45ce2e66eec59c15425663e400eab create mode 100644 infinicore/build/.build_cache/0c/0ce03dc3ab55c9ab50da09511c741b58 create mode 100644 infinicore/build/.build_cache/0d/0db018060c11e1dde1c84662a77b70a0 create mode 100644 infinicore/build/.build_cache/16/16813d01d7491e13bc68a1fbc2d8485c create mode 100644 infinicore/build/.build_cache/19/1930a2362e3e9ce56aa85bf8a34c2b84 create mode 100644 infinicore/build/.build_cache/1a/1a590afb820efc0028c6ce22d2f45895 create mode 100644 infinicore/build/.build_cache/1b/1b7287ad34df1a6e007e32a18b9e4e33 create mode 100644 infinicore/build/.build_cache/1b/1b94fbad3a7729ed46b7ea4d26793b85 create mode 100644 infinicore/build/.build_cache/23/23bc33f6571fcc76ac0b30481c73592e create mode 100644 infinicore/build/.build_cache/23/23c1737cf6b3e7dea1b42e83fed588fc create mode 100644 infinicore/build/.build_cache/24/242bb4e0e02aac1f8f82db155d128974 create mode 100644 infinicore/build/.build_cache/28/28c56532e9af3e996ae0a1aab0ddb43a create mode 100644 infinicore/build/.build_cache/2a/2a21c4abd4dbe8a5372e53ece00edecb create mode 100644 infinicore/build/.build_cache/2c/2ce267e5c00137d34b82c828463350c0 create mode 100644 infinicore/build/.build_cache/2e/2effc204d106ca3ade254203795dbe48 create mode 100644 infinicore/build/.build_cache/35/35799ef56d08df52008278bd80e12134 create mode 100644 infinicore/build/.build_cache/35/35799ef56d08df52008278bd80e12134.txt create mode 100644 infinicore/build/.build_cache/3a/3a9735854fa4306a2103db5d249eff67 create mode 100644 infinicore/build/.build_cache/3b/3b6972275901f725ac6af975fd524301 create mode 100644 infinicore/build/.build_cache/3e/3e1e5f5a7f4018ca799116595dc2a371 create mode 100644 infinicore/build/.build_cache/3e/3e9968804b582619978d551bf7c2fa71 create mode 100644 infinicore/build/.build_cache/43/43fccb1f0e3dd562b18da2496db32ad3 create mode 100644 infinicore/build/.build_cache/48/48715c83102b0bd0122a823a18462ed5 create mode 100644 infinicore/build/.build_cache/4c/4ca9e7295d0b02bc979602a9798c8636 create mode 100644 infinicore/build/.build_cache/4f/4fdd279746cbccf1730367960f8571d3 create mode 100644 infinicore/build/.build_cache/51/510b11b3f3e77f76008cf051c8f20a9e create mode 100644 infinicore/build/.build_cache/58/5877c878b49eaaf94554ff6f97170d72 create mode 100644 infinicore/build/.build_cache/59/5967ec2dba3dca8078c4305b8c7c31fa create mode 100644 infinicore/build/.build_cache/5c/5c547ff9d8aa359f3d9451d41da67e35 create mode 100644 infinicore/build/.build_cache/5c/5c547ff9d8aa359f3d9451d41da67e35.txt create mode 100644 infinicore/build/.build_cache/5e/5ef16d3b59b6ef16a6ee10e14c4abf32 create mode 100644 infinicore/build/.build_cache/5f/5f12bba0194596c745b7f8e0f7e7a933 create mode 100644 infinicore/build/.build_cache/61/6164ba502ffcbd8246dc051f0a1ccb4d create mode 100644 infinicore/build/.build_cache/65/65f33bf7911e944c233b098ba0d0cc6c create mode 100644 infinicore/build/.build_cache/6a/6ae6d73efbe28109fcdb1141db5f0daf create mode 100644 infinicore/build/.build_cache/6c/6cab78194d91d5ea32648c1790ee34a0 create mode 100644 infinicore/build/.build_cache/6c/6cab78194d91d5ea32648c1790ee34a0.txt create mode 100644 infinicore/build/.build_cache/6d/6d5a8ac3b00042fce3bf4280f36ea468 create mode 100644 infinicore/build/.build_cache/77/77375e6c2de02c7da0e7f38d9b1e356f create mode 100644 infinicore/build/.build_cache/7b/7b0f3e8e44aaa0db4e3a755e069c5090 create mode 100644 infinicore/build/.build_cache/82/82253da2a9b6dbe64995385bf3430976 create mode 100644 infinicore/build/.build_cache/82/828e126938dedb7e5e4375af7de6483f create mode 100644 infinicore/build/.build_cache/89/8997150f6a1a0f99f60254427275cc3a create mode 100644 infinicore/build/.build_cache/89/8997150f6a1a0f99f60254427275cc3a.txt create mode 100644 infinicore/build/.build_cache/8c/8c7dde4af721c0f837f053165ef4e436 create mode 100644 infinicore/build/.build_cache/91/91165646f8d377c87978019a0be2868b create mode 100644 infinicore/build/.build_cache/97/9748c47c72b101ed65cc7ceb4f172447 create mode 100644 infinicore/build/.build_cache/9e/9e189a5ef986c457cef19e935e2d747d create mode 100644 infinicore/build/.build_cache/9f/9f9df8837828a6b2269d6b824b68f479 create mode 100644 infinicore/build/.build_cache/9f/9fe4409b33f035700e8fc666a8d25af2 create mode 100644 infinicore/build/.build_cache/a0/a06cbe05cc4571da34aa734953a6af89 create mode 100644 infinicore/build/.build_cache/a2/a24cc3806ad790f5f2b97faad2072f1f create mode 100644 infinicore/build/.build_cache/a5/a5b3e124e86ade460f75a7adaa0b2871 create mode 100644 infinicore/build/.build_cache/a5/a5b3e124e86ade460f75a7adaa0b2871.txt create mode 100644 infinicore/build/.build_cache/ad/ad2e31d31efdbda13c270e2f178d2bc5 create mode 100644 infinicore/build/.build_cache/af/afa0c9b9f5721f7730071f22f8c9c3a2 create mode 100644 infinicore/build/.build_cache/af/afa0c9b9f5721f7730071f22f8c9c3a2.txt create mode 100644 infinicore/build/.build_cache/b0/b09e272a8ff712039da5f3f3b05f0154 create mode 100644 infinicore/build/.build_cache/b6/b6029d0ce31f67c9fa0cd7f280f8e550 create mode 100644 infinicore/build/.build_cache/b8/b854f6eb047932c12a12c4e6184c81cc create mode 100644 infinicore/build/.build_cache/b9/b9355f93dd12452230454ec21480a787 create mode 100644 infinicore/build/.build_cache/b9/b938c1086c7d7ff86ec34645aa82b8dc create mode 100644 infinicore/build/.build_cache/bf/bf9b14bd9d6fe3479195fb29eb2549be create mode 100644 infinicore/build/.build_cache/c0/c00777ccbe785ec6ac87fe1b6c78d549 create mode 100644 infinicore/build/.build_cache/c2/c25ff230fd6c08c038e68655b7f2acd2 create mode 100644 infinicore/build/.build_cache/c2/c27134bd84820557efa438dca1573592 create mode 100644 infinicore/build/.build_cache/c5/c531785beef19fdc65a1d1ce4cbaf75b create mode 100644 infinicore/build/.build_cache/c5/c531785beef19fdc65a1d1ce4cbaf75b.txt create mode 100644 infinicore/build/.build_cache/c8/c89f5db66fe2c43876a3898c335eccdb create mode 100644 infinicore/build/.build_cache/c9/c902b460be650828fed61b8fb134426f create mode 100644 infinicore/build/.build_cache/cf/cfef3dabe2383e3a7ae9d0a0fc659f06 create mode 100644 infinicore/build/.build_cache/cf/cfef3dabe2383e3a7ae9d0a0fc659f06.txt create mode 100644 infinicore/build/.build_cache/d3/d369e61bde40654d96cda75ca8d26421 create mode 100644 infinicore/build/.build_cache/d4/d4964a0b57d348c3160657520a75ac7d create mode 100644 infinicore/build/.build_cache/d5/d5831114e2166fe232eba81718436b20 create mode 100644 infinicore/build/.build_cache/d6/d6098f785e367a2d837efdf44d097bdb create mode 100644 infinicore/build/.build_cache/d7/d715d807a8a3e47239ab84ac303ffe99 create mode 100644 infinicore/build/.build_cache/d7/d715d807a8a3e47239ab84ac303ffe99.txt create mode 100644 infinicore/build/.build_cache/df/dfa56d0f177c8224b99a7e7f2bb46d98 create mode 100644 infinicore/build/.build_cache/df/dfd463214c3b5645cd204b724feed178 create mode 100644 infinicore/build/.build_cache/e0/e09a324f71a18401d5d306e49772a72e create mode 100644 infinicore/build/.build_cache/e2/e28f98de25d79bf4b5cb6dc4a9ed0f78 create mode 100644 infinicore/build/.build_cache/e5/e54c931a6b89d000abfe3009a63e9b70 create mode 100644 infinicore/build/.build_cache/e7/e71e476f674a806a21e061e52c3ed8d8 create mode 100644 infinicore/build/.build_cache/eb/eb4f29fb20cb1396e66ea71e61b4575b create mode 100644 infinicore/build/.build_cache/eb/eb4f29fb20cb1396e66ea71e61b4575b.txt create mode 100644 infinicore/build/.build_cache/eb/ebae43a3754c4b5dcd2fc487fa865ceb create mode 100644 infinicore/build/.build_cache/ef/ef0c63f4bc5d819fd4c13bd7e7355359 create mode 100644 infinicore/build/.build_cache/f4/f4b7bb59dbfda1c1b216a74713ba8848 create mode 100644 infinicore/build/.build_cache/f4/f4c57700f62e99b9d9e7b6be4d6732bf create mode 100644 infinicore/build/.build_cache/f7/f7978159b8ebb0680adf4a687e7a1b98 create mode 100644 infinicore/build/.build_cache/f9/f9a10f14dde71fd2f042191ac85ae61f create mode 100644 infinicore/build/.deps/.DS_Store create mode 100644 infinicore/build/.deps/infini-utils/linux/x86_64/release/libinfini-utils.a.d create mode 100644 infinicore/build/.deps/infini-utils/linux/x86_64/release/src/utils/custom_types.cc.o.d create mode 100644 infinicore/build/.deps/infini-utils/linux/x86_64/release/src/utils/rearrange.cc.o.d create mode 100644 infinicore/build/.deps/infini-utils/windows/x64/release/src/utils/custom_types.cc.obj.d create mode 100644 infinicore/build/.deps/infini-utils/windows/x64/release/src/utils/rearrange.cc.obj.d create mode 100644 infinicore/build/.deps/infiniccl/linux/x86_64/release/libinfiniccl.so.d create mode 100644 infinicore/build/.deps/infiniccl/linux/x86_64/release/src/infiniccl/infiniccl.cc.o.d create mode 100644 infinicore/build/.deps/infiniccl/windows/x64/release/src/infiniccl/infiniccl.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop-cpu/linux/x86_64/release/libinfiniop-cpu.a.d create mode 100644 infinicore/build/.deps/infiniop-cpu/linux/x86_64/release/src/infiniop/devices/cpu/common_cpu.cc.o.d create mode 100644 infinicore/build/.deps/infiniop-cpu/linux/x86_64/release/src/infiniop/devices/cpu/cpu_handle.cc.o.d create mode 100644 infinicore/build/.deps/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/add/cpu/add_cpu.cc.o.d create mode 100644 infinicore/build/.deps/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/causal_softmax/cpu/causal_softmax_cpu.cc.o.d create mode 100644 infinicore/build/.deps/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/clip/cpu/clip_cpu.cc.o.d create mode 100644 infinicore/build/.deps/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/conv/cpu/conv_cpu.cc.o.d create mode 100644 infinicore/build/.deps/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/gemm/cpu/gemm_cpu.cc.o.d create mode 100644 infinicore/build/.deps/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/linear/cpu/linear_cpu.cc.o.d create mode 100644 infinicore/build/.deps/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/linear_backwards/cpu/linear_backwards_cpu.cc.o.d create mode 100644 infinicore/build/.deps/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/mul/cpu/mul_cpu.cc.o.d create mode 100644 infinicore/build/.deps/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/random_sample/cpu/random_sample_cpu.cc.o.d create mode 100644 infinicore/build/.deps/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/rearrange/cpu/rearrange_cpu.cc.o.d create mode 100644 infinicore/build/.deps/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/relu/cpu/relu_cpu.cc.o.d create mode 100644 infinicore/build/.deps/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/rms_norm/cpu/rms_norm_cpu.cc.o.d create mode 100644 infinicore/build/.deps/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/rope/cpu/rope_cpu.cc.o.d create mode 100644 infinicore/build/.deps/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/sub/cpu/sub_cpu.cc.o.d create mode 100644 infinicore/build/.deps/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/swiglu/cpu/swiglu_cpu.cc.o.d create mode 100644 infinicore/build/.deps/infiniop-cpu/linux/x86_64/release/src/infiniop/reduce/cpu/reduce.cc.o.d create mode 100644 infinicore/build/.deps/infiniop-cpu/windows/x64/release/src/infiniop/devices/cpu/common_cpu.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop-cpu/windows/x64/release/src/infiniop/devices/cpu/cpu_handle.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop-cpu/windows/x64/release/src/infiniop/ops/rearrange/cpu/rearrange_cpu.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop-cpu/windows/x64/release/src/infiniop/ops/relu/cpu/relu_cpu.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop-cpu/windows/x64/release/src/infiniop/ops/rms_norm/cpu/rms_norm_cpu.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop-cpu/windows/x64/release/src/infiniop/ops/rope/cpu/rope_cpu.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop-cpu/windows/x64/release/src/infiniop/ops/sub/cpu/sub_cpu.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop-cpu/windows/x64/release/src/infiniop/ops/swiglu/cpu/swiglu_cpu.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/libinfiniop.so.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/src/infiniop/devices/handle.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/src/infiniop/operator_descriptor.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/src/infiniop/ops/add/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/src/infiniop/ops/attention/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/src/infiniop/ops/causal_softmax/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/src/infiniop/ops/clip/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/src/infiniop/ops/conv/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/src/infiniop/ops/gemm/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/src/infiniop/ops/linear/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/src/infiniop/ops/linear_backwards/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/src/infiniop/ops/mul/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/src/infiniop/ops/random_sample/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/src/infiniop/ops/rearrange/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/src/infiniop/ops/relu/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/src/infiniop/ops/rms_norm/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/src/infiniop/ops/rope/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/src/infiniop/ops/sub/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/src/infiniop/ops/swiglu/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/linux/x86_64/release/src/infiniop/tensor_descriptor.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/macosx/arm64/release/src/infiniop/devices/handle.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/macosx/arm64/release/src/infiniop/ops/causal_softmax/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/macosx/arm64/release/src/infiniop/ops/conv/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/macosx/arm64/release/src/infiniop/ops/linear/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/macosx/arm64/release/src/infiniop/ops/linear_backwards/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/macosx/arm64/release/src/infiniop/ops/rearrange/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/macosx/arm64/release/src/infiniop/ops/rms_norm/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/macosx/arm64/release/src/infiniop/ops/rope/operator.cc.o.d create mode 100644 infinicore/build/.deps/infiniop/windows/x64/release/src/infiniop/devices/handle.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop/windows/x64/release/src/infiniop/operator_descriptor.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop/windows/x64/release/src/infiniop/ops/add/operator.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop/windows/x64/release/src/infiniop/ops/attention/operator.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop/windows/x64/release/src/infiniop/ops/causal_softmax/operator.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop/windows/x64/release/src/infiniop/ops/clip/operator.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop/windows/x64/release/src/infiniop/ops/conv/operator.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop/windows/x64/release/src/infiniop/ops/gemm/operator.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop/windows/x64/release/src/infiniop/ops/linear/operator.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop/windows/x64/release/src/infiniop/ops/linear_backwards/operator.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop/windows/x64/release/src/infiniop/ops/mul/operator.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop/windows/x64/release/src/infiniop/ops/random_sample/operator.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop/windows/x64/release/src/infiniop/ops/rearrange/operator.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop/windows/x64/release/src/infiniop/ops/relu/operator.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop/windows/x64/release/src/infiniop/ops/rms_norm/operator.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop/windows/x64/release/src/infiniop/ops/rope/operator.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop/windows/x64/release/src/infiniop/ops/sub/operator.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop/windows/x64/release/src/infiniop/ops/swiglu/operator.cc.obj.d create mode 100644 infinicore/build/.deps/infiniop/windows/x64/release/src/infiniop/tensor_descriptor.cc.obj.d create mode 100644 infinicore/build/.deps/infinirt-cpu/linux/x86_64/release/libinfinirt-cpu.a.d create mode 100644 infinicore/build/.deps/infinirt-cpu/linux/x86_64/release/src/infinirt/cpu/infinirt_cpu.cc.o.d create mode 100644 infinicore/build/.deps/infinirt-cpu/windows/x64/release/src/infinirt/cpu/infinirt_cpu.cc.obj.d create mode 100644 infinicore/build/.deps/infinirt/linux/x86_64/release/libinfinirt.so.d create mode 100644 infinicore/build/.deps/infinirt/linux/x86_64/release/src/infinirt/infinirt.cc.o.d create mode 100644 infinicore/build/.deps/infinirt/windows/x64/release/src/infinirt/infinirt.cc.obj.d create mode 100644 infinicore/build/.deps/infiniutils-test/linux/x86_64/release/infiniutils-test.d create mode 100644 infinicore/build/.deps/infiniutils-test/linux/x86_64/release/src/utils-test/main.cc.o.d create mode 100644 infinicore/build/.deps/infiniutils-test/linux/x86_64/release/src/utils-test/test_rearrange.cc.o.d create mode 100644 infinicore/build/.objs/.DS_Store create mode 100644 infinicore/build/.objs/infini-utils/linux/x86_64/release/src/utils/custom_types.cc.o create mode 100644 infinicore/build/.objs/infini-utils/linux/x86_64/release/src/utils/rearrange.cc.o create mode 100644 infinicore/build/.objs/infini-utils/windows/x64/release/src/utils/custom_types.cc.obj create mode 100644 infinicore/build/.objs/infini-utils/windows/x64/release/src/utils/rearrange.cc.obj create mode 100644 infinicore/build/.objs/infiniccl/linux/x86_64/release/src/infiniccl/infiniccl.cc.o create mode 100644 infinicore/build/.objs/infiniccl/windows/x64/release/src/infiniccl/infiniccl.cc.obj create mode 100644 infinicore/build/.objs/infiniop-cpu/linux/x86_64/release/src/infiniop/devices/cpu/common_cpu.cc.o create mode 100644 infinicore/build/.objs/infiniop-cpu/linux/x86_64/release/src/infiniop/devices/cpu/cpu_handle.cc.o create mode 100644 infinicore/build/.objs/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/add/cpu/add_cpu.cc.o create mode 100644 infinicore/build/.objs/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/causal_softmax/cpu/causal_softmax_cpu.cc.o create mode 100644 infinicore/build/.objs/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/clip/cpu/clip_cpu.cc.o create mode 100644 infinicore/build/.objs/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/conv/cpu/conv_cpu.cc.o create mode 100644 infinicore/build/.objs/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/gemm/cpu/gemm_cpu.cc.o create mode 100644 infinicore/build/.objs/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/linear/cpu/linear_cpu.cc.o create mode 100644 infinicore/build/.objs/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/linear_backwards/cpu/linear_backwards_cpu.cc.o create mode 100644 infinicore/build/.objs/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/mul/cpu/mul_cpu.cc.o create mode 100644 infinicore/build/.objs/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/random_sample/cpu/random_sample_cpu.cc.o create mode 100644 infinicore/build/.objs/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/rearrange/cpu/rearrange_cpu.cc.o create mode 100644 infinicore/build/.objs/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/relu/cpu/relu_cpu.cc.o create mode 100644 infinicore/build/.objs/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/rms_norm/cpu/rms_norm_cpu.cc.o create mode 100644 infinicore/build/.objs/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/rope/cpu/rope_cpu.cc.o create mode 100644 infinicore/build/.objs/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/sub/cpu/sub_cpu.cc.o create mode 100644 infinicore/build/.objs/infiniop-cpu/linux/x86_64/release/src/infiniop/ops/swiglu/cpu/swiglu_cpu.cc.o create mode 100644 infinicore/build/.objs/infiniop-cpu/linux/x86_64/release/src/infiniop/reduce/cpu/reduce.cc.o create mode 100644 infinicore/build/.objs/infiniop-cpu/windows/x64/release/src/infiniop/devices/cpu/common_cpu.cc.obj create mode 100644 infinicore/build/.objs/infiniop-cpu/windows/x64/release/src/infiniop/devices/cpu/cpu_handle.cc.obj create mode 100644 infinicore/build/.objs/infiniop-cpu/windows/x64/release/src/infiniop/ops/rearrange/cpu/rearrange_cpu.cc.obj create mode 100644 infinicore/build/.objs/infiniop-cpu/windows/x64/release/src/infiniop/ops/relu/cpu/relu_cpu.cc.obj create mode 100644 infinicore/build/.objs/infiniop-cpu/windows/x64/release/src/infiniop/ops/rms_norm/cpu/rms_norm_cpu.cc.obj create mode 100644 infinicore/build/.objs/infiniop-cpu/windows/x64/release/src/infiniop/ops/rope/cpu/rope_cpu.cc.obj create mode 100644 infinicore/build/.objs/infiniop-cpu/windows/x64/release/src/infiniop/ops/sub/cpu/sub_cpu.cc.obj create mode 100644 infinicore/build/.objs/infiniop-cpu/windows/x64/release/src/infiniop/ops/swiglu/cpu/swiglu_cpu.cc.obj create mode 100644 infinicore/build/.objs/infiniop/linux/x86_64/release/src/infiniop/devices/handle.cc.o create mode 100644 infinicore/build/.objs/infiniop/linux/x86_64/release/src/infiniop/operator_descriptor.cc.o create mode 100644 infinicore/build/.objs/infiniop/linux/x86_64/release/src/infiniop/ops/add/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/linux/x86_64/release/src/infiniop/ops/attention/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/linux/x86_64/release/src/infiniop/ops/causal_softmax/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/linux/x86_64/release/src/infiniop/ops/clip/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/linux/x86_64/release/src/infiniop/ops/conv/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/linux/x86_64/release/src/infiniop/ops/gemm/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/linux/x86_64/release/src/infiniop/ops/linear/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/linux/x86_64/release/src/infiniop/ops/linear_backwards/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/linux/x86_64/release/src/infiniop/ops/mul/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/linux/x86_64/release/src/infiniop/ops/random_sample/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/linux/x86_64/release/src/infiniop/ops/rearrange/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/linux/x86_64/release/src/infiniop/ops/relu/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/linux/x86_64/release/src/infiniop/ops/rms_norm/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/linux/x86_64/release/src/infiniop/ops/rope/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/linux/x86_64/release/src/infiniop/ops/sub/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/linux/x86_64/release/src/infiniop/ops/swiglu/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/linux/x86_64/release/src/infiniop/tensor_descriptor.cc.o create mode 100644 infinicore/build/.objs/infiniop/macosx/arm64/release/src/infiniop/devices/handle.cc.o create mode 100644 infinicore/build/.objs/infiniop/macosx/arm64/release/src/infiniop/ops/attention/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/macosx/arm64/release/src/infiniop/ops/causal_softmax/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/macosx/arm64/release/src/infiniop/ops/conv/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/macosx/arm64/release/src/infiniop/ops/linear/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/macosx/arm64/release/src/infiniop/ops/linear_backwards/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/macosx/arm64/release/src/infiniop/ops/random_sample/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/macosx/arm64/release/src/infiniop/ops/rearrange/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/macosx/arm64/release/src/infiniop/ops/rms_norm/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/macosx/arm64/release/src/infiniop/ops/rope/operator.cc.o create mode 100644 infinicore/build/.objs/infiniop/macosx/arm64/release/src/infiniop/tensor_descriptor.cc.o create mode 100644 infinicore/build/.objs/infiniop/windows/x64/release/src/infiniop/devices/handle.cc.obj create mode 100644 infinicore/build/.objs/infiniop/windows/x64/release/src/infiniop/operator_descriptor.cc.obj create mode 100644 infinicore/build/.objs/infiniop/windows/x64/release/src/infiniop/ops/add/operator.cc.obj create mode 100644 infinicore/build/.objs/infiniop/windows/x64/release/src/infiniop/ops/attention/operator.cc.obj create mode 100644 infinicore/build/.objs/infiniop/windows/x64/release/src/infiniop/ops/causal_softmax/operator.cc.obj create mode 100644 infinicore/build/.objs/infiniop/windows/x64/release/src/infiniop/ops/clip/operator.cc.obj create mode 100644 infinicore/build/.objs/infiniop/windows/x64/release/src/infiniop/ops/conv/operator.cc.obj create mode 100644 infinicore/build/.objs/infiniop/windows/x64/release/src/infiniop/ops/gemm/operator.cc.obj create mode 100644 infinicore/build/.objs/infiniop/windows/x64/release/src/infiniop/ops/linear/operator.cc.obj create mode 100644 infinicore/build/.objs/infiniop/windows/x64/release/src/infiniop/ops/linear_backwards/operator.cc.obj create mode 100644 infinicore/build/.objs/infiniop/windows/x64/release/src/infiniop/ops/mul/operator.cc.obj create mode 100644 infinicore/build/.objs/infiniop/windows/x64/release/src/infiniop/ops/random_sample/operator.cc.obj create mode 100644 infinicore/build/.objs/infiniop/windows/x64/release/src/infiniop/ops/rearrange/operator.cc.obj create mode 100644 infinicore/build/.objs/infiniop/windows/x64/release/src/infiniop/ops/relu/operator.cc.obj create mode 100644 infinicore/build/.objs/infiniop/windows/x64/release/src/infiniop/ops/rms_norm/operator.cc.obj create mode 100644 infinicore/build/.objs/infiniop/windows/x64/release/src/infiniop/ops/rope/operator.cc.obj create mode 100644 infinicore/build/.objs/infiniop/windows/x64/release/src/infiniop/ops/sub/operator.cc.obj create mode 100644 infinicore/build/.objs/infiniop/windows/x64/release/src/infiniop/ops/swiglu/operator.cc.obj create mode 100644 infinicore/build/.objs/infiniop/windows/x64/release/src/infiniop/tensor_descriptor.cc.obj create mode 100644 infinicore/build/.objs/infinirt-cpu/linux/x86_64/release/src/infinirt/cpu/infinirt_cpu.cc.o create mode 100644 infinicore/build/.objs/infinirt-cpu/windows/x64/release/src/infinirt/cpu/infinirt_cpu.cc.obj create mode 100644 infinicore/build/.objs/infinirt/linux/x86_64/release/src/infinirt/infinirt.cc.o create mode 100644 infinicore/build/.objs/infinirt/windows/x64/release/src/infinirt/infinirt.cc.obj create mode 100644 infinicore/build/.objs/infiniutils-test/linux/x86_64/release/src/utils-test/main.cc.o create mode 100644 infinicore/build/.objs/infiniutils-test/linux/x86_64/release/src/utils-test/test_rearrange.cc.o create mode 100644 infinicore/build/linux/x86_64/release/infiniutils-test create mode 100644 infinicore/build/linux/x86_64/release/libinfini-utils.a create mode 100644 infinicore/build/linux/x86_64/release/libinfiniccl.so create mode 100644 infinicore/build/linux/x86_64/release/libinfiniop-cpu.a create mode 100644 infinicore/build/linux/x86_64/release/libinfiniop.so create mode 100644 infinicore/build/linux/x86_64/release/libinfinirt-cpu.a create mode 100644 infinicore/build/linux/x86_64/release/libinfinirt.so create mode 100644 infinicore/comprehensive_demo.py create mode 100644 infinicore/include/.DS_Store create mode 100644 infinicore/include/infiniccl.h create mode 100644 infinicore/include/infinicore.h create mode 100644 infinicore/include/infiniop.h create mode 100644 infinicore/include/infiniop/handle.h create mode 100644 infinicore/include/infiniop/operator_descriptor.h create mode 100644 infinicore/include/infiniop/ops/add.h create mode 100644 infinicore/include/infiniop/ops/attention.h create mode 100644 infinicore/include/infiniop/ops/causal_softmax.h create mode 100644 infinicore/include/infiniop/ops/clip.h create mode 100644 infinicore/include/infiniop/ops/conv.h create mode 100644 infinicore/include/infiniop/ops/gemm.h create mode 100644 infinicore/include/infiniop/ops/linear.h create mode 100644 infinicore/include/infiniop/ops/linear_backwards.h create mode 100644 infinicore/include/infiniop/ops/mul.h create mode 100644 infinicore/include/infiniop/ops/random_sample.h create mode 100644 infinicore/include/infiniop/ops/rearrange.h create mode 100644 infinicore/include/infiniop/ops/relu.h create mode 100644 infinicore/include/infiniop/ops/rms_norm.h create mode 100644 infinicore/include/infiniop/ops/rope.h create mode 100644 infinicore/include/infiniop/ops/sub.h create mode 100644 infinicore/include/infiniop/ops/swiglu.h create mode 100644 infinicore/include/infiniop/tensor_descriptor.h create mode 100644 infinicore/include/infinirt.h create mode 100644 infinicore/infinicore_comparison.py create mode 100644 infinicore/performance_comparison.py create mode 100644 infinicore/quick_demo.py create mode 100644 infinicore/scripts/.DS_Store create mode 100644 infinicore/scripts/__pycache__/set_env.cpython-312.pyc create mode 100644 infinicore/scripts/__pycache__/set_env.cpython-313.pyc create mode 100644 infinicore/scripts/build_ntops.py create mode 100644 infinicore/scripts/format.py create mode 100644 infinicore/scripts/install.py create mode 100644 infinicore/scripts/python_test.py create mode 100644 infinicore/scripts/set_env.py create mode 100644 infinicore/src/.DS_Store create mode 100644 infinicore/src/infiniccl-test/infiniccl_test.cpp create mode 100644 infinicore/src/infiniccl-test/infiniccl_test.hpp create mode 100644 infinicore/src/infiniccl-test/main.cpp create mode 100644 infinicore/src/infiniccl/.DS_Store create mode 100644 infinicore/src/infiniccl/ascend/infiniccl_ascend.cc create mode 100644 infinicore/src/infiniccl/ascend/infiniccl_ascend.h create mode 100644 infinicore/src/infiniccl/cuda/infiniccl_cuda.cu create mode 100644 infinicore/src/infiniccl/cuda/infiniccl_cuda.h create mode 100644 infinicore/src/infiniccl/infiniccl.cc create mode 100644 infinicore/src/infiniccl/infiniccl_impl.h create mode 100644 infinicore/src/infiniccl/metax/infiniccl_metax.cc create mode 100644 infinicore/src/infiniccl/metax/infiniccl_metax.h create mode 100644 infinicore/src/infiniop-test/.DS_Store create mode 100644 infinicore/src/infiniop-test/include/file_mapping.hpp create mode 100644 infinicore/src/infiniop-test/include/gguf.hpp create mode 100644 infinicore/src/infiniop-test/include/ops.hpp create mode 100644 infinicore/src/infiniop-test/include/tensor.hpp create mode 100644 infinicore/src/infiniop-test/include/test.hpp create mode 100644 infinicore/src/infiniop-test/include/utils.hpp create mode 100644 infinicore/src/infiniop-test/src/file_mapping.cpp create mode 100644 infinicore/src/infiniop-test/src/gguf.cpp create mode 100644 infinicore/src/infiniop-test/src/main.cpp create mode 100644 infinicore/src/infiniop-test/src/ops/add.cpp create mode 100644 infinicore/src/infiniop-test/src/ops/causal_softmax.cpp create mode 100644 infinicore/src/infiniop-test/src/ops/clip.cpp create mode 100644 infinicore/src/infiniop-test/src/ops/gemm.cpp create mode 100644 infinicore/src/infiniop-test/src/ops/mul.cpp create mode 100644 infinicore/src/infiniop-test/src/ops/random_sample.cpp create mode 100644 infinicore/src/infiniop-test/src/ops/rearrange.cpp create mode 100644 infinicore/src/infiniop-test/src/ops/rms_norm.cpp create mode 100644 infinicore/src/infiniop-test/src/ops/rope.cpp create mode 100644 infinicore/src/infiniop-test/src/ops/sub.cpp create mode 100644 infinicore/src/infiniop-test/src/ops/swiglu.cpp create mode 100644 infinicore/src/infiniop-test/src/tensor.cpp create mode 100644 infinicore/src/infiniop-test/src/test.cpp create mode 100644 infinicore/src/infiniop/.DS_Store create mode 100644 infinicore/src/infiniop/binary/binary.h create mode 100644 infinicore/src/infiniop/binary/cpu/binary_cpu.h create mode 100644 infinicore/src/infiniop/devices/.DS_Store create mode 100644 infinicore/src/infiniop/devices/ascend/CMakeLists.txt create mode 100644 infinicore/src/infiniop/devices/ascend/Makefile create mode 100644 infinicore/src/infiniop/devices/ascend/ascend_handle.cc create mode 100644 infinicore/src/infiniop/devices/ascend/ascend_handle.h create mode 100644 infinicore/src/infiniop/devices/ascend/ascend_kernel_common.h create mode 100644 infinicore/src/infiniop/devices/ascend/common_ascend.cc create mode 100644 infinicore/src/infiniop/devices/ascend/common_ascend.h create mode 100644 infinicore/src/infiniop/devices/bang/bang_handle.cc create mode 100644 infinicore/src/infiniop/devices/bang/bang_handle.h create mode 100644 infinicore/src/infiniop/devices/bang/common_bang.h create mode 100644 infinicore/src/infiniop/devices/cpu/common_cpu.cc create mode 100644 infinicore/src/infiniop/devices/cpu/common_cpu.h create mode 100644 infinicore/src/infiniop/devices/cpu/cpu_handle.cc create mode 100644 infinicore/src/infiniop/devices/cpu/cpu_handle.h create mode 100644 infinicore/src/infiniop/devices/handle.cc create mode 100644 infinicore/src/infiniop/devices/kunlun/kunlun_handle.cc create mode 100644 infinicore/src/infiniop/devices/kunlun/kunlun_handle.h create mode 100644 infinicore/src/infiniop/devices/kunlun/kunlun_kernel_common.h create mode 100644 infinicore/src/infiniop/devices/kunlun/kunlun_kernel_dtype.h create mode 100644 infinicore/src/infiniop/devices/metax/metax_common.h create mode 100644 infinicore/src/infiniop/devices/metax/metax_handle.cc create mode 100644 infinicore/src/infiniop/devices/metax/metax_handle.h create mode 100644 infinicore/src/infiniop/devices/metax/metax_kernel_common.h create mode 100644 infinicore/src/infiniop/devices/musa/common_musa.h create mode 100644 infinicore/src/infiniop/devices/musa/musa_handle.cc create mode 100644 infinicore/src/infiniop/devices/musa/musa_handle.h create mode 100644 infinicore/src/infiniop/devices/nvidia/nvidia_common.cu create mode 100644 infinicore/src/infiniop/devices/nvidia/nvidia_common.cuh create mode 100644 infinicore/src/infiniop/devices/nvidia/nvidia_handle.cuh create mode 100644 infinicore/src/infiniop/devices/nvidia/nvidia_handle.h create mode 100644 infinicore/src/infiniop/devices/nvidia/nvidia_kernel_common.cuh create mode 100644 infinicore/src/infiniop/devices/pool.h create mode 100644 infinicore/src/infiniop/elementwise/.DS_Store create mode 100644 infinicore/src/infiniop/elementwise/cpu/elementwise_cpu.h create mode 100644 infinicore/src/infiniop/elementwise/elementwise.h create mode 100644 infinicore/src/infiniop/elementwise/kunlun/elementwise_kunlun.h create mode 100644 infinicore/src/infiniop/elementwise/kunlun/elementwise_kunlun_api.h create mode 100644 infinicore/src/infiniop/elementwise/kunlun/elementwise_kunlun_kernel.h create mode 100644 infinicore/src/infiniop/elementwise/metax/elementwise_metax.h create mode 100644 infinicore/src/infiniop/elementwise/metax/elementwise_metax_api.h create mode 100644 infinicore/src/infiniop/elementwise/nvidia/elementwise_nvidia.cuh create mode 100644 infinicore/src/infiniop/elementwise/nvidia/elementwise_nvidia_api.cuh create mode 100644 infinicore/src/infiniop/handle.h create mode 100644 infinicore/src/infiniop/ninetoothed/build.py create mode 100644 infinicore/src/infiniop/operator.h create mode 100644 infinicore/src/infiniop/operator_descriptor.cc create mode 100644 infinicore/src/infiniop/ops/.DS_Store create mode 100644 infinicore/src/infiniop/ops/add/.DS_Store create mode 100644 infinicore/src/infiniop/ops/add/cpu/add_cpu.cc create mode 100644 infinicore/src/infiniop/ops/add/cpu/add_cpu.h create mode 100644 infinicore/src/infiniop/ops/add/cuda/kernel.cuh create mode 100644 infinicore/src/infiniop/ops/add/metax/add_metax.h create mode 100644 infinicore/src/infiniop/ops/add/metax/add_metax.maca create mode 100644 infinicore/src/infiniop/ops/add/nvidia/add_nvidia.cu create mode 100644 infinicore/src/infiniop/ops/add/nvidia/add_nvidia.cuh create mode 100644 infinicore/src/infiniop/ops/add/operator.cc create mode 100644 infinicore/src/infiniop/ops/attention/attention.h create mode 100644 infinicore/src/infiniop/ops/attention/operator.cc create mode 100644 infinicore/src/infiniop/ops/causal_softmax/.DS_Store create mode 100644 infinicore/src/infiniop/ops/causal_softmax/ascend/causal_softmax_ascend.cc create mode 100644 infinicore/src/infiniop/ops/causal_softmax/ascend/causal_softmax_ascend.h create mode 100644 infinicore/src/infiniop/ops/causal_softmax/causal_softmax.h create mode 100644 infinicore/src/infiniop/ops/causal_softmax/cpu/causal_softmax_cpu.cc create mode 100644 infinicore/src/infiniop/ops/causal_softmax/cpu/causal_softmax_cpu.h create mode 100644 infinicore/src/infiniop/ops/causal_softmax/cuda/kernel.cuh create mode 100644 infinicore/src/infiniop/ops/causal_softmax/info.h create mode 100644 infinicore/src/infiniop/ops/causal_softmax/metax/causal_softmax_metax.h create mode 100644 infinicore/src/infiniop/ops/causal_softmax/metax/causal_softmax_metax.maca create mode 100644 infinicore/src/infiniop/ops/causal_softmax/nvidia/causal_softmax_nvidia.cu create mode 100644 infinicore/src/infiniop/ops/causal_softmax/nvidia/causal_softmax_nvidia.cuh create mode 100644 infinicore/src/infiniop/ops/causal_softmax/operator.cc create mode 100644 infinicore/src/infiniop/ops/clip/.DS_Store create mode 100644 infinicore/src/infiniop/ops/clip/cpu/clip_cpu.cc create mode 100644 infinicore/src/infiniop/ops/clip/cpu/clip_cpu.h create mode 100644 infinicore/src/infiniop/ops/clip/cuda/kernel.cuh create mode 100644 infinicore/src/infiniop/ops/clip/metax/clip_metax.h create mode 100644 infinicore/src/infiniop/ops/clip/metax/clip_metax.maca create mode 100644 infinicore/src/infiniop/ops/clip/nvidia/clip_nvidia.cu create mode 100644 infinicore/src/infiniop/ops/clip/nvidia/clip_nvidia.cuh create mode 100644 infinicore/src/infiniop/ops/clip/operator.cc create mode 100644 infinicore/src/infiniop/ops/conv/.DS_Store create mode 100644 infinicore/src/infiniop/ops/conv/conv.h create mode 100644 infinicore/src/infiniop/ops/conv/cpu/conv_cpu.cc create mode 100644 infinicore/src/infiniop/ops/conv/cpu/conv_cpu.h create mode 100644 infinicore/src/infiniop/ops/conv/info.h create mode 100644 infinicore/src/infiniop/ops/conv/nvidia/conv_nvidia.cu create mode 100644 infinicore/src/infiniop/ops/conv/nvidia/conv_nvidia.cuh create mode 100644 infinicore/src/infiniop/ops/conv/operator.cc create mode 100644 infinicore/src/infiniop/ops/gemm/.DS_Store create mode 100644 infinicore/src/infiniop/ops/gemm/ascend/gemm_ascend.cc create mode 100644 infinicore/src/infiniop/ops/gemm/ascend/gemm_ascend.h create mode 100644 infinicore/src/infiniop/ops/gemm/bang/gemm_bang.cc create mode 100644 infinicore/src/infiniop/ops/gemm/bang/gemm_bang.h create mode 100644 infinicore/src/infiniop/ops/gemm/cpu/gemm_cpu.cc create mode 100644 infinicore/src/infiniop/ops/gemm/cpu/gemm_cpu.h create mode 100644 infinicore/src/infiniop/ops/gemm/gemm.h create mode 100644 infinicore/src/infiniop/ops/gemm/info.h create mode 100644 infinicore/src/infiniop/ops/gemm/kunlun/gemm_kunlun.cc create mode 100644 infinicore/src/infiniop/ops/gemm/kunlun/gemm_kunlun.h create mode 100644 infinicore/src/infiniop/ops/gemm/metax/gemm_metax.cc create mode 100644 infinicore/src/infiniop/ops/gemm/metax/gemm_metax.h create mode 100644 infinicore/src/infiniop/ops/gemm/musa/gemm_musa.h create mode 100644 infinicore/src/infiniop/ops/gemm/musa/gemm_musa.mu create mode 100644 infinicore/src/infiniop/ops/gemm/nvidia/gemm_nvidia.cu create mode 100644 infinicore/src/infiniop/ops/gemm/nvidia/gemm_nvidia.cuh create mode 100644 infinicore/src/infiniop/ops/gemm/operator.cc create mode 100644 infinicore/src/infiniop/ops/linear/.DS_Store create mode 100644 infinicore/src/infiniop/ops/linear/cpu/linear_cpu.cc create mode 100644 infinicore/src/infiniop/ops/linear/cpu/linear_cpu.h create mode 100644 infinicore/src/infiniop/ops/linear/info.h create mode 100644 infinicore/src/infiniop/ops/linear/linear.h create mode 100644 infinicore/src/infiniop/ops/linear/nvidia/linear_nvidia.cu create mode 100644 infinicore/src/infiniop/ops/linear/nvidia/linear_nvidia.cuh create mode 100644 infinicore/src/infiniop/ops/linear/operator.cc create mode 100644 infinicore/src/infiniop/ops/linear_backwards/.DS_Store create mode 100644 infinicore/src/infiniop/ops/linear_backwards/cpu/linear_backwards_cpu.cc create mode 100644 infinicore/src/infiniop/ops/linear_backwards/cpu/linear_backwards_cpu.h create mode 100644 infinicore/src/infiniop/ops/linear_backwards/info.h create mode 100644 infinicore/src/infiniop/ops/linear_backwards/linear_backwards.h create mode 100644 infinicore/src/infiniop/ops/linear_backwards/nvidia/linear_backwards_nvidia.cu create mode 100644 infinicore/src/infiniop/ops/linear_backwards/nvidia/linear_backwards_nvidia.cuh create mode 100644 infinicore/src/infiniop/ops/linear_backwards/operator.cc create mode 100644 infinicore/src/infiniop/ops/mul/.DS_Store create mode 100644 infinicore/src/infiniop/ops/mul/cpu/mul_cpu.cc create mode 100644 infinicore/src/infiniop/ops/mul/cpu/mul_cpu.h create mode 100644 infinicore/src/infiniop/ops/mul/cuda/kernel.cuh create mode 100644 infinicore/src/infiniop/ops/mul/metax/mul_metax.h create mode 100644 infinicore/src/infiniop/ops/mul/metax/mul_metax.maca create mode 100644 infinicore/src/infiniop/ops/mul/nvidia/mul_nvidia.cu create mode 100644 infinicore/src/infiniop/ops/mul/nvidia/mul_nvidia.cuh create mode 100644 infinicore/src/infiniop/ops/mul/operator.cc create mode 100644 infinicore/src/infiniop/ops/random_sample/.DS_Store create mode 100644 infinicore/src/infiniop/ops/random_sample/ascend/random_sample_aclnn.h create mode 100644 infinicore/src/infiniop/ops/random_sample/ascend/random_sample_kernel.cpp create mode 100644 infinicore/src/infiniop/ops/random_sample/ascend/randomsample_aclnn.cc create mode 100644 infinicore/src/infiniop/ops/random_sample/cpu/random_sample_cpu.cc create mode 100644 infinicore/src/infiniop/ops/random_sample/cpu/random_sample_cpu.h create mode 100644 infinicore/src/infiniop/ops/random_sample/info.h create mode 100644 infinicore/src/infiniop/ops/random_sample/metax/random_sample_kernel.h create mode 100644 infinicore/src/infiniop/ops/random_sample/metax/random_sample_metax.h create mode 100644 infinicore/src/infiniop/ops/random_sample/metax/random_sample_metax.maca create mode 100644 infinicore/src/infiniop/ops/random_sample/nvidia/random_sample_kernel.cuh create mode 100644 infinicore/src/infiniop/ops/random_sample/nvidia/random_sample_nvidia.cu create mode 100644 infinicore/src/infiniop/ops/random_sample/nvidia/random_sample_nvidia.cuh create mode 100644 infinicore/src/infiniop/ops/random_sample/operator.cc create mode 100644 infinicore/src/infiniop/ops/random_sample/random_sample.h create mode 100644 infinicore/src/infiniop/ops/rearrange/.DS_Store create mode 100644 infinicore/src/infiniop/ops/rearrange/ascend/rearrange_ascend.cc create mode 100644 infinicore/src/infiniop/ops/rearrange/ascend/rearrange_ascend.h create mode 100644 infinicore/src/infiniop/ops/rearrange/cpu/rearrange_cpu.cc create mode 100644 infinicore/src/infiniop/ops/rearrange/cpu/rearrange_cpu.h create mode 100644 infinicore/src/infiniop/ops/rearrange/metax/rearrange_kernel.h create mode 100644 infinicore/src/infiniop/ops/rearrange/metax/rearrange_metax.h create mode 100644 infinicore/src/infiniop/ops/rearrange/metax/rearrange_metax.maca create mode 100644 infinicore/src/infiniop/ops/rearrange/nvidia/rearrange_kernel.cuh create mode 100644 infinicore/src/infiniop/ops/rearrange/nvidia/rearrange_nvidia.cu create mode 100644 infinicore/src/infiniop/ops/rearrange/nvidia/rearrange_nvidia.cuh create mode 100644 infinicore/src/infiniop/ops/rearrange/operator.cc create mode 100644 infinicore/src/infiniop/ops/rearrange/rearrange.h create mode 100644 infinicore/src/infiniop/ops/relu/.DS_Store create mode 100644 infinicore/src/infiniop/ops/relu/cpu/relu_cpu.cc create mode 100644 infinicore/src/infiniop/ops/relu/cpu/relu_cpu.h create mode 100644 infinicore/src/infiniop/ops/relu/metax/relu_metax.h create mode 100644 infinicore/src/infiniop/ops/relu/metax/relu_metax.maca create mode 100644 infinicore/src/infiniop/ops/relu/ninetoothed/build.py create mode 100644 infinicore/src/infiniop/ops/relu/nvidia/relu_nvidia.cu create mode 100644 infinicore/src/infiniop/ops/relu/nvidia/relu_nvidia.cuh create mode 100644 infinicore/src/infiniop/ops/relu/operator.cc create mode 100644 infinicore/src/infiniop/ops/rms_norm/.DS_Store create mode 100644 infinicore/src/infiniop/ops/rms_norm/ascend/rms_norm_aclnn.cc create mode 100644 infinicore/src/infiniop/ops/rms_norm/ascend/rms_norm_aclnn.h create mode 100644 infinicore/src/infiniop/ops/rms_norm/cpu/rms_norm_cpu.cc create mode 100644 infinicore/src/infiniop/ops/rms_norm/cpu/rms_norm_cpu.h create mode 100644 infinicore/src/infiniop/ops/rms_norm/cuda/kernel.cuh create mode 100644 infinicore/src/infiniop/ops/rms_norm/info.h create mode 100644 infinicore/src/infiniop/ops/rms_norm/kunlun/rms_norm_kernel.xpu create mode 100644 infinicore/src/infiniop/ops/rms_norm/kunlun/rms_norm_kunlun.cc create mode 100644 infinicore/src/infiniop/ops/rms_norm/kunlun/rms_norm_kunlun.h create mode 100644 infinicore/src/infiniop/ops/rms_norm/metax/rms_norm_metax.cuh create mode 100644 infinicore/src/infiniop/ops/rms_norm/metax/rms_norm_metax.maca create mode 100644 infinicore/src/infiniop/ops/rms_norm/musa/rms_norm_musa.cuh create mode 100644 infinicore/src/infiniop/ops/rms_norm/musa/rms_norm_musa.mu create mode 100644 infinicore/src/infiniop/ops/rms_norm/nvidia/rms_norm_nvidia.cu create mode 100644 infinicore/src/infiniop/ops/rms_norm/nvidia/rms_norm_nvidia.cuh create mode 100644 infinicore/src/infiniop/ops/rms_norm/operator.cc create mode 100644 infinicore/src/infiniop/ops/rms_norm/rms_norm.h create mode 100644 infinicore/src/infiniop/ops/rope/.DS_Store create mode 100644 infinicore/src/infiniop/ops/rope/ascend/rope_ascend.cc create mode 100644 infinicore/src/infiniop/ops/rope/ascend/rope_ascend.h create mode 100644 infinicore/src/infiniop/ops/rope/ascend/rope_ascend_kernel.cpp create mode 100644 infinicore/src/infiniop/ops/rope/cpu/rope_cpu.cc create mode 100644 infinicore/src/infiniop/ops/rope/cpu/rope_cpu.h create mode 100644 infinicore/src/infiniop/ops/rope/cuda/kernel.cuh create mode 100644 infinicore/src/infiniop/ops/rope/metax/rope_metax.h create mode 100644 infinicore/src/infiniop/ops/rope/metax/rope_metax.maca create mode 100644 infinicore/src/infiniop/ops/rope/nvidia/rope_nvidia.cu create mode 100644 infinicore/src/infiniop/ops/rope/nvidia/rope_nvidia.cuh create mode 100644 infinicore/src/infiniop/ops/rope/operator.cc create mode 100644 infinicore/src/infiniop/ops/rope/rope.h create mode 100644 infinicore/src/infiniop/ops/sub/.DS_Store create mode 100644 infinicore/src/infiniop/ops/sub/cpu/sub_cpu.cc create mode 100644 infinicore/src/infiniop/ops/sub/cpu/sub_cpu.h create mode 100644 infinicore/src/infiniop/ops/sub/cuda/kernel.cuh create mode 100644 infinicore/src/infiniop/ops/sub/metax/sub_metax.h create mode 100644 infinicore/src/infiniop/ops/sub/metax/sub_metax.maca create mode 100644 infinicore/src/infiniop/ops/sub/nvidia/sub_nvidia.cu create mode 100644 infinicore/src/infiniop/ops/sub/nvidia/sub_nvidia.cuh create mode 100644 infinicore/src/infiniop/ops/sub/operator.cc create mode 100644 infinicore/src/infiniop/ops/swiglu/.DS_Store create mode 100644 infinicore/src/infiniop/ops/swiglu/ascend/swiglu_ascend.cc create mode 100644 infinicore/src/infiniop/ops/swiglu/ascend/swiglu_ascend.h create mode 100644 infinicore/src/infiniop/ops/swiglu/ascend/swiglu_ascend_kernel.cpp create mode 100644 infinicore/src/infiniop/ops/swiglu/cpu/swiglu_cpu.cc create mode 100644 infinicore/src/infiniop/ops/swiglu/cpu/swiglu_cpu.h create mode 100644 infinicore/src/infiniop/ops/swiglu/cuda/kernel.cuh create mode 100644 infinicore/src/infiniop/ops/swiglu/kunlun/swiglu_kunlun.cc create mode 100644 infinicore/src/infiniop/ops/swiglu/kunlun/swiglu_kunlun.h create mode 100644 infinicore/src/infiniop/ops/swiglu/kunlun/swiglu_kunlun_internal.xpu create mode 100644 infinicore/src/infiniop/ops/swiglu/metax/swiglu_metax.h create mode 100644 infinicore/src/infiniop/ops/swiglu/metax/swiglu_metax.maca create mode 100644 infinicore/src/infiniop/ops/swiglu/nvidia/swiglu_nvidia.cu create mode 100644 infinicore/src/infiniop/ops/swiglu/nvidia/swiglu_nvidia.cuh create mode 100644 infinicore/src/infiniop/ops/swiglu/operator.cc create mode 100644 infinicore/src/infiniop/reduce/.DS_Store create mode 100644 infinicore/src/infiniop/reduce/cpu/reduce.cc create mode 100644 infinicore/src/infiniop/reduce/cpu/reduce.h create mode 100644 infinicore/src/infiniop/reduce/cuda/reduce.cuh create mode 100644 infinicore/src/infiniop/reduce/kunlun/reduce_kunlun.h create mode 100644 infinicore/src/infiniop/tensor.h create mode 100644 infinicore/src/infiniop/tensor_descriptor.cc create mode 100644 infinicore/src/infinirt/.DS_Store create mode 100644 infinicore/src/infinirt/ascend/infinirt_ascend.cc create mode 100644 infinicore/src/infinirt/ascend/infinirt_ascend.h create mode 100644 infinicore/src/infinirt/bang/infinirt_bang.cc create mode 100644 infinicore/src/infinirt/bang/infinirt_bang.h create mode 100644 infinicore/src/infinirt/cpu/infinirt_cpu.cc create mode 100644 infinicore/src/infinirt/cpu/infinirt_cpu.h create mode 100644 infinicore/src/infinirt/cuda/infinirt_cuda.cu create mode 100644 infinicore/src/infinirt/cuda/infinirt_cuda.cuh create mode 100644 infinicore/src/infinirt/infinirt.cc create mode 100644 infinicore/src/infinirt/infinirt_impl.h create mode 100644 infinicore/src/infinirt/kunlun/infinirt_kunlun.cc create mode 100644 infinicore/src/infinirt/kunlun/infinirt_kunlun.h create mode 100644 infinicore/src/infinirt/metax/infinirt_metax.cc create mode 100644 infinicore/src/infinirt/metax/infinirt_metax.h create mode 100644 infinicore/src/infinirt/musa/infinirt_musa.cc create mode 100644 infinicore/src/infinirt/musa/infinirt_musa.h create mode 100644 infinicore/src/utils-test/main.cc create mode 100644 infinicore/src/utils-test/test_rearrange.cc create mode 100644 infinicore/src/utils-test/utils_test.h create mode 100644 infinicore/src/utils.h create mode 100644 infinicore/src/utils/check.h create mode 100644 infinicore/src/utils/custom_types.cc create mode 100644 infinicore/src/utils/custom_types.h create mode 100644 infinicore/src/utils/rearrange.cc create mode 100644 infinicore/src/utils/rearrange.h create mode 100644 infinicore/src/utils/result.hpp create mode 100644 infinicore/test/.DS_Store create mode 100644 infinicore/test/infiniop-test/README.md create mode 100644 infinicore/test/infiniop-test/test_generate/__init__.py create mode 100644 infinicore/test/infiniop-test/test_generate/infiniop_test.py create mode 100644 infinicore/test/infiniop-test/test_generate/testcases/__init__.py create mode 100644 infinicore/test/infiniop-test/test_generate/testcases/add.py create mode 100644 infinicore/test/infiniop-test/test_generate/testcases/causal_softmax.py create mode 100644 infinicore/test/infiniop-test/test_generate/testcases/clip.py create mode 100644 infinicore/test/infiniop-test/test_generate/testcases/gemm.py create mode 100644 infinicore/test/infiniop-test/test_generate/testcases/mul.py create mode 100644 infinicore/test/infiniop-test/test_generate/testcases/random_sample.py create mode 100644 infinicore/test/infiniop-test/test_generate/testcases/rearrange.py create mode 100644 infinicore/test/infiniop-test/test_generate/testcases/rms_norm.py create mode 100644 infinicore/test/infiniop-test/test_generate/testcases/rope.py create mode 100644 infinicore/test/infiniop-test/test_generate/testcases/sub.py create mode 100644 infinicore/test/infiniop-test/test_generate/testcases/swiglu.py create mode 100644 infinicore/test/infiniop/.DS_Store create mode 100644 infinicore/test/infiniop/__init__.py create mode 100644 infinicore/test/infiniop/add.py create mode 100644 infinicore/test/infiniop/attention.py create mode 100644 infinicore/test/infiniop/causal_softmax.py create mode 100644 infinicore/test/infiniop/clip.py create mode 100644 infinicore/test/infiniop/conv.py create mode 100644 infinicore/test/infiniop/gemm.py create mode 100644 infinicore/test/infiniop/libinfiniop/__init__.py create mode 100644 infinicore/test/infiniop/libinfiniop/__pycache__/__init__.cpython-311.pyc create mode 100644 infinicore/test/infiniop/libinfiniop/__pycache__/datatypes.cpython-311.pyc create mode 100644 infinicore/test/infiniop/libinfiniop/__pycache__/devices.cpython-311.pyc create mode 100644 infinicore/test/infiniop/libinfiniop/__pycache__/liboperators.cpython-311.pyc create mode 100644 infinicore/test/infiniop/libinfiniop/__pycache__/op_register.cpython-311.pyc create mode 100644 infinicore/test/infiniop/libinfiniop/__pycache__/structs.cpython-311.pyc create mode 100644 infinicore/test/infiniop/libinfiniop/datatypes.py create mode 100644 infinicore/test/infiniop/libinfiniop/devices.py create mode 100644 infinicore/test/infiniop/libinfiniop/liboperators.py create mode 100644 infinicore/test/infiniop/libinfiniop/op_register.py create mode 100644 infinicore/test/infiniop/libinfiniop/structs.py create mode 100644 infinicore/test/infiniop/libinfiniop/utils.py create mode 100644 infinicore/test/infiniop/linear.py create mode 100644 infinicore/test/infiniop/linear_backwards.py create mode 100644 infinicore/test/infiniop/mul.py create mode 100644 infinicore/test/infiniop/random_sample.py create mode 100644 infinicore/test/infiniop/rearrange.py create mode 100644 infinicore/test/infiniop/relu.py create mode 100644 infinicore/test/infiniop/rms_norm.py create mode 100644 infinicore/test/infiniop/rope.py create mode 100644 infinicore/test/infiniop/sub.py create mode 100644 infinicore/test/infiniop/swiglu.py create mode 100644 infinicore/xmake.lua create mode 100644 infinicore/xmake/ascend.lua create mode 100644 infinicore/xmake/bang.lua create mode 100644 infinicore/xmake/cpu.lua create mode 100644 infinicore/xmake/cpu.lua.backup create mode 100644 infinicore/xmake/cpu.lua.bak create mode 100644 infinicore/xmake/iluvatar.lua create mode 100644 infinicore/xmake/kunlun.lua create mode 100644 infinicore/xmake/metax.lua create mode 100644 infinicore/xmake/musa.lua create mode 100644 infinicore/xmake/nvidia.lua create mode 100644 infinicore/xmake/test.lua create mode 100644 "\346\212\200\346\234\257\346\212\245\345\221\212-Qwen3\345\222\214Qwen3-MoE\345\234\250InfiniCore\345\271\263\345\217\260\351\200\202\351\205\215.md" create mode 100644 "\346\212\200\346\234\257\346\212\245\345\221\212\351\231\204\345\275\225-\346\236\266\346\236\204\345\233\276\350\241\250\345\222\214\346\200\247\350\203\275\345\210\206\346\236\220.md" diff --git a/.DS_Store b/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..0a4b76d4543d5599d763f03946d2c1e6a4ed175a GIT binary patch literal 8196 zcmeHMKX21O6o1zS>QIm(QA9>dEDVWA+dvg{F(IrB7(qp;LL8gcs_hGLjE5+aCnVqt z04sb4_y#PnGV>8wm{@rCr(h>>(k@X&@1(om#rK}?-tWCT`#AtWnvuN@PzC@EE5p(% zHd6{4XLX`1$t#x;3F-p`;6fFiLk!M@w7Q@ePz)#r6a$I@#lUG`0H4{cXgSY)F{)a{ zfMVcGGQiFU9V!&=(+<0yh0({zn=730 z{o~)ij-!)~0f6M&z3V&%=FXLlg_G!IVq)*-vFsX)jSHFB=wdYVGx6|w;oG{5+j^@} z(W~{X8T*>;TP(Tz`V9_FwJqaUkG#*5OZ)P}kDchP*w0h%nfBj2*^JL#G8?9^bIea% zP&JByMKZ7~CM&c2|1O@H>N~(9*;jKZ22KY9q|h>3O-TlDy{Sgx&OeTtP8CYK20 rD%kLJ9FeEvh=2bu#J-NAOvjLz2*fK`7XKlDzXPe~KWASxXO)3JG~ay< literal 0 HcmV?d00001 diff --git "a/README-\346\212\200\346\234\257\346\226\207\346\241\243\345\257\274\350\247\210.md" "b/README-\346\212\200\346\234\257\346\226\207\346\241\243\345\257\274\350\247\210.md" new file mode 100644 index 00000000..2931e66c --- /dev/null +++ "b/README-\346\212\200\346\234\257\346\226\207\346\241\243\345\257\274\350\247\210.md" @@ -0,0 +1,85 @@ +# 📚 InfiniCore平台Qwen3/Qwen3-MoE适配技术文档 + +这是基于InfiniCore平台的Qwen3-1.7B和Qwen3-MoE-A30B模型适配项目的完整技术文档。本文档集提供了项目的详细技术分析、实现成果、性能评估和未来展望。 + +## 📖 文档概览 + +### 🎯 [项目总结报告](./项目总结报告.md) +- **内容**：项目整体概况、核心成就、技术亮点 + +### 📋 [详细技术报告](./技术报告-Qwen3和Qwen3-MoE在InfiniCore平台适配.md) +- **内容**：技术分析和实现细节 + +### 📊 [技术报告附录](./技术报告附录-架构图表和性能分析.md) +- **内容**：架构图表、性能数据、代码示例 + + +## 🏗️ 项目架构概览 + +```mermaid +graph TB + subgraph "应用层" + A[智能对话] --> B[代码助手] --> C[内容创作] --> D[文档处理] + end + + subgraph "模型层" + E[Qwen3基础模型] --> F[Qwen3-MoE模型] + end + + subgraph "框架层" + G[InfiniCore统一接口] + end + + subgraph "硬件层" + H[CPU] --> I[NVIDIA] --> J[寒武纪] --> K[华为昇腾] -->L[...] + end + + A --> E + B --> F + C --> E + D --> F + E --> G + F --> G + G --> H + G --> I + G --> J + G --> K + G --> L +``` + +## 🚀 核心技术亮点 + +本项目聚焦于将通义千问系列模型Qwen3-1.7B和Qwen3-MoE-A30B适配至InfiniCore异构计算平台，实现在多种硬件后端上的高效推理。InfiniCore作为一个成熟的异构计算框架，已支持多种硬件后端，但针对Qwen3系列模型特有的架构特点（尤其是MoE结构）需要专门的适配工作。我们的适配目标是在保持模型精度的同时，充分利用InfiniCore平台的跨硬件能力，使Qwen3系列模型能够在CPU、NVIDIA GPU、寒武纪MLU以及华为昇腾NPU等多种硬件上无缝运行。该适配工作的意义在于，让用户可以根据实际资源情况灵活部署Qwen3系列模型，避免硬件锁定问题，同时通过针对性优化提升各类硬件上的推理效率。特别对于MoE架构，我们需要解决其在异构设备上的专家分配和负载均衡问题，以充分发挥这类模型的计算效率优势。 +**将Qwen3系列模型适配至InfiniCore平台面临多项技术挑战。**首先，Qwen3模型采用了特定的注意力机制和位置编码方式，需要在InfiniCore支持的各硬件后端上精确实现。为此，我们针对Qwen3的旋转位置编码(RoPE)、RMSNorm和GELU激活函数进行了专门适配，确保计算结果与原始实现保持一致性。其次，针对Qwen3-MoE模型的专家混合架构，我们遇到了更大的挑战。MoE结构要求在推理过程中动态选择专家并平衡负载，这在异构硬件环境下难度更大。我们基于InfiniCore的现有框架，开发了适用于Qwen3-MoE的专家调度插件，实现了跨设备的专家分配和同步机制。 +对于内存管理问题，Qwen3系列模型（尤其是MoE版本）对内存要求较高，我们利用InfiniCore的内存池管理功能进行了针对性优化，并为Qwen3模型定制了KV缓存压缩策略，在各类硬件后端上都取得了显著的内存占用减少。针对不同硬件后端的差异性，我们采用了抽象层设计，将Qwen3模型的计算需求映射到InfiniCore的统一API，然后由平台根据实际硬件选择最优实现。这种方式使得适配工作能够专注于模型层面的优化，而无需为每种硬件重写实现代码。 + + +## 🎯 应用场景 + +基于我们对Qwen3系列模型的InfiniCore平台适配成果，这些模型可在多种场景中灵活部署。在企业智能客服领域，适配后的模型能够支持1000+并发，满足高峰期的服务需求。对于需要代码辅助的开发场景，我们的适配实现可以部署在企业现有的多样化计算资源上，为开发者提供实时、准确的代码生成和补全功能，而无需专门采购特定硬件。 + +在内容创作领域，特别是专业化内容生成方面，适配后的模型能够根据企业的计算资源条件灵活部署，满足不同规模内容创作团队的需求。对于文档处理系统，如合同分析、报告自动生成等，我们的适配成果允许企业在现有异构计算环境中部署Qwen3模型，实现文档理解和生成能力。 + +科研教育领域也能从我们的适配工作中获益。研究机构和教育机构通常拥有多样化的计算设备，我们的适配使他们能够在现有设备上部署Qwen3模型，用于学术写作辅助、个性化学习内容生成等场景。特别是对于预算有限的机构，我们的适配工作使他们能够充分利用现有硬件资源，无需额外投资即可获得Qwen3模型的强大能力。 + +## 🛠️ 技术特色 + +### 模块化设计 +``` +infini-qwen/ +├── infini-qwen3/ # Qwen3基础实现 +│ ├── src/models/qw/ # C++核心代码 +│ ├── scripts/ # Python接口 +│ └── include/ # 头文件 +├── infini-qwen3-moe/ # Qwen3-MoE实现 + ├── src/models/qw/ # MoE核心代码 + ├── scripts/ # Python接口 + ├── reference/ # 参考实现 + └── include/ # 头文件 +``` + +👉 **推荐阅读顺序**： +1. [技术报告](./技术报告-Qwen3和Qwen3-MoE在InfiniCore平台适配.md#3-核心技术亮点) - 了解技术亮点 +2. [应用场景章节](./技术报告-Qwen3和Qwen3-MoE在InfiniCore平台适配.md#6-应用场景) - 探索应用价值 +3. [技术架构章节](./技术报告-Qwen3和Qwen3-MoE在InfiniCore平台适配.md#2-技术架构) - 理解系统设计 +4. [实现成果章节](./技术报告-Qwen3和Qwen3-MoE在InfiniCore平台适配.md#4-实现成果) - 了解功能特性 diff --git a/infini-qwen3-moe/.DS_Store b/infini-qwen3-moe/.DS_Store new file mode 100755 index 0000000000000000000000000000000000000000..c82967c1c0bc54fbc4729f696d72b9bf63d40280 GIT binary patch literal 8196 zcmeHL&x_MQ6n?YqwkbvEL80CPg6yHz?sivp@ltC&Dq=+sD&3}uyJgc%NLs6uLXM(G z@#0bJb7L44-h>1C-~+^Z6;|H_Ts|Mz|2eLeQ)OLHp4qGz8Rx$6 z>Q*bD6*!X$u=j(BmD!fGk+A&LfelXq(B`q57re$ifXA59wxo@O<)JvH$R3CcB{Ibz zA{_lDgQIOp8wo4IfrxM*a%Lhk6vAi6IFrMHw1m~IRzNF|R)FR11*pSw7(qXszwdrb z`bV_Wv0Q3lKR$BNw>ukOZ=^ByKQaDc{5wH}L%=#%Fo7YsVB6uqwJ3EH-O9&>`t5US zP@LL_`Nfy-u5fCxVMlaHY&AlmRsY52FCrxm21siHcJc$q?%OmX&LbvKVw^aaY?Q(W z_8Ff7P-HR&pS6l2;A|xQ8C;YxGMFeBl?T_Q1OYhk6f3h#At%0UE8GfMviBXyu0X{p z-~IXJ(|fUuE0N3P{5w*iPkbh05kjOK;NtoyX1Jiv&Xq=K{zf)c^+8H1y@8_u2B@VK{Cyj>%dszg{RBMWbvy=^oN)cibOO!gfD+#P06n zLu4(k`__Y@GwoH*-=yBSKlGe{2MlcoA&(ypJ)2J3bmG||kL?PFQ8G%sN_94C?lf!W z*7nZ4X3qAu8Z~pfd2K!~85b^Hy?(EEW0it_~`PibIJXuQ0}8~JM%f-|@5QiRU^ zI9_@(1eS3b!b-0qXr8aV^gQbZo=d$0oQDxE<32dV{Se?|-ostd$Ez{CP*{n-85zh7 zmYL|~?u~7l47Bnq1Pi3Bi&o&o3W%2rR9XIiEBpKZiO1K0S^=%V=~sZ{JKatjc`i=v zBxROsn^^a;vchnYusj4Co{ppC={VXyClp?r2xT!XX(M5=2g`r?5TMV0uDx!~t^&US D%4viB literal 0 HcmV?d00001 diff --git a/infini-qwen3-moe/.xmake/linux/x86_64/cache/config b/infini-qwen3-moe/.xmake/linux/x86_64/cache/config new file mode 100755 index 00000000..e7ae6f7b --- /dev/null +++ b/infini-qwen3-moe/.xmake/linux/x86_64/cache/config @@ -0,0 +1,7 @@ +{ + options = { }, + mtimes = { + ["xmake.lua"] = 1754418356 + }, + recheck = false +} \ No newline at end of file diff --git a/infini-qwen3-moe/.xmake/linux/x86_64/cache/detect b/infini-qwen3-moe/.xmake/linux/x86_64/cache/detect new file mode 100755 index 00000000..2b070a22 --- /dev/null +++ b/infini-qwen3-moe/.xmake/linux/x86_64/cache/detect @@ -0,0 +1,294 @@ +{ + find_programver = { + ["/usr/bin/g++"] = "9.4.0-1ubuntu1", + ["/usr/bin/gcc"] = "9.4.0-1ubuntu1" + }, + ["core.tools.gcc.has_ldflags"] = { + ["/usr/bin/g++_9.4.0-1ubuntu1"] = { + ["--spare-dynamic-tags"] = true, + ["--disable-long-section-names"] = true, + ["--trace"] = true, + ["--library"] = true, + ["--copy-dt-needed-entries"] = true, + ["--image-base"] = true, + ["--architecture"] = true, + ["--no-keep-memory"] = true, + ["--split-by-file"] = true, + ["--print-gc-sections"] = true, + ["-rpath"] = true, + ["--script"] = true, + ["-rpath-link"] = true, + ["--version-script"] = true, + ["--gpsize"] = true, + ["--gc-keep-exported"] = true, + ["--no-check-sections"] = true, + ["--filter"] = true, + ["-I"] = true, + ["-h"] = true, + ["--enable-new-dtags"] = true, + ["--kill-at"] = true, + ["-Ur"] = true, + ["--no-print-gc-sections"] = true, + ["--entry"] = true, + ["--no-isolation"] = true, + ["-y"] = true, + ["--out-implib"] = true, + ["-Ttext"] = true, + ["--exclude-symbols"] = true, + ["--cref"] = true, + ["--subsystem"] = true, + ["-Bgroup"] = true, + ["--no-fatal-warnings"] = true, + ["--warn-unresolved-symbols"] = true, + ["--pic-executable"] = true, + ["--default-symver"] = true, + ["--minor-os-version"] = true, + ["--verbose"] = true, + ["-P"] = true, + ["-g"] = true, + ["-Bsymbolic"] = true, + ["--no-print-map-discarded"] = true, + ["--no-allow-shlib-undefined"] = true, + ["--end-group"] = true, + ["--exclude-libs"] = true, + ["-plugin-opt"] = true, + ["--enable-auto-image-base"] = true, + ["--print-output-format"] = true, + ["--minor-image-version"] = true, + ["--section-start"] = true, + ["--force-exe-suffix"] = true, + ["-fini"] = true, + ["--no-omagic"] = true, + ["--dll"] = true, + ["--output-def"] = true, + ["--warn-section-align"] = true, + ["--whole-archive"] = true, + ["-b"] = true, + ["-qmagic"] = true, + ["--ignore-unresolved-symbol"] = true, + ["--error-unresolved-symbols"] = true, + ["--no-export-dynamic"] = true, + ["--print-map"] = true, + ["--orphan-handling"] = true, + ["--exclude-all-symbols"] = true, + ["--disable-runtime-pseudo-reloc"] = true, + ["--wdmdriver"] = true, + ["--export-all-symbols"] = true, + ["--start-group"] = true, + ["--require-defined"] = true, + ["-Qy"] = true, + ["--dynamic-list-cpp-typeinfo"] = true, + ["--reduce-memory-overheads"] = true, + ["--gc-sections"] = true, + ["--no-strip-discarded"] = true, + ["--no-copy-dt-needed-entries"] = true, + ["-static"] = true, + ["--no-map-whole-files"] = true, + ["-nostdlib"] = true, + ["--large-address-aware"] = true, + ["--enable-runtime-pseudo-reloc"] = true, + ["--eh-frame-hdr"] = true, + ["-Trodata-segment"] = true, + ["--enable-extra-pep-debug"] = true, + ["--no-bind"] = true, + ["--fatal-warnings"] = true, + ["--just-symbols"] = true, + ["--ld-generated-unwind-info"] = true, + ["--as-needed"] = true, + ["--auxiliary"] = true, + ["--trace-symbol"] = true, + ["-Tdata"] = true, + ["-Map"] = true, + ["--warn-alternate-em"] = true, + ["--format"] = true, + ["-a"] = true, + ["-EL"] = true, + ["--accept-unknown-input-arch"] = true, + ["--warn-common"] = true, + ["--exclude-modules-for-implib"] = true, + ["--library-path"] = true, + ["--add-stdcall-alias"] = true, + ["--warn-once"] = true, + ["--print-map-discarded"] = true, + ["-Tbss"] = true, + ["--traditional-format"] = true, + ["--wrap"] = true, + ["--warn-shared-textrel"] = true, + ["--version"] = true, + ["--no-relax"] = true, + ["--strip-all"] = true, + ["--enable-long-section-names"] = true, + ["--tsaware"] = true, + ["-l"] = true, + ["--print-memory-usage"] = true, + ["--nmagic"] = true, + ["--no-as-needed"] = true, + ["--dynamic-list-cpp-new"] = true, + ["--allow-shlib-undefined"] = true, + ["--discard-all"] = true, + ["--discard-none"] = true, + ["--disable-auto-import"] = true, + ["--no-dynamic-linker"] = true, + ["--enable-stdcall-fixup"] = true, + ["--omagic"] = true, + ["--allow-multiple-definition"] = true, + ["--no-define-common"] = true, + ["--enable-reloc-section"] = true, + ["--strip-discarded"] = true, + ["-flto"] = true, + ["--no-demangle"] = true, + ["--no-gc-sections"] = true, + ["--force-group-allocation"] = true, + ["--no-undefined-version"] = true, + ["--export-dynamic"] = true, + ["-dp"] = true, + ["-T"] = true, + ["--no-ld-generated-unwind-info"] = true, + ["--file-alignment"] = true, + ["--enable-auto-import"] = true, + ["--dynamicbase"] = true, + ["-O"] = true, + ["-R"] = true, + ["--push-state"] = true, + ["--stack"] = true, + ["--heap"] = true, + ["-A"] = true, + ["-plugin"] = true, + ["--unique"] = true, + ["--no-accept-unknown-input-arch"] = true, + ["--disable-large-address-aware"] = true, + ["--version-exports-section"] = true, + ["--minor-subsystem-version"] = true, + ["-L"] = true, + ["--print-sysroot"] = true, + ["--compat-implib"] = true, + ["-c"] = true, + ["--dynamic-list-data"] = true, + ["--enable-extra-pe-debug"] = true, + ["--oformat"] = true, + ["--warn-constructors"] = true, + ["--dynamic-linker"] = true, + ["--demangle"] = true, + ["-z"] = true, + ["--no-warn-search-mismatch"] = true, + ["--no-whole-archive"] = true, + ["-assert"] = true, + ["--strip-debug"] = true, + ["--no-seh"] = true, + ["--section-alignment"] = true, + ["--map-whole-files"] = true, + ["--no-undefined"] = true, + ["-Y"] = true, + ["--retain-symbols-file"] = true, + ["--high-entropy-va"] = true, + ["--disable-new-dtags"] = true, + ["--default-script"] = true, + ["--output"] = true, + ["-Bshareable"] = true, + ["--warn-multiple-gp"] = true, + ["--defsym"] = true, + ["-Bsymbolic-functions"] = true, + ["-m"] = true, + ["--disable-stdcall-fixup"] = true, + ["-F"] = true, + ["-init"] = true, + ["--no-warn-mismatch"] = true, + ["--discard-locals"] = true, + ["-u"] = true, + ["--relocatable"] = true, + ["--undefined"] = true, + ["--disable-multiple-abs-defs"] = true, + ["-f"] = true, + ["-dT"] = true, + ["--major-image-version"] = true, + ["--sort-common"] = true, + ["--nxcompat"] = true, + ["--target-help"] = true, + ["-Ttext-segment"] = true, + ["--dynamic-list"] = true, + ["--major-subsystem-version"] = true, + ["--relax"] = true, + ["--sort-section"] = true, + ["--no-eh-frame-hdr"] = true, + ["--emit-relocs"] = true, + ["-debug"] = true, + ["--split-by-reloc"] = true, + ["--warn-duplicate-exports"] = true, + ["--major-os-version"] = true, + ["--check-sections"] = true, + ["--disable-auto-image-base"] = true, + ["-o"] = true, + ["-V"] = true, + ["-G"] = true, + ["--default-imported-symver"] = true, + ["--mri-script"] = true, + ["-soname"] = true, + ["--stats"] = true, + ["--pop-state"] = true, + ["--support-old-code"] = true, + ["-Tldata-segment"] = true, + ["--help"] = true, + ["--forceinteg"] = true, + ["-e"] = true, + ["--task-link"] = true + } + }, + find_program_gcc_arch_x86_64_plat_linux_checktoolsh = { + ["g++"] = "/usr/bin/g++" + }, + ["core.tools.gcc.has_cflags"] = { + ["/usr/bin/gcc_9.4.0-1ubuntu1"] = { + ["--help"] = true, + ["-c"] = true, + ["--param"] = true, + ["-time"] = true, + ["--target-help"] = true, + ["-pipe"] = true, + ["-print-multi-directory"] = true, + ["-print-sysroot"] = true, + ["-dumpversion"] = true, + ["-print-search-dirs"] = true, + ["-pass-exit-codes"] = true, + ["-Xpreprocessor"] = true, + ["-dumpspecs"] = true, + ["-shared"] = true, + ["-no-canonical-prefixes"] = true, + ["-print-multi-os-directory"] = true, + ["-B"] = true, + ["-Xassembler"] = true, + ["-o"] = true, + ["-print-libgcc-file-name"] = true, + ["-save-temps"] = true, + ["-print-sysroot-headers-suffix"] = true, + ["-E"] = true, + ["-S"] = true, + ["-print-multi-lib"] = true, + ["-x"] = true, + ["-pie"] = true, + ["-v"] = true, + ["-print-multiarch"] = true, + ["-Xlinker"] = true, + ["--version"] = true, + ["-dumpmachine"] = true + } + }, + ["lib.detect.has_flags"] = { + ["linux_x86_64_/usr/bin/gcc_9.4.0-1ubuntu1_cxx_cxflags_-m64_-fdiagnostics-color=always"] = true, + ["linux_x86_64_/usr/bin/gcc_9.4.0-1ubuntu1_cxx_cxflags_-m64_-MMD -MF"] = true, + ["linux_x86_64_/usr/bin/gcc_9.4.0-1ubuntu1_cxx_cxflags_-m64_-Wno-gnu-line-marker -Werror"] = true, + ["linux_x86_64_/usr/bin/g++_9.4.0-1ubuntu1_sh__-shared -m64 -m64_-fPIC"] = true, + ["linux_x86_64_/usr/bin/gcc_9.4.0-1ubuntu1_cxx__-m64_-fPIC"] = true + }, + find_program_gcc_arch_x86_64_plat_linux_checktoolcxx = { + gcc = "/usr/bin/gcc" + }, + ["find_program_utils.binary.deplibs"] = { + ldd = "/usr/bin/ldd" + }, + find_program = { + ["/usr/bin/g++"] = "/usr/bin/g++", + ["/usr/bin/gcc"] = "/usr/bin/gcc", + git = "/usr/bin/git", + ping = "/usr/bin/ping" + } +} \ No newline at end of file diff --git a/infini-qwen3-moe/.xmake/linux/x86_64/cache/history b/infini-qwen3-moe/.xmake/linux/x86_64/cache/history new file mode 100755 index 00000000..c9f99e68 --- /dev/null +++ b/infini-qwen3-moe/.xmake/linux/x86_64/cache/history @@ -0,0 +1,69 @@ +{ + cmdlines = { + "xmake install", + "xmake build", + "xmake build", + "xmake install", + "xmake build", + "xmake install", + "xmake build", + "xmake install", + "xmake build", + "xmake build", + "xmake install", + "xmake build", + "xmake install", + "xmake build", + "xmake build", + "xmake build", + "xmake build", + "xmake install", + "xmake build", + "xmake build", + "xmake build", + "xmake install", + "xmake build", + "xmake install", + "xmake build", + "xmake install", + "xmake build", + "xmake build", + "xmake build", + "xmake build", + "xmake install", + "xmake build", + "xmake install", + "xmake clean", + "xmake build", + "xmake install", + "xmake clean", + "xmake build", + "xmake install", + "xmake build", + "xmake build", + "xmake install", + "xmake clean", + "xmake build", + "xmake install", + "xmake build", + "xmake install", + "xmake build", + "xmake install", + "xmake build", + "xmake install", + "xmake build", + "xmake install", + "xmake build", + "xmake install", + "xmake build", + "xmake install", + "xmake build", + "xmake lua /tmp/.xmake1041/3.0.0+HEAD.0db4fe6/modules/private/utils/statistics.lua", + "xmake lua /tmp/.xmake1041/3.0.0+HEAD.0db4fe6/actions/build/cleaner.lua", + "xmake install", + "xmake build", + "xmake lua /tmp/.xmake1041/3.0.0+HEAD.0db4fe6/modules/private/utils/statistics.lua", + "xmake lua /tmp/.xmake1041/3.0.0+HEAD.0db4fe6/actions/build/cleaner.lua", + "xmake install" + } +} \ No newline at end of file diff --git a/infini-qwen3-moe/.xmake/linux/x86_64/cache/toolchain b/infini-qwen3-moe/.xmake/linux/x86_64/cache/toolchain new file mode 100755 index 00000000..f1dd6d28 --- /dev/null +++ b/infini-qwen3-moe/.xmake/linux/x86_64/cache/toolchain @@ -0,0 +1,22 @@ +{ + tool_target_infinicore_infer_linux_x86_64_cxx = { + toolchain_info = { + arch = "x86_64", + plat = "linux", + name = "gcc", + cachekey = "gcc_arch_x86_64_plat_linux" + }, + program = "/usr/bin/gcc", + toolname = "gcc" + }, + tool_target_infinicore_infer_linux_x86_64_sh = { + toolchain_info = { + arch = "x86_64", + plat = "linux", + name = "gcc", + cachekey = "gcc_arch_x86_64_plat_linux" + }, + program = "/usr/bin/g++", + toolname = "gxx" + } +} \ No newline at end of file diff --git a/infini-qwen3-moe/.xmake/linux/x86_64/project.lock b/infini-qwen3-moe/.xmake/linux/x86_64/project.lock new file mode 100755 index 00000000..e69de29b diff --git a/infini-qwen3-moe/.xmake/linux/x86_64/xmake.conf b/infini-qwen3-moe/.xmake/linux/x86_64/xmake.conf new file mode 100755 index 00000000..7fba5237 --- /dev/null +++ b/infini-qwen3-moe/.xmake/linux/x86_64/xmake.conf @@ -0,0 +1,23 @@ +{ + __toolchains_linux_x86_64 = { + "envs", + "gcc", + "yasm", + "nasm", + "fasm", + "cuda", + "go", + "rust", + "swift", + "gfortran", + "fpc" + }, + arch = "x86_64", + builddir = "build", + ccache = true, + host = "linux", + kind = "static", + mode = "debug", + ndk_stdcxx = true, + plat = "linux" +} \ No newline at end of file diff --git a/infini-qwen3-moe/.xmake/windows/x64/cache/config b/infini-qwen3-moe/.xmake/windows/x64/cache/config new file mode 100755 index 00000000..e69de29b diff --git a/infini-qwen3-moe/.xmake/windows/x64/cache/detect b/infini-qwen3-moe/.xmake/windows/x64/cache/detect new file mode 100755 index 00000000..e69de29b diff --git a/infini-qwen3-moe/.xmake/windows/x64/cache/history b/infini-qwen3-moe/.xmake/windows/x64/cache/history new file mode 100755 index 00000000..e69de29b diff --git a/infini-qwen3-moe/.xmake/windows/x64/cache/option b/infini-qwen3-moe/.xmake/windows/x64/cache/option new file mode 100755 index 00000000..e69de29b diff --git a/infini-qwen3-moe/.xmake/windows/x64/cache/package b/infini-qwen3-moe/.xmake/windows/x64/cache/package new file mode 100755 index 00000000..e69de29b diff --git a/infini-qwen3-moe/.xmake/windows/x64/cache/project b/infini-qwen3-moe/.xmake/windows/x64/cache/project new file mode 100755 index 00000000..e69de29b diff --git a/infini-qwen3-moe/.xmake/windows/x64/cache/toolchain b/infini-qwen3-moe/.xmake/windows/x64/cache/toolchain new file mode 100755 index 00000000..e69de29b diff --git a/infini-qwen3-moe/.xmake/windows/x64/project.lock b/infini-qwen3-moe/.xmake/windows/x64/project.lock new file mode 100755 index 00000000..e69de29b diff --git a/infini-qwen3-moe/.xmake/windows/x64/xmake.conf b/infini-qwen3-moe/.xmake/windows/x64/xmake.conf new file mode 100755 index 00000000..1d5bf485 --- /dev/null +++ b/infini-qwen3-moe/.xmake/windows/x64/xmake.conf @@ -0,0 +1,22 @@ +{ + __toolchains_windows_x64 = { + "msvc", + "yasm", + "nasm", + "cuda", + "rust", + "swift", + "go", + "gfortran", + "fpc" + }, + arch = "x64", + builddir = "build", + ccache = true, + host = "windows", + kind = "static", + mode = "release", + ndk_stdcxx = true, + plat = "windows", + vs = "2022" +} \ No newline at end of file diff --git a/infini-qwen3-moe/README.md b/infini-qwen3-moe/README.md new file mode 100755 index 00000000..1899e666 --- /dev/null +++ b/infini-qwen3-moe/README.md @@ -0,0 +1,8 @@ +你需要参考reference中的qwen3_moe python代码 +和src\models\jiuge 中的jiuge模型c++实现的代码 +使用inifini-core框架来实现qwen3_moe的推理 +jiuge模型的推理代码参考\scripts\jiuge.py +除此之外，你可以参考qwen模型，那是对qwen3基础模型的实现 +所以，你的任务是在src\models\建立qwen3_moe的c++代码，来支持基础推理 +在scripts下增加python推理代码，来支持qwen3_moe的推理。所有的模型结构细节都在reference中，你可以参考。 +编写完成后，你需要在对应的scripts下增加qwen3-moe_infer.md说明如何运行moe模型 \ No newline at end of file diff --git a/infini-qwen3-moe/build/.DS_Store b/infini-qwen3-moe/build/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..79b6da37ceaf6d825d04c947ee673a63bfb755a8 GIT binary patch literal 6148 zcmeHKy-LJD5S~3xxI>GSHcCou#Op2M5nG#k0f`ZRNK8NuEcX$7D=Wb_@fG}LX9%~6 z(e4nLf!QyaoykYOESX&*a?5qOAWDd624|eiF??e@&pxn;M)rX&Jma~#TdZ!Hx*g#(N*rW^ofie3;An3GbV zlsY|PI4MUw)x1p63rab;z>Mc8v(pob(;bydLpr%k&_+|h6o?hrk=um(|K<4kKQ6L6 zQ@|AXR|>dkzRH()CGD+)m*ZZW!mr_MoaY6*5?q8Bsr+fj-6ulrKFxv_5TY2#jRjnidNfz8`h1m35u2NX#zwe0i_mwhU5X!ki_H#aKQ~x(hy6n zT58dv)r$V8qP2=UDhgIxv2N6rTH8{wEv>eq)%yRQGiT<`efOTcBoO%Z@9%veGw(a+ zo;fpf=FFMp&bv4{?u0IRdHD(X{xZT>B>t*|>m>Xw!q+7JI>I+3{-%U)A$(in>kvq5#gUCz5(H%CH@}5jS_z!;a?>FSA>6)_$CSej_@B6{{Z2K68|T{e@Xl!gqtP) zF~UzI{%?ezO8hek|AX*zi6^?C57B@0qb>;ZCEgWbH;L~c;f@HqOME8@cSg92#CMf& zHwnLlu!qEZO1L}1FH3w6gnLSSFA4XS@GA&=Nqip(_eJ_zVea5Y|e(PQr5$)=RtrVMgMO2xm%smV~nr zo+t715ndqi3lYwd_(cdWmiQ$IFO~RZ63&(IatY@lyh7qvBAhSrCJDbIVY7q_5VlDC zDumya_(Fu=llavLzc2A?5MC?s>k$4x;)@XeP~z7k{E@^LBfLT4HzNG8#FrqvN#Zvn z{E5VGL3pdge~R!ni7!QXyTtE6_&*X~CgINz-YN0B5H6Sa-3afI_|FmEEAd|-yiely zBU~Zz2PAwD;X@LC7~vxle^kQ95U!N?;|QOS_>&T@lJF^nPfL6?!Zi|qM#5hrd{*Mm zA^erZe=Xtj2-iyd1%$tm_=^Z%lK9IAUy=B$60Sq|TZz9W;p+(BkocPtz9r$?60S%1 zJBh!8@b?mb7vUcy{zrs=lK2LMf0p=r2scXneF^`9@UIg8n}nMX{$1k#K=^^gKScOX ziT?}XM-tzR@MDR8BH_O!{8YlvB>WG;&n2FKD@Pt5k2MJMCEgWbH;M0na7T%EmvAS9 zJ4<{Qgu6<7H-ul3cn^dRN#F7XivM@oDY!qE~>B0NFjCn7A8 zc(H_IBpi!yoWx5cJPF~+5-&w~ip0kwER*;I2`5T;D#FtwJ_+HsBz`)=Z%ceK!ZRd( zCc>1&&yuiQ!YK$VBwmRyE%B)cr%8OegjER7miRdmRwJAt@frzh5!OlkT!i%!Z$Oxl zc%y_f5zdnMYzfapc)r9hknlo;b0mI|gcl>cMB~CfOS}o; zcO>2{;R1v$62A)JcO||M;rAqdHNx*p{2B?bMR=XWe}Hh2#D9qJdWru?!o>)0kob)V ze=PAO65fRHW{Lkq!doP~72!`MejCE262D!-I}rYl#Ft6fA4K?-fZh{PY2@G*odCH^?VCnWx)gsTufCGn>bu9o;3 zgwII)ml8gU@HvV9O2S_wd|u*fC42$lZzTSrgfB_>GQw9R{;GuQ5dK!;uOWP0;%`X! zri5=Hd|Tq{5&ll%?;!lW#NUR_I|3QzheKN7A=?^0&loic=zo4jT z&W56<#`Q%_6W0|rClO!00oF26)ZC4OmeQUEHbFC}no7GfGCY$X_X}(*Yx|1I}%oeT&URD!clR;;zx@YF1{^T z6KcrR;liT1s|vn7V)7YJF0{3dDj$w2UsGAs-1VToTp!8Jg)41&Ol_|zYNFa+0}+q~ zg!9{_3H25JzP_jh1#SV}HF(A>5;ylL)NBW6d>?5WCJv0?jwXI8h2g+Qh1H%%0B*5@A( zbV81q!pFKJOQVeFz19F&wT&|lR&d*vsF$XB!>P0_CUDCBcu{j+QS&KCPhJ9>IKQZQ z1p4ABz3_b@zAy3%)GXBz^bJ?PeL=_kz4~1}?=0jWI z+ypJAHYxkYONv^OOGa?32ij!Xx23c%B^G&^P*XKcTmyYWIIq-eP^11W6E7*6`vi*T zOar)f}&o_NQW;#BU zwGb;`Rnf%U%z`y}N{Sk#)^~$T}V(1M9i!lDLMVQF3*13EG9a z4|fVL-=^f^xpNlh9>^94wPA-lTb6A$keyCNu(?tV8c!>b%p1r|YxOkEAE@H$q_GF2 zxrZfJm=-HRt4&-2i$U75HoF_TW`~9Zj6&uDP0#sJr<=GOOwiDNcLVPd8uxbfyIU~s zkv};*uIXs_vT8>Xhw>+_m>(K>gC4ceYhida^JvWmVgYVPPgqF18D51;FZRd!CY0Qi zIGkfB((bFqZ409_-mMs~EQahW_A3B0W;%cqLszP_LG{~_g_m+MN@LE zS3Zams8F=|Y$29-Wkvo)$QO`lIw{0JA1i8})II!tc+aAiJ@b+CnxdgkL<)4FS0KkY zTA(0_Dy9j%JUNQ83K>dDfZA7GaB}5tw=tz&g&xIzR6W!H;0!l0`@N{<>dvlzqe z8BYfDLgTsmlfTWFjw}R;FNvY@{oBYBL2gXk)VK-FwaF_GxyV4lI*#*dDo6s4fXK-o z$YA~J)x*0Rx{2a%lg_$$FMfXY@B*Rs46-aw^tgIu zR=%QWX8Ib@7q07we&p0B%$TTcZSz0Vx&}hK1MQgmJ~~)<2KImH|K0V&n%q*_k3(J` z*<6IK#$!8mjqqv|rS;;TYtjnPN5#!D#)R3d9^Q{=Ji`bVrbC)TY_RWJT{W#{M!F_b zFr%>{Q!pi6P+m}1J1bpZP&>6CGpja{tgo-FFBnx@nJ!2aRMixmm8hzjT2)h3pE;pE zo$gzoo>o;`b7qJQjy zItEDpdln&nZC*F*OWmDE2QmbYgy8h~kM_^ob2R8D9Nm(aw=Az44zWC$hws$7ZA{UL4y4SC&gmgini+}6`fUc1cC>-OC| zob&-lWV>@|HyrTcGDij?HvDA1VSih(|1p|t`bx~`8AGeuN8=#z*mEeaRXn~qpS%|q ztGmhK*Vah4Cr8Jm$Z$d7nlwhPQeq_c^W^y4m2mU6u+;Lrl`)v%`njofW}l`y^5g$& zB+Dgvu5`Z^{CL)L1sH`f<CNOb!ZEVL1bW<)J+8&nZB;Md8@JvCl$P@6hNeE6?U8XWNyPa&&bK1Z&s9YZ zFDBNhM&I>T%qyWp$xu% zHyrt_K|U*^`K&=cE5rF9Tday>1#?>HZ(6ZIx%T{w9hrJaCadO!4uOdtvHNY!9W?P&~su&^_ndp-yRwe%naTjlOZ=Y}a*TVMr5mS^%zn zR8$L%Nls^}-ZX3AjD^e5S`*Wfn`m>X>4`RSbgfSf?DB7mVqOb_7VQVdvfAF1#RrMn zGg2yFL%2^9KX1`Nrtz#-tB2FBboMGiXtr#&(v4-H8Ti~en-ZC?@YBYw_*xy8w>P8C z=H=Ae)fyN?+Jxj+$`+CoS`+&29|E9F^vqaD4dW;SAljFQN0)av6`4l2+qu|H$@LrW zj4LVnbI(v|BRZb_R7l?cZSrT!ll|TH8Lxd5uafQDo==ggaSb*n9_6@g3f53>T5Hu( z>%Y7&ObsOdNvw7B#xodEvLaBU@OmFgL~zh%y7wffDXhd;E_K@RLQ&IePOGxKDbc~` zSl%)3&W$)!n7JzpY_>|T(J1=qHh!&fUf{skqPD=d$@W5dSb~Z4`%W8t49II7(`bF5 zU&2-9e58N6`VBG(CAaETQ_EUg9^dwJ-#-h|ECtn)h&0%2VmB(tnT-XAJp?|N*^#7U zKAEY|hoyLGrAPC0@-f<$T;i&dNAz{r$U=?IS&5{j#ZAt(p#s*xYkaJzr37KD7q@Zn z)5(?SLfj&75(8;?@m+l`hxbQX-`6&cDHMBjXcKEIgMAdw(d%&^^{Nf?vAT;{L|=av zG57saFkFG!T^^}j-qu4cU?+~51*L#xHQlO2l3d4Tt!yr#WF$3){7u#z8kQQV5tgIw z&@utjmXhaDt!p{2W;u3ci^a3EO(svnG;4Q}wuhLXJb`)c#-3oc$yE)_e4FRc?%}Zb z-9ywvmX(Pj)B;6-(WZ^iAl|M<jyO&_~6D6gD9v6s$3et zdDOEqgcT`Ux|3ZmNY^$sgXSc-ULUd1me>8=PK6s zL`eIlUk=c61 zBT)AcY+8zy!rRZKVo-0~)%p?Zit?x@ARkvM9Hdw;UWS)Fj$3+0xVay4q%2kzHT~9; zaedL;WQGJ9O6G$Ky=xM)2n5w4@7S{eX6G2`uKR^&TKu%%?fL937&MpmLhpf5aSwtoj7h8BS8Jm`kNR01 z9`dry0&J@q)lNQ;Spp4@=jb^t)U2LFG`z$6X(rdg?{Wb{>hg#3Du1l#!3cmJg;|+y zSV_=0YwA7G)0yJf39g!#bY|GlUY-O2ipgd0tQ}w2%VmFjbGS;-NM!L{E%_Fo1oTT8 zS_TSv5xJ)B@_whah}a-uPUoY`Iq%E??QW5BS8jB+jh(vrbzv z7{P-ku{LQPx*Jpzhn1R2Ld&&$ta;AbL}n6f2n{u$vL&D)YbCj7q^M?1bnd1+Sdx7W z9gm$JeRA!_pNBaBxe3zCB*C>Wh)_8($3~6m2{ebw6k4;QNpVDdS$T}mA)m!4cka|+ z#hqKMsSD)8f}G1IkOGR@lM=IM=l=Q)*)p?I3S|h2%w{g44`7C-9o`pD&;fagXn;mO)d$!r=x^}f*+#)pYLZ7J_DFIs*gge&kJ66dRPbcrn1Gb!I zvSl1;f+=l2W08s52Wk9U=#78#lXqP*=dJ_~g3B%cjk|^)N-p1c3seAZ4(YvXJAQBi zNPWTJ-tpYlCq&oi4H(a3R>*TP-@>m2>vbFF80I3UksCa$+|=hFx8LX5!u1filWW+W zY@yx;e=a_IrPq<8=2|Bh65@`EK|W|AtT-=Fh586mg5GIX{7b?r5=VRS7e4jSQ#h{%DXi@YhLHIPI%J>l@&U3IY0R-7`qlXlg*{x zQA5z?#`RVbN-`ak)D9BTk!~K>eIbyvPN~~cK3W~rZp7w$Gjtv> z5=Cr*HrUEAYZseo?IPRW>$XuwgP=XQRYNTVceSV2)|+j9rjB6hAc#9^>E|3R0W^ZP zj%b*|td2B9s?W2cbKmcW+)yg`eY6_YRDupc|7nK9zMv~H`-F4nY))i$VrI<=IYAqi zA}l@Yr0cV&DyN&aoSs~vn#_?t;?xlnjW$yOpaFq48;vm9nLXb`$^5~T^n)Dqryrv; zQL-u_+U_Z6F)i1UZ)jEiDSSomv0qmiF7hBR^+F=^wWMj&xTZgIe|XDcM80t>RfUJR zCyHp9FS(ZPpfQbYPQ&I{Kj{aJ9&yHqGd-I@7JW8qXDMoD37U5ins*`k`uyy^PBTVt ztrsm#rV0h5Zicj$*##VuGf&@}heNMQ=A+(ND9GEimNW-*BlOm_=9d!${NAw|| z8ejd2S{QtF^CGSRdZj4*Y!hbT%>hDwkeeuBQj8^%CRb^nM0&ci%|KHr|ebx%H7Tb}CG3g`OcPK$&9QAWyWGm_+4{;v&mfMoV4}9d;tPd4Wb>)X=y;GDm+O*4F z2l@-e>WAur0%Af#+Y79WuCt)6*8|3X8d^wC0~f4`l$#f5k!01_suV_HUYOsJU)9*y z7Y}bTtA(5K5dPD|HT)32+u=zT8FQ%@^d*gfu&6X%%7a9br9I{JwEr$;uk8H^+moDr zWb1~n6yfop=uo_w`5Fqc(i|8^FCm9F=zOg_N^4KNQ$tl4)GmC2JOJV{*^&%RuCMWKJ7t26Sb! z;I=8Go;lb1tX8l(^n_M$C&&8~>WMv3a<899{9p2W!6@h_zMmS(F>c-z8eqbDd(nZ+ zm=*_`@F-r8y>=yfaevlk5o{y9V6ut*4xfg4A3A{PEF0hO{9)3@=}?wM(8l)8zF{Mm z;|cXnmVFcV)k#p~+1!z&-9GzaJ(p z&}+)S2%O$h*px9y?RPpB`L?Gz+#$&WWH@jP`Kx(|bc-WWAvN*QQz z#`YzMV%BCnuCXg>G{;HsbC181Pp^&R8=Y6<$&FlG(?ZKV+@_REnTKx7-wVzIyovvTrCAhH#atTP^PT^(I{eYWjyr2Qg111Un!CO2=qd?8P_e1FMm zKxl2#@%QYdv$P)5Q0;hj*ldHE_lScTp+_BU(GDTOIw2}W=6sk%U?mJ5M_`)Fdk1TS zYNX+1UGNpP%0s}0k|wC5JNiL(losDw(V_UQLa)|H>xV$?keiu{_^tw5)X1EA<49uz zrOipwmeK+)4=I;hBW`V>3e=uib)HSuNrTqKuXrbEcd3Ke%udf}nUyMwls!u;gFY21{&U31Zn`ni4 zjTOFi&5kTjM;1R}gs3K5e{WT%=p-jCMYq-;)G@)vc zW|QGt^-xgIhI3)Q!{N2?COQOw2SaL_UUXex0?gMk_1j~4_r)wQQ*7KF& zeFFXL$^)oz?^&E(xedy;2Koof8>lp&U5Q=Bq`@rP!`>Y45z|2sl)DSE+~CBuyT2o8 zqu29@3uwsli@ZfA-aSy^K^sJUIO|1F=Zd%MmV!GmRyMdBx+42U%6Y4wJH%Ydb%n zRcmqj+3&%HW-DoO3dhzGh8sCcV)Vnn?0A&cXPt!XS3u5uT%mUZ*75y! zNOF^yB`)4XKE_%vgK;f0P`%!9ujVtr_^sexWO%vbn*KVj=}%nY8~3M%CP%pGEsKq7 z+Ti%Z6$w_9+|*~1xUShc=+`1jZR8EK1$8|fg+-f_X9lM)%NJ^b-Y_HuwTH;%MMd}8 z0bTXnBI-1`?$4jv8;Jvj zAbD59ckB)#ilYI0daA|>j=Vj8ubmy;8OJvHhniV{qS~g_Rt-FFo#upO4V)=*GW&Yh z(-zM1_Vmr~*W{wkowFg4*%@_2Uj3%bc?LL)s~ycq-J9kiExVeQU?}kS^K4tT6nW|9 z5wln|^zla8rG|~A7GQ-!JYWne^NM)*C~!X%0o^Qo&B+Sv6w#*#+I$0;hV|U+TOboR zx^q5pE=9EZ9MUe^2knbDd2x#qW*BfXo2m7B+eG%(bf*m%708M`ZVrIip}<^zZl&#P zrLy@rRp)LSv0kKqC{yTrR4?Mn6O)O$J`?+E z0nmr19~%xcZW`Rv@z(Vvd`vJ{I&zic4246?dNa&0CZ|o*IZM`Km?bB~>8L$B>UF6m z@|dXe;XyZ;cr6-J18q&UqkmuO`w$D-PdmH5I3W{hYQza^mag!SwWm*yYxTJbC^XV2CUFvncAS0@b9PBfwHoSi9vbS$pP;1bpqF3_; zY@w;T(84B1zC?6We@i`Yms%Etaj^j#CmSqT355_5c4VT#)2W%yzInJvgR|n2+78Eh&P+?l? zclw%KA3mJXd3_XPHVCyLoQM$k@+wCfIHaq?AGiXG9;0)Y&{M41#Y;f1)2P!=CbZZP4KKw z>0Z=43Tjiv%Sw`>HsjogH(BiW5hTI;NoS#3_w6nZ|;lM-Du)-w}n4{H^e!{{4>ZH}J^*hZ>xp?TWv2!&f)_ zKk=lA3H;2!O2YE-pFVqUl+%QM>Ih z2?Y2`X~PJ|r;z^G^yxwRWWw>;oBr7Jk|6y;!ttS>KXU0~g7n3N<8vtevCp3sq(4kJ zKF8A^n?5s0e}iy*_N6~I{gfd6Q^N5%nEu%GGlKN}==~jhj-Wp_y)sB2Ch1!FCQJI) z{pU{&p1)AikMh%}2I-3>{bWDA%Lz=dAFzidy|1613Z8#M(tG>qUi*ETBYi(83w<>C zhDo|szR5Y#FU*m?I7j-!Inv+Ak^X6p^!=ohXyqH0BYkp?^b2#OFV2zvaE|mha-@Hn zBYi(I6!@G^f2{s@IoIjmFiF?yXL62oVw-n<={UALJN}5ie%0Ea1M!9a)29)7%B-vm zCoagJ+ciI(2T2N$K8}d+q2D_X(dEBRBy6iTa6-H;7vS0MWFYY=r#i9u6wS@gFU{xj zk}UTU0Y3B_^`U%>Jc}ta>X#FQ7jNiTUio*kECoo%H>@r>2YmSfJZKozJChhp|W%-VkSg$i_jEh(_9?nGkgam;&Uwhv1P3f($A7~ zO^)u;2kD1CLULpR9=GJkV^xkks@sr9_Z_JK`0aN3W6R;$>b|X{7vK^4EoS;0jsKx? z)CA?I79OmxHjgf&n1I@RfuxrbBbR=nOJ64G!~JyV4OKTi|D2>B=%;@>$p3vw=X;zi z|8YV3t~9XYqm8$ROZwM{$UZ+Ycz&FuAMB@l<7u^|_x97h@n}Jg^ktH+$@iS3PxJYA zIg`3ry_F6q=~T=}{0cNsyM5RcMUcu;zsq-%A!P|~$J zTrTNa9j?uh|0YSl*e^pD<1`fj>$x)};CGVgk8R_uG7c0Pt)3@Hx>m-Fq-$loM$$FE z>)ssvUy}3zei_aV^z=hXKh#h6{K`Fdp#Z2V@8PE#w~%%4B}pIZr`H9~e<>EQW+lCH@=LDDt( zGC9((k@PA*|1PG1s1MyM>6+jDlB8?$eV8NvJ$Ivw6A4ZJfs(GtH$l=h`7$~9U(-r@ z0puDceRu}qp|PfmnVe7^trH&F+-$R?YdYNPOJIUeoQ@O#A0p|Yu?k~Rmyv;g7Ku>vF#yX1ND<#%y@ z+%QQW@AIcR7#)=1ETP#o#3P@1kPMeddR#qI`jS@CX`C;TzWhAmAzdIj9uMT$Abjoz zovZ6Ece=XhPKyNijG;faEzGdSHl?qmYi(5|>ElAr2fod8N#`GQv;4)!E5I}JB|YRb zqYU}WxT;u&WkM5IS5y}hq+hpI7nIkB!sivlU0JEV)&%v{Yj;Y-cus$;yq@k3k@Ro+ z>5akjlX5&iThjS-u;YJ^YmY^e&dp}iJsZD5();`Qdt<~pN&kkQULMG|IS2n;{ z&5#`FlO#Q^y-BZUOS)FRMLGDd$dSG-NB*00Jl~5Kr4os;e*KAYBl(BqNS`F>abrLM zm0!{y@%fVu-W9abVxiIesE2c;zmX&TQ%OJEm%*4v?r;0a;-F?9he`Uk{O7xrIvt)Y z=^FnFb3DIT(oYWYH=UpW*u#>3gKt~OW_&lW8S8~6t{tcy6FeD8B(yp%kaSiT+YiN; zBN>KE`d54z%zD6%Bn9PBDKsY{?rh0LK|5cbLxx*gk%9QH%8~xAq|@Bgw<1`cgXKS zbxmWyB%xi9O-p)lQJ@!_CC|8aF94rjU%?-I()7o+JN3`gOuZBU9wsy|XVa_>>h3I| zDf4N%m_=^lafzhA=ck+b)@#It>h~6*(du`Vq{o$oo_|-;ulMul@;w%k(!27{3C8<+ zlzyP3kM^HOUC^wt0N7|r-;|Z!<(;4&rnka_@|Z8_NBcZH9bGEvaki!a&#aMjt-UCZ z_l1U=-nOaiM-Y!)_rV{0TC>UhghulVN+n&J+tx|C)>aE8{Rk4o%INJmEYI=$T1jWu z);^yM+F_HVPxbkCF^e<>zXGVsf~0FcZ${F!I=V*E zd3dric>6;4O8O|j4A4h%zo-n8rEedDc*tIurEBg}n{)73(gq&GzDfef5Mt$&^|S)u zlO_FBBC_dZL(CE@%W;X&#Pw~mXMLsb{sQq(S%}YVLA$PLg%7pK21$<_$0(f#Duh~k zUn0UM&W=%1k)&(=V!EUk5ffV%-a7JpN!NUZrIJ3}e}0x%eo3b`cl=RD)gYmy4U(>v zulxS=pza6sm2^$#izIzGF|perLm2}N+tdaC1{rtT( zn+=k#mY*57(*8Npi&{x9fK1aRz0Q};8cSoUHoDqa{5~ zw=7GWurzgkVgWFErYiuvF3y0Y6vgW@(@6XpN++`Y-8&{pZEUrS@1zjfPKK z+GAjELmOwS3Xn8G((flSt3PC~Zx3{Iw$Ny8x=7MBd%8l>)|yHomPN!Rr9LP^*9;bKXTvj@cgVM*8We$4>50!cqF&*PI{ z$3~LUmr1%-mgjP$zc1-=?Lqu^Wx?Y2eF~6%xTLq{&z1^}CS#qXYced9^jnditIIBn zoy}P;=~bchU~gz`j{G-CdfZw}0eJT00Va`{=JPjWJJ0-yW}witRu7YeMyrR}lCIUm zB1zZgKr1Aji(_@p+iPE!OF~!~ ztzDTD*p*d6$IBa*?&Ki-T}dDA^MReU=SI4+af|QwRE`5BUGqgoOS-1(m90Eq0RERt z`cz*&(-u71kqk?7hT&$ zKhb|)_6rJt-7D!+{B$aVam{%?v`%Q)#kDrU+gsn9Lxx_gMEaa?h@|t@g5@u3U({bG z<#>L!r1O;CKJWQ`izJ-`o9?Y=tdMlA4%bP#)`puUUDL&0M^WYK_F#ylYtK)T^td*p z`k9@B{~}4(#XxA+I$xmFR;Pj?Q^pl^sTEVYT@aq(Ok%C{Z;P#D)uuS=h8x{H` z3VySK+Y`~*{C}#@FIDh66nvS2->Kls75p9rzgNM3q2Tu^`27n0fPz1y;EyQyV+#Jb zfa4Yz04G!7o(sixm74 z1;0$e=PLN+3O-N4uT=0R1^$2#_!NiLMI1gMxQg@SPNV7X{xh{&fZKt>6bJ_(2MOh=L!g;D;&r;R=4Ff*+;eg$mwZ!3Qe%UpP=9q75r2MpQPZYEBIsu=bK+5 zrY}EnmO{_B(!|hDQRpibe5!&^SMVwYKS#l<6}(2lYZd%l1#eLBMg^a#;IkF{JO#f% z!7o(sixm7~1;13m=PLN+3O-N4uT=0R1#edH1qyzZf-h9?s}=kj1;19of1uz$RPY}u z_+kaWQNfoe_)QA_69vCT!EaUY+Z6nE1;0bVmnrzq6#Om)zgxlYQShHD_`M2#pMtMY z@COzAVFiCw!5>rb#})iZ1z)A$Pb>Hu1^=ajKda!sQt)3Z_*w;jLBU^C@Rt<)Wd(m# z!PhDHYYP5`g1@QYZz=eC1%F4u-&OEGDEOZucu$P#U2v{u`EF)-=kq*>4j240!R570 zCdyB2i149={KR_-{uc%RtAhV6f>(0R(dpf6k%$BD&jOC`;2|b@8wcGZ__K`c9KkE)Vch%*hHq_2!|tW0b3A z?+`{`&SqR*;`CB)Vw_KRc__|pG48jvNj9FFya4AtVKiS}CiC#)`G6yxGz))jW?Wtl z@{*3|Mql_en};S6=AJ=$fy_%8%uQS1Atrh+P`gpjo?Q!E(fE%)E_f0hlM@-H;l^xS`Y8TDk}$1J$yOiM|g;d z?y?Joa`@0oy_9h|wCCY}X1t$=nCL0HQpj(`gwjrDdoezQ(U<9rmqqX!8J`frpJaSe z1pgP~--_UaxQm`1!7pZ<-yjR6y~g;NA&kE4(1ZT+8$lFtpV^F0ao=s^BgQN2R~OD? z(=gS2w~+@Jud!cUxEq^{I``d1&SU&C`_+XvGCt3Jw~-BuH`%W)Je^I?)$Y5EJj?iX z_NxnzWRrA*`)(swF}}opb>Zub-{ijA$boFajMcRmg-OPj`VhzOWc>CB-jhw<-4Xoz zj6WX1f6w?U5&T3pk?SJ(T*lvy;7b|*TLeFfP3i{`{O62+7Qqi;Q@a~mdq43`#=jK7 zf6Mq+BKYTw?;FAQWs^NHf=_0APz1k`@#7=-D#nLL@DthepBlm6W&B$ad|3?B7o3?B7o3?B7o3?B7o3|>G%mSB z{tOZ2C!+q0!4Hh^FnH9@F?iI^F?iI^F?iI^G5DbrL_YcE&E^>Vun0y#ej@7U82ku@ z{zwHsO2PXnc%g##SMUJ}K2X62DfnOoKU%?uDELqXKSsfiRq$a7ew>1TQ^Ai{@ZkzR zLcvEW_$UP*t>8%oKS9AyRPZ7NFIMm|3O-iB$0>M;f}f<|Co6cVf}f(`;}yJ2!6zvA zL7}R@E<7nA_f1U zf?u!TKT`0;3Vwrv->Bd}R`4YXev^XVtl&RU@LLr8Rt5j5g5Rd#OBMWf1;0bV|3|@> zDfrJ6{7wbGOTm{b_}vPAkAnYP!S7Y@Unux}3Vy$WuTby@6#PL2e@MX}R`5p@{80se zOu<(w_~Q!xgn~b*;Hwn;DFuI8!B;Ez8U=qw!GEdX&nozH3jQkv|FwcYui$GH`~?O7 zje@_Z;4dlo%L@LAg1@TZ>lFOA3jUgczpmhKDEONS{+5Ejt>Ehw{C5iej)MPQ!QWNz zKPdPg75q;MzCpqNtl;k{_(lbPU%~&P;D1%{zbW`81^>H(|3kq)Q1A~G{GST`F9rWd z!8a@T#|r+5g8y5=KUMJ06#PF5{<(tV%7~axej-o7yC`_Rf_GK$ZVJAGg72u{-4%Q% z1>af0`4?XzHIbj-8?9sT=)MvW$-f*+{h2Pyc$3Vw)! z_f_yi75p0tZcfoe_2h7c{s;v>Qo)Z>@O}zjsNnq-e1L*S_h*g1MfYb7K3L&%w1N*& z@SzHRjDjDl;KLOBI0gTvf=BnAO?jjH&ITW$@G+;Vh{pfqCq^msqZK@<;3p{fi3(n% z;Kd3)M#0A__&5bGQSg%#{A2|$Rq#_3e7u5}Dfk2hpQzx`^9Iz;`H9mM`bi2NJ)dCs zM9(J}{M!nj$qIgkf}g42DFr`E!OInVih@@tc%_1;6@03KPgC&e3SOn)XDj$Q3SO}T~?6N#G=vmL#S{0?}Z#Ci5>qPINoPvAZ6!(O<{ ze)0Hzz~lJzWBdZgzfgFMR`^U*=&Kl?jzpv2$OTqWupXy<*qd!>aM=^f6!;cWWTEVYne4eAX<$ap*D;)lHD-YvW zI{YZXcl{d4GvDDS3x2qQk7K;a(f1Yla~S`Q!+Q&U8SvhT_4)EO(N7}RDfB;Md_jcI zDg}R?@v9uYE!RJQ$LZ%MOux|4+xj^Skv@qBBKT0oqx5GpeszSth4K3xewdW^QHB0@ zj9=sE@s&THDfm7GBsME|P@elO!u`s1PSeG+#>=pRw=V-F|#WfA%Z6#Qry#y*KV z9sT*DC&Q17$BSVQ`XrV+dRsrM6+8(8)hBUJg#Jm!pLh7tlJ9dEu;S>)^ryRVGmgGd zVkzLgy9Do{bQeAs5S|F$yRmk10q{6GcB_Ivso-xb_+|y~I>5-kNCbRPvV;Fs*-P&C4ld@sRMz~jnWt>7~i{7wa5q2M1V_#kW{#L3fj5Xt|llOJY{Kcj%h(N6$Q z`k7sy(*?J0Y}@hlLczZyBYC-m*8uO6_^r#=`cd5m6aKoxG3@f^Y{uVoxb=UY0Ujsk zn+pC{=JS^0(_8qDLC5TqSRcVJXZ#(9TRC?bLiARf zF_iec=kQa6ekE|K=l6FofVFh5GyVIH-ttKuLwx?~a9eM0Grq~;w%#UVV2Z1^Sqi>T z!G8+8PvY;6e{Ye0_hE#8;BYJFb-<}y<99ZMXNo?o1Rf{P^UUX;j?cxy|0Bjfa=5Lx z`55^6BtCZdG@*Z*@qatq)=%P_ME{w?ZT%d__9V zobu-}#&>gg55eDMJnG;6gYk!)JhuM#8bNXf_lQDW`BThzl)jqrsQsVMcyRCNmxa%x zjPLH`#PH3ZcNmZQpJPUnJbO5LTMu&>-^<~)ybm)T+(&Bh9YztKsGmHP@u>gV#CUK| z>E0sedd9!v@~slQWHj-Q`i-wMZr*|8Z_*gloFdp1%YVnsCkNUSqoj`n| ze&el-7dZY_{yrxX{ns3B<^Kiay&Z1z9f$$FPa^71E@r$>g#N3=L?87V=Q1AkqZW)I z`U4!F8B(7=9835?4u`wNpZhV;_elizsrDB96UO^GdaG~8l@NV!pK7VlKhAh?pQ^=& zo<#IVI6i%Zei7qGMewzZ7dqUI7iXVLdYYl<_kiZsmEF@u|xBBoK#w#3diQT zZp*uO1K~G19O}xS#~Hub;TX2~Q z<8SL>%uK?UI^34)9>(u*`1K;s?`9GGGKbst>NT72pE(@vC4Ww5{4R%Y68s9r?{>JQ z-@y3K9d6rw@AHWNFC1?B(HO?>cet(p4;T;bxwZYJ=zQYyprg0;WF6xVJG@ln{Kf@D z|ER+c7QCGC;GSDsZ@*&vaYuiI(0{`ClMc7^V=g5As~r9%pvE|Amh<-|82&fbM&^nU%r_51oz+WA@Yo5JUU)4VElQ< z2h$V&umZP`*<@byS_vl*uz$@bW`vl;9 z64CMKeue%+1wZIY;l^NG(N9d5_F3mMzF<|9v$=@;`6@aW9xG+IImZ9u=xu*&Y9>B^ zbGYqCKW2QB!)>{$7Z81L5Ae$(>{TsOW6L|}TEc^SiEaHq z$arus@mGY;_16*oPL9t9g1__w!gq1_VS-OwMEGtFUnO|g9}?cf;kF+&GrqgSQFi`p zWIVW+*y`bE6tGWXPe*Ud^$O$B@%>vrBKo}}^lvgA+*fSd<%GpV-^Nax>8%>2OPb z8skSf{0yQ0G4Qy0`zg~0_Z#;S`ZpObbbM?(j{FJbJJ8`)&u?ISu)}Tt+RXS6hg&{# zZy`Rxy~mc%4!05>+KG%+6Bdu-|V`6hg*C2AI78O*OaBir`XZk@;=J=mt%uJT5AIJ+37_4T5q)rfa-rbEe@6I(2%o8pN5`FJ#)Es6tvp}5 zllYwK_*nhFjd8mhmqj>oWl6! z5&G{lKF{G+{@*ixrNgbC@Xbewf0M&&B;QqxH#^+c|GtkAeQ-asE$^F*U*+g6{_IMk zU*K@750^hqcyOOH+&}(IdxG$&-DqL_8po%X(C_(Vyj~rl-~)j7NlbEl?6`0e>yPI^5Q0!EcEER}SAx_>?dnt+)3W z|FxsHd5n<8RyRuZ+L! za4Y9&FB6~O9&DTM8phW}=y!P~zFeajf6dX`_FBmJ8xFVpKVdw$AKS_~`c*oo7u?RN0r_jH_^nY^nHm^+zefM>cr;9I%iSI>t z6iytW;EB}fWtnk9QmLxis+y_{z78x*r5ZBjnW~BuAywtoRp;YdZB44YrZQETZphR( zR`BOno|?nw~DNOxFiVkg_tJDX*#qO;LHn^i)l4WjZyb zydfP!)4yM;w7zOaI#r*ZPPBEk)m0Vefp9A2UD&ULB`vM3tsXQI!k&Ya(wSw2sTor^ zeZUA?gpnTq!Z1AAm87tL>eO@v3e4gZgHw5NS-(^=86;HAsH<+n13^&+&t#4svmkAV zt)$|Zq9joaRCubmGF_9Y%A8k>@5!>v!0L2EgNZ{pK*_=ZmQ9MJD=r(98aJgdnGD3- zI{5^04`@hd;zf=TuY-@pNk0R{uPd*rFCL3n|Df0vwKWZyR8=NjPudzQSrS!O*e_L< zu10edHe`y+1_!k{BS{SnAX=9r^t$YPX@aU7GNZ9NQ#GT!uDJarw>o1=8PwnalsCBz zvhZ6{x_)Mv!h40yWPi4>4$PA zxl4GZwN_|*i4s#Qg@e6XVdGoeNtNL@yRC?An0rumX=YXgGLcq0VRhTz($tBMP9-r+ z1!IiYV{Oar%}DU3p+ z(^JdQ1x;clr`n{L0g96fkM^1~Q(s<{X(%RTH2-0s{oYqNps}X59s^~%GDY3I1Dfs^ zCsI^YjbFC|!+n2yvCAu~@k(SQZM8z2j<$c4LvPwX1$DX7xD~QITXz8SDy~!WapfU@ ziHh(=HO{Ug*xtD6kW24AKRr17cB@=UpXYG^7ot)?+mF?)95fSUBIU}_aIWn`qi ze$$GLuzWE{_H*9Bod#SCQ?Ua2?>5oofCSN}T?vJW&xp{ zGdc5c2EENLF}MtgyWwf(iS70wr$Dsg@zCTr=KWL2@l$vnkvw@q>f{OiSbn_);HBhl zrf9F(dR=zQG~CLr)P{8@J2F&*bt%|7*DF{9%B2xrky48PB-$0sUv&ouh% z`l6dT$-LOK%*^uYMzcm2Od~olLkaA3=Q&D&*1Dn;4osayYm9-Ju)5okGT12`saa`c zrD6!yi)g(emmaIFH@3BYVXT_cb;$UU^`hjqOf<&hfrvG;{VPWv{-dzug`)I@g{cXo zU}hn5+>Cxnr)f#2VjV3+S*k?SL$HZp+J%&g ztQU7PbU~+9*`_UdGK?v8!&V09oHb$gI6BV3D?}?@!%DhEI)^qNL#mz}gSv*w7S>gz zE7G&78q$~)V11nSds3OQjIF29b!dxZ*@#r}7({88!ymXhr1pUb)JhP9M1eOI&T>3f z)HqLa1<|_wQ^hrv>Deh6nkF?EX_95dAfs>ayRwAqfX_jev237pi1AoogbNoca0hwL z2)KGqVY^L+3|pvLq_vme>*GYrwaA{&8NW5s43Hju0s5emtu)xJ5( zWe}M0Y)fYOU_=dwVwQL-s*^Kjn9@2rW#6&C*?069DujiG(9b%alB$FDY16hka)=OA zABkL+88D-&rg#P{;)E0|p%}xMebN4*ebMN?X`!D!Hp-wnL-yOuc(V3Sr7b3$nuSx9 zWxvXmnoVDkaVyKbZ)=stEFjx&tt{w-A9p=R%`Z?(l_W?Y zk!*#q8I0e14oK`w8XXT4X|mWselclI+^sa#34BVkpli38Z9TSq^~$z4^&^?XZ`mwN z6KyM3K~UYDD|gW3TPL4@U;B3MZOdlb_n52oc0IPWm11-U#c+=Wbr09TcjgCI;$olluBXNdl_UshSX*bl?;C+(Jd8wKGu*ipnKW{xTk-{mba(loD zr@QS7BW8uQaBx)(c9W~{#z_6N#u<3E0lHT@wn9slDw#khqau$=UYY6jwX=A&(UYU_ zXnB1leNuYX`1CZqvXaDdAqzK^-b=|6EV_VZvYuEos|s(-j88W}+U!gqyIDH+YameP zD9=GtAW4l~>y4%lNY_-F)L_r424`An6Vg&+kH;OPNlqC*mNx5R4&Pu$fR+f$M)D&4 z*nnRgv6?|8ElgL{OvOt?^_de;Z>R#SJ5;3W>p519cTqBhg%#7w>+!k|pK-3>J)gn> zIE{<9k7}{=lrEo9T!GytyqRK~y0*F@iT93(7m8e6T+uczrciiQgIc1Dn3SF{zO1&) z>?(G!jCFNX<+Kc6aZajY`Z=kocsq$Q2(@ZV?-*@XqC<3pSnWvF=zXb9o3j-bR@cr- z<3**~MrcfM?4;8&bedU~NuP_YfcWDzTd+=+#I~TRw4rX#$?N?ANJO2Hf1PYefpB%%1IZl} z*$ysN)ok>W5u#(A)=!R3V_T=Zg3egs{Ka^jLBQs<-d>~YWFOkMHAmOuH8nIh_T)@; zZx451AZ$Ka7F%Q6Lk4Gb{LSF>xm$J!%^c8@fiw2(&xF(@9y^m|qeWc`2h=y>Elj+K zih-}TKIF56lRT4^^^n*tx53*<$I`4BgU)S-PG*dWQyX8a3kx&#=Z&f@$GP;1GZ8a6?ROY<8G>;zbEu-f%(~ehtdge9~v=ep9brU`=T$ZUxlS|d1#v@;-oezqM zpwlDevm>WR(3{CMuxl3P1sVITmB(XVkddc6*FEM18Oz!n+z1wGG&zGQ zSRiqBQ3l}TO=a3n5^)~AUnbUe4uAM9ht;8V-kaH!MM&$^TIZrSyxU=xRt=Tz2>+K{qrXsdkraPr`u?09%J2o%9N=7HmDXJeS}%FtRE)aPWi7w&Q+y(Cwpo zKtn-CuYzcY;?lci9njJ(lg_D9oHlPEYiB&SWed9kZPT+Rt@NPIS8B>!?`ODcgiN|V z>iH8MSrs?-&ow_S3AS_GH}53`8jZoFbMp~p+fK!FTGz?N3`UYK`u>GA*kANl#|Kql z{ifb5b3+HU@&1aFA#ThYhxhI2y7c| z%h515!U0X*v%SDAajj^MgAKmAjHo|mr`h~t=Mve^(Ui!Z4e$KXI(rQ~ zc6nU}ZC|puqO^4$mYt_QmuWUAJlUkkG7TuXLm|))gUeOry6Uj z&TUK=qlcPTyjp$D5hPRSeR?y+wWRGVC|;f}Z0CFX)QoZ+8Fa6_&fWQEwVzAgPC57@ zIc&q27bx9rbJzgzuAV}P;l-J&rw}Jvc#APNPw5v=QA+&1Q?wWCZf&uParT{M^ExR1 z<=qf|L9#4!^wiq=S>^SWRDR%HFl?qgJrcB7re?@70iI?B8;K#4QYC>4tEyt+jYsA^geazgcNme^Xg;V z$}bEqz0%6OWdx)PHiU&g$B*-HK?Xk-E10qo>>J8f?IcEy|5bJpdt`05l3T~;h;E*3 z9loyDqA|)VXO?68s{LC|)leTku0)Ovyx^@PnztFYjR!~9op`$9sVRGJ+Gviwi$F`g z$zF7m$1by)E(y=oo_)|ZdC>17*RADTaH5+d1J+`(nd=Mv~kM{9;daPzu3@j z-R^vm+S>ycTx%w>S|Ll?^Re&p$jqQ;#sw=1Kl!gf0u7~GXivz+I_DIAWf znpQ&>;iqRzX`BjYcngktP-6s}5uPqZ^NCyjPKKU`t=-!7VqQL>cSZCn^(sKW^ptnR z;^m9E^tAJP0T1AO0wPu3IGgV2ZO-8wh@9#+Byz@wrs7{J=R%B=Jubm#I{$7 zxq2^EQQqM0J8NgrjTwt?*FZGyXM$oY5O*s}w zMMuz2Iq|qkyIOd+0%h@5A#k4DKWr{1%Uj!Mzyal8X~H+tag2B~$g5?1r%jybk1_Q_ zrAURC4X~3SYDj#2CjmSa8(DS&-B&o?twP#uvN&@TCqA~=utB!Pw+pvw!>yHS&5~)d z^>5$io!{>9HaRw<*0e#}tZH?23k?>i5U=rB&r@8P3j5n;^s^*(7-L;(rx}RZQQVRp zM1BW`tsref!b;&4HhRZFEZrB}O@$+hv$uRq1$`NJlcH|=o73U{H{D{bp?AZ~0v4A2 zrD}qEUgdW2Lj88}ZFtiMZ|d4({wy=^&g8xzmo9e2rf(5$p@qp#@?_5L(!gE&g7*=g z%7sUM`n?jo4{zpjq4f^Bh?;Iiw{Jz;gBP8%6H0!`OuKy$_&VFb<)xe}uoaj|TSbGd zmwhp1Cgy3>wiI&5|jlY2~&Feu)E9fK0`7LF&_TWd+|xfu#bQ%2x*1&e2Klhtfpc98aQSLPk%^ z{GTm|pCe-;j)OhHT)ZR4&m)uR|68~;uLzJ2TUC=uPpikz=kQsC@|sMs`6UnRcUb!Vq+q4p?Pk6WbYK3>E zL%+WyuE6q+bm+8299`RM+Z%bErm-F7Bd9V3Yvt$L>|zf5Q={YUoJ-OOK*N4~=14#< zcCT||+)7WwuYtG!01T$AG4F-r8X(v@`Ke|a_n14a7!5LJ4w-NHnH#J@44zU|Syhi; zd&bRA-j9-{GUd}ut-0dUZz^|ug&BqPeq>hrUFkyO&hT%`Vm4_U8Z^ST6k=dQ)%n}! zg$FwG5M2tE+cknQBj&dep{sljs8yoSX`ND3Ztxz%w0Wl&LV-)@YlYib58&%iW5y@%SjSnaNGIhOmJrccpY62Izge_gF3d%gW*_jrO{viXu8HTO3TJk-ffLrXu}4!bvv}DzY{c+cY?Hnc*i`l4oKj2*^$|=%WjXdgq8zY zCLF`(B|+b@h`&LBw=HVYIDA~6rcoT{8SG->7kk8`_Zw`Lj2|m?5*YLU1}_cc0hXa> z-9 zt|6V&XvmP;!v_cT;{^ZLbh8yAgJlHEHFiro;xR8iIS)u?J&!-(eu4G{NpXOhiA-r07?z=y27v%yjds6ZW^0lnDY|WaCfE+%|YL zFL=3SsmfaI6cn^TJ&tC(#P%K*>>;T(3<~WuxMUH?gsUd0o&17Ta>sF7oH(Jg$c=UQ zwKn;UsJKxNgQv0pZI7r}<3=hGtYpjao#L_%4x3n;=l79t?EoDO$FU5W9tW2!ZOe-C z_E&QMh=#2WI1B2o7Y@^I`%dEBIGg*|B-&L9UEABj25)ZC&%9Mu2R~)x)QCp<`1|~` zJs#f?j@wVbj!T<<`hd5Y$;D+)qKyk_Q2i`#E=~lf@X_{Q5jLW6u!!Ct!7`Q6h4^)l zHnYxJ(g0%|Sap-|ZD)(OQU!XN(|&i@uV~m2s_3EZujnllA*(?k0J^=M?rw+myFmAb znkwOyIBSQj|GMp#t9j0wbux$F!{;k8cy3!bu&$O4y}Qf!=va!kKM?V}l7;=;6SbIn z$-REK!Y{{j*b|Su@vz0)0qF49#KzCF*yDHcT|Ad;C{$2QDAWLGl6|2rcggH6FjkwD z(VTXqN8ugY7#|?kmc<!tm4>@j?>9!nl z$-QVmuP1v)CAR+Ih3Jc8=I&(ez^Aw5;TAAui!f6@1t-1IQ)*Mw)75bQumDLv80q$; z+1;Z1o+>a9k_wwsO!kG{f<8p9Psp-pG|TB^8c zNK2PjntQ;<&FF`|Q96^4x?)CX2OIp>v9@iGUS@BaZ9|{zM^}YM8qw-$hnJF)_>C7# zr)X`BySKbGMRb#Pnx%uE?92)(?Z_=)gjMD`ytq z)qCEO*;s^$>QFv>SMPvKV;%Vk)qDuJcoN)LD?DzQnE;1TGW31{ICfqWox@r_9okU| zuVD^#`n_%P_U**cuoJN8*&&y4lCUXTC6yH-54mtXs6^L925 zQ{ zm-KP$Ja$RnGg6Q%ZX>}un;ojHR?L(?Zu)O{My*ur+%IH(bmc;!vg~dwH*bZQeomYm zN>ZvV=bPmf4m6`?d-!gdx(2I{`0b+(m21oL&nmxo4SM-Ib*>Rq#t5%6pdyj=&Q290 zYq1m;5!3r9L+N!goX)MCnse0@+DS_dZ1!@mL}GcdGiI})HCDYrbFH$v)8V-bVt+NX zVrGAyl*RpET4ZI@ zEXTww4)VE0<1vz>XvRa5vi*}bqxAkSy>0oujIEzjHTxp}pRI4wYHs>m`$j|H{iB?hKv-;Ibj}^KX2U?Z_0|OT5O{re)Va^8bnx(UF(DO2CnW=UD>`yz5-Wq+K>mI1GobzYw; zi$S5sK#WR9CZgD{Tt#3%>_IQ%d`ulKa`=K#fy%WdRQ|{@*4XtzJ|gLC zgj+T6<;pTwv~1uW5Tk>gw2j)dyX|wevBz*KJi+!22;~avcAMpoI zq6Q|m`ort(vT}fDWBX5=dHTxIpZDA7NSr$p=a~eY_%TJ_1`*f(w2zeJT-P9#<0f*N5J}p#xtV> zitL)vt{VCfehP(dl~3^&T=AGm4aR9pmKB4HzHw*C(*JFFWxDP-wK7yjjl&j==vm_~ z?5VU<2bPHGQk=4Za?*S}24t)l)9-T8FP58X-ZI`83oUnWb~|(7aV0b&Z0(a4!En|t zzb}4B*fL)-o~J(|4B!!gmcXM|4QJc?t!z`XPL_Kc6YvJDhIPn}wz3IOYHd5;M? zb{pEfdq0+bg3g?`RnIrY81F5yiOwu{icKubxYG~#m2L9q_45p#8KzmWJJG;D9)x;q z=Se{8w4niF-J@Kq(B2LtuO^{gcr@^D&B(;sh%OMjvP& z+@9W*gNRrPqMv_)Z-MN+i-YLk&KOJ*;c1X}VHBgidM(FjJ3bkSju^h9u(ew#JlXei z%F?b}ohhtU6N>#`5JOD@7%+lsmcbN`^FV>)#BF-2@#n&BbNaeZ_lpEB+Tak3A-YklAlO?`nmahLvRYO_wZ9Z z`Z^CU_wd;sey)el_3*hKevOCU>ft}~@Vf->f{1*bN-%WJP!Jo@Dx zzTU&{_VD*T{2mX_N1Q&SZ+^eq!^8c4cYufc{ceDV`~B{C5BK}s1P}i?;?^$Jd-%N` zPBxG}B4|w?59`4(#D?Qxjf0Kv%{2%ggpZ_x+?%U^= zJp2?-p4SDZc7M>r|KQ;ddH5!ek8g)R_wa{3`hB`lt(kgz#KR|f_@f@)`2X5_8#pV+ z{C|A%I1+^-gz3>lGSlNkm>!0io{T08;>ZuFv)SaNh^; zmjG`9ya@0Q0B-{PfPG@RF9SXmaE#{~z%icJ0DlGI`4Dg{R|EHr>8^tCqXB;n@JWFG z8t`hsF&}4XoU0k<*>y@Td9G#yRY2z<&Yw8-U*h@nHTw4EUQ6{zbqsf4>DB z^YP$RWiS^P^z%@_--7sa0mt~y2Yd~Lp9(nk`!fK?{`Y#oF(2Or{5KH)Uo`If-|rv) zu5tD=_K*L9c(7mFd%u_;upMrvaX)^)f9we1F`grh@b{=_Q@UHiA7_Lgsl)fPgdbvr zKV63(W(l8bgkPz{=Uc*`YlQdv$A~5TWFvg3j=#zhe!3CRN;MhMl0sc0Wt1kh^c4@!1%3!m;3;_Hch-VDoSe~x~d@Y2(6Y$>xz8vs( z0bdO`*2n(>{vL#HzkkgBb%6H*9OFL;aJ1(vz~6^>DgeiJ=32mi2jOo89OHi!@ZUrD z7Xe=n_(v9e|8}wTV)!EfM|*k!j^W1uj_H~RINDzeINE z3OL5|g~pGgxJN9|8GN2zW%}{|0T2(*F{Sdk$%20lIJXOYalPIQ_#h8ecrD;) z=NEvZo&Nwl89TQ?c#Qumz|qdH0Y^J82OQ)1mj(aUg8v6_OxJ$_&jA122zXz>w*l^t zTiMUp{`a$l$MS~Z{qa7>^Bu(B2jW4zJK*0#c=XQ?fFB9rxls7!3h_38WBhvnj_KMH z@H~irFTjfd$Mj-6dqeoMA^bjo<9ag2gLdu<;g5xQQUS+uwIARkA$(iFG5-AlM?bd% z9Q}C!;26(=fMH?zaMbMuIO-k;IO<}%in_-`c+~9!IO_76q@Q2dZf5{K4(!ha9K-hm9NWDCfMdJ& z)t)ik9Eg7)gvWdw1h}oe?E>)(hVY1I0p1_N4*?w8|C0ds*Bb~*IT>(UJBINOh42{v zDS(d$drk#>GT_)wqJOd>{748t4Dd3*hXY;@_z1vn0emFjj{}|q_`86g4*1^y$MAc% zQD0EdxPG=LgwKU|uph_%Xefl|wRFEcp9J_Qz_C9X4LIiG7{G6Yc*X*Lug19^jDqs~ zFoYim;nx6uCg2+Y$Mw=L0MCc;`&j(e6~bfxhuE9A|G!>Ov0hAqbe#hMx-YV7pNPcrobW_{TqIL_t#! zzaGL@LOizuj`2SQIQs2)z^6exI3BLH;5C3_`&J7$#xoso3_k;KwC7U5hl8EhdVTJmv3V2X3m`nUhYJDEg7n@7crM_#1AZak z<3She*R2r#5(s|>;0=H;0v!AII|087!s9v~&O0z&FGKjdARatliu0O}Aw0(O4d8b} zJpTn8=Q}uF#re*`P~LF9gYDHs@Mi_YzZi6}U%m(M`yl+ifMb7w^UxO{{CyA}{e$x? zoX;+S@VE|eKj3{Jy@h~d`-A>3g7EbaJ_Q~2dHOkkKM3()|MeB*JC+Are=CM~aJ*3qIF2{21{}v5Hvx|0jYk1*1UugZ z9Q}s<`$G_Z2!wwaaGbwmJf}c-Om6|;zk+xs0gisC1|0o>?FRM>*e~Pw>O+VJ?d%Tt zBVZ4fpGN`5@S_07aTM0C^C5n0|FJ!M4C2A^c>&-9As%d>vENw+;c*=MIN<(1GC?U% z06rDs$MCZ(IF`3RKskTV5+2u!&`#{%5&sn8!TdT3a9jt#dWSf#1M|;cqN3#_OQW%-wHU!^CaL{PEG{-p95W7 z$G~-o^C0~55FXciaX-BQ!eczY0373a67ZD}|7yTr037>;7XkkVgg+T@+d6+5gntRb zV|&#EIF{#^0Ur$UyaG7RM=*SU2>&XCM_q3G{QVcyT?IJmz6LnzVtCa3HH1f99Cx8E zj>A#+b%+Psnbm+}fA~uCLU=51Zvj3X!mj}w+fT&rhwwNqeHd`8FW4To zf%vgK%m&;xzH)7!)4-m$A^t&tzXLd~|EvWZ@E-t<^Yl9a$8_O*8P~^f zzKrAKKSDeh{|3OZetigd4%mtFH5?ya2!1Yw@Hk(?@%gzBKd%2^y|@wL!FqxI{0QQ~ z_UB{3dq6zRfG-6c^YKN%G2c-a+qa$&{!b8pU%+u3j`@i5N6bf@KNdnf*gj7I9POzG z{1dR}r+{NVVmY4y;c?!C`H1Zm<|FP4U_PdS?na2;)}CNKehT5y|2V$F@f+r2CFo*4 zUJW>gM?d@(ba9-3;a`JzFn;X!F??qj7ZpPMxL;Wh__+|?R{l3Z_%yKRZ-8Sv^BLgS zKYkAQj+9%RzwAi4-3<2pSmhS`wZDTsIRCI zF(3B>d?>_^bV}H65aI7zB zfaAOp{m>3{ab9*JgyjmyMc7`o zhj>u81K_wWdJy1aK=&f>e+l3=KOYS7*ve-%#DnD%=XZxdJlJj{o&)h9j^(o>#M2YP z+x&_BHTts?#Do4k6!0{NAIqB!$9N8dco6Rlcrx+e`bJN%=SYYL%Ny25ELYebj(~WY zpkCnm0O}S&_%0URe2548$NK=!hj@;Hc+fvbTkvB5?+NkXycpAq`_!0TY|k;hX%PQ~ z5Pw&|rviQ>q^kjNoR4<}yc@*R8}JJt9<=}WfMY%C3GraR&m9M@IEWwZIUaDVUwr^aKV$tudrpAx z7=K^DaXfP(;61<(8GvIvnSi4`{Q*b&vAp59XaI!AaTKl};=UfPAL6`VAjDG)_74L5 zD!?^1su!6seofVr&(}p$IuVi5Z=`fsoFn0%o;}V_*V!0hViFq+~Qxp zos7Q z#Dn4Ypby_2;3Mjl<6i>lP19d<*nRYSe#zGF8Q)jG=a(Y=o^d}d3^ni{x)|Tjd#C;^ z*6$f_tKaiWlYYjuCN zg7Aw0KN#>Pzz+eu8SqYkdxq0zw*OGT(-@$OUk>x$sXy6(`(=#3DFXa(2;Tts5r8iS z++W+_ubTk(+YrW^0r&T+7*Ekfhu8}Ky1ivaHmcmv?w0ACDvcfgwf z_xnWtx*71E5T4hN`C|KfdGFMpG{F6_1b>qaxZhSVUIch=2K043;J0ZT_^ZW$Yu9>= zhXD821o(^9fOB~NeClrj=kR{n;x$9QI4u7^Uu*N6moFdZXDjf3>@!Dz)xU6UmF1T`(plT8{qyP8RL6vds&n1 z^y5kcJc9%3t2f{rK2zf(0rxTfsub{kIxN4;0o-3hVEpHRbNqfk-3YiphUf6B0r$f) zz6o%DtjPF30q6LKFi2NB-6*pk{5>qj`v9KBfWF26?#IDjoeen0e~J$8Kc9``_vdc> z)qDtlvJT5H%K#q=_*%gIxcTcXfcs-e#yjZp#k&5QGvgV6XEUI$e87kKAiY)tJ{<7* zfR6xt8Q>!UUki8+;9CGc9q2an4iC{2i%`iGrkb;0tWQ80`Lhw zNU!SwKO69^fS&_+=l#_Ow%=c4XCw>oLKVdOTL}1hfY$@=uQl=43jx0X!mj{)BH-%* z_ty~k>#cxa1mQdDK@Z#S?=|z+S%4QYpszx}i+zw@>j9qx_(H%-0AB%kDd6h?j{?3G z@X3I8)`K&)e+uAPfR_PY2zWW*^?+Xt_(H&^0=@$93cz0m{CL9sHev(dl@NX#;8lQk z(Su7)*EGO~0$vSx5#W~q-T-(F;EMsT1-uFHI>4I&pALA69_+H6GXPHm{8GTP0j~$V z6!4jV&jEZE;7b7a_X7CqRe;Zi@V^0^^TlryHv|3?2;WW*);V351KtPlD*zt@cmv=S zfL{stJixC4d>PuS@aqBJ0Qfw>w*mfhz`N*45!-(Q;6nkw5%7tC&j)-q;5Pxj2=JQ$UkUgx z0N()k0>HNc{!75S9HKt3{kH%<6!2RCp9uIuz-I$~8{mrpza8)w z*v>lvPXqiez_S6r8}K5)7X#h^_&tCx2K-*Yn*hHL@Qr{k0er7c>I2(O{cq!n|06qus z6@WJa{w(0D0e=qgO@KcScxq?$f$dxgcsk%O0Gj3WwILG6sIUDfzA$$?wzXN;@;J*ia3E=Aie*|#W^?k7h@IOHK&47OZc)Kp@1E=ec zfcF7>1K?u-{}AvBz&`?f9^fAXz7+6gz}Eo&C%`uY{t4jij#3}k{yziW2k?!6j{*Es zz$*a%3*hqr|10220pA4p8o>Vs_-4RA1H9eQ>I2*VIpBQ&-wgN|!2b?-1>j!*J|FOZ z0KN?HEr34@IF~Emck2QFCxqV$_?LipK1O}ubbSSQ7T{k4J`wPL0X`e>ZvejraJF-s z{;m=5Zz23@z_$Xv3Gjado|>jUu>Jo5JRR`=0-g)FZwr522KaXj=xbh@dc*O5uQ7gE znx;N5{)2k!{aur$J}{orM!o01o72>L#@hhiu4@eUe;bIAPJpxD{PVb3fbR+M6au~% z;Ijeu?|tU@7Xj{{0cLz9;QqbgjBfxuwGAWc`V?@sbAbO{H}yTIcRwGa_i2Fp&otnW z*?{j4;fnxo2Y3VE2LQep@B;yF0=zxo&470RJf%BZrPkXI0z3_H{~mb`mkszK5WWcT zj(|4+-U;x?+SPiz`FrH67cSTj|H5|XO{l18t@(vem>wm z0bd4qFTmFVek|Zy08a5|TWmHk5?{7}G80(>IiCj&kk@S%V&0{j%fR|0-2;2Qux4e)J% zX9J$rJ7&NCj5tQJ0Ur+GivS-1cmv=g0bdMw4&bW+KOOK*fS&<)>TxkUa{*5WJP+_( zz()aI2KZ>e=K?+k@TGu{1$+(Q;{e|b_?dvWEu51Nd1GehlE_0j~zU z0Py*MPXK%w;AaEA7VvWb-vYRQ4-VVep%4F6_7_6<48YF=JRk7$0j~z!KReCw&If!V zgkJ{ug@CUG{35`&03HGS-+*%x{j$~h1Xfq+DuVD?fENQ^2>2wx>j5tTd?DbafUf}D zf5r>ju>tVO5PlorQvmPMms!ew|5=Xwc_`rJ5Pl-y7Xv;U@Tq_=0=xq7m4H_Qz5(zm zz_$TD4e%}}vbD!%YOd2ExwSudvrk7CQfJcGor=W*K>r0~5b>-8lM)0rUCVV}qxHMACzXhwlW zf%T`-3i{>nNCG!=U}XHHNL@`dNN-rj~eTEzqXEkVb)^4ouv{TvyFc2yRy70uQF#m{a$iGKmHdp1b<9- z;Bd;CsbnbSKxU+pN{a8c(TTw->L|W~3^S*UANy54jVmTbh>?jU#SD=ll%>pmk#nLY z^pnEgulMGW#l?AKO-@eii}K3q3L$py*P&&uU~NI&pc$SHNQl9*{PLM6{rEQdld4A~ zuQXa!S6(+Wk3Q!V)D5VJ*4C=`WIg}%bY_1)nh2YpS1>R#W>RKOj+l;}BPph)Kh+yK zA%f969z#g-D#=y+^Wk};>7C>6a+90shouXm6_ksawRL#~gJPbq%;8c)K%(D=CFS?l zH%#O6OU_@z#zJ?7W`3I#uY>;UxOArFDN3^6gZcWF_SVV^*LTulYJ2Rk*#XkSDAyrS6&aB z*Gew36cOJ&yU-BPBf3aJ2LVuY|#a_Z9uJJr{RXA?j+1%j;_M zxZtY))Ja(CZ)X4LRnux{$P_J&aGS9m6$De3kZST0x_N-LA^9;3(WS&2$)@z3Qc@C8 zW7LS}#$bfQW2EAGV~i9xY-wTG5I^LQ+K*y#y4;7yxI6@fL55ho6UQQAu$?n7u#H9J z1d@MZW0$IFrBQMC(5fSu5MQ)@XAK@uqc=T@NY-OPJ#z2Xsq+8m$(Ib^jv6LXJG;l0 zJ=?01QYv|f8MMv2VZS5#B(bp?I4L@%yh@nOB?j#uttw3jI3yC8QZ+qNQeR(1O(pkZ zK9T2FdJ2`3Gim&29vu#j(A>VJlKSWJx@b)?%@y+|L?-YYv0ymNfMbzL4690YHBQV& zZ)Yr}-t^w+Xj*M}pP5-P?b0YsnWs%B?~bRzO~N`$t7-cgs+SZB>Y|s>`avD}++WuT za%qQ}K^k5eiE0g%YHw93F5gJTaca72Gk+k}f!gTsfmC}kGkI1lc|I$WV|b57MlNds zri$Gqyxy^6T7$}~YNIvYI#i~a={y|S^~2usda6~!5^ZUD z#mfniLa$rODHv&H4z3et4u~i3xLMw^htQP$IF&~=9dZTo`+3R$>gKr*A5uM?*48H# zmrNDcp=r?BT8m^*=PI|e#yK^qx~hz(2GeLSsdPBSWR!Bh54N|_(o%L#!2rV!>aVM4 zD~HOXV@x@;A3=qhirYZH^QSEyZ()P*oXmlx<&!7VxQ;VsR0E)3oKS4A&XJCxpZ z!Fkb!768W5Lb103%UgH()KtYb;y9gFD~AwsO0+IAp*m)OIRWCC^vEb}+D@D4O@F8z zoZ)ThjVGUQ7h<#z+*T&JEEDe%DkK@fxO{4KChhrFRu|Vqxy$kPti*=F$P>F?lQl}5 zR_>g|E#6-B2E>#{2`*AY(^%Qsd6Je1u#`_bGz&7aA~ZiM{k3C1(Jq;{W0rnKZp*YP7Bv7ww=F3u}eel|QNqG>pPSG~2uSH-PtigshA8 zH?7~um}`OwlMEho?7BA3pQ6MqtqDE!4z#^FJk**|Tv1(@SZ!!%3!6ChZLAie*tPyL z*W1KZJ+BWlJQ~YXX)MR<>B*c|GUe@a3v*0m4=LHh;~DF6v9yr5bCO5L8OBD`Dr01N zRrw{;qj_WnAJp*g;U#Gi=pyyC6O_Q#2EFaaGuy0WDyD0$YOg6i>CL0ydNoQo|<4U9=yYw9lO z4YcY! zH;Ot(UYpOE!EHwVgb23^LT{9~norFNpVs$ebcag~k00B(kdZox(s-Osz^Lslp{Z`M zm8Xu$Z{>|z8XHCYNKG$oFl<-D8#7!QO?lmY5~tBy+i7zeQa^*jAB~#cVCmQ zpt9$k@-SLl8rbg~gXKMhUA!_BR92p?g|7?+%jHJ#&a~F+%nVq?_A0Yl$;Jqb)o8sG zos2!y*d}#WkOuw6y*vIybVo{cR-_X(EhwU6 z;|kM`q?C3WXbq%hTD6z9@Re?8W{sVY#2gEHVrm&(XAb6KGrGC;8`|y6|JFBb2TDon z8<#{`id!n@m6Y9kL>0GupHRR3h5B80=SC(kTE&`muT7|>MnO7)vwP>9ZS6Qe-@C2D zym2%qZi_eQPRQfW(-8V800@5H)W9$sLVD=snkD^=Um z+t4t=B5e#3q5m5cp>6o`DOG&WOtf;+^vS&Sz1{a9=5R)d>&-1nN`1dIDsry061gK* z$DeU-sh%*Zf)l7ZKZ@#chgm|j?U7%dvs!m%eh_!hbvA;p=jgK1P2 z>}jaojB#XpTzc#)y;PDcO=-dDwDS-dFD>_|c2jPjq?FLs3GH^!S!Q!n!=LFoMAcH; zFXWa(%t;v7Hh&adu?W#yeN+HBh|s zLx;!cP6s|IYc!|-H$|XlkkAny?_>hCccoFjeV`wo7>w&O{?AWdS>ihe)ItWWiY9ug zpl)E@^y&(}6@qpLc!?ozhT0URz=lV;7tnR7b#@xm%2ix0f~KArdx?pSE4OD$!>ef_ zV!LpL$-`X#@T*E4FXGqr%M_j-6sdz8@gpNX zq(YA(Qd3hkZ`TEDFg?SBpKD>ZCr}bwr}Xm>ZpM#&u8EB(p)IAls9zGPXVlv%mI#_~NsLq}+0cb#$G z5`?Z7UWKrYQgb4fycf*(M_a9P-xZHc6CPrdrK_xAjl4=gs}gjx#Ex|!EiEVP-iv^- zW8pk0D6guEPN|`L$-OCXaaCQOnw;=RBv`z{!$CX-X9tSsy^+7Q zZ)uO>&ob<^Ex}|~jHNlbH>-r@7W&L5`w|~um&{tQeHB5k)>w0DM@;jN8U!`vnjx#W{i2=$yy$fnZQXcxz=O&?kG!IS$i&Hy!p@9x_?NG&~enjVF!mWy+2XAy+PVAtoYXEfxjfAbw1 z9Z7ibHlCjd#x|8|^R5!-S9lJ1CvVVm)hI44_3jAs2TgceU!t7UwC7+ntulyhaTTWZ z`U2~zSGBiiTWukO_`W4!iX(cD$2BeA$0Qwb5TmNB6dJjtj71xy=?In3*lGM2H)~RDuxvvnxlEfF zC+uH(JI*ad1h;p#6cKr{g;pAFM~$oLUk$Bhr* z`Tsm*3(mc6Puod1EKVq?=ud|oy&tup(Mix;(c0(aYTny9lj=t>?&In3|CnQWKf9J5 zz05xwRZ0&;v+j5%7NOmI@Y0XhIQ5pcJ|dW6(!yme}O}OWavW|Y?CvG0)pSLp_VXCZ&Cl#{j z=VoX{SM8)hg-Z-++i#P{#(eS>=VIQbmN8jK)S%-D1(DKewxbi&4pUKju^+3I z_fJ;1wegOPrntZuG%34aAJ#ikGQFC95>5Y}gVg^}kH&CdSFEWMkHoZer$)hc_GNT) ziFZ2(-QCO;gny@srnusBK>{;(f%{@B+QEEWod>a(gP_T;Bwt7?zm&9stRJ_Bi#-l9 zNLJWWpSL$zJLV|5HLg6?$zii0c{bTKb#LQ|hoIv^7s-tAlcZ{li;QMeZh#XWv#}Kx0q5jQdes{z<1`pVJ#= z@#ADls4wJFV*X^u|K3WXx3TLz4=S!37FI;n52UuzZL#o+ue2%eyPP-Zxq{QQ(fplw z#;w2in#2=Ky)lS)_>*pc=f0b&qcQKa+%&?AM7#Gb9Els{{JXc4XX564znHtzztiekry9L3fuVt^=F* z>Cz@g!oA@?lI3`Z&9J2RH2D#Yv71+|x8wXsX^QKf^Jzax7*X!#?F8>qsohNP@h>zH zYH7hM)=mV&-InUUYOjN3boDFqu}W*Xe=L|vY|r?xrO^(J-~Hhe_N4SvV$_J_abrhq zo%SVZFV8*q4Ja&i7fyZUV zD%);8nDM5$`d8AD-VaV}>V?IBzG*Vpj?hP5eYf)<(t3X;ZyWMrTT8!LhG({oF$nEd zv@}xil!Gma2+e7!T~aro@rXp`u0F-}+x<=?R7F_e*`J1$`iga+U`s7vfSmPN6Z zH2U$39Zl$tnPND^>kPwlv{k1LG4BkMy+h?GVFj;MEi#t>%%`Q&y{CNN2pw&mHaT=1 zkV-Yr;e#$6E1%go#w99Y`Bg7$^!XHtXq!E>yA z$4>H6j%O4cc%-p>{PI_&wYzd#5j~1;g~j zRkfDxUlA7Af|g->Ci}FX<^3#F4nMBVdSgZG*#F3CS_#OZbM$$m=q-=ljlmuD>bv<8 z6wjsnduW0mtArxwc__k9pynx;w@#BYFyg)6uJxLn@jMpMi@mhX%(-G@ihUfAIH>ny z+P*cJmZpI%jFEP>8Mq{q#Y-|YwdCVVYM5>;$F%19M@t7Y#nmbA*k8~%m~(i0M)DLq zpv8V`g&a$FZB_BZS9DKA0co~v4n8Q{fCx;{Bbhwrpprzb;E-rl@g%;TZgO5>Z4EUN z6tcWFGO4`o((+pVp?B{oO&o-FuxI29oWZ~HMLVLwZY(|KMg6nb+kr5Url=AI`Mh24PsEdqQru<0BLHUIjXMIM z_b=3W39lxqdLK0F+cQeYk)Hz^H|B-ug}McPN46i6yP-;Y1fA{PU#^>Fv=i)Yy&K)* z4*4cFg5n()8b|R+iuz-JZ&Zz#_&l07Wa6C@uTjq9r$k5egN_Sp)nF>8AdiCaC$9u2 z*cqE!?YdK0-o4i?4QN^zL+|YNJFbf}8WDBmd3%N}{A3GiD%6qY@zm?l9!i4icPx79 zXncX-%(QM7%r<)0bP#dQgRah)1SMvm+{Ev+ai1JLw($ME7M|CZTYO$-P*qx|MHFwe zg2#E#$Zdxjz!tXgXg#%6_I+u>OgEz?z41HR`@N}H=Vm`@(?Z(|o7MUu4fgG4Xi<6M z8Eilgj_YF`m&VlUYg}z0FUUjlxSE-`kQ+D5!L!*EVf>>L!j^LLz5TG*$ibGKETlf1`)#97u(z6F`D?b8ekE$ntiZ{Mq-Qkzp-I5WrdNGRi|dN{xix)j zruB`SKjECrjIq&5dWvZNv}qOEa|uEE596QSDW==XhkG||N)hE55l!=AA%z0ti)cl0 zUqc};GeH693Ey!H(zj|_U9|5R)FVcLp zf4qbJWdgszf$!%Irhkru{j&xAvmNy32>fdt^ydlu^BnXS2>hQq=r0oZ6CL!I2>c5j z_)7);MGpMs0{>cv^sf~7#SZ$bG#|^~TnGKNn%_~sa{a%~fxl7E|Cs}Si@?9$fxk`Q z&vW3XQdqvwe?NENcNF+HIPlX1{*4a&J_3Kf13ydP-{im_De!N0;O7hcUpVk53j75Q z{4#<6O9y_Pz`wMSZH_Pur2R?^2`L{Xnmk9jZ9r()x{v8hdm70(J{~ui1 zPeFf?gZ^59f2RX~gTTMbfxk)M-|fKvTHr5s;HT`7Sbpwt;I|X__d4)93;g>W_&o$Z zm+8G!vVvFJ_%}n~-|xU5D)1k0;O7eb2Oanm1b(9fzewOe=urPE1b(9fzh2-!;IFA(g1#(}?B;IDAtFBR;6)`7oV&|m4mZxZ+~IPlj9{1+Yg>jnNx z4*ZP*zsZ5WMc}{ez~3hDUvc24Qdqul{JYA5-%;Sd=D<(We5`-JcHpNA`fob$2MGMP z9QfG+|7{2U7=ize1HVw|7`tSD+{AGgvdIvtY)uw)D8vG+4U|F8p}(`xd6 z<-l(i>}M?5^lujQ|Lnm3TF~F_zh2Nk*nvMs;2+|^pD*xvOho^Z%Kt?Izmo&MQQ#lyz+W!#4|Cu*X+GxP zGY;*~YC-=f2mN&d|7Zt(vta)U2m3b({O%6=TLgX&2mV$e{XHG{d+npjDdyj^4(V?v z@Q-uQ?1AnN%@8`hJ73?42z|R--2RiU43i{7E zl%G;Tf3SmowZPAE;LjHLLmc>Xh4eq~kpB6C{>cvdiv&K;N$Fow{ZFI7KgEH+T;QMT zz;6=x!yNc)G#}fqYzO{2fzNx&6h%`0+aT~qIOuN_($8xG$?9(r_%ArvzfG|JMF;*~ z`zk-7|8pJeZzu5c9QYkIAKU*?4*WEMKiYxcN8t0C4E;;0{A3CImmK^*Qs6f^@bd-! z%MSdB0)L!C`pX3VnGXCqfuHZdZxHxrIq>HR`~nC5LV-WtfxkrHPjKKb6Zo$<YJq>g1Am?7WBYS~1Al{{Khc4|N#I}T!2ep{U*y2|H%_qqi#YJxX~!YI$bsKk z;1@gaduTr9-y{cqA3?vwfuAMtOC9(l1^c58{4s+5WCwnsz@Os4FBR;6)uH}Z2>fyf z{W`(^iyin40)MIlf1Y4}g#&+qpkL|0Uo7yaIq;VX_E$Ubmkate4*Vv8U+chMBk-p? z@Yf6c84mo7ny+6{`272&4*bo6e!T;KtDryAfuFLU8fRkrJxLBB}gU+17-q4`*T<~i`| z1pONv_zi;ojSl>|g8qC5{sMu2lLLRT!2h*F{w)>wyeCZmlA3?65csz_@K3CiWlqI`BIR{QDgE zX_}AaXNdzpUC_VZfj>atKj6U67WfZ3@W%-J*B$DAp}>F0LBCY<(f2DDDk2>hj73_b^fxkfDFLU587VLlAf!`?TKk2|cC$l=>N@u-zezwnQ;1-RR6b3(ErSVzf$0T?!aFy@Hac~ z*9qzWhXa3upufd|ze&*lrvrbBp#P->jT15`QD(Rlk8Upw$SXg;=I|8n4W z7WBVy;P(*t-#YL!1pD81=zp>V{yPr*Y{C9*4)%``_}@A33kCl74*XIf{XaPHD+K*L z=u9$2kyQItFYwzq@aG78zNaNw{rLiaPY3=Y&Byw)w*!BPpwG{gN!I>l0{^!T{FMTq z@99Zaf3=YQ1048k1^ojZ_!|U%dk6j|!Tt^o{4Ik1K@R+Fg8sn{{JjoT^$g43ArAa@ znveYl-;s)XY({kfLu_QjH${U8DJj>UPxsLO&MG|Cn+lz2h-ec7Uiig?p%w zTXZ~p{*#~S>bHlTb@*QN%KG$m-1W&(aUX*Z92T!W^nv5F@y+n&=S<=hkr? z@aE@23eMO*hGRBr`@8D+{1+owN&$W0`Oo+w_h1G1eq#U4Bq7s(ml9vn-{)wh$mzq& zkV&bbFD3mVt)HREtbduHe-nK#=?~WWhoZO+i_qN`{#eaF1Op5FCoTLU&F^jS&GbJ? z-%IJAFZlloA^qzt`aje9r-kVAbG9Y_y{Y*}gz$S2S>kWhe6C-3ndv{d=NfT`ZF7G4KXgp&Ru%|8p;Qh%03ze&*NXKI`J_gM4` zwZ55u6g=))WYJ$Qv|s%EcvJsXi~d}#A6EWewCJZEqAbPsd%mFmgGIkFN&Wv=^wR`= zetx`}{u8KOCpe-1TAidmKPOZwe_4Y50zrR*MSoM0`e$17^96l=e!Q9f*Dd;~UCAkQ zh2_8hOxSFd-HJqi%LILXCcLTNjhshtg8$Q#)IW;&lK&e7eSUttsb6Z*&rMR_zbDd6 z|3X2ZpZ{*^^O<0R6VhLnq`rS9+tgns==1a6O?{pdN&0h>)c5CYB000a z9U<+;fbW0alg{a}{U*P24;6yv7kII)cy-czbN*b}Q-w0ee@1jEgDm_-ntwcs(RxboErA0qO>*M&5pC4xGUvJUhr1j4VN&inR`nf{- zd9G&aKWouHsh1MY4bfk2(XZF~KA-C^KYz#6Z?@=PrS(q^(f^}Gze&*N=f{}(1++mx za6F;n@V)^6q_h$M(w&?#ZN&OEj`sso`pMN&>$G4Z$|Bcok z98&(r5?@Mxw${(pd~SdE{F>bE~g_J7;nN~TAgT4Mh{!NTvW z`JF=ecU$-)HDA93v6$nJ+lepbe`oyzT-g5qK}i3R2g~WN(!zQf*s6ah@g@BXLI00} z{!cCX^R&M1h5~&v|F>B9PiTJEfT`Y?{`=fwf39HvM}qz5(gp+DWR5@AC8%vM z1pQ_~f00H18?CQj0xy{9zty5&FX(?F==Y@`D3#KmcAS!&5-6)TrhaeYOX;7l^>a0! z+s}=H{_Ph1xmut1x9~Fce`(QQFX;b8(Erq;pK`oX4fB7qML+FuWhvHQ{!Kx%{knu2 z7|H*~Y5lV>@$}wIe+BU+|7U4^^#5mqetX)Wk@ROLnf|uKm-HKi^lujQ@3QDWp!LJr zuiGs83kCfz1pOX|$?0FE^~2h)r565Yn%^}f{~jd1D<^|Ac^OVIz+qW`qk534`T7X9=iR82$u zZw37;510L)dV)%LSo=Af_>%vt1^s^u`fpnFtF?Yu{dvuzze4Mid%TO=@Ba$=ZE1s= z;DqvbgVqnLKjSU@<(jWwv=vG#_&{&Q5?}J)TEYJB1^eH$=zo%||1A2O1buG2&H8%~ z9iWl?m(o|IvOgx4-g8*I+7Vw$f69@G{STjUG4-#r=nv5PVeQ8pi+*Q8e=kA*-xmD~ zwSHLn8P-Mi-yF^Fiq4|<$}2H&DDfr#WeE1~E7*UfMgM-SPs7u=%gnzf3x9>?hpk^c zZ?QjDu)nQf|2m8Qdad6t&iX|Eb)a)jQu%GG7j9{K9e0`j&$jT7(EK5B>_q*Oi7(|} zk&ylah4jy|=ugo4X8xJ>KX2hr)_m>8AQfi*KW(wUUay)S{s7Id)amzM$OUoNYT`@&Yta77)uGvc-HB|LpHI8X_TQ`Z^-G{>>NgW#(qAo< z-=2d0ezeae)xULGn6fACGWEw>_}^%LSo^=&!cWy5XMUV|V*hmq@ul=P3;ypdr2j-3 zAW7*jOj7?i;!FDJM=4L~YT(`fe!QUnghjtm>vs!D|05RtLamSO*9n3??=wm1e=|w_ z!-y}Xzgo~gQP7WA^fzn$u=T6+E&6k{ev$Sc*S}0bf3I{o{iS;1P!Zz)A1wNt1pWSk z{v3<`TCE>e|Nd;@f2sNVbLPTDCd6GI5nsx`TyE^?TCYR1{{{*6Uq%xgsr-CDKoQ4l zzW-wC&mg{}pLUE2nG5<^g8r8l{rrK7*e)dff4AtDY5gsrf0CfT2dz^{>7O`A5f9dU z|HUvTCYdm1fL1f zFD1UDU({9U&)0m`&ldErrga8MKQ&7cM}_ELPJBs!i`K{Xd$^!~8%^zxjrj@@fpZ%%Zz8Rhm%mYh{@*P6D^5}xVfDW^jgut*y{Y+O z^_R!FlK)bAsFY&;9V^)XiADeRla>9yi@C~sXZr6$i+(B(HtAaL{Y~-4KW7U1BdMR4 z(qBDP5%V?Qe=+q>BfgaWX01=z<6WG8X9@cETJ)Q={;9sOel?0y%AFSd4!xA6s9zxH z|I4EPeUj zX`CeKcREe^JIw#9E&M*3ALhSTi7)xDUhv;! z!GDvfpOx&t*kZq#e^=20wZw1G{Lx1FGxUSZ&q`zL! zpDO5INd2s&-#J^(e>44mw($FE{tF@g`-u3G{i*3nX};!j`&T8{e;4&rlKq>t{?Q@& zM^FKk_-%(NVi>;@@g@5gYWuPMzC^Hpo<%=BN&Rar`pX3UT0#H!7X8z;K249}E|n`W z=v|BcYC(Uxpr1wUl;r;!t=}~!<-IfYAF%KjX#RO2{CkKmEJ6Qri~bSARl>v4KhvV0uJtkfvjzRVvgP#GY5g;m8NpwQ zBJTRZqF*TJUoPmMlP&A7()xFD=D=m@pOr29f3DVV(0p!x8U+2-#Fz4a(nuw0_MfKy zcNTuV=7+UEB~;EN{aZBuU}Z+|S0^X<{EUU)sQHH(1k?T}h`$%b#%m*ZrSf7kUF$W- z{=0^tRQ=g6*VBf}_BZ9IfM@9N{)-`#axw8G{bsF?<@Z`a{}qdV>gh@`%>M_CknQiK z`C95fGVeQXS;!FC4g8qC#|M#RL>8I*~YtFyS^uJ4d zNxxdqzgf`VlXN8gep>&SIPWL!Uwv=UpDXAu5cDs#=#NiQzlOp|>0d19-y-O5wCHoZ ze068%z-9XXV~hR@thx9DeU{iqj9?2n2y_PLhAN$KCD^^0^+_Wxai zeqWlPN$Gz;>%Sk9^4^*Hy@@aB=VqvoO`yM6(0{|Cf0>>*78v?QaY$KZ(NE>gLb`M@ z^2*=6g8pP$V3*SWw$}g5&{wg>z+&P{>Ce~tjXEg%e~F;qbF8etg5u@tIQ0_zHTBbo zFX?a9`dI!R5cHq4=zpU1!^;1o7X1$Wl%=TODCqYZC#OG!;^pgh^%DFw)8Cc&Qu@=i zevxLf{~s3gFQ5)i%KzK7e%ScqP78mj=8p_4S8q)F7ZP8xf3CJ4#}AJP_J3y4-=g)m zgy?^2(O;qUeLnl|F+sm;zU;qd$``)O^qcj!3-P7&H}YmXT^ZhA)qgxL=-*(`??LI} zD{TCBokf4W*2nVuq@e%0MW4$RU*`PLO#dqu{VhWJpAz)DyxEzDIJI}+25e;pRa>*`F&2X zf0;%93$3qT0zp&%R~G#a169Zj&|fL&r&0$ewLhnlU-{~#UV^`7`59&5mukM*ew+Lp z;!FO^*7jrhe@U?aWsCk*TK|N=a`nda-%5*qp`ib=p#QZ+f0-ryrv9i2vj5-K`~wa9 zP5-@T;cw9V$szo=i7)xTS|~rS3F$xmY}x+dR4(~?(Jnph!;!jkCB}zMHc=<%^wk?dhbm8FCf0;zd~(4j-THV>~FT{FV_0B zy%Kks`hT?OFV_0le*RX_pI9jS?}H@Me=hN*^sg1t|DK?~2X*kAC+7Oi(Nr$@3QPZY z7X8jcl&3KL?+f~OTl5RHekWx{@K?ned`vmyeA)gg&DSph!Q>xEe93>g+WsQVBTECAktY4K^ zW8hO3{SGIqkezilf+JT>ifl^gETv z>2KEhVeNMZ;!El8JXBeV_3vvzKgXiqmii^W!pcv!ML$E(|3=XN*rK1V_4_I_g1>Eo zck0at7X4g7f2*KhRVw@c9IYQ#|30_y%aY`uMhhgICUg8YD@p!b3x8gc{C6$hoE0lChLEo^~37lMZ}l%8%|M{`srl- zy#)P?a#???)(>lcjwimPzfkLA|F@5z|GY)N8;z^@I!wI;e@*{CaItLvNt*vvz*KKc z{yoH(>|ZU|zn@_L=Tl|v;7N{5AFeLVQVotDwKXpnqwFtbe}NUlk~;H>Q3K z@g@CqzFCJZT)#L#(9f)t^^c};4PUwHCHQOV_aVNdpK+QBiRGugp#Q2xf3nu^7AUJX zrv8f-{UWW8_2(c#zju|K{<&H|Z2a?c3;!<7H|xLI{>>%6A|3LH2{?p`lrw&GH|2J#?BqPBl|7hY%>7TFDzeVRiw|||5^jBXZ>yIl|#IW_Z zsl=D`8}xxHEWbwx`nOy3>$Lu{A?5cq3;zzy4;w!58BDgTTqA2xsPKzvDm z(=cTz`oFuNf1X8ui`IW8B>fXC`YFSe{(Q~n`qNX;KfF#(e`ARvhPD44i7%ypzShV3 zbF84h(4xOZ>*t50{}&ehCPBZqpx=4Aoc`=mC2Y1IX8E6H;ZN55VIlTkOnk|ITeSVY z|2hBq2=@QOqW@r$_MbjOPXF7QuZLklDpaWoKBf#KzGQ#a2o)0Z|3tz51s452G%w*R ztpC~TQrZ4fHNTsRCirXG|AWQ;GHpNB-+qGqr(5)kw7xn1Hp|a2;!FN(5cCHK`b#bP zKhygCf|RQ_rv8H#{e^=5AVL3tdfERglgz*SEc~~V^xs{?m;AR(uz!ePf96ct{;f&+ zuMhDh{ndj0$%6h%7X9OBUd&fm`FYNw-z?~#BIut#OHTg;tv^bc5&Sjl&)LM6(!W*f zWBYZQpuf$cf0xz|8$W+*(a+|a<>|uuJ512ezf4a5O0C~LNV$4rrhhc?rSumH`XdDW zPb~V)TK||3{SPhr)q;MGpx=46ocLr)_auI3Xf2%F}DLKkg^k2T9zv&7&{r$?6WM-hO-k9nCGx4SL z=W2ayKgJ9CU9Xh&OSOKO|HoSRS80A&`TyL)U!eJ6-=cpYtxNLtmU;>Pn)-_^`b}DYy=LUXMtB_d!CJOoo&5`v}E>?Q{o5y&W`t67>=`Yjz^_tA-zev!}yGGWZ zN$a9~g_YksEd1+}xnVkQV1^s6&`lr#lG+$x)pF3B!f4t_K z`Dd2j5yY45Uo7~qT(JLXi+;7%zh5OI_-pEKyH2*h6Rk`0rJDiuf&QBQ|CacY{VN3j zR|xi>@H1Kea;76Z%qEh#Fy+($yFh-{7e(lY|(EN(qAj+Uw*xue+#s}nSW;b+s~8v4`_Z^ z{b@^lDgU-;`?3F-A=p2`qQ6S(AEMkC{58wZGZy}O&F>X3)fF2WlzE9Hr#TNd-w6DTfnEiJUU-EyNVE=5v{$4l8_GfARF#o@3;g8mQv;3Il?^)tY z_Gbz9Um@6k=8dxb=+Cq0&(-?vRW=5H&GNU-!e5vqzw3NC{f(L* zR(`q=UrK+Okp63g^v|~FzpeF;4^pn)nCZXNqTitPGc=#u-)jZ^Pb~WX(fYa?23DB* zA6oR6X?+~OUMJ{R-6Z?JtXfHiwZHFJ__H)Wtp9kE_)`9*^TP(|!v1rfVE->}mhE4w z_09Tg`tJtfOZwGXAKT9x1pRrxkoCVvQvX`wOZsyK{rQ6a`3q$Ic9*F9yHULaf6ert zO?*jzvDRvpKRD~mY-i)^i#*EkVTr$`FEF~pL?5} z{*!7I(e$6GztX}Vqxok0W%~aa;!FN3)An};`|lC#|EER&Vy*AfTJ#qR<>x^`|Dro&|4pKOeZIoV z&qEe|o#uy?pC!bX{I^x`-@}6a-51IB-)ynp%>ObAe~IRs?XSr%A--gP2Y%QuU4HrH z`um7r|LYe0ceVa0UM#Ucrv57y{dBFLp{cC@n4mxSPT7AQ=v)I|$Ela#uUY>lTKFev zzFB@u`yaFL$0W&5y-T+LBF#6;pK1S|#Fz4~P{_aKg8v#U`m>Yt-(?p4Y9arg7W5Cg zTTcH2mh_wcZ%2G7{fmY4uMqTeE&8jpeqK;=)Em?PBP{wWv_6iXo)h%9So9B^q4d^; z=zng}PZ_5yU9b6Ee^v_m*DjX*|B2T3>%VV?nf`Y7$o&6kezs54uO@$A;!F9rUfb{M zv;8j#_D{9wpG40ty+7e%~f0nfzNce^29!t(xz*-#$Opf1_V-BEIDRYMp*v4ZY{zeI%s+ zS&M%9EF}=8zucm~N$`KOp#QN&zfkL&{!2B||A9q6<1FRSTpgav-zS3pcUpfB!~fM< z{{TbJ)IahemH%e`HE4c^5dNXWm-4S(+wc3I?f+D;f2h_s?VoSaH|@EQ_y>}W6m{IS zSo6*F`k(#RLf`&4c1lX4w%^Y`wttggf0ef1^j~9$zAsnq>l^vkB{2 g;p%*x^3+M|s-|S zIq&nJ)2(IJeeVT4==0vm0+*^buDj@K=%br4v;gLWG4%h8j&H;FPZNhVPE0+qdxNnh zr!hH|D`2qhf`~Aj$g$_AdXn;_C;)gh%S04o6y+j5>YYg61n`GQM8M;!0-j8>EX@(i zry}KeGo4CA$&1P-dCiM*TBN0zIVY|J9H5Hs)c&$?NjKy9CF2lR!t8eoWz}75{Si_& z(O1=zlQsP`uC)l7`N%^hNG1Ks|G;j>HgRn&pL)imm7YpMRJ qNOKiQr=6gEh2uZZxjQP|szG-YPX~McJuKLFPuzP&ywgGzZT&A2IXQm- literal 0 HcmV?d00001 diff --git a/infini-qwen3-moe/build/.objs/infinicore_infer/linux/x86_64/debug/src/models/jiuge/jiuge_kv_cache.cpp.o b/infini-qwen3-moe/build/.objs/infinicore_infer/linux/x86_64/debug/src/models/jiuge/jiuge_kv_cache.cpp.o new file mode 100755 index 0000000000000000000000000000000000000000..ba4fe6cdbb9d8626875ff08b49232c0729992e40 GIT binary patch literal 1088 zcmbVL%}T>S5T11F?@7Uf;3b|^uuEF%K@SorwHFaD1#cxuL#s7EG#jxOeFYz>ujCWx z%yt)-P3oZov$ONf_svc=$!GumuHiVK&4F9^%QOnGG(ca%KJ3FDz@XQ=anFZgnTpcA z=3VZ!+NI3eA03Z-UEaA^;ZoJa^#FYfeef`Z2DT=Qq5o%ed=tKZ>Nu=+V(N+AYmBWq zjmfF10tVYv5D|tGRqW-toW!Hq+b9a6Y0RTM2jImb5kZJi%vW)$cOrflz{i0IfQLl^ zJj#+JP7zBp5p%qvRw06Drt(Qz@_e4<@mwq_M{fe0pos3)enNjuGvhOmj8j|*v)`XD ztJTHUpCM%zeN+7wf_j2R$qR literal 0 HcmV?d00001 diff --git a/infini-qwen3-moe/build/.objs/infinicore_infer/linux/x86_64/debug/src/models/qw/qwen3.cpp.o b/infini-qwen3-moe/build/.objs/infinicore_infer/linux/x86_64/debug/src/models/qw/qwen3.cpp.o new file mode 100755 index 0000000000000000000000000000000000000000..357df4d28598153c0f9547cf489a903d706a5b5f GIT binary patch literal 1003384 zcmdpf34B~t_5TDYly$YDT-n(y>%%m-n-@l*ceZTkJ z^_+X|`tEykW4QeMJ&KC<4Der3;Osz=oC`~~2PPSCH@omFY=dx z1LUs&w~@aJ+)n;);D5+p1MVPy9ryo^T z;OEGrz-8ny;OEJ|09;NU2Y!(}LA(i=B>xid%jE6AuaJKg_%-sIfgR*u2Y!Qm1@N2X z-vWM{{1)I!^6vn@OTG&DJ@W4Ze?b02;E%|E4EzcCYT!@Fe+K+H`7eOClK&ET8+j-2 zSLDA2{)YUwz%}Hz1MeVT3%ry3F5undzXNuW-$T3?m?B>XTu;6MxRLxm;{C*K;3o11 zfWIgI1Morehky^0{}I?j{s{3;z(>g+1OA!(FTlsiHv^v_e-ijt@~41*BmX<_Y4Tp+ zGvv<#pCf-B_yYNhz<-c$A@%`ZBJT%oCI2VzU*s|0TW+e1|-+H~Mz@LpQ$%a8L5RfFsEF2JS;X61XqbKT19tcr5wH zfFsl;i->BLKkml0o^T;OEGrz-8ny;OEJ|09;NU2Y!(} zLA;5WBz_6_W%735SIEB#{2KYqzz*`S1HVDO0{BhxZvnqeehY9V`FDWdC0_;n9{Kly zKOp}h@JHl72L6P6HSnk8KLh@p{1?Dm$$tsFjl2{1EAn3he?$IT;u_-Zz&ps-0`DZh z3wSsA?|@z8_Wq{C;3J`6l24L*x$w|47~ge1!Z@ z#7BXTk^h4x0C-H_#g7uh&zCISVo)(JeT}D;v`_0{CwaA}BCjT11Dr`-1FR*VMVw8X z1FR$e3~?@T9&kQ+J+Ohik+^`^1Z*a60k)Da1TG?9OuQC&9ri z8{}^S-y;7n@NM#UfPs-_y;4Nn1Gp#oUceFLdjt0&A4%L7xF7lc!~=i_k{<*Nk{1IH zCO?FDDDW`y!-?+$zMp&)@B`!@Bz}l^1n|S;M*@!`4*`!R{|N9H@{a;XlOId`7_fx= zIO6fdj|0b$p8!0O{3Kv0`B>t~#8Zf;5>F$ZPW%LM9C18x0`Ls-Gl6H3pA9^RybL&z z{9NF9*U`6t|0#=@LS~H2Hrxx68IhR?*dnme-HS5@*ekI8=mTuuH{;?ICTC;tWTR`OpGZv%Fc{|fkP^4|b|OTLD9JMj+STJk%Acah%> z{2h50@E-Dei7DVZ^7X_Gz>Vbh0q-a825usM0Qh_IKM)@TK1BX7@Q>s@#7BUCB7YS4 z82O)pe<6PyxS9M3;FIKk1wKXoH{#!kPZN8A&yYV0e2)Bi;0xq00{=n21=vUa60o0q zEAXG>{{p^DJ^*}$d>ipq;CAwV1OG$*8gK{s>%ce2-vquz{$JqRU39KQn z12mB2AT;M$N`M`Se24Exk0$>w)Gq8oc6}XUm5pXg2wZQAhKMTB`du_^4o!Tkgo;aNq!gbZt~v&yU6b$-V02TuLG_p-vHc5ejo6D z@^0WJ@&|ywC;tQRLGp)y50n29*hBsZ@K5BA5+5V}8Tc3cJ&wQ4y!Qm~N%Fq}pCbPo z@bBbL1AEDzAwCOyj{JGz3&a?3~(*iXKd_)p-!$X_N75MKdqBYzdRo&4Xx z|B$~%+yQ)@{0-onXAsW> zo<)8(@f=_o`9$EkyC^oO}xKBJzuY737}+UP3;VIE^@+ zcq#BQ@)^KSlV47}0(d3)RluvsD}hzy)xc}WX98=;Yk{-KX9MSu*8x96K9@L;I3HL~ z-au>wE+B6LHj}pyTY(G77XcTOUrW3W_*wGniA#VtklzSwBVS7V957103>YK-Jn;*_ z<>YbV7l8@#n}A93FA={CY$yK;@T=rs1Kv#D0sK1oH;5~U-voY({M*1=$X5ctL;hXj zD&qHu-zWY6_(Sp^0e?*X6X0s{p8|hI{&V0j$ZsY75_lVVC-GOnUz7g^_*?Qdz}w01 z0Ins!6L=T--NfG!yNLG??**pF*8$g)Zvbv2zYlmnc{gzr@B#AQ1OGt&An+mbhlzg# z_K-gU{1f@3z{kk{O#BP*aq`W;C&-@!{+0YG;NQsq4t$!t7x)bMv&83!&jVi|f06hP z;uc^Z`Afup;#S~4$^S)s88|@x3UC|wtHAB#{|5et{59eZ;OpdX5Z?s8MgCvl+vM*M z1N&oqkv|z@fP0ef1sp-XH*p`}Nb-Gw`;qStJb?T_;6da;U@`f@z(dFn1s+CzIPrbJ z_mhtTet`Ugzz>lhLHscANb;kAA@ZY%A0ZwC{3!Wo;IZT%BbE@410GNQao`y86No1Q zPa-b`jwL@CcnbNcz|+W2Cw>Arj(j|D0{I!hGs({)o=rRlSVlgPcrNff@=3rj`T4*L z$RohX;PvE7h&K>#1h$bcC4LSVC0_=Nk$)ce1@h&4B4UEnJ6?*YG0{sZCxLEIa-5BW&o zzU2D><=D2Qa{%!`;z7V5c`@-|;vv98$qxe_PX0dP`+=j#KLGq7`G<%{06$EAB=9Km z5b$X7j}VU`ev~*Gcr5wHh$Y11fX9=6oHzz}0{Mx+lgLYnV~HmdPXV4vej4y}@=p-Q z5yulJ0M8&llXw>JZ1Qt}W#kiy=Mv8&P6CF>&j(&W9wAO9UP$~Tu$+7f@FMbyi50|8 z5ibEwC7(u|4!o57GT;pIPZKXEUO~JPcoq58z)JEeU^V$Q#F@kzVl8kM`E1}E@;cyW z$mbI05$6-@feqx1zy;(@z-ICmVk>YV`6A+C;I-t}5kCvOo_q=L2J#z$ZTMS?zt8br z6u68$2K+qv7l_M&aq=$`6Tq9ulf*9pzl^_j{C$P@z6$&r`OUx%@~;EGLB0a`P4aI6 zzfFD%a3%S7fZruwMf@J|`@kQN|B(12;E&0F0$ff0Q{d0Ye-8Ww`K`cTlHUgGB>xrg z*W|we{+4_V@OJV$fNRO`1l~n{H}H4lUBG+D?&Vv=Hvl)1-$%S3*iF8P_yF+t z-uN{{(!L{4wI6fqx-?9JrbM3E-3DenS22F3i&qRtK{2(|0e$r@HO%sz}Ly& zAiha_i}+vQ+vM*M0|#QfkUv@P5Jv#_#^3Zv{Dr&3BO)jA0kKr;h`WRZZ(o1ahF60} zz9B*4JC=vv3f>-mOV0A6MR-0Fr%jD4-P(zpk$BgPciwp?5?@mRcr|!%tlVHS0<4K7 z_8MKH!NOa{rXq=$3EaI!5{wQMElaiRcen69d9NqUKN2_LBJuT+WVkaD|BHL4JKUz_ z*W2M0kp%wS(Bb82?w&|;I{t9qO9DBcXl^l+1Y`-XM$EQI@HUV*h(-WkW{YG})xPR~HCP;whBXJi zIX&mWvpd!(3E*FMtc46l+b(xtYweLU!mEPO$J8~w-I$cyq?k*1c_O@gR1hRT*D6H67T;jhOY7Ed%JBGR zB($X_a>@bPF}9rFe~$FG5siTk%7p4a)Vqo9BQu!@(&24Wx}_VXJHBG#m0m9`eY6P5 zK3Z&Ono4(YQnXS)=Nyd5k#1ThxJ0kFj0spl0sUi>?VjZ!keA424g|u-SArFbiR@J{i^r4 zntrrPTVCFtS3(VgPN`0f!jWsxPKiQFcM|Q=P9?995^5L-Ik#{j?~gU0%z?DNa(OU1 z5!M_@Wa}$ON?#$Z30+h08!4FW8MxCSchJr{%7NJDZAqT(-l`Wc)P?la>+A@F})^ zg*SLcqFYPjJ7&ehPub(-jKmRzMYJ#gP1i8vOC9o2d$;;rjF5Pn$0wcRKlqoU!hjoG4Lgxi7+L)zPp zbYUNGS9oR}*DH&fcvs0=GE-U;0>nFxP)D*DV9Sq19BWe-Z>{gc=xgGc(Znq2(z=Ua z?9V80BpD0**PDw)yCjQ1aQPL|$0lP{@Jnd6-i#5qLlPy?A?3(esKm;|Gnv#d$-|*o zh+(a?Vr}Z*QDuT@XrF(|c$i>W;`fSwqC+%-Y&5hl{wfn+gJsFT!}wEdccu-g)WezK zBx9O@=S)emp0myGEb}`UDPim~REN!dy2WP5B zpr!j-qHi9oI|#mVI`dfHCPF#N8|{*7{`Iwl8V0kxu=E~|?&5_^Wh+!jx#kyALJb2U zCl@Xx%&7{6c2$ya%t`M$wq~1?FL-xsjwnX@4J`U3q49dB`C{~g;^o?}+nojP2n6p) zu?l7dZ`k04G?wayjE{;$HMOGt?{T&mn`F|IkUoo*u?w@w*qM3!%fufW9{_);EJQ7u$eLSHug}EXHnk- zT{#%t$5M;;c{LI|KGlD?Wc|oX(cYrSzU$@IdK4JbZKv>8Zwi*U<#%)&5h1!`ivfP! z$erfPP)|Z~>E%k*h3I4&LLn>Sr!`wi+tHck6V-f*`oHNIP9cRisc7MK8V2)8#bcUY zG8xbzMF+>bilrWXa4qi>Evr1IX(TomRDt-zbCs6DlOs{=#PI9| ztw%x^^(P#^v{GfyL#5QVJX9&;*esR?A&bSF+r?Ssl9*DBI~P+L{m#WK+o_n{vE_)T z71OQVXAL+;eN~#%p;~nlK(8nf=01v6_u&qi^(H0y0xiR6Y# zPYHMi!LHX!BF!G8=5;UBwr^}Sy6N@$(=*!wUZe5bYE)$|>f4Mpy<>om9sB3@l3n`V zXl4wPIh&0mo9kSqMRT3n88ElLFJL@$7kNhS^s*Lzv6MAuwC}g&xL=mzekjNND(imV z6R5fTMOtM=$yKDjFZq+Q^w58`mqnF*{}bN*-W*xj+P1S!S?QpTY_;3L0#>pGET&~i z>D!COayk@=FWC`^w+_UoZ;iyK^kO5nHvyVx9f)9+>P%VMz2&KFU@ta{)74tcHKuqS zE8vPJZ^m^}T<7tV;>nvNL?S#2)9*IWv{h>&CZQ9p#Wa+7tP}2Nmt6DD10~cTou|K@ zq&33b=e*~ZA3G(R=TpY3-i64QQjuWsAEfkI?@hkFr7Yz+_k&c~RfXZ_1Nzv1{}U zAyh|?G|_!{LX;isSS!L`=12`m+b-R z`*xttz4^3Vt$j!4_S!SImsmrN1m&bw4lmid%Hj2moL>DmW{=W8JIDQtbKJL96t)KX zkMhE6KTAd0ONBCjJ%{|xfZy5cubhaZUMF5TJ#o#P?dIBd_^OA9;@G;tC!@rp!K#je@9w?dh>N;7YjW!|dDyoFBQoW@B% zl%waO(3km->X`FS$qqRqSG>`|4vcR4@Cqss^$G`cxEk&Ji$Z;%BgMNt@s zw?>#jCRQc6Bn&t~u&tN%t|{1%%C`mO+v4V(t9&d>S_xNcHJYPo37$}M1b0d(N5I_V z<2$24tZl=+ZW1|5XbO{7!jvgtqtTW^h7Qy2?t~6;3N5@je+f-t(n@HCdOb#pT3+ne z7Hc`}?oO1j2PN#uUqVxuv=W-h2F`uSY2I}0`EKn*!7yRKq`^rwXXTnQq!rA}a`b?* zd)5yP+9*xVCo}v`SzQVROSwtpESM=nTEWcp3TH&+oTyg{7Rg zLt4Siye|ESx1+HYZmK(_u}&1M(@i30!Au#_3T7sh;Z;atm79cHmG-Tj(Aa9Ey4p=O zXTeMv`~{Ou5vKy#FL#%XM5hvPdhSHq9G%ws<8|v@f4Ii1;pDFT6Jr_daLHv8FP%8u z95D$LBm3w99}QqfFBhMoQ_rroJ<|KG=>)_(ej(T7YFx|Xu~h(U^xXnL&c$dq?S_Ux zmm69F+Gc1e=wd_1fi@ai20F*k2xt|kJ)UpUw-mGtiHcE)m}DIpzrN{%P#*7w+`GYO z4^(k4WaANGIm8=l-z+CtB%-(av|=MfDMDDIu;)1no;{EIJ(IBkxs&WrkM!NdSh1v) zOqwu!9A-b>tx*3C%`%zP(7qLIqls^VWr^Qu;$+l6}V*NIV=3mV!u{Ul=21H|y=P;|+tKrnVtmNwXlx=u_V zU`pAILvak)9g0z zFpAtA8wnV6Xw`6u4hff6!qj^sAs!RTtT!GjHZqL@$c9q(c`v358ZOZx;quDVqcU}v z`7fu)kKpXI{PGp|Xq=g|JIQVI&9>90z%qw+CfJEq*`U!o$z}+YR zvP0>vZMIOIb|~Gc%oZwThiZc%WRlw(*SV^}2seEd-i%dW&oa*aB1@>{|s zIwYKJSQT2a{)JUq3M-SpsXG%{s+Uz#==R@Wdm4@|jKqCU;5o9lDYeMil&a^!W78&a zMou&@uOe*Ae2+qIt>1?CD7FQQbxFG|c=bBx;4+kKI*fb%sruH>MzE!O+$}+HSr8w5 zV=02MJ>c+_B>gxv4l~Lo8vb>wZ5cy#YF+oLyjo8LaLN(E&07VhC)5#19u|}MsC3yF zj}-M_hcNTpHI7Nax2S)Ge4y6p0zCl`Teh{cMs>Ch9F<6JssZ^1lp~(JHp{!B!;G-V zK=)g+?K5VUHo@Kf4r}8xk|@JYW@cEQezWD0B}!&6-x54)Hkomp@gR4hJgp2Oe{{=; z=#GPeUx-3)@swMi^2^BBxZ7)w$V?naGW(!KJ!ikuo1xqO&t~45lX+`;=B;xxZ=Ezq zNwWyro?%Y~$!uPsO6%1yv)PP|ZV}t<^4PJBS07{e^U^(|TZ(Z0SE4rW8%cgx;`i@E zf{tB0+$i2rEqOx|`<7X(BdTc?^)139ZA(1Uz|cfw7u7rjf<=yBAB-M=AJf7;X@6N; z&>$*}?l>TLGhPx3F6%HSmn*D^1{Bf=Cp!gg3(l^fC5If{&{*hyHL^V!@g?UL(cK3Q z-iE-j)zUlqhsW|H9@1qqd5Nmwp*iQ%QTn{gi>z=1TzA4$F}6mUVKdUN6D*@ zPTAcN9XUvlqW+C`4!$W~-^uS?qg<7>Q95IN=3pufQ{z}BRanE^fqpC6CD;5mE1|3m z-m8XrW#Kls$ypP8L@#|Cn?4DsH9Fjwqi8Yi60lLMZW~@!%V=%$kvnrs>{*eM77P|W-rDj<> z4j^@sq3boO`b12_B|0RWy{t8TblS4k2k=&fus524MaET5Hmrny-jVf=J9V1(E z6G*GiFS&ZS`KPx9Esp2gZZ!?(kuauPre}8y&8Z7HNn|u_?xd|#@G&|)#xXgbx>Kyd z%(>X*jnuh_T=RE%5^5N9dC|hVyzbazXdsZ8ha1ML8{GktNGP|OxVJQxVbYqsH#NHM ziZUnK0nR9EqHW4LsZZM<;T$yoxfN@#pbAyy?iEZ`hsqR;oezgonvC`RNaGVZ|f=uiz{YL`ODE*-dOIQyi zIn609^{H7fa;8G$v{bg>J(J^)?LH+)Lk(${=rr9t!r4t1ChKpy2WZokYyPGyp@u=z z)%B~b=iFQ`7(!qB6{;nE)2};^VC& z#N~!@t#YshzQjmdN9dNcwT0+w66v0J5}KoIe5zvh&!i*5(N1kHPP1~$C*N!zujv>` z&cRlNq&hT7Bt}u+LjQ)LzsdCNt4-X0)b8DG*3P8T9+`HgX&<3hFW3B5FQJBk)vqnw z>e;kIDx_TV3n`(7fsnr|T*z!qyEsSl86OiLuW5@bvq22DMIbTKnznBMvBts6%uhb( zWOT@D!3wh$oRM!WsNoVF5^lGw1%1LmH5Jz22&eWn9BiF6p7%L%UmuXx9A#Q4%MM-6 znI{x9&yT8k$~C`vN~mFAo(~MgJkwRG^9`Jz&|@Olg5e6wJ!fI-*iktjLC3QMH;<^- zm`BuRy&qq97kcb>oC%ljT)n>zk-jMhr(^!EJ=!H}j01wPbyP)i%c-Tl^_{cM%Ut<6KC(8| zZY{BRBudtX{AW!J*55ZJ-W!QOfoG3w*(KVt<(j``OQ>Pcvi~-OmW{6)$~G#hM!+k4 zJT^uJmur5(C6pJ}?GBs<`EES}Ov~t=XNMqovO`SH72G%H#A{ySM*|tL8VzrkEU+xA zZKhe^a%7oqe5G0ka?M`{5^5OK!5%VWaF(*@>j!x(i`l#gjXKb(@CqM~ovMP%HNW5z z%8Q$36aO>kJtKx7_z;XQwX4QgQHb%iXEDBUS_g8?Uk4Iu7}UYYA=E*lLxj#|d}pfQ za?LNegc=5d?>_{=hhTi0ch&f+3NgMFS&Xkt>p-sg>p(&cgE}~92z7vWjdB>@Bo$n) z`309y!$9zZhah-<<8zl}?wgp6vckn@8@x9$;i@u=%Q*lZxA(6(oJXx14j)D@U~!oo z5Nq|nlj%N5@vkb-ZP?irSms7k=EeEStl{wK??TEP`18A4<~CF2<@w62;V_^VQs&}! zxyD- z>0@TM4)8czA%~AR4Tsm83+dKNceYy}dht6=@z>@nzJ|l= z%7qlaY-fw_Y;HtT73m$@WSE+J{-i9^_db|&Z6Cb%e>)@+-H2}j&(L-w*Zl29LJfmz0_jC)Gfp%u;R`qA6(~hq>i-ELOx(cn--)G|UWLgJu&0hx+Y8ceP4MV7d zM285SZGtsV1($1n!6no%5Ii~r!3$`Z{w)jNoAK>~pXwhZX6H`F?Ahs<13MkFSRXjr zFWq(8k01w;vvly7-LX;Fl|rkQ7X>!K1W9{H@wy}y{cbcExyjz2qZIN;lX zjW9zK-17I{SctWVe}@Sryr>7C&jiB!?Vdv>V%_>oTnm-ZqH+T*p^3oZpIWEP+;kH_ zXW;(rExeDfyIuGozl6iLWeS;_R_NT6CkNK+uUNY01JY_iQ+H-3cou1OgF7Kij}c~b zK4CN*-a#!yn5vx?rft`R={3S^$tR44!*^l|5vFmch0!`?MM=%z3*u&TR!?BK>o)&Z z?SK(s2Vl^xS`CM9zZ4)sSDP)uy1T?^VPWoR5q~33GcT>np$ZZBBP<_>8Ucd}q2X{$ zQG)0Da$5T?2%&Z86JkdGnkY6xjLIj3hQs$H3K3$}E(l>i9@>e99x~#T00!-~G#tJn zQh+$gPFtM9do8!yzK?xw8@hkI+9pR><(C%ad#u*+iX3yn`|+)I?a#Z!WaY&nNliaW z>zKA}%bv(V{gRUQR1STdBge<M)Ae45a8 z9$x%0uC@BzU4BA3R%UE40$?v1(zzWMG|4sU#Or|)4zCBM*`P6g{|l0QJ&*>KF%;Nl zS5SMw&yi!-kuAg=ER0o{V$T3TN3!!+;cq(XOqg&u$(mN|P7UlYcCu%XrlWq2CHg0+ z8F~8;)um6|Vm|%4b;RMCbtJ>Je`4s4;f zm_@`OV@tNS(r9dCv1cQVej8CI49wVwgv&Ega;-MAXK}4ICEbi`wVQmMPMaywa<4=c z{t~GZAH6jtl5lp3+OxFYu3}LYKk8y$UvWOgWn$xkRbr|XAsT+L-VnsN+JO+(j zYram;;G#E1m1;s>6%IN^EmOOeYks?yP{Y8kw++FrUAvLqimA;ef7V}h`TzU zxEc-*)(R1~LUfQz+`+^!6@_||nXrCiLDgxJLeDOnE%qD!GQrlHZeQRJFFP_B;q@kd za5u{+F*z;|xU!8OrUj~CZp(ApsMi8}(mGLJ3!D<)l0H#fu8mNx`5U2x8U~GUPrS9~ z47c*NutCoFx|8R0wIqt6<;7Yz?iZWlv71yGa?P&{2{jCqF`{r~pgpL>&S`tEovRTy zLK}mvBa`xLH~+hZq^=&#LXvZJ&H_74cd^#%F8rO8Gl|!5c#vI4chRWbMM0Z|{>*#} zNtl!-l-4{=lg`#~3!5w!4>qV~Y0~X##d6JW#S&^5Sn(l+Td~`u)76r_N#Cr>kZXQr zNT^|;jKd07MgdJ4lMG!Mds7cy2#q`yb1gIdj?x`X1A} zYzA{|eX?uBbO zJXI@1b~T#a^}|m2W%t*xS>>1-&LQ)Fk$FcxnKc|9!xbX4wi}KPTAmr%n%@N0%3_z;XweuJX` zTqp%`dow8U}(l4ngn& zj1P0xtX3mqY;H@-=94kTwr8=lwW@Hr=2y6c8U_k)8G^#`RpT6%_B$0^uK5L*P{Tm* zYlk3s0hZ>xn^$Fh+$H2p-5fN#b)`O&bUsKuajiU!^iIV%abT&wd8WGom^n|Bq}s92 z86+w3c2jnfNWd$sa@H~S=2)7^C-9SV$u2SLtOvnz+k48_bex>h z%(oNOl>V;3Ca68FB}KdBW`0wZP{W|cemI00!)h~!sXho{qFr*$FSvvn27<31g5U+1 zs&(?iJ^T}pQ*bx|E89@!5LA!o&oSj>hdL8U4VPFc;Ue-&>rcynrvrQ>B0u0vo5ZBd zIx>bJ`s3-b3X{7M(Ru?HC$q)gtId;y|*tTK1TPS*WgA4%aZn_-6EEt&%jPxdGJL~q zO-i+dZF|4&e0T2=`4+z`mg(wTr$6z*zGvy| zQza{^D79roOy9SIkhw45^ER&++Eb4-?{zrHsda|lTB&IoOo=GuTUT zMbnqGH)}hPYyNg1p)3*7b@{*gcHrxUw}WA1TBhE1w`sMn_S%$HHMy;85cAPj%&iG& zZtGf}&ARaGU=QG1#;BY?Ae*He~n0}VNfGS>`aZwxABJ31LtWYRRz_^ik+#E=e0)Un!iRQ)G(-# z(S_HD)f;~d&vM;?{)2E}*!E8+-ra$|OzUuW^~V3(vYQ^^|7O`u+j$Sl?oRb$QuP?8 z^f#)V@yb4R?5v{Q8K2)S3pZ0QOR2G)&$2rf!sHEon0#(YXygn%G!l%yjMcq!52q>k zDWGSg(sA)|6xZt5WMlegFM0l-nIEfrHi{6xJYz zYl6{j#v1gVXAM0P1p636sZWig*09;cZri!oy%O88hW7S?hhd*VbQ*)uk|dHQ$$SQ} zRSiO}`3*us4FiL?YDfklm0Z;K1`NU)RDX)YVeX)M3miikRK4Cxn*PrBRznyIy)4DP zb|uT~MRt4e?8&>{{to`#op_jj^*tB>cf;<8=>>LI*}G$-^j$X&sqd0{E$Z9nZ78BUHVzKj zfKYcZqYdai&jxl#5FE|Tt^drajmA+M2uQU5tqsWc)?foB=cH2rkz!CA*seAp*Zejh zp@xACd~HZJAoc3nKp7lj1Ncd2cb8b0eu-)OL9%!kr}d>J>q|^KAby6w^(Cga;4kB# zWirx;eP3OY0=QpY(g^AAT^39J_>vTWPt^STz-E&(311!BR^K@Us~@Qfm1};5N~mF=P&qo`G+RB7s?UzBZ*SczW372R zg>TBlQj7Q576~4o>OWj!Jn~Ysw6%wbxga?M|% z5^5M!=);9qXg;m_PnP+VZVyhSS+t#NN;XOrD%bo9l~BV#p?@x1p^2ET8H-}Nj`YvZ zpzKCfl)JjKCkF7fU&xQ@Iu{%L12Th^?}r3h55d0ZXqj(g1pASUVEqT6-~T6Ew~EWz z&q%h*hsz6l-Xf!oe?^>L$x<$BNt5Pnoikd6%F1YDD6yXS1hS!Jo;x~1EkdsOEkZ&K z1B-a6aEowiCDGO>dzrCRaM=~IA8$6F$`h;QHh-5yj)u+=HO3#sB53nqsy7ukNX>Mc z2YcS;5`LOv;##pD`M~Xy?ss?fsaZ&?K(6_#Ktc_JDtLQHRbbYKrs2h7A60SXnqOQA zH4MZp+Na8Z(R~< z7+BYl|0k`hRL7;=^$^N^I4vJC7|Q5riedVeH1kP+88gk>vaTwb`wl)u%^W$$sHMp@ zzokj2VPI+F3%4|9zo(!lnD&-gx?dtMbm&%9CsU z%9BvTpz`VpuRNz#hV}kHnO1>Z^H+g{8U|IcXh>Dy-yfKy;>tC@xDsj@hrOwScPtYeutt#d`~ z;wH){zs49xl`g`0I$E3~NZ7&%So9_@3# z+L5J#Y;F6z&734~y-Qu=TVcJ5k#*TR+CeG|FbQ*NT1t_M2`aTb;AmyGYa9-}_64lLjR`7-t zew`YzGg4n{+CP#DVFIO(TQb5lnlOpL#Q4Sx&pB=aB}m{_J+Ot(&{{^N2X0g-oOo)} z6HNArCkWJ&HMC_7_Uke!SYje6!eneDK?15~LVW! z63WmcHJ_d6kekq*QI6ZAL=wd&R3v#oOya<>sp6AX!E@?XnCH~#m*>>^fajD> z*{#Toulm?WTArBX&L#QA-}WO?+o~z z1NMl8BRzg+qd(sYzq8cu^bLM#BPZIAKjl?^XS?6o=65#womGBkh2L4`cb58{A;3rCdf2i&-s_mUBUx@AXAYRDi|l#e@J!ac@jh#g`<+?u z=PprpAF~$d%+Qm$-)P;J{?T@N-`B~|#&do1-fzrer<3nHo+*f7xkZ1j@6Gq^YQlH! zatf8ZoWeo7oI)3NLPhU8ui$yRoWh58IfdUJwyRQHyUQt*>~ac^ADT;wzFTv8O^!bm zNv_lmbLoa&^Mp}d<2SS%6p8j@WpBEj{uj&R&-G1bLOavXTFvJ_huB4}u-bF~#W~FB z6W)Da_@li0opw6f?HjYS`-{HM`m(m8^i}2bGM#yr)b=hYie2Ee{;O}}!A2En9n7t~ zZSriiB;mhy++wUjsbhVp_Y z!0khia@EkI96$6Zula}a-B`0L-?fy_^QUZe6n1;Gn{!|B(4*X3&`@aS(m&m5BRNWC z_Y(c@^TONdr1f2JR#>UQY4>1#r}^hayRioU>YsA##*~-7Ybk&1T}$~Pzdh_eZS*hn z%wRVzy8G=tl^mtbHImv>DcIY9HMGHLPn_T^^&4lg-#Or)r1kinoqlJ#<+R&!nZE#~ zmeURvvhJf`#eQeN?;Nn6M%an;TKADiV~!$KnHL$UE0s{pNuB+Qeh zNIblGYzn1X5lKvMkHp(InL@5eOus4K-a`lc+TbT&z?Pti>C5Boy{=z}79`#-+At|~ zXnErLGBNK+u1KO7{SohILF4T^yjWV^czeJLqs5NvOM8*-a3_MA*wNG+=W$VFe7H0C z)zq?7a9M}r?drPIQnGyMBNGz!=eRXHj3d z(fkU8^2!_!uZ={Pti^-A$4!~x+){+fSgp!&%d~n3<@be6u3at(C&MdrQ#-s`e`GD3 z!dH&U<9$eV%Td^E^2z5b`TVs-eV3!;ZdWXdqgbB(5AG|^G&j4?G+RcP*0OY?ShO~n zt2Ss`@!AR7!Ln5AA&4t22Y2p5@MWo%ee{>$`bVsVGjNsHOfH*vsc9jlx0Y)1ePy$YFqxMv8hS;(XE68n7%5KXjSdMvqr)tmV_<~$JfCh z?aQgpjm6qZl401Db!;Qtet(-3OY8+cuD!IQ&XF(OK0maz%v2)D$P6z7oag%EmaYTT z;@dqv?02^)pzr6nK&<8Ei8=UfJ;ACn!AN4pIOCpCYOtcjU?e0sDN!~`Fp^jdVZ^<7 zr139JOc|A!j+?D3aAUP9b>h^CmrVTB#0se5^(pZISC;5T$qZskw+(|1W_#Iujja{XGrLV~MypPT94#rCV2vSn)2?MXfRYTSo+@6HiXUwLBghmuYX21iRu6jgZ(clDK?i zB#!?xPz@JC`%`Zu5-ZgzEIX0rXq2BQt16Gb z5`U5fh{xtZfzd9x=2y0a8V1UKqi|(+$7)0}QE`VbqrdU%gUJO@UqXE&GkZ*%az`Wi zf|z5g6>}KMHdJ}GXabZc7rV`M#3tn9R&5-u(Md$7ty$Ay6=RI`0=G$VP0{?BhTEJ# zj{TVNxLiysxuh+UOiJGw!B37U=#+NirBm@UcVZEj0!0azPP{CVY;C`EN*pJU(Pf{8 z{Je5UH==`dx(?G&C6cdC*J+@cFx&RPJ5Ki@(W0p$oKFSm`z*$)?(lNS3v*wuBDBkN zGZdGMtbV~|(ur=wtst*}S;Hq6my{=dG($=a&6rsm<}uE!?Vs4vG6}j3%1~Dtp$^D9fm2^&WZFWl|OjTT=8Yv*AVlbNBFOQ4wFLJhFYLbvuYLgRAHUO)581fZQJ$>rxpx%W_ZHL+ z40}Uj%1t^Y#OgzTy%i&sIl2;wU$xpUVAE-AholUUTe+O%<|6GCA6I*?ESi^km9n_xI1M#HJ1{)e>IVt;Kp%~7~d`~61_USra(rH0K~-dSNq zH_o`i)46O##FxAkUGi2S*pUj}o?-#y1`b0d&ytdBb@d-$yV@TxtF!3VknUu}TL-YL z>cz5Z3-ku;-g3yx=p+Y0MBc0=qmB>onO|Mn|872j8elI?lZ za9Tb!5pUfA*^63FUb(OX~wS8XVA zd?XuJW4em17zx!tJ=pdczrN)})e{`S+XH7F-*T++(=Zs&5XS2bVyRZ&ve(Pr%3#2% zYgtO8qS~!nrvA1d7w9~^L)#Q;>>zW)w733a)YZKXl3;$n-ZZ@UupI-zb;a=>Z@0km zDJq7n+(pc;O-5dG${MzJ;H{bT+O-ujrtz|EEbiSDUb< zyylY@EW-cV`cqDv+1MDE;1+OF?ZUd5wNtOHpE;+gp&r?m949hA`>j=>OD0})K{#~T z#7i!ke9;9HLZh2QbE=y|wToMtYUkI6kZWsg^Mw5Zq0p4NdU)542~DY9EM&g`ZeCP< zQD|mEYkdpu)Ify#S)Mag0xcPdLk+V+qpxc{4sqmPUcHIFDjQpxpuvWg>Ul#^*vxs* zv7@BA`o?+HGiyz*8Top-WU09!T3K`T zLf5crYOiUXT{&y)>49TH)6{+^z~D>HuNZrJXjVhheAw!K@R~ZOy0I2Miv!}CCh7{+ z&V$X=w>0DSF`+57)i4;uoY_#nu(qkCw#LM3sWIMSP7+M4i4}BAXc8(uR5GissktRI zwzSkuF$9~OQ@c1+U0(wObJBxdmM%sP0UA~)G!=r>gtV+G2bC^%4cl|S%Fb(OYO0;tQdd7aRMS#9uVHpwOS5U` zqie)K%oVDhIVZHJwr=*E7Abz<^4g{b=2SharMAf{NL@pH=!DSBIkhwAhO{d<$z&|b znpHQuwW+!VK?05d*J#Sp|4azgw=~Q>3+1kFs$Fna09{!_)3u?-hK6~ey5>-Q>%4h^ zz`};Q8d&hG+9pdg=Smt-ZeMvtF9j zCoes(8sZ<1R7|qktGX^L6Ti3XAuZTcL?znl(PydX=z>DW9vk|I=TSXxUc*dPW@uJ* z-MreGk09Rp%B14xT^sUZT+=!Wic<^cO*Ho$8Hi7o9CD{G5tzY(N|3-@y)M-2g35Zd zSKfB%Y#Br(XltWFgI*h(8m>{&RNIWvPHC2GRg3_$gvHg>H$%L-nRRIPq4~ykFiO&V znlW}rjAw88T4>_Dc}@>#3f$~SDvhXhCK^feELhvM2#>+WGQ-y9>e;Z<(I=ifD>UU? zGa`iQYinz2Yn-mr6if4&B;8lY^r&bb(#*u5O~bA^2|ZEWth(A7jWH>F?sSZKXURVd zzZKJ`g~C(L4Nsbcp)jPL#*fJd7=Cg+dTW$pE_z3=vpNpl*yz@f6X?8FG;t^e=H@lP z2@q{Qn#qZ7tCw+8MpySns6-5G9#S(NW>!!+Y1*eN!j}38P!hfbEmpOPK?jBqV~J z5UOZEE3a)rW{o1#38p)5%(&a)mPiAM6KbihM{%LV^UE)uI1PhM$;qK>u5GDpJ}z)> z<}JN*%E_UUQ=LGQFPi3aPtUk<(I+pu__B*iMx&pXe4P<20J%r2ol?&8q#G60d7@(A zGZh{h-HK6%|D5rJ^S_o_%skvUzCpSPbF>o|p=S~Q8cqn+f+`zk&B7Rh`5&i)(j)0u zjTvH=6eW~QEwj!Rv-Z)_p3dipYKdf~b{yEv@G2%;x_ETysf$A@;h^zg*k)Lle$UKd zr3IXLI_AgxH>iHx3Cs|5W3tIm%!z|4We_P)VEN}S=fq~rlS>*KnlU42mX6Syzw5YJ zNiSCbtKwvcI}YWO=95`Rm?)x}F|)3ouM;-K&qEl#n`JVe#b>Hc7ZA?0 zIBPy-?BD3k9r&oVI%uN?_L{qvo3E(5ih*jhBHg}FG%2f!jD?Y3jc)x-Rn^+d981}# zXSVN&I(Kp#D#q%DxvhYcT2akh z^LdTc-1haWd7CFz09ZI<+1T15t)I-*o{kbuf6>r*o~{-ynKJbv%(KvNW;WF|VzyCY zjC~FkEAy~kKDGuvIu#q^O$$*>+?9^pn-qHY*jTDtTIx086{S~so;rc3ZEUWa*HCZP zou(cd8ZW4AVPP(7XqwyHh^4PITC8&10vx-@@zF8gAK%DxLGAo}!f3%3j0x$S(<@m% zQ5IBUn)w+^7?VBQua;jj&r*0N!((+sly)Lf9l3{AZgk8#OD??Jb9b%W$Z&U*d*PUn z&g%;=_rhH(H!>Wo+=baa7a`5)lcX{DoRzX__QY8Llax?tx9!GYqFWX;#>}UE%^}S$ z3M=PavsiM<>2szPXm0*;n&^d=bI^9{)0n?T9Qz+MzE6_u@!5|xRD|<%VoeQ=wUtga zNoDE{tJW${@zv*`-u@XDdqb^kBv{!}ea*aD&o=b_Ow7rA_tW~1p~>M?qjU#@7Br9d z?w3RA6P1yrA55zNfAe(hK2Lm4@Eq}n)}J=GV_03iS7ExX;5qF^+>oPNvdcf%IDicO z1$Wv3>%X?TNq1N+W7h2b1M=L3l^DKF$mT1s`E?a$fU<{FId`E~?zs-l9B(dhVh^4V z8i^Co$!0-ii<2xGkwcGj0>@!FGre)x-W;g!O>~o>N%yu~!SORF7-I~`TJNn68UMkgA zN?Ms^#_FVF%6(oncC_YAZJ5~Yx* zCs!EO2GY?fs24e7q37roat^-EtJDJ8qDnWE)(;iXo~2B(VL2w`JRk5%XSZitG zX~2?BQgY%K*uQ!Ou!V8-VA{teSZ6oPpQ?}6^I4e&8pG$B5_rOdnYxHC@$IuuKK;}& zq0cUnKT8zzG}IT%-Ks*+8f@4}l(AkM(=EGtmwK}bo6SmDK)XP4Yxbh6#+J(>=9B;AabchqHze1l*%SHqPIGby3 z2r)a9)h(A;7z?Fguw&qs#iA^hvzg#85t8(Bw2Q>eg!Dpcr(Lq3VF$3y0(s*u@n!2@~@OT@~L~27~ z7&w`7oq4{jD$DL=_dkAnOdDWGAcFOQV{oaHE8DDC!9U^iaH#r zxXRxNHdf=r*Sy+#dEzs7;T(BfF$Y_^GA+-dCbwU7SB^u-f8h}FmsLz*`Rjm@WrgZ4 zS+U#Wdjck_)r_PaJGJ`Pw1wOt^^cIg9CBt|P>n$QV^0j76lz}3)Ka3Akv10C_UcBtLdnsGKWf0Pfj~!Vci;X z)}#%dQDiS+fqKBk&LF~b!f8)YoJ}TLq&+?0CR2LXtJ4E%$vD{&Dh-_jYbg!k`Im3P zVC;G3%y~Gpa!GAXE2j3&V;Vd6ECrJ1zUHB&H8V6*%k$1Fubeu43J0p04fE%FQxw%2 zo`TiYw4!o3i?Zs$JYk}R$^cP_qI@%_EM=v(WED8*Xz2vVESF`+5=d5RgBHwM%CW}m zMTFC@Smn|Py0p$(kN#To^p#dFSnv0UDRHJA(`CpE;WR-hhMv;FB%=-$X1Dh=y_tpmAmzwnRGArD)#+_EWK~|CW~p&J zBeeAApT~(YORzyJJ;U&@YL$lTJ&<)rkATs>SKl)}~tCaSYY1U-x{yb4;MI)jeI1Q6^+6Pls`=9!7rp zsXUE?LS;Kfg7fve^v~$Csbz9~T?+3Br(-oR#zl2FS1HH0Lw2L)Ny}zaBs(UKE^AaRp;LaI zUvtlwyQMNG{5;FbcD8&PRGxW42dd3DRF2c$lVBntJ?|&Agi|QaSt1-L$Ib@MRoBgw z*A0#dm5$%+R(h}rrEWjlC$%q&(d8}du;>G ztLtg>=IU9s=G;HBO3z0Y(k*y86&hbU+7Yc9_pX!k-Z+VdYG@IuTVW+lI5o3~=iQMG zPS#$k?`VXo8ym5#k<(r{d5$B@n58v0sGT_Z8uhDo&3Qna@o|srtIn~U3HfiTq@qQ$ z`Q^MHB4c%E0VXSTE!RRta?lsXs3)P!QG1Oyv%0>%0Y_j_l$KiT$J9)iuu#wA8bzY8 zA~c?Y=nfECknx=`p>j$kU%ROkvFdAVAEYLSurZ1?thS!7isl< zrMrp|FXFz&tuLPGYgh7mR>qr+c#+F`1um*@CjO6diNsrbvj~It9(yNy(u8(i8os&* zx$&l8v51e)-{5`0=Dp=s|6X#(d-Dr#bmBcmywi3}*1P+s-9^CH@B*FlUYYZXqVpOY z-fG_-{|n#SgGf8#dq{RZT{|zdiAN%51bw4C%?9MHQF$$|Y!u$hON4uwPh4Nj!`ro^ zyjSX=O}w(_7A0$t@LQSXl0w~eA(Vg5NT~r=l58?)*u9DnHh@dcu%+^H-BW2*o(Mk` z4{uF`x31e#JZp5IZB{(|lDgiH7i*(iZj6Tq;2bS&E*{<{S0 zxjwXX%67E2z}jtS$R7_6#MeDKf4d1YYya>|{d-9~yz?#XhVOivZ!Y=tm0K${C(kNj z)3DR%jzfcAScBB!cm@Bkm?RrX>RY&pz4pMj75IK`QP0`$^lEp}mqZZ#I8A{o1_*X0H|N=&v$jb3f37as>q(1pY!}T+-{#&A>Z9pfZz?;DUF_b8Ausc~DLQr0twr{~ zJ8a%$wqoKfUsLJXa*%h#t*}{M3NIj*n?si9lDf~DP5Ct6j2F=TdFhCh)|~Mo`hcSn zB#pvPhf?tYl?gfq;obJ^=I&FS$sD~k;^XTqH>W&{7S-!Jx~LfZXktoJ^_QQo@E$w` zulV9UUjz)6Y#JlE#!*OmIX-}dC=|-XcO1SC;rTvo<`p#Cr_J2-tj~0)5ovzWjf?pm zy?5RzI=HQ9#3Mz4TZ=}#Toibz=ooK0=4fpWw6+3TYsYsSpuH)+FP@mMK(<3wq8|Cq zNsp=rLnPD_zReK~s=6&xRh=2C@*+Sp3H`bOJ`+?{iU!pdS@+DIk)lnJ1YGh3#wm#I zXv+GC3JPO?KhUQ(FZpb0bJNSBHhhS2;~Y2K!6pi?MQQ!%mj;zn&$ykif;ps=5;%mw5@tF8B?eEo4d4BG!FS9sWuPiOEoARKEE~bUQ?)8 zrIy5r-678M_||bhR$SYq?Fgh-k_#~|AL|%lc{SBy-6jhP%y!;P)P@) z9@gM6#cdd!6qE{UDl#fkGPKE1FNiXl{T$75N?wjeIW&}$Qc{nezwg?6?Q_=ov(LH1 zAH^5*QO-U4yZ@}c_F8MN{rA)v77M$mZgP$4dzk8Asj*Bo6`SnUq(fF~D_!$Rc3yq^ zNK`jYzr8HPHAG2Vkb6|Ck#s|u9DPMFto5-!&A7c4 z&T|>}cGvk;)`f>WkahoF;j~q_zjU4T74CM|`D?{D(R|^~FV&1oYpsn^O^iw39ivn? z{?%>9<2t{WfopZI9-9J0#hWnpMqP<_DuJ5A{Ah`(!Pvd%juq!93*y6!Hg5T zr_b2CC(UQT4fjwm~+tw)=* zBaQp*2s|IM-+SARBy)Pd9q~SoLB8*PJMt>?N2cz#BQKL3N#^u^JF*Au2+gBn!m38O z!FS}=Ui>+J_96xmWwRI>QzMvS>If#j7;j!pjk7kXp_CZasL-~X)c}j` zt}b_8O;M)25ye#8-VZCHTsAZPyEjHPrxOSLQ4PJ8GpbQikT|LdW}G;xX}6AQEDf+1 zJd`1FzfH}i7TWjPNGE}Puh>XOHTuGhWC%hdoxZGl|1jsrIrkgGov&25JBB&8Rk*(& z<~-@TrQy!4!`!EbJHHE#bkyg%^f!nS7g?h7Ot21-1()5LotyotoM+lY_P7?dHfWE@ z`m#MHAfSWscG<{mH_Z8_>waOlv)y&SJKR|rw5nqMXuo~fW1iUj{q|w+ zjOQX=S2Cyf+lT*i_CfVM-uP|^{=@v}=*`lP!JXN=Bemvc8XEtb4AvIJ4%T?Sm|O_y zy6u4_sIZAhXZC(85w;Rv9_Do9+<{@v-*fK9Va}2Y_qJiqO4nU7-1(>L{&2YS{h${S z_>{8xI%A=zJ}W0@b<_?xJ+7xYoJaf7MJz`9&%l@gpOoQ!Jo?jk6oz%$@tB}N9Q@Pa z%{dgYqivJg7YHSS7MUE!7bd9CDtt=^;hXA7$$(>f;ydaWSGC+vp}e86wKTu#g+kS_ zz1UZI^E8w-@J9N))uVsH-t58w8wlF)LeH^J0HaR-D` zB{SGlIFjm%4S%DvZ7P+w&-LPBPHg}(MLXC9KJT<$$!6dxk;g&vPPXQuca{@5nI6> zbupwYycu;FC5djj>I`iai_Kc?WSdPz(r#4xmakVds>;7<=M{1@doBapEZ;M5y1y4H zKzrO?C@&YW7fOi+_^zl3eC`>GKZkW!9PT*3aow(A&f;P2Z-zOK4~y(oHD5j7ud2}J z{i?cORrf_zwSLU8z5ZTfTdAm=LwYy7frW!zPx@i{{GXNNmaW!!Hc?!1(7+YWcWk#%2qgL7}zeew;?Q(5=E zH#jX76!Y5^?iY@5?yGSBa=7#R3isECJCC@BBm9ox?l+Hc28RD3w{WzxXoUOIBb>)a zY^&Ic$7{ba{Lv$vjiYsmU&~z3f;FNiD%`cNcLpomYYuh3?b7dcu6y&L&U5NF?0eUt z&M${i__ks0FAsHoGMs+*40r!>sPnZ-`n|T&Egt5)q<$|yz+HKm^Ya5Je9Hmuy2G5? z4y51R2fB|R=G^ca`n~!!?zY367uD~M5$>YbJCBT@@aqry(y$e;cm6g?(|6#N?u{Ae zXPKeCtv&dK)$1=w?MrMnt=Jy{7dIu3B}`FhUz?g8#Ccx*hty%rDD^6!YS=s@>( zc-(lP`v4yQIuOwxJ;?1vgg+mI-`fv%zw`~jRqpc$`eC(u7v3JMcJKUuIj85f?pC}# z|62D^Jc_S#Z|KZ9-+7(;bzt|8bT{Db*^%z|@mPD9yXc!a=huh1+wk`7QSJkHyML5> zD<0P$?k@foq8#oH;%(t*_aVG}d$fBy9{q20mwY?teCbW@cDy}ur2GHy_SYlbJMp;Z zD7Rx}&gnbK-GR5xqus~ww&7^^E*bGIV&4{z&n7vlAX zce!7o(0Bc%;zxMiG0EM6$IAD*H?2a__qs3P?Uu9Lhw%3Ev)tS9_}%-i$^8iVy*Sz3 zj@KJL@PkTvU3{L~_C2)Xc|RL=3!YD1;I7Bxo74YNu^G?XXSvsXKj%C#%Y6}VYcBp< zX8r1%^PAc34umeg#C;Tx?_c8n6pwFx!hP%qDEkxc5Ab;56YdRo{Qi^f_mJTFIqo%h ztexZj34z_0e)n}$^mW%{-0cYdA)cG@ysstWe)oqt=V?5z!}G;O8TYq%E^E!W8}R(W z7cy@9t!Q&R|AzOkelc_R0o#6_bFR5QCLpWOQp@k@_p+{Y1c z3!dxn{L`bEUswD9zza`g+#B%xaxa(#;4i(IpDN%{9I1Wcr-*>(y?FlOnT-3DyK>HR zcrL_q{T4JAo?Gy|56_?dA#>x%J03zU|B;z+^Ure5t^drtwH42s|0lDuf`DJXlzDd- z0=k>CKgdzQLUa#X*MS&3H{jX+g{=F1JXhj*J)YnDV%8l%g3=eW$NUn{g+)%dv_0!C zx*N5*HtYTl&t5!#g6BW+{05%SUYGqv_GT*KhOFCl51I#kl8p~iD5p8s!UcHM9)a0`s|QwV7PVb1r>|8heri!kEod-s^w;k*(J;=T3 zVCVZ)$c0>^Mt4!xxi&MZC+oo0{47iPe=Fba{>WCD6@QL_+;fdF09GFa%JTBIS-^#dak0uZ9pGT1g_smh`!98>od2qKM zMIPK&5uSCgIf^_u#Qb)Jd;gosgZtr|$%9+=X7b<=4iE0X-|Rd${E-9Ok9HnAVC}m zFTbwU>vBgrG)K6i!s#NJjG@|E{Tk$VuKQ7^NvErPyL*Q7k>uZ$zl<06Zt5rF$=?d6 z@xa&Km^tJ)ly|m!e&)XpY|S_~QG;P5qW@Bh`r0v>BW}x_2|>@gJ-G~Y*>V1oJ8%d3 zFJgtsh}$GgLiio@9FsZx!xuTueHphUo7qOh+?YMEGwa+=if^(jKOXe(9Jf15TJe>v z12aff67S4VRP1I6|)R*8#A zV^miyJDG8}3TZ@Jnz5@Y+Z`VX$wxveBO!xmfnYwpk&sd(WK|@jBOKCFnjgrwr8Fb- z-WYnX3%%Ed-t(dNYV#g@HvzYhzNAOB&tO|)1+l8A9%wst-W@Xl>N!|wZ#1oloPTXH zM+$=S+`u3r4An9Odc#m1L#SF9YGkM}3e8|>MijzIwYrx9ybWd8p;|!6hHi^6uG&oD zy7j`iYGsA1Uy=NkTqs;iNI1QcB!#PAk@QOT6t1N(oL)(*!nNFnaV6gh*ODH_l}0FB z>qZc7>zGZlPk7P%ga~xZ^+Wk5TDwE(o61Z53e*Kxn?&sa#ce5h5~<7Knfti-d)H#$q;b+`zH^Y@o~$w3))apE|RajpIM+Pr&W3%l>vj_d-`t zc;d6CWz6fDm4(n^nLJ~hjK=m!l()XWNg0Ts6Jl-<2vCw9+stDp-wPEiD&BsvGR9KA z*X>oH)=89z_KBr*U%+>l+xtQtUhQqLyY#zj$l`D=N?qOM;gFX8=0J>=(roh%+4K~e zRFob5synNZ*I?D1&iKeBWVO^?%Wj&k;@nhE-jwq|{c~_$rcjqXSjcREC47#qKr23> zdIsDKhWprt7h38EGXsySII$N18j8*-EZnA@u)8bOzK1(7pvCrwW%~G#x*`Gxt7fPT z4(+r<0NEd1HCElaQZ;sRbJf@>9aUq`@2na-YfaVIOG{N_=k-*Ly;|PAscwEy7YV6M z+Z!v@9ef3BtifK6pH`z%cTvp>FaEXgN%EXF6c#??3=$c?FO20i_P& z7Fw2!*-?e)16vVu;WIDccVX<~{_obunyQacp)LKDg-l+9&jV$-GrFX}0nGl3Jch4y7CO4}Vo3;^8%f$N2WxggGXMD;JOsTnxR zbEpF4MHW-dT@~lY-NzTSGDbGk!YE6t9^MWnU2$sGQ4_p1$fIyGvDs3%k{LLDxFS); zm{8w;)f)nl5c28aU2xQ$0eyAj;`(QsuYT5PIua@Y-qv^D1I$-e=KjDDfV*|x{$Rk< zXeg7fL5S6hE61eQi5@BMF{6kp2WhMh$2PVsweN_=7JflsukifM^=>uD5}x$Dl2mlH zsqK6swPZwxGHkDZBAB5-$pr@*a={V89<^ZdS;hO2Gjbm#3!tm9M0pKGDAsS~ZC1p) z;=3z55DQl6WIcFj3f%AJ7Z=@iBnXIzWy{7IzS}L)6ERF}D^!Ar9VCWyw9%^}*sXs{ zV;gff;V$K4{3fF;WU3zmAZ=wmS$c?;sgJ5D zU36LB=LeAlf&nlfKs+YUt-TaV%|uteax-~XzItjM!Pht#)(jn!~Lvc$qu*5Y@ zJyMv};1b$Iswvc=UCQ=-s00#oOVR;bB!I>Dn|I(%#>6Se#6&Fq85m;f>aqeS%9P(N zQ;kNM#=vkG>=o=L)1ZgL9msZ6Xf#^mr7KdOtkU|eXX+DB&X%kDoTftqW#ArVbcF*X z+~}y#FSgA?YfeT+Nco8B*(Do90X`c@m3-SfBrm#+=evYtXFXNX8t1nMXxw3VpC&h6 zCPUF%Pv1y~)D@?)9rMI+DP>Aypb|}E6yal%3^u( z$`n=Es?}cYjaEbQ6*Gn+InuJ}>0Uv-ukWfX&#`e-cfS6aEHp60X&&ohyRm8xJe{f& z*IyNT!@JCs6AUIM5O&luFDIQP$d5FCvQJgUgSZ7-;I#oNMHZw-2aCQ^`9U(KY%SlK z_}w^i5L0+;U#YNEC#fGvbr)fbwt4EdmDGmri6+g2m6?a+%HJAlR*AADWk1bM?$-^B zG9Oc_;q%nA4O=U6hu1%YD}sgo6n8AfCDeD;@AOsbjB{zU*@v}!y2fKC4UX=}<565T zTD`Pa;~_0O{D-PLX<^V&!`gX=6>zjhMb&b|gIzfuSX8QNz3{a(jB9oNL6rNq$_&2dqsMD`c>}HBU*}mJuD19k8u+1Y-239*62-hObxct<4HpHbX~Ku|=oZ z-dfUtjuw-?kMOo-#1RyuVh7T9>hvlaU#c(hWg6#d`*xtcRRPAVwzsMb{MQm?>%TS! zU+aRe`QYme@w)Xf8h9}qY&`6w1nuq3>XFI=64%Qv%O@sY z>X5BUX)+>*zqL!}MMcwHH0l*oHY%?`^nq_j(zokq%T%tA4N4T?c(6Fmb#$u|{~&7u zqhT6Tn(VZ){#0F26Bt+)ThB##EqBwuPSv7cD88MlvsP*^xZS4Wk}s|&x+40i(;3FJ z@2%H+I*CeCOfm6{k@DMdcd4o*{hMA%inR|^a%K6IytwMQ5}j+S1LfT00ZAYqi16_# z_(b5FS|{OO5`ayu1@Bh_|DuwfM^5wcjw5@`91P`BLOel*RJFWQt>Z2u*b_ldMGZA1 z)@@2TyfbakuMsV(*v!^#AR!he`Z8NdQjrs|^J-dJiQIFOUfqm9G!jjlS{LJi%EHSw%|szpEhQDUj9#M6 z)P|#q{P_*t$)ALtvCOD}`9U=5WNT8*rfEuasb!{t7Gjrvei&k1-A=VIEvIVKqH2`c zdc2;qRial8r*uY&Q_3#rvzRK&_I5}2ThuM-FKh*DFa;OdxV!X385ghLI}A+JR>qs8 zAZDuEf)X9eqf8rJ2TK`u&_msrB&5ZHOUZbc@KSX$TRC@m1W~IZiqt#Nq((;m4$X( zd%DXh$xQ58nPen(A*DKYjHH;OnLqsb=s9Kz+5j5^lWpwH1^7@-&3ZGT73smPTq&R! zdwb%$TdLbgjzK2b8smZ1w*huq?xvEQ&mYiox9W#J_r`eP)6p}2Ln{@mqK(6jQ*{%m z!)m#e0>7Y&{nOWEK!u;J7q#3RMp*W5SySy)u5CB|L8lpCQ7x-hzsylhi4a$fBEN z_yvAGO2|wVPVmF~Wc=YGTr%GD%9@9=I7cKG%XhTO6rC1axRIuJu%eeTJB`Gc-P45) zWi~D_TNkrxft7tdwb;h<=l%rE5K8ni*_F9xo-;PJGkN*u1@)yyppR~v$wf@N&Wl+& z!>Sp}G&Du*cmwyaEpswi#;q}lp$DES-LTNTJF7-^S!;| z=+%%bA7$sRhIm@DCpa^nRHsnXOJ}dr-4&+a8{&?ge-oXq6}?1M>h-0to1vxGB>#z&yLJKU;~0YFeVN) z)vK7HT>!EUrf>J`sqe@sYaGd|BCsMFXBk8evZNSp#iib-xEVwLF4}-o4u5AU(C2J=ZVs8jqA$s zR`y8g#JB$oQW#^OlzgfE&yY3G6v7mF9Tr;xtfWxKwj(wkO` zNv9=v3Ork#QmjYYJkTLq=Ae{&_9kYBZ=gpDXiP-EhK2OYSVXOnP(>chx1ny>ybGp0 zP=})lLMS;b6UP-qtGctUl2$2Y64?jHK7!)v+1CT_r40#Yh))9s8kD|XWrE)xRhk@I zelniY>KMlF>c>w0wRk%teQdC}F_tpPSd_)Fe0$)2jAC!|1P!OLw%N!k0G?JnmPAJ25Ntt|9ZxU%&Z;_wz^?%OAw|KZZ+<% zTGWSKI?7aFBi*7NV>h%R8K9;zL^FB;a(00^A z3S)8T>7l~fj&wui}a}r#=;I; z8Ery+k6ON~T6A}8h9ZAxEq3jy#J}-N#k{P{PqiH?sM-!sSleKYc8Lu%ai*@ZP&v}! z3blEe>)w7|kHk@5U zW)P>Af>fwD$Ued!%A^~Usd@f}l&=}G|1?#Teb_5BYvqqv_wtPI{`zmzBNj7;GMX^~ z=XBMw{usy(oHx&r8>IOW736us1LFg6D5=^YbCe2J0?E6j!zm$ZNxi(3#G$toj%Mlg z1(*x9GT;$hP^UV2sE;P=WN!YTx!B2lh>JN|&uIAzNh;hJ4%j39sShDZiH(i{Q4q0g zh%iY5cS&dM+Qq--EqtkJAPjbz-0@TLFFPc%c5Y zC_+`9EbM0$Hz7n7=!Kxq)YNz(IB}~<^+GlxMAd^sl(WJ|&5cIsv*zYwCSE8XHYfH7 z+Vn~E)~UhJe%{I%2*u`#`sSDJpp)iaTQsz(FQ(#vyUt!6`o~rfm4TtLWn*mtduk}N z_9@L`??h?l)ha(u^3Yqv7V~jHyCRpEZCq&C&FOW;hO6+r0{y=8BsG1fZpssi5|DQPZmve zRne7>=t$o8Bctu)^}ga%=$`L(h4(-`*A6w4Ry615BC~1I-{6}1bue8K@`yd3PR9%V z=|AFjztQWU7|Z7#I8f)wT$A3!m`;H@)c_Sm_mhXVn zKzU7x{7(D|-sFm?gFGK4^~j?`+}7q#RpH7vJyit}@nkwIL1D#hxq8qsv-Pa>zH=2< zw#Yl2SF&~yry@KFNOP2zVe0j}shlw@Xx*6ojTnil?TEf)$=!()>U*nh3T%5^^dy!( z^xj~zngM~DOdB0XsQWCNyvskNm*Z)IkPpc>LnNdZd+^j=BuMgJ=!bkM-LvZ-#RiKY-pN@)`MvY z_LX}63s%RQT|+ENf~#j#p9rY2a-cStHY8ZoSt4F&3LPO_mUap}ek zd!$35!bPNRv^z zdG*#$bs`+6Cs_mKU`HM5fEA(|4b>AB zCWDx!?C5c@y6bn~P!JV+^$w>A$H+q#i(;zfS=i^0A*LCQ#;~^(_J@TlZ%j&r?^>i@ zW^Z>ly~3D!N`0^P^qmDb(0&0a0Pln2&7ei_bBg#ss_Axj=Ll=ZE{v;-m^$@$PhvCg zt7VK`35KKZl(e^x%s})ZnY%sOjc(A_=x5>WN0%^YN%iKc5#~2H?8@s@Ex6Ok(6!z>==YnF| z{}t!~^g7FK?H$}xJph)nU$s3zuCQK1A%j-WIIQpQ-^ab{pH}~!~S&#-*;s%SoEO5@d$(i{b*_IwT-^R)bD%V6u=8Zs4V^LO1|C!t|UUIe^x? z8~QMe1%Yz3?9of5(YGn!j_L|fKu<8BdUDSTRnzERv8o#U`tX#;5ot;&)l`*DWaU}3 z0!g>lfkEN>u*{1Pqc6l{EDN`m=7o+;@Y#z4?>!oG8)o@I#r2r34p0QnZW!Gv-?|`0 zpx@nf3lLYwq5A2>js3LKm}baUJ|J5|;k%SX7Ok+hl3Ko++ePq(q!jweCl=!UOEXIJ zY)NneftafdQJfJ{-It$`I2WEy3t0V!zQvDS#vKX5v&35bwh%sIe>(fCE zRk**9p>=o^iCcJir?sqx+UX0z=2q(Lk%s2yIXhh<;P+K4Ld9`N@F?Lpq{@WS5R=s% z5A7obNo46jwAzV)(O4W*9qGnF2HD=SyK%D_FVgQqEQp!?VjYfJ{NHaKj{He(YU#(n z%+}80)vDQ5g^_52P|tVMuUbSU-B5e-s}R+!_W1SFC&`q_P%{t|!w9*z^~{$P$_0gr z(Lf~c$O^J_{S6iH-9#Sop=MOYa{(S*0s68s_m?9@lmq)%h<8Q^%|6?+_C;)uC{A9} zP^@3oP@LR}bA>P)bk52c$ol1YuNChdc&{<<>fBU4q)12C2aQMhuQ)Mvb+sBV=o8AF z27>)qjEq)^_h!7KpEL#Y;VrqZLZ4X@Cz)TGxBR7f9j0atZIc%_wACY@c6E%YZJ9oW z8okI+=K;`64uxL40MAffm>QPZDjp}cx1!oMEU|Op> zmZtcrzPlmL#E&A_;lF_PgBH zx_VwyYxC#Tmd@mvdNvgQfqNE$)vV?=t|_ycac!Jd%d$jRS7VjW5E2JbE8>q=E7gXo zUXzPE8j62yDE=wFS|q}|+AK=)chCqAi_mGpu~PXCKvxPx#-BPzz)6O{mj0?_~?)xhKO)Rybyq5;wW_Au8M;AJdHRxM=gY*!aJ-&u+XY&weea$G**CF3% z6MHFhKP1i=iAL5&NSK&FyM`_;DiWoT4}UIPUsU9m20H;|v8;HJCNStvyu}v4vu_n3WH+?1B21u!%SJ zUY@dRIh30D3N(~dyUvfmj7#FokBO=F^ox_*l^!4U$!*VnQM@JwV+LBN5=1BXrsSD= z(c=(z{SHic8;SSr_baeO;@88}ob*oZNuck@qgr5>Z!-p5wVoaj3bGl#x|XHmMq9~0 zqTYwfAp=9A3UfveGB;7F4VI;*0#i&wl8Oa|uNL))EWCH2$EUv|xkK$Wij4O`-0y(@ zL8{LnrJIT8g(RP&h9yW+67i~a*tHo-s&p;FmwW~glrm_61De}O}a5W+5yu}a&ylKzRs$|B~|g41+s zU{3Hq%^&zanqKEbYML0A!M#WgV?+j2PMDGu;MUGQoJ^yI!(P392%`PvmreD37;M0W zkkj3C9RdnwX0+96Z-9ns@($8I?D^~?>x>Ky-$E&oFD>)of{jp0#RJS;x1C*GIBrS_ zSF%=s<0!zVOh-YjbP>wmR4z}Kt`SuZHKGRyfN{c)xHF8iIJBB48iyE{@|2O+GMmTO<{Lhq)J?}zmCsW}Bb4i_{~tpItG2xjyptZ0&9 zMCyIPBgI)Zs>lx2JVE|=9?+iiNaIvoFiEstk*V@kD1l!QwZKWD@_inIB_I$hosDlk& zi6hw`N3xwG!oVv(a}cFK$&8sJI|F`I2Nj({5lNYs0HJ+#@OC}s1`$1Io+Iz$Vv3?3H}6$+1qwcZK2w4_w)yC12#6Q zrLuM7Tm&H*@Cx@iuO_S@<&1-a!e~w5TWQaF4XO7PGnIih*xUV_M{Us7>V{M>l(?TY`HCj@V0vVY?@iU=Mya^)AgN_G&@S1pS7+X6)4e?V zcu?K-hmq@6TyxVoum~6`syR1p8Sst? zBJiQi1pr?{Wcw%gWXUn@x%U1e?FotS1wwC7=vu|!3?~M z`)6YuzEU>8t#}FKubcf_NMNPznf(iK=rV5bQ)4jF*a3*uhESIQ2M;&ym{VD%*a`K0R~!;386j64EJFq=@P;|t*t?+v zcPC2QJ%nC9w3!84{Ug)NZIgSTC0GE$Qi7VlWUjScO>H{SvO=-9rztGDmgp!*9!2#R zm4nvn%1QSkEtS20bCh-Qu82>yDQP2ot16VW3nIB5vc?hD&l`ye776`*IpUi>-5X0= zecZs_Q8W&XuINCNMPh_?_VeBgv4=*X#v|C#9ZP_?%Zku38V)S3^T5 zXs4JTBOs+vNYYb(H9<;g*)gdv80C5B0TU)k>8F($LX$T;C+t}+huU}zGIfbZJw|%0 zHSvvB;5*1wG|fZ{SvYoE41Mn-#Tf7Q)OEarFk&XTYjciRJaw!@d_Qx?{5!ut^HxF872TfJM+ap5%8G>ojU+`1-WvP~c z!j*q>nkj_(r6G+%$`Pet=T9jV^O$iwIPsKiKR3XgWN!E^U7l3Q)+I4|NBv;a5u|xW z>AYu4{f_E^Cq;vqjT4@@YddHR9}U3!up8}OQekbF{9{NRd@#s}^1s{IKEZz-rTJz? zL-iTUx@JanF zQ8nH$fV>(M(9M%^D4j1;n$<3zQ1~4jUwFgMLIpsNE+X8|A9=v z=4w24=;0LJc)DLt`_ntam|zvIK+w3L=`_v4hA0OD83z)1-GFrVv~2kGrYQiTi)@V1 z6(RyLc6l>iwIRo38Q+^4qlH%VNKNmt0?bH~4v&bT^wC5Jng@M!QSI4_^V=lfZ4MN# z6wm@EvXG7}m_3BGD#0VoRV8kF)PagGBCQXr4fJ82&Wau{*30lLikE^aG$gCL3$ju^ zRsZ+=x_FL9$M90BcZCakN^_>@Kljz7Iczhe&1M$v8wNSlHZVbK>xGW)f}gv0%n*%T zRR2@IL!J3690gkx?At^Xa$E|TX*!2`g zpKcVLBwtBvEFk$4Q_%se!p-g7qu_2o`Bs{^c=GB#1oY(b6VG0{(*5G9=Sp;;c^9@I zsfkj0SqHSFE_|n}5}Pgp(W?xj3q`&iS9_tYrTN2OC{!I=8aR~Zs2=@OOJAmNzy=E1 z@Iuv8Xo0*d_F`nB{*{~QDo zovIVpSKUymy0e6v&YMPo;s=2!V$~?(;;PS;ikt+ah^fJoI^{@=FZI&Kl#?>hhjUN! zCTe4!UF?f2tlzO1EFI{DdD^R`l)$k)R*H4xR~g@HmfNXo7a%a{Cnr+<1~C$Ce4v0C zszoT=J3S|vIJXkWMiCjfhA(0=VlM~?*ce>{(b59-$(JpyQ|wJIAEzW#$MJ1te?V;#w`_Ji3capCW&0x2HCJ8I6-4vVuy- zpkFDBG|gMC5eOPswZ2WW?4)~erQ?Oflo3FKH9+x!)kdz>zB$85O}lF*EZIkkHl-Kh zaVs?EbUt}2eUXmKKRybW=11yMqR})6Euh-Up%J+mpr6lr$?wHcCb4I*V;g*rpu`X@ zD%@pc zrp}jhaMUv^Qp7QXX@+0hF_pF*#(>R@xW6EAHe0oysm1B9TrD=${~flheahdVX3jvRmO~FaU{Duhjip(NLND!H zKyZ8gt#Uo19s5(;r$mBS!eHLlaH`uxs)qJUZv#u44O4JuXhYj9T!8(@dwbC3uvx*C zfX_D+|KeRWSj0um?e)lqyJ8f5kLsaJ0Nn|qt_-13L6;Y=Z3-%sMbeTNq$;7a=|Eg~ zt|#j)8cpX8Os0m%MWEtB!iM5NxMY*nz;KF=DF@9M<0V8);tcN+(12W_T_T>K4l~h0 zE|7FhMuGG`LkMj!oN!;&GJHHD=-YUbiU|I0HzW-kV3KBe#y4EDUzFNz3?Iq9P#V~t z<|$CnsvE0N6|f3<(2K0+^-1BQl@xs8PqjA9=$C7T3=a zie5#hFMQ1taC?D-%3*EO}oXUEeqSnC1^1f3pavI>vDx zc+L;#Kg9eN0_53H&!ibQFzAHD!tP?IiN`%k&{CQjdap6>Qh`&-D3_?+X;m!i>QnlX8at4YJJwaAf^klQPl8O3pn2s4_9uw(sXxmf(0y4I-!^zn6$6V}#-ctYvYmq-%bxT>46Vdq z*j;~mHORT4&twq!L!N6f)I^PHoOX!~Hp!Y5y@@+?9MRmw9nXE?k?nCd4c)hpy^zPR zq-^LSO43^zQ}~3jZISQ~gIMXK^vune5)_Fv><-9oQ0l6U`7rrd%ko>z)BMP| zR}TL5JbSOnx6r0GR`6T~f@!t~=bh<`C)k18UjMAnFa7FuJ^Gb(Dsq6@FG5`-3#Z_> zg8Ks}kdV<#4;9x^LWRAIL`wfB{3Cx}iRXvI=Uj}w6K6NoRjSc!uq!u`MlnR8qLnC* zV47&i9x+u$$y7^snQfQ8+Y#~?XaI%PDc{(EN_@xzjHBR|4Pc;W2g~>#v7Az(uR591 zz;K+i5{IhPM||*^pW^TCtp!7sM~AKoQ;7}5XVu%l8>uQ3ulPF?)=Y@s-KwhE>930D%=f$EfrNbaNjGtfL+F z*uY5NkKJg6aWse5Pn*(Uz=wA|wWjhb{K0|3IG7;#5nNVrL=R^6x-(nVHwdvy?)OB= z!mE2O!}LUeE~rD0E||qwx(1Osu#Pu^R6*CEN7rB~T_91>)x&ftYMbZNOb9cE>Ia56 z%nu}hbm?d_Mi4)6%G~f_M|^oYi(-e$mXdB#Gc%~aBR3fgRYxx=(S|o8>9+n%SY41! zRv4ZND~Eos4ZYWe-Wx;jGeYn41MhAJwW}<_T@?u_MM6kygQ;k~JQz|bs;lba<|85F zBO#5Ekoh9S#x|@D5O>-9mKur=s4bv(jZ;*;SSUXFLg5DuT+Lu6d1XP!>rf{HgfWU( z+Tfh}FjHNbANPi4ca7<0H`P^x5^OWAqwS~k%+oU{Pt`;9YFI*G<^mEjHSivpk9(k5 z@rDvV(EL*(L;qCgHf3F;Hvk5g5HrGHs!pL8WpKzoC(LG$RL3@qV8;XZR+_>MXJe8? zJduc}hjqphDr>AHeTclLX(B>;4GG%i>i1vCB$j9|uUyiv@2#YDkII3&fYSHE#wYKx z^=gPOqVM_|zRxISsmmJi1>7jG+V3EZuqB2_5ojrmH*66jN#hxTWi(A!?z~Mc7}saA z^PiscY@fFg@`5}(&RPXmtA=o4Wt~`x0;t^}lL0FYQ!YG%q!KoO4iq-drm|jj(h-RQ- ztA65KT&&0b6j>;JjH|!uCsY1@F<5Jg%nm$LR*|t<72gXC578fx$h~N@Q7Jn08EdoA zP~p81QjSdR6sv8E$bN03z9GB|(?`Pt#SRfNUF4MGh-lDsh0G6G`N-sqOc5tyN<9Ph z2~;3ub|plCScpV77)y+-x#+edrZ*&ow2egj=Op@r5dA@?NlBsd)0GqsD#5RemBN7F z#P(;N?E$}g+ni!p&*%$m{cTgW3xTf~*ix3j$MHQ--0s${47Glxt=Y>{exh_qRY+1h zCAu^eL>r0ygpfL;m=uFhZ=OhIy)U$Otb|jQq*rGYo$O0X3H@)35^@@kue<2O)u3o7 zF55NlS%weCqV1*iPS}9u9#x(@LsjA2130${j|il>j<%L$OB_oAnma^8;Zqz}E}Fo1 zffUD%gOuU>lzwk-_Q2l{?**QWZ4W%d7_879c;WT!^&;Zx749X1iZpr1s^)Vss>WXX zN5@aWF}d32ufENur7~89lS=Yx5nyn&QavTLNqwi8DXu~EUNuwf-Cd!{X>%GbLVtdK z56u^s(AYel-guV-&vi(X#x*=??1@3Njminx8~(rKujU+Su(+*3eB>&&fxJ3JB>{SB z?2W}25sAxewzBMMl$p|`w3N{1nlNF(cG0fphiq3tT&(%?TY|{%y1BPV{{_FHP;0kO z8w!DMqxc5vKeT`5-yscSM$G})rl}Oszml^Xjutjx5!MplNQcnam{Y*WQD{&|-r6XYLkbDD9xUCSl0Mpcx=+`2#Gb2b<-T6m$~{n5 zGnMLzml2(iMJ!J?O=97c3s=drGM>iS>~CPdokyCVo?PL*`R5hsRVZ$D(Xb!AMYKkO z1(2g1_cVfDlK}Bw-oJC2PC9w|Idi99boso?nr6+MKDX(L>6ctKcg}?u%{psb?c`aP zUVg=A8ZW%OLS?_Cn&sAr?f5HbpJ?pZQC(mlSu<`xlPEs%D&AQ~{vzxAL95-Rotg9}$XqE~* zd)Aeg&%5HHS!a){J?pZIXI-JPKYQFs>gDWNpPD!8%BFFrTyojPmt1y(3cK{hTwVpY#6d zXPxtIywP(Kp6{8CY|bf6uUCJGTtVKaN&bKA!YgN;mBQsRWiY%Z%%;99nd|}4LMBa# zem|(DJWPkCH-XCXf{s<-96L0-4sKE4%3*yAUANpkr0;5eO3bUkdBpyN$xmT}6gnE= zEYiXplOjL#+g^{b9*YI(-=q(u#{QA8t} zoJ3huMvNC%1Kg<|z&Yx|9NRmC3Xsa$xS&9HpDJHdz$E&9lliF$vZLAb+mBq;<_ChV#&T8>r<%_!)PHNz( zNkdy6ufCSFjl!!?5}oR$dCS|TEIpelJE?7Q2lN6uF}c}0J#xlNzrtpoyw$KLGj%^I z_nN#?++fj$K*}A}WrUJ9X#azMKY{%FOE3K7tm#+IzVHf+v6?RYxIIA#LHcpiC(mn| zHGg_JLxWIiyuSC$GvA)Cnes6V@8;zvoN&qswQn6aPrZ%%+)1@3)Shy}Nn>@mN?U<{ zY0n{zYTx^TDGs*nmMAQT|LJd}emI%Wp6z7jS7(kGF}$(^&hKc1)Bp50i~gxUxcUp( zGY=Y;orV}1n@HjKJBQ!AKy&j__N_VHS8QtPm_L%pMEqA0`RByY8w0;AD_thu@uSR zvwVCS<2L#iFg{79|Gb}mMGC%-@jRtB{5#K2zgd=V@DKU;PR1{m>9gmn2%`};8p1+< zHhd!Evni6vKRew^KaFw4e}iA_m%o5b zN@A@)6gJj=%-)6c#X_oi;uD&D;Pgb z;+OjA*D?NXiGRk&H#4sE)0BUqkMCssJefZG0Tn^XznThkoV>(^eobWDT7SlE^06Q# z{fZQP9pk4+`abE`Z!_aZOI+B`os8S)8$FB?I*u*RHa4i9W%$?V^$63y zP@9jT5d0klJEWSzw9#4Zxg`}XZmue$DnoH;QkDMRY>Qj%$6cVKI@OlA+5go%qI$35 zxTixG0yH#Um%Fy2-CZ)QWq9r_Igbvdt7Xzrf%54ujb@^wmg61{BCK>|Dij?Qf5s5w zQ=d+M<>p`F_{S&F^Eyorwa*~OFQ=}f+3r4{#Pi>1o>N^Lh8TYV$KR2Z|5JXOuIBh( z4##g`Ib7h!?JYCz7|>mT{OIpq{2yr3`j%YwM_LB+z>ep5q(^~x!j`T&03iO}ME@8) z5xz<{<2GIFXPkUNFa3Kp4P@&o4N?1ExCJ&PP>1&|$+cAEZg5c) z#T%tS$9d=QP`pA*E_eKJMG(=moXh(IetYy}H!Ga-nZfyd;^0s|y1eraMoh}5kMlWI z<}*iSqy5|05WqgB;j7?#J_WB!!DldjG!QoyE+x74@R?aEf%x3beMAxQ6w6HP?ds>F z8u9sfaa~0!p^&~pya$oc?eh26zUAsr~i{dRWGhV68c>QI>Q}tJO z_S^M$aSFbQahra3GkyZ<=&#pyTEU&FXfjth+2=$p-W9Tj3^NVM-_ z#^c(T%3qa|e|HMrpMqCb5n*;aP?LfeQt;V~+xWLQA^kSc)yVeylh20ayGi&dP47FQ z`qC4e+$S>n`Sd&V^PO4!oWU$ItV7 zIR9U9d9(T1!SPMHuW-8aIs7t?_e|ciC(Bhw|SNUl#SoQdA(icHBaY7zVm9%>*R1=lLc=t%A8R5(G#5PH*`9Z$%$&uCdV4K?%&If%`?qXG(lFvn)&zVT$ z*+!OiDrY6*$4FdH!>RNe8NXEG*_ldEiH;>q$G^gKP#?a{r(*<92HI@+@r<7+%MtTA z4UDU1F*cs(WAaEdhw*Nip8Dsze4XmxIFn_Z>?9R|^nR+bgR??63?x8L#U+{XFX#tX$!_am99%Ud+EOU_6dbl>dqZoXVWZ_WSer z9clMZX*%-2mT2ykC}+eYc80JnKop2N}2Vy!s84aeSXgboMfxX*}199?9~%fXlM+d+!xauB^{PFq=r7``=-yt6Xp6fC6?_Ue95>LmdL4o5Ohu@xTe9Ui?NsPzwG>;^v zU-=+@2g(upUdwhrNk&(5Ij15Yj}9KoQ+=8^k52p!ltXs#dQC6!dIiTjGNlipy62T~ z3E2QKZmDDZ9f<4Egz{*bn$pi;e7eN5dV+!Q=9KiEj9(Z?KUGzR(w7*w*|Px_>POG>NNYZD@qUMc6faApZHiQ>a@YyN&;uL&U3f|4QjSu|^IPqZ(_ccjAL>}>a zIbM>Uo~s-T%-bTqCo`hoZDQOelS>(|lj(&_u4UXNlaDiQlgaIj+uCxZVxoOr`xM5H zm-Joa%jCI?+sbcZ{4FxQUUqdzmXPN)(->FmHgt+P>jjL*)r06;k$@A;Gr2!VswZW!gySX2Y~@ji zh(8-GH3>M;GRS45>CY&($4Rs-<+9#N1R1{R;hd6>wT#=^?QzE6A=8UKX*=U_{dpc~ z)-jE?{+ilmyfVn~J^|&S2*tye2hl!;GQwXH&nU`DF3ToI8yUCF9d2QKJQZg6dAYBz z70SVK92>vKFiyUMmtOeBlNh(P$4tg;`n)Km{FRKqN0xuJrazCcId3FDw?zg0Fm)6@ zT0iIUpAyf$M`c1hRs%dd&m`Q;w|&m%QH`ox{CpgxFEAb_Q-sfEJWjvzz-!qi+vJ7f zoXc^Nbf+hUj=>ZuOGyZ`Hb86QKtaMNupyW z$4TNxUkV+L>V)F$9@Vv*%SdXMZDr6go6AvC2u41&_$gDdI02_7TgCX{IlP$k@4q(J zUa`b&8I~JHE40Mdo`m{PKFjm?2Y;W(|AF=m^^K#HvO3Omifqage*1XFRn-kH`e27` zY+554y)N9`M&mU-Yz#nll>Yx|K7gZsPpn`eWkl zM8#moX~ypg{O`#El}VFt2qM{D$N8KjkP-MT}1k;Q#6A_{s!2hz}bXkDC`Ed<)}o^V3v?)!av<)#G?I z;ECr4mD9j@Tsef#VVvaAYt!siUOktj~Ff@uW(;CKYGTp=YM`amVJ)uH0 z4l-`jnQGN|xZ7E#7yWcC;}ayVmok*BH!^OcZ!Y8NVd^K=N;()<-HpM8F0Dz)zlU)f z{eua39(Afdkt*gmt@Mv6pX`p9^E-v(*zE33?q`zpDUa;Ou>H3AR@7ddT-GG~_UQR9 zUtUU#+xXqbxT?0PkKngcOBvheI`WKHQ6!UI$U|KMo=1`yj3@DP5y!Fdvw`h&8b3|H z(ui^c_E^~TK`zV2&+2hVj6WMcYZ*Jg%NI$;j%jCHn$5zLt6nrV;b+Qbhe`^`H`FW2s zPIBPUC}dzeZ4netiMsas36QU%|LCCq|aU zyu&)iKTAXyoG17w{bt5(^<2q)!9$1}&=+dYb$)w}IfV+a>%=6+ZFJ0Jyq+>KbT07o zU&Of0=C4dBKaVWuuuY$d-+_8$_27Yc(a(9<^rliZLVUkRWz;a9WXBpf&LqV5>M3mX zoP;un#w98EYQ}AP^)Tac{!JeFFXg_#Hm5)|RG)@_@OLQvW7klGJa1eeRhp64@)^toZM0wYZ;HTQG`FvxJ{n6r{q8KZA6@17EWQ@ zCJSRIG27WR8`8{q*<_}Z@xv%j!%xwllo+@9i+zmS5d-dgYVrttpT$YVrJ&fCAXOMB5>{P3cGoBw*er*cg$aowt zDE(Z#j$=zdp7FSLB>JW@Zfm3Yj4Lt> zeXyfuEV-Or1>+wI!?S9Wh#%(e_ zjd2@&3mCW2w}NpSed`#v@ozKZHu`olZliDXyLA0+^i5>kM&C5XZS*Z*+(zFD#%;EA z9pg6oH#2Ude<$NM`bVFx>u;lfBI7ptr=_G{!1(K_7^7eM_^FypS;4rPCNQ|zJF|}Q z$ud3fl_mYy%(!iww3G4k0_ny1*E#}V%jh3depb&*Qop`MG0t(mjo*R(n(WXr?a$?b z&8)*e_zR53C=Smb6RnHuC`Krb7_+UaQ?PYh+Rb=U8QVBcoJ}PfMx3F`h?AE*pyL@& z^7~A`lH}`9S@StBn;)~B@i^Hd-p^p${SbafcrV6oeJOQv-i-|KmsF>RxvaE0HE>?G zI?Z9+R;N=YYFg6jw3hRUlWXG50?z04*(h&v2k7;cJmRh7cr)=kz(X-^%&Wy0yFXZ$ zfKwSW7`O44WML79tHk-aC1@B402N{o(L28TYdSXmM{ZsJ96nt(9-jRZ@Nx^$k@WBL} z=9#MBqpI)FI-Vzs@*LCw`g{e&oyc+37k^FvAlEX#>BpVTan+}pOk9y~_V2pCpnR8d z+;1T+-b<0-W70oyG_apN2Jj?~8}OsOy$QM} zqO7I(w8=GxvFfcX6DL$~e@gin)Q!97hd#O&t)2&U28g z?_^v}0~(z7nNd58E>PzBeug+6MwDTej?*>wj_@< z(-^-@rq?`CyAr4knmJCuZxMb^C*wAsrNp?+UiL9QMbePf6M9s>^FI2=zV|K9_-Pc$ zwCkl>PbhsIt6wMM$I9|A^vf?XZlkY{aT|Tk`zd3`vE`p< zTutYh`ipZdb&T8kyBUlhOof>A^*;U0jK4|ZyxyBfSSRB)c_$uk3 zZ9J}Mpu!!;mVON5HW{15xQ&N18NY-IG4y5iK!)gFl#+fW5$hfVIw=ixS=T)cy zlwAkLFm9{gB*ty^o0*b+5#zS{tz`Ve0sb|5I<}E?f}H00DL#7{hp+ zO{Vsn#JEj9W~Su7i19c(O8Kv3T)DSKewws==YelzTwU61aP%v=@9X}F^62L{)F=2d z=-+i#`JV*X{rs8~yui3}uLTYBKqi$xn{itkEN0v$1FIOf$#^&8HapwTxXlJsst(BR z_tr2T@Fm0^>H_079huGe=ZVRNKVt9CV#bxIdbmFKOZ;2KxJ^d784vhFV&6+Y|A=#mF#DRv@r>K* z*TA@qzB!EB=v$If{%Xc;GW;;(HW}W=xJ`yfOwsjElHpp8W0T=V#!scJOdIq1H`%ed zjN5qB!T2FEJ?~AS^lKQm*|{FZZFX*u@sni!V%@CzLj>6Ge5_^M)&`A?+uC3*<8l5j z)vtqb8-Ld@eku_e!`~jp50$vkvB8w`tIwl^_IsLYQ}9N{ZT;|E#$Qi`82ZJWe+T1- zNnF_JHH_QZuZMA)d<~|guU3=&c3-fTaa;L~De31jUM1@%_7irbP=| zH3$4382gJk-R+#m>nNgW13ni|GB{E>+V*k(DU82OrWfO_a~VHG;;h4Ygf%gKnZ!f$ z0?Rp$tsT2k%IGbljGY|Erc0w$N8vcO{%9iOHX5ffewwVOm^)a&`1umgp6iWkRxobs zZ`LtBUZxj5>1M`l^zCHa=97+AM%HmA%lyUq+C;`}{F%nMt^5Uy+xWABagt?EPK6w= zW8B7{&5Yalvy*WfeWRyQ!TPL}U%tOze=aFOu-vA*xIT3 zA|imd?;1M9e19$DHs7%^1)s~ftsOcTj~k1jN9b2dKMLI|7J-o8W~qhE(V7V(p3_9 zB$>;2l25mk<4lreAP@39Ne9+4ZsXD8jN5#^?Tp)Oz{rnN!SOy}9^t1jK8GTSI%dUL z^Qjyst_;d!KI03*d5Cf3iiA9f59=5|L*_vTG4M|KX2y?}xVU?KC*!t0WAtoF=s2ID ze@s2aocBb=ZMraxaht3yVB99FD;SUCp{hUQHeK4Bg70M9W=lsaApR^Il|PYjn=PHj z_(aJ+p(_g*xAAWU<2L@SW89_(n;Ez9cV|lZqd!4}+il@Q#%=tc#<-2X1u5xQFm9uN z9pg6sY-apes)UhW;ZyEp+(!TCPf|krTwd0pm9LS;4rie(M;w z>F;L7ZTh>D@wZSV4gI38AFT|Wy)T%^xUGKE7`N4L0pm9Nw1V-oBz|DHLU+n^V$vrlc>Wr0-+g zW`i6x32C=+dB$!2Y8~S?{>)(9CZo-a+hnAZ@xvtl#C&*(@o^Ft{(2wdHXG|)P6_Qc zF3-5lhSf3tR?5V*uh?5LgK=y98MoP}PR4D0e2HA- z>(98g{)}7e&$zYzjN5Fq^C?|_8-01kZS|{T`~=Fzw7(c{&0yR%?r&z?);^t#+w`l% z_(`&SF<$Cp+}3{16}tYm^m)c@`dP=gjlLO-A0p}FwGi^5np5!31U!%QCB~I2Yxqw# zUfdD8h2sQlAmWg#O*B+oiGT2CYlAT<_@oqkCgZQ8LJW;!y=f8SHW^#V_#raA;NeEb zFOqmxJNi_=EsU#Sk11b#=dq%R658iH#xNcy17nbG+odYGlQ#DJ(aHd(uWy;oy3{Xd4z3ayiwvLE8-kd^*jRYvQW$TdnuBkkp{PTC0;i& zK2GAIKbp(^qoJe)UyEnB5UObmZ` zQ#$cyF5@;??qK{#nf^mQf7dW>E5C>hC z(bUn~@1L7VvyD{OX`EM_JQBVj1z*8<9pZZBKwjwzG)ljYaneB#7klY9Gk&PV#a@V= zj29%1{6EY2kDgC}eNWs(#%;1Zjq$0JiQyBUA5`^c+{T|3jN96J9pg5;y_xZVzVqIC z;@L{>7jDMyKwFYcxXzckF=_y1=ffn%Z8AHP@llcwLgyDTZj;%SjGrgdi~Hj?rsTgR zCI5=gQpWZ%&=|(2Qlyx6oW!`T9cQNGzld>LJFa9rt{q8^HZpFrxmy^w+3bq{qKYN- ze~jNmk-Yj}s|19~pTxM$rp;vBCLfC!ub26Yy{Rix(r-*De@jaGiUm|Kz7J0SnEHu6 zZ4BeK`b}coCf_p|x5@V+#v5e$S^qmND^tqfm{R_hl=3UCrh;*QA^l_OFUlXoxV8R_ z+vIm9MU2PsKM#^tGX8N%pWZKY9tq)Sgm-fsTO0H=}d`3_1MU9 z)Hlmb+vk3wHhmDjh4J^{wjBY6ztHveoA)Xo^4C{cslIy=<2D^x$@mpAeO6EMsroZsFLB{NZecvG9aR0lKnd;r-WbMDrbveV zPx!hwiSf9;KaV6c8COGZlb-Zf%!@4MI5t^Z#kft@x)W%iI`*gFm0v_8{KeIQ($_E^ zXJ7Kb3ygmpzdas>{Es;t=QtTh>;+uHxXr& zYZf_a@|@2gy4aPtxPj?NqsV9^?@|k@2KFW^f!^-`>pl zN2wsgM{#bsGl2$bhZ5s9KJ_tfYX|3ADmcE5lz%=2uS>yaFdo-$QvS^e>GQxl8UHwb zd;AIYs~b7aTV))v2ET=ITN_kdM+xm?-Z6~ZbY&9bHeH;_xJ_3UrR2Yo@r#H^Pe%Op zy^SgOmX!P}Iw)huvB}^V#^dHCs0}AEZsX5P#@|AP82ZINqeYCz*}@tmX<%Eu0KY3_ z9%jrVe9@j1S_V__>Mv6Pc73R2+~(gjGH%nixheT~Fn$6NVR$9_tu-m>dlsN4Xu0C+YiIj#EX18XgJX;PDh1wli+?Uq;?Q8RK3W z`p4v-)fYn%k4|Cy5Q&TR_;VR=kT`4>b&r(3iSeT)uGLwM2bZSgzcvMbJO$s*xUCIF zE+xVoC#em_a~vBFrl#QYQ}E@C+i2`!Jc-85DP`V$4`dD1&I|OTnERDPz07=NXTyV;<@17*DFhOpar#!=e-#R;J(^ zQ}8W}+v-qp6LBC}-fK9HtqujoZL%^u1z((kuVOr|4%E)wjMq^`jGVycn7so1jN9n1 zTuuq?a!|v#&BrOEq@T^W%|~0Dl71EAU!0|=VceE~f$=kCdhp-7|A}#BdQJY>FL`s2iy6OIrWbp(RxxhtGrAeKwNF3e zwtm0zt5mRk%wNN}O}-0^$N4Sf!_8)VikxE=_kZS2SJSn+|4S)rE#o$S^6?aWdkQ}C zYl;=tGjOM*;OC~`O)2=&6nt$8{&)($Jpre-8o5H(f25_Yvj6MPQ`B)>o7~M{+$J~8 zDd{^?(w9=w_obwFZq{|R(Vu7BMt@yO`WY$dn^V$vrlc>Wr0+{f?|eO_{)}7epOSt? z3f`Q8cP5livRY#N7>lgts(ib$Ev2jx-$>@g@hNyi3O*+VUy_2aPQf2e!MCO0BRW&+ zpMp1};B!*&B`Nr7#t(JyXZldFpYh=od|L`W;+vGQUE^3T+Vqk zlF$29d_P?IPn;&_{4x^%(<;6n9*qC1Nc``p_628q6{%1@0bC&SGTEbtjg#RrP{tfUyYyD&(jqnIJJzBbN)xZiu`v*!hgm2RC6pt&kG#>+EDmwvqb;X6{zw4_U&`vxI-x68;fO_%uuS1(xvXk?<+X zp8I~LvGdw)&xpj=;TKxMFS3M__Z$7oIUl!#Ut$UWq$T`POZerM@K0I7ue5~EvxI-z z68;%W_z43NOZc}U;h%th(KzU#*K5r>oz=_RK`{6qL?HvD8 zIh{{0$2a*9oO4!2%GKewSi)CX!oP0`Uu_A$)e^qO5`McSe61z?PD^-~CH$^P_-XJw ziT^Wcz=c1<&l@;=)Yu^ZHQ&}n^3mbqb0oG628?EK41y| zvnBjFOZcEA{I8bq7cAj_`~TG433!cH_c#7S)J)CPET~x}2vIXZ%o4MbMg$>(h?%OX z=CP(~O*PN+P_wF{Rb#bgH7jZ?s;K|H*V$_&vd-6|zxVgN@Ac2stN5IC)?Rzx`=0Nd z1N+z zybAEDz^efd0A3S#An-cC>j7^7ybz416f?;lM`#9|e32@Ug(h z0Ur;1BJjz;rvRS@duo*j4&;5mWk0-gtWUf}tG z7XV%ecwyj0ffoZ_9QX&oeSnt)UK)5A;QqkN0WS}{0`N+}D+8|ryc+Q8z-s`n1-v%! zy1?rLZveaz@F3ugfj0%-9C%CM9|3OzJOp?s@V3C)0dEhy1MrT(I|1(uJQ8>m@NU4n z1CItC1H2dTKEV3|?+<(+@Ik;o1wI7$P~gLW4+lO1_$c6GfR6<}9{2>{6M;_#J_YzR z;4^^F1pX!PS-@ulp96dz@cF?<}|KZ&C1pW^2#K4mPPYOIa@RYz)15X1y zE%5ZfGXl>H{5{}VfoB7r9e7UQxq;^ao)>t2;O_%32)q#R!oZ6HFAlr}a9`l1fcpV2 z1Kc0@hrlZUuL!&n@G8Ko018)Mn8SoasTLEtkJOp?s z@G#)vz&iks0NxpR7vNoicL&}Bcu(NHf%gU8ANWAvgMbePJ{0(H;3I&K1U?%07~tc8 zj|V;x_$1(;1D^_fI`A)me+hgR@HxQe0iO?i0q{k@mjGV|dj2=HUTPXPZJ_(|Zu0zVD>Eb#NdF9N>= z{4(&Xz^?(n4*VwYTfpxC{{#3v;P-()0R9O0W8hDL{{{SS;4gsx1KgF|^Z8do;E8}I z2A&jna^NX}rv{z|csk(efoBAs33wLZS%GH*o&$I;;JJZ&1J4Kiec%Ow7Y1GwcroA~ z0QUi25_l=#rGb|L?hpJ!;N^i=1YQYv72wr?2LP`LycY04;B|o41Kt35BjCZn8v}0w zyeaVJz*_=u1-v!z5a6M}!+?ha?*KdkcxT{|z`FqN3cMTeXy7rxdjjtbyf5&6zy|;y z2z(Im!N7+A9|rt0;3I&K0zMk}Sm5J27Ej49l&=2-vxXR@V&tI0pAb& z0Pr7x9|C?D_z~d8fS&+<68JB`PXRvz{2cHLz<&dN8Tb|8SAky#egpVT;J1L^27U+l zUEueCKLGv+_+#Ktf&T^kZ{W{?zXbjYxGM!XuLPb5cw*p5fF}c<0(dIm?*dN?JU#G? zz%v2Q0z515?7(vX&jmaWaBtvwf#(Na0C+**g@6|UUJQ6~;3a_j0xt#J4|o~i{=h#3 zUIBPT;FW+^0bUh&b>KCC*8*M}cwON2fHwf%2zU_iVBn2`Hv!%Zcyr(_fqw+N4e*bF zw*}q~czfU-fp-EP2|NmTH{j90V}SPp-UoO;;QfIQ1pW!|LBIzC9|C+B@Xvsc1U?%0 zSm5J-#0K5|LD!{7(uLit2@EX8t0S^RT8+aYy^?)}3-Vk^r;6cC} z18)Mn8SoasTLNzd{3GCPfPW0UE$}el?SO{^?*KdkcxT{|z`FpC0^SXHH1HVUJ%RTI z-UoO;;QfIQ1U?A(VBkZ64+lO1_(3xO{Nz7+U!;46Wz2EG>fdf*#@ZwCGi z@NK|%0N(|CH}G$P?*+aO`1imM06z%)N8pEm9|nE|_%Yzef&UErB=BE>p9X#w_&MMg zfd2;kGVrUwuLHjc{1)&#!0!UT5BwqUKY>31{ul7Sfj>z@`-_Ci z%{RfB{mT-&f1&RVmUu4}Z(j0py#qWc@MOS~D|df0S`J%5>>uuL5X%`0_HXWQ;EMY# zweA-d(>JpeQSsNviz;`214{+v?r##QsN9b}F}SkwcI5o6S+Os>KjFN3EZD!fKM^`W zd3yTfr5eiJpTJpDc|VG;rQH1qY=O#ur})~+6Vq=5*H!NRq>*~c7f^hC-IXVxEk;k}f#m&^wR9=Mq zOXX$Bzf#_me3tUAH5TjdML4=Z0wenj~Z@}HI8CqJn?Eq&qkFUkYRe^nks zeoFZ`^3%!}k)Kh1i~Owev>6OtR9=bvH{~JZmz0kqzpQ*E`4#1t$ge5yp3%^CSovM@C(3hVHuO|^bMn8G zcO!qMd=~lN%J-5#SN@#*h4M66482rdj{G0xt;rK6i2uSTzfSI_{Jrd^h%(A6kbkJWKY0b^bI2_fz zzaVd}d@XqkxgyA@8hw2zjLPMdV$T=g4jP)lK;n^6ttr=P~io$|sWd zQ2vg$iSMU;Ciwv6JIOy$euaF9@@#oc{-Mf)$%iYSPX3wli17!?6BIT6Q29CXN6Hrz zGx2{aZ(iK^W95&?pD6G5fr)>r{51Jr%JY>l@z0cZCjVRcD)Q&buamz}p3BGNf2I6m za_7=wyq{s@36(D;Po(??c}nH)`I^M3l{Y6(t9&$hF6A@Hb1OeY{=V{ack6Y{ajvz0Y@#wiaYpQL;m z`BdfW$fqfPKt5A>Hh`6BXvlozjR^1S0^{vY?JDS0yGlgX1SKTH0u^1sN_DEF^sS}UQvCwVF5tI5kM zzf4|UdHU)me}MAZ1?wVC6^1+bDlZ-d%ZA4U@lz@)hL$l^3XK z;zucOMgF<+uH-9~=dNY)tWv(6e6{k#fhK;h^3LSnDgTcAd*!~hO`ZeF>ysZ;K7{-S z<>$$NR9>-;$$wn=Z1VHUw~_y*`~>+WELX28o;5;&c#tJl0-bTaM1=_J^L(@Cudr;|<(PA8EboKF5cIGvn%a5~xY;B>O( z!RchngVV{42d}KmfR~F-GSq8cE;7d4g(?N#^r-K3yPCI`O-b9%J zFIQ9G&44!t-U4__;H`jv1iUryHo!xGe+)bncw69Mz}o>22i_id2jCrnM*!~xyfg4f z;9Y=60q+XD8}RPHqk;DT9s|55@Ls@s1MdU8FYtcA`vV^Ud?4^ofDZ!xDe%F-hX5Z6 zd>HWIz&`^%0{BSaqkxYFJ_h($;NyUg2R;G#MBtNvPX_)u@F~Ej0-pwaI`A35zW_cH z_?N)H0zM1)Y~XW%&jmgY_}9Sa1784qA@D`O7Xx1cd@1l{z?TDG0emI!RlrvRUjuwC z@O8k~1K$9ABk)bYHv``S{2Sm~fo}u89rzC5JAv;4z8m-+;NJq@3;a9a`+$EBd_V95 zzz+id0r-!=4*~xP_+j8jfFA{Z4ES;2CxHJ9{3P&Sfd2~o6!6o)&j3FQ{2cJ}z%Ky5 z2>dtTmw;aeeg*hd;Mag(2Yv(iP2j%+zXkj@@H@c&0Dc$vJ>d6&KLGv^_#@zd0)Gtr z3Gk=D{{sFD_}{>v1AhVhCGdZMzXI<0LceqUKY^z=IM@Gy>*u%4`@LL=K)img+a0f; z>vnq*kS8f{{T#SEkA4o^?fN-zx2FX8Qvpv6{9WK_fTsnX4tRRt8GvU5uAgsrkAr@` z-R)UGp7(%f1+JfmcjwOz;`MX#?)aP_J{R!Z!1Dn22A&sqKH&L*zYn|s@PfeYFC&{l ziSy&-Dh%R_051w${~m#9#>=IDkHGC8fIKCD`vCU^UJ`gI;QIF(+;!^TYjC^%y#}|J z1^NAfmjkYUKf;~AJc!r7H{p)22;wUNuMAxOK7~7fRS;hdcy-_bz-s{4zpvr0rxu70 z1g?LN!=0xNh_4I09`O3W8vt(zyb;cLLrScqH&Hz@vb71>OyKci_>$djO9C-V=B) z;JtzO0p1sQKj8g=4*)(8_$R;z0sj>EVBkZ54+TC9_;BE#0UrT;B=AwdM*|-Nd@S&B zz{dlh0DL0wNx&xq{~Y)f;8TH513n%24B%e?p9%a+;9mis1$;K}Il$)vp9lPF;PZhm z0KO3TBH)XGF9E(3_%h(jfv*6*68I|MtAVcpz83g8;Ol{J0KO6UCg7WaZvp-d@U6hN z0pAXM2k@Q1cLCoGd=Kz%f$s(W9q@g?zX!e__yOPtf&T#fN8pEm{{;Ln@FT#F0zU@) zIPep|e+GUM_%FbJ1%3+nY2asop9Ovn_<7(LfL{du8}LiOF9W{<{3`Hkz^?PfWHSkEAVWrkjyB_`f@@_8(@)QDI7(3K#$Cm@~9|A8AyaMowz$*c-3|xP{fxG^yAif&#>c9hl z>(5Vcw_6j$>(5ni#|MJ=+Q91ouM4~$@cO_T0B;Dq5%3`3!N3~>Zvwn2@MggE=TW%F zUwt$=?7yfyGPz(at43_KKgTi{{9+W`*;-X3@d;QDhs+~XVp;yVHF3_KEe z7vNFAy8`b9ygTq{;5~rH0PhLB7x3P|`vC6?ydUuXzy|;y2>cV^gMfbud@%4Kz=r}K z27Ea1&w!5rJ`(sS;G=<$0X`P^IN;-fPXIm<_$1(yfqxEs3h=4GrvaZ1d74C9|L~^ z{3-CifIkENH}L1cUjTmz{2$=2fIFAsJ^VP=|A8k2?gcy%aQ!)T?(^2fAU+B3q`>v( z-MRDY&%1Mb3Xmry@KnH41J|FY=WbVjo}Sy&f;{Pfrw5(^ct+rvfM*7t1^9cwvjWcs zJUj3lz;go61w1$KJixty=LMb*cz)pT11|u)An-!K3j;3#yeRNuz>5R_0J#3#Nca5r z0r9@TO9I!QKk3e28pQhnF9TeEj-@-lKZq{}{6paS^Dy1{D}eZlz$*c-47>{Ps=%uO zuMS*)o~OHC`tv;9UK8Z01w0UVZQyl)*9Be=xc=Nxcl`}Od_&-kfCm8&2HqHW6W~pO zHv`@rcnjbyfwuzw5%AW)_2;#^$3Fzbe+)bncw69Mz}o>22i_id2jCrnM*!~xyfg4f z;9Y=60q+XD8*u&ky6)u|4dQzMj{)8jcrW0+f%gI47kEG5`tyO^{T%?}2Lk^D_#oh) z0v`-~2=Jl6hXEfB{4?MqfR6+|3ixQ?V}OqZJ`VVJ;1hsP1U?D)WZ<6zp8|X;@M*xO z1D^r>3*a+>e+m36;In|w20jP)T;TJ7e+_&-@CCpZ0$&7tG4Lh8mjYh~d^zwHz*hoa z1$;H|HNe*bUk7|W@D0E>0^bCDGw?0IzX84#_%`6%f$spm6ZkIRyMgZk{w?smz`p~& z5BT@M_X9rw{2=fjfd2^m5b&RX9|nE|_)*};fFB2b0{G9sPXhl1_^-fE0Y44=4DhqS z&jCLV`~vWcz<&dN3HW8;SAbsyehv6_;5UHZ1pYhlTflDvzXSXa;CF%F1AZU)1K(@E5>e0{;j2E8wm~<~$?*k8}MWctYS_z!L#~2Y6!O zNq{E>uD@Ty^wG`GXl>9 zJTvetz~2L&6?it_*@5Q(o)ftK-V^sY=LYe4fO`Ya3p^k2{J`G_UI2JO;QISq-2K(x z=i+w#eJ*a--{<1?VxZmPz&`+90=N%wU*IKymjYfIxF7H`!1edqxW~sI#Fqp9A@K6R zD*&$uyb^Hz{X6dZtAO~bz^ehT4m<#O4d69_>+cD2*RQ`P$nCX3o;twm0 z8v<_xJP3F&aQ(eV?r~@W;+q0*2D~|N{k==>c3Xn@R=__3-Wqrt;QITX-1X@1dvbdy z$kP^h81Qz$_4i4++tuGE<@OFBPea$ z9}fI8;3I&K1U?G*Xy9Xjj|Dys_;}zGfKLQI3HW5-p97x)d@As1z^4PB0sIT#Gl72z z{43zIfX@a#2l!mz^MHR1d_M36z!w5v1bi{@CBT;gUj}?R@D;#U0$&AuHSjgS*8*P$ zd_C|Dz&8Tl1bj2_Ex^A4z7_a3;M;-k0KOCWF5tU??*aZT@V&slQ(mcw%jNAlC?0Nh zfX+`d$J4m~`r7}y0FR+OKT4iV_FV(W4~eI@KAij~@y6B{k?(qk8|`BMa)kW6#OJnt zn*5@8F6&pwuZXv|{?c(z{Yui*jKdyjmp{?tQ^!5UPons1D*tlw8{!>pyZgw07yrQe zS@K)rnXEr1zpeb;W~RkXczj2`nxtA!K8ep`KWS*5<9QOklgc*dN^1YIgW`XYc+UR= z`5E!`HvS*-Q{o(-;v>(?y{hBhuAd~nmd(=mC3dzwF;C{F>&<8Ax(6?tLtG}aH1 z7Za~!@3_1mKPR5k)>FH+=XSe0?(Hfe@!alQ@{-Drko$>eT_1bS zf5iDnHBarP1@W02_jWZ>^_L-UEWXR-{%PpAr~Iuc{x^yL%Er$J@lS#KgnG_j(Q$8= zuBRD!Gif)MZMPf9GXcaeA#W*pa@st5K%UDK|B=LVyJ^~*Ixj2FL9XNLkhhU|9@`M| z5apd5Pv_c~h}+9*>-@yJzy0?j`_A7f&rQjb&AwhJ%=0)`aNOJVqr~S*;C@3(^54bl z+C063~Rj+=RP(Y{h4iMybE*0b9^;Wb<5DeDPRyP=c)Gr&18_SEhm$Gu(u z*z7JBf1}u8a#v!T)>X&;C3^=GzhB}HSno)lP{nU@+*7-cC_a(I^L%i1H2IT?^Zdz4 zo>-jYi;$-jFKFwjNuF9ful0`|_f*gKz%KzW9N{_7c*nh6$7Fm8*mjqbzbp0dy!sC0 zISl+1@GPB7yJ;jppP!a<+*5yt03Qo{7x24|JIC|VZY|s2B%R&!wsUqjl(rt@xVI~< z#6PhoNSDY%W!!lA7Kk+Q86}>*Ho1h`6#J?c#E_u2nbT^!>o2e(S4x_J2}_R zw^1qGZ%EPI^Zd^N+!uIF$Gu(Iq+OoRL&q^>atD@9;{LN> zw8>vc;+?}7`sqYoM7*x`IpoE}8(Ke3{(*R9>wY~s3wp>?+ z7|-Kb7V z=fa4$D_F%}Cm*WfWBQr+#uA^&j>8M`;S$f=v26WK{0Q;lHogV<7;!#Mo=iSUoa2v@ zj}_&Nu&!3-wr~K5^^R>k1vh$%R@Vbth?RlFL?!eAB+<&fd zJdOLWul@g^Gj7HxlRl>|Vz`&65Mb4^y5dlBcN6e-*^v zq4=f}&-D}=Vd`ln{=Usqo4mPr0qY&eTZr?x#gMlW=jAn-TwiaQL*82Ac^tNow-GOB z+x^*bPvdhFc*2pUUmr^z&R^1TPkCwpZx8%4%D+$6Up{X63dDZ{{1@P9Mw$9UrT(0D z9Ksy;RDXZqGk|ZT{9%%xm*X$w9mIJ#CLe9;2^Z)1BIKQwS10c(&gi&p>uKY-x9f<^XCD9FApR@hzfhjRQV-AD{Np{h zTN-#(;0=I>0&fp|0PqRGXFKlg8Y=xNY?p8G3C4$um$lxEe1v#0>nSIC9)~=Rd%H$S zd_fyufqaZO&(EeHPjBE;fv*St3-Fh~b4>EwuWF7v*Eghoo>$T2*VTL*LmsH&pMm_T zC!74^B%a5)l;h@jELoNWG?M&@a@8q*ti%_v;~YvpUc83&;pCIW8(3fNcpCfnxPP3W z_=yr<#J>2a1@RZj`L2S?#qqDmr$`>I zr`QzJu8yxqzC_j29{3>gWh#Cz`D|&I>$wWz3r;oltWbH{kbfn4xc;v}{3-HPDo>JW zCckz+@--^HJNY7Mm+Kz|;=dzbr}8`?Um$t7{-V=OJvzPx`399|5b%ZMn^gRH@>S9< z*Z&a27oK72*`o3^CSM_Wxc;#qeiQjtmFG0@7ml0tVs;Mqda*Z&d%dXpg{fzgw42q- z?GfZ##JOLi$+wCRvg_SW^6law*6)*lEzaAeTr*9(dfrwh-zo9jZgcY8;@s{~@^8ht z-4*1^#JSz$jbN%y7a?N2VpyqIlj8%o}M=dCBLuQol1UJyrJ#y&*Trp+gpD|{#2anFS@|A`&hiWjju)i zNSxNWa{)1=XPVre~~;qZwHenlK6LRo`vM^C_hMk zM)It;@h`~}OFW+kz|P47w7Sx zO|GBA*-TzY;&~iykrx){@l3JO)KgSElWn)GqvYr1b(}o2IM2i6YfZaVBoA)~g2{FK7;^nQ#&+^*l847F$vTt2x;T$p zW%3&0&Ss5%#*)_(=lU0r*A=g9r2V| zh-b8(dxwc1D8AGBKJs4TZLODc0=!-NI_eVgK@!i)x5O?J|EV}{2NsbJ5$AdG;cgQ@ zRGjl1CD-={&yx?6cpm2jdrTgE-!K>XaEa&TUY~q~I4}1e{z5tqUntJ&QMT_*yUWCRJ?cw7U%agyx5ec8zT{!@6%xHOZSIIq{;$@Tra1V5Viy%N9F=9xmS@9XU#-zV{D@&=lL@p#IGXP*HQPAAC){jK7W%R6EAK%o$ZLJ z^DC*Ri1iTiE#iFK@+tWV$;16UNdB|#fNji8r_2m;8ZvGwajI zAB)eo{wMh}@rKsl{nga-RCx#T=i)qWTghLFbDm$wUx}Bt`BR@V`4c3y|8(&<*C5|2 z<2=X)^dnCw@x!g}B2OgV-umCxUo>9Djt$!zZCh^P(ZNJE~ zh_|!x1-(qA$g@j4kDLE_Q%`Pj9=ERK z`NRWl{cFg*#d&*jiCo`*OmV^Fe_!HxoEwlA5a)S5n7oj9Cfn{F^1{mRk{1#8wek5c zns$pRZ%zJz^6BLIerxC7JRhI*bKLp8a>>K}+DTqk`6Kc&%FA3bdHj@5ATO=_8hI(@ z#V?yYC6z~$`zqg0?xVcx6_cm2IDd}8#H+@einnyRe^y@eJU(96O?-2SkFfF8$yXI6@jN!Z26ZbI#vQp3dS;Y&}EByNDOFewe(gxSw^` zEt99aI4?&Za(&;kDR~cx=W&}#9wW~EdPv?&ys&LI`EAo~A91dyJb6EHuBQvRzHT;| ze1ODr{$=Eel5jV9UL7WXNBKkY#NwPM(;ZXiAj!|$AAj-`63_h|K>n%3r?v$gA|EWC z(|U;`go@3-+it~6rBlnR!JZ=RZn))ArC@5h}dpD*z|ZYdv`c1ue=4Q-yHhM`TjKZ==*4G$d^fej$cZ?LY(XQm3)wlK(2s%eU}LQ%^V5?mY7D%Cr1q;-kfR zzI_Qi=PMI`TH59Lvyfch&&uga;KZMmcz-)?3&{2TsrM3?`12B<$;MA0?=N-oyvm=@ z#9xs3)HePA`9<*$tPk=s@xO`l@qkw%>*Bn;29oRhV(Ai_ zczvJCkNlR(Q-l1bI4}1o@;l-@{&UIiiu3w=j9lLrdqRF+;(7iQPh#qPD4x-dLkzjT z&$WVF-_N>A{#f$x_#{he@;?&ie$^zOEX$YY;TYgY$e&6c9*4}yO#Wx$JPvKh_5HFh z$e&9*&(Gh;_5HF;$xZ&363^q-j{KE4_je+>zW=qGT;J!qPd-PDL(vqbUF{vo6C|^X z$;IQagnXXl;d;)Hdr3U^_q~)RzrN2^ll&ct=k>7{`4Y*`>%~6u#1hZzW1dtde-d$C zA19F~73cNwFnMxuULTXBHhEHtbGr@5Q;TzddyuCQ=k5PR^0eZ-{m=id$)8@FxBpS( z>!iQFcEIP6XOwvMpUF3>_$1B;_j$WAtN3!{`uQFjpC!*Np3#m&j&vq}9&yg!kUX#Q0p$9A*;?}aD!x^E zlmC5j-p?k?V7#DsW?N4g@xdsoIFsR*V7i{`hLVD@_G`_?QSBkBOYPv{FA(aIPYJ|JHM~!d|rWgYa2h5T;IRg zNggcmyx&Zj-Q;N^&d2`^$(xCDokPi6h!3*uo+WQ3Ue9`Z=l93FT`6S0%=5V_d5FYw z{x;;T#UpH<&&fl@c^rro~0-r_B6d_Qu1 zpJgR^Hnm)SBkwPHxW6vv!npHXoOp!I?@yjn^6+-45Ae<810@eH$G^!3iSuz|d2iEh zZpqK*_0i;mC7zF?HjsNud?Q=uHS(bn&+BQ)ye9u}aqe$R@*znp1C zKMmhE@#V$0TCZQgxQ?$}(0E0OcMd!0CtD%o6Qy0={%;`H`PUUT@s%ZybDB&){fZc$ zBzbtf*h{Y4on6$#SCu^1ZJwZF#y?kiMw9DygNmE@>XPSIo97SmsVYz54?MSffxL#~ z*>3ZMmoV|uRi4%4y4~77Ccc*BS!(kvCI3R@c}%X`9pG#7)RsIGY@Yq(UrL@L)?bn9 zcH5LRdFo1@o;H3V`7D*^5V>wQb19RjzU1NK|CZ!)RGz`)y4?fh4J8k6w_T-8{&^}- z0dn1LC-NZ4!}D!Exz2wU#Aom`?ar5Wd40@W#`rRE-oAO2HNHZew?FmB_5H;0&5we@JsTRG7h}I-bKDi;^>uFulw0lCFkLSmd>-(5H$@TTp!{k4!JTJ*Vk$Sk@s+CN;rzJkZ*3*gn ztax$jGsw@2^EmG%zbMYfSC7aqiF3caE1PyNi}QNakX+w)`qCA=oam>Os8GoKv*U0ZlJm*PO z)zqW!XH_DD(c_v$R$z{Qrpa@%AWkS8^NX;^!+|0VY17IQO?2c_MK>4j)UNR6N2a zyhN_=L#3%<^6UFemB{t|ritXqC4VFrT+`%9DPGum0djpGsym3EL9XvNJs?jl`3Kni zLu;9K_5GtQ-m#Bop=?Sr$L~}lU_Wp^(Ex`{?T3XOcLMM#!stl@*I`&r{>tl~Tm^VczXvWfG&4I#@|oA`X zW(qR-%ZPJ)2)VyF*FTs1LvgO>3V8zA{_ycq-e8lzyu=r_^QSU-MR9I-FnJ~A--0|B z$*V{_&!3EqO}o{^dHysY4-n`1Gm^ZfIQMHWd0OcgALl1)V)6${Jg=u~$?J&o_WTlg zJ>_Yenmi4}xz0-Djl{XmPUOMjyggY=-dH@)j`Ml){8B&9w-L=uyGx)tFP;JC+{HfJpOaYBgFam^8tBhaUO@DRwjR>cskqPDDp1iyd1wE?<$_j#&0F> zCeFv1|By$E^Kz{Gk!iP{EZ@R5PdIsu#B-hV$o2Ku-^qJQJZ~Qgw>J5Ei}QA%GkITe z?(ar&eckmYc|VCSZ0q!HWAYCW=ly1X@`2(!&ySN266bM95n}R$sBtJjK3L*;99oix zN<43`z91hW@jMR4$cKsZI3)YnwELMjk3(&8ecd&Ne1yaow&Spne3Uqk!(;N%;ye!3 zLQT8l#kt+V+7fu$Y)5rudV-c@-M`Bxf~(?Qk>({w=?-?iSv3FL_SBH>zPPCSDeRl6Zt&l z=gGfTo;cjJtFIr)Bp2Bvyujpv%oG;_t!TNv*<2t@dC*w<0 zo~)fc`y+CFUF%|G0#_c_z?8N*S3&!i{as8ReZ4Dfl<|d9C+~M!k}nqL_2nS>QswV; zHF=gRuSu@2Uo|6NDe*j?2a)URS8K^vNjz@{E|9Mg=lP$#o2h55c$giBX!7;q?X9mQ z-zd)eue{w&o^QmfC2+rCH2IIRym&qRiF~`nr?&ZXM4LQ2#Isr-OukK=_ovGp_cRX= z08iP&b9_DElYyTAo+rk0{*J&`IPP5clYa5?{h551ctbl5-aSqHNf}RGA6q+~$2C;u zGhb((3gVBF|0;R-`f7?^CjTC3m*=xT`Cjpwww?~;`^0%ZEO*?~a#=_5`z1b$&2xcV zU%yJ++tj14Q~8r0kUYF!>p=d4I3G7oCqE?4*F7$i9~S3&Qui_K>g!H^?c1V-qemmy1ul0#d$rgPX3EH-^c7qeoCD4FCjl8&h^|OKPS%n{l@)F zyNSgM+j{1ZUl8ZxlN;o}i96k)pEmtXp3CAqpOX$SuCM>}A-|&HH<4czuWakdHqhj` zA)e8Cck-L!wXClozb>B7dV)_(9(|ps0{Jb8=kXs!o>BRJ^4sD(e;$$R>oeH~nRf3; zd~;h*d-A*DJZ_iBbBX)f_=rzUo_pfF-0zY<5a;!`>tGZANSwDP_sJiN&$9VP4KeXg z#d#e5BL7RA=i$dgP5j^DJPu>Yi;8ocgPEhb3J*-n*92FwkdfbiRXGI zkQWi>dajY{^UzG=O#WgL&-FATFD}mWa2$CF@j$!Wewd`S>uLJW{-hZTAj&H*sEHKAL6nM2i=;@qNhqh(EP{n!KMl z=czT@cYw{@ZN;XfLIVQf7IIl+$7lzj-N-ak8@6vkC%9^ zGu3?4?gVk3=VizzDGw(9Tzt1ZkGM;oMVx=nBH;qh>sM~ayzzfW z-Ie0Jyh@X=7U%pY$k!+zz1ZYgtGv+?;~T_zJMfTvoj9*|OP88>y?<=E%(&j)Jt40l zd3e36yxhcZ5-(!MVHEimao(?;CEu#N@(Pn@yEuDL&D4Tx=`M2UcZl}q2i}UrC zA}dY)z2dz8YD&IOocBk)$q$Kh{wd@?iE}->$&ZM0{5kTY;tg#5RacpMj)`--VdTff zInN;S6XJDjo@?Yki*ugDt4;os;ynJf$$u5+`9GHYlsLD$ll-(e&-27zBWF3y~%S$ocnu* zT<_POkYAN}K7TBKuKwz} zV(CpL&oznX{q8{!pgp7Z3ti}U=+u-WAKUYy6R1o>_8IyTQV@`K_$KR1&P5$|N< zGjH*{zi8sP^SK`?ek1uEajvK0HzxiMniXUerr~e2&CN+i@%5xToXr`XIh7`2xwq?e+jZjQqL8^K&D+ z$X|-{INu`wN4$=$GjxaP*DG;ezLUsZvfgo?bL0t>58Y|yxLFc&=wO zd1`Uqo}402BhKq%#l0qfT5--ZkUX6@-~T#6zF(FL@8AC+&!FNPe`oUR?dJsYOcKxS zt{~4W&hzInx!&Kk-e>ab{n|eA_aqOWkC*x0#OwXq3i50c&-G;4Z{o9y^YxSQb;yf~^L}tDd2w+bpX1~u#5qrjA5DHA%&dYHaxnAzC$o29H_}S#CC3(2c zN#udb6Pz^hwUtMb*AwSD*OJ!}=Q{KJV)E#IjUjI!@m$Yo@`lRu{c7?wQa*^hi8$A@ zfjmf@>v=`qSe)CfddlQ)CSJrY_rB!1zdOlWNId6xO5Rer?`f02mGZIVZN$0$-Q*vM zbN$cBKNjaY^PDmH^*A&o*Zmqt-d6H({hP_dls_PEr@Z!A({2ZGu74nTxH#9pnLI+A z>pxGf$06-GlVA6@40&hC!}W)fM=GC5-bMK%@^0dl?Q(zrylFQ|oa+xEj~3_phm-4Z z*g~%Rdx1Pg@^Jl0FPL_Fit}+!74qKV96yY_k2oKvZ70|Lx<{_-&v?r?l1~<|X8khx6!FikSGi>JOcQT!eXZkp zTqDJ~o-Zz&_(9@)KWNAm@jR|El85V=4B{68 zKTLT>NS^w(zqxOi{G-LMT2Fe@b3M5n&*K^=@uh5hGjdld`%l@ojs|(YHlC29WB*(Y z;`amp*>UH2SE+~RNul43&lKnRKZg7(aXyav(eXU4m*RYEaD(D!i*ueGw@m)I;yllv zkK@`F!|FI2w$sqt;%8SJ<%|I7Feab7O7 zpBdjJ&f{70Z{xegIlkO;rIeuIM6Msycrt+#OO#Ch73&`&%e?op&dCin2&wb_7$sa1eLH?KW_fnZWPm~9e z|EatW`QPFh?WTAk`E%tb$X_V;N}W))`LSDbm#dKDd0ej~p7&n? z!2W$4@UMZpLRz-0>r*x)q(e-{=q1MCugmny!>JT0h8s(JS zu3bRGrV$ZN*4N^k20~-Hc8Lf>5x%8ELgHr%=@AjtAu>Fmq0Z;j5StIj`IhmJwOvGy zZlN(@9X;gr^?5CCyvX2~y3W#x=o%3fVTLVk2hAJXg|`pw85u+49MY|8WJFls02AgH z64I?lOr(E!pYX7rF%ey(e0%xSE!iWy`x{>E78;FT^>>Ln8r3es zR22@`DLgnP zG{*UdFa0AZ+-a^yj87@2PtL+|zbzytA|j%pPe~`jv9PenYF)c@3HVpN{C`~Q+m2~{ zpAzm-?Ak6ovRbKthHn_v_+u0t;LF4)kjr9@V33blo_c z1Sg`dv%%#Jfp0)SY(k$B`n_>qE*b6gq-z(aFJ>i)t<2~37vd!J4QWbCF{Ev1kMO`w z&Z6uT5a5(xHKI$mNEq$(##22`j$q%A|E(;so#6TQADWGQL%gx}DN!%HOV{YW4ZC)Y z{2vW?+yrq){q;o}H|ovJIF$y~2`F4a$oZYXFkB_gjyq)#HnexuAGBC_}hmW&2jt+^5cDBc8Fr5NU{)VQB-HFe+6 zMf?8Odgqi8e=yxENcHd@VQ(JH*fDsm?>;4B6X3qYje2u4Z|G3mB)V5|SFxIP55M{KWfOO> z?CvGDZ*SZt*=XGh^+Iey8Z3`1mf4h*c>7J+8*}KA-ZDIXp>d95{%w68-s+*OH%7&) z#yJdd?+O29De8T?4j2;ms3mq<{o75O(~&BX&dEwh$B1^}(Xn~-+U0Dd-5cN7V;;Ne z>37+cCH4iMQnJGj>)JCaCh#p+6SIYoF){y$O{pt%wwr!(NE03%5*g7W#(9fVcnME? zmpGf=2GX^EJATzl2Ri$%0Oz1FEOswu4j*C*jW<{9(M9a*+>!V%ga)>c+xPZyQu15c zJM*o5fQjl7rs&-%%w=j&Bj3%xST%PmEU>c zwPoDUr>y$~ylXde@E0A?%{j`5eS2u2=KyI;@_URkwfVS*$+T)`5Av8JwyAg<4v)v2u}i~qG|XmK z%-DCCgEE*+?G$<1#jf8TCu?NfJ(zFFF5%HV!lRM<$r5Eso0L9f>3=@HrAn6cEnU*Re{i44`#M+3BHDM2?h@i$bBzpf zF8Mj!G^_c(k*4Wp1Vc49pFT^h0 ze|v-+rxK3`xwtxy;k<&xmc})|WqACacsRn14S8)UJL`Cy;5XDAC+ZF5zA-Ac1R8I3 z{2Tjvyb>i6*C!#kI2_^(f~QNWu^YVD?V6`JZqk_*W@Br&WAR2-rc7)u4}%pyg}O{D zY0M5H-X&Z;ed33hO5^Mnd_p|#8(vS#?}?M$eMaOQAP3dx^?(0Lb!}%`S|&Q&xz-&X z6&4=SJ36#mw{Yj8kaHj!6zmrgTs9;)FeEV8KO}(utZKM=NKiy@gAnJxpx_!I0lfkO z8Z-{6-mO7nd;VayC^EFNgksOBLauQrt?CIj0VmOa# z$DS_6w*JP_J)e5UzL_R({4-EzOXc}0Z=CfsUJCFAvjL-%)c?sW0y$%ML}oJ|d)+;@ zxOfhS5;VX?+vV-vEW}w>o+dI%CDYqk%}ic%Rl`nk_jzudyz!=Z>>2EvVz_fS=D%_N zC1d@6xmoV5Gu*GQegBP4n&oBAcVnA+`7~&KaZ{3M|F(BS(dAcTm6L~zte~a1g`F%W^YWjK%H}lUN zF~)Y_jq~q~IqbHKFGj?sq#5+aw{p>VkY2l5@D_vfdOYV>gYsIK8H|5(yCm+_6MnCI zE#hk(@y5VoYp^!HIF@3v*JMWVV1+Q-| z`5YhDf|vFi&i_50%JJy?{7)M5m@)PUApUM3?g{n(S-Wvg5M!Hj<{;mbl4-7%vxNK~ z6(Cupb-YA850bO=`Jt)T7rw`EF4lQh?1}Unnv6R>xTQEz|FVJjr_VCpxF+=$xc0`6 zG`0P!Ev@I%bUSu%BkTEHajNuiY8~gXDQ5@*gPb~hHKbylvuJw}(8Jr}t_be;l`!u! zThKSZ@3n{5&^t@V^z0TH9$35K|BGu7&I;$D!)8c=OPe$K;9AZVh_cSV%%9HjTX3}y z=b^jc0H^ED0M<4aA)Jd7K{mK{NN^qJAJxo7ibf&Me?ca)Zd)Joe2R0Cg2&h=q_OF9 zh}i`P*6mUv;Qx=EH+`@DA3LwE&^Qn9yxx1C0B|8F?reR#XD6L0`TAv0g4btZ{O1Eb zJws`ht#fweEME6zs45|5nYu4N)p0LlPZyt5*Lv#pMQz23cNGyDA5O?-)vTIM$#=#0p@Jy#+i{=4e7E#dA# zy!~$p`)s*+>ehJ>(mu8147_>GF7- zM4rykBb>_YwO0FXXAokqwW{M-KE;uPN3-j5ze8Q$_S9Y-cySjTAn^$jY|uGOqwBuT z^c-I`90gU^YEk6il2APh-@jE?U#eB++;RpZZ@8>h~kF3DE?U$5PYDb@X#kJuDHYvMbzhp57+ku z+?R*TzpA_IobS}_I``bkOcu=VM|-A0dBD)zoa( z8Uqjh70DFI`X%I+Y>QoY@O{`{I6)TQtY9VfbRs)*+!KliUA7E!x@!p` zE%}zHbBmG9Gsv=zE@^FuH}1IMAzI&064%0K+RQEYjIJMB%SkTv?Xg|t6!{oI|Dq7? z;(sYbKw&Hnd!0a9!Zv9c<=^;lVHh=v>A3xgk7zCJFK_QH?W8tQeh~$B7E=r9Fwp(G zfc{Uu3CDVcti9H`{^RqoOSaG(w1c-dXtpc|Wh z*NH2_Aj)w+B-5DhGczB z#>|a@@sMnD{My{Gu~Bsv8QKBj%yh`z-~Wr?!K#Dx(H zOgvei5){pbY!`9r>!lo9GwMVd%qMNl)YmY~LBpD5)6Jj(ZkCJ`GckVg z!5ImGJf+~<=V=>hu(1)`wmZ}w=kg{yp2TM_9+)SM;*ch3uRjL%yBc;U+9?p3x@>N~ zj*bsk2C5C3Pdc3+e{wHj+f5U*!jaeDW=`)*K&i;egaq_ZAA~1p3oW8BMUxgKnJUB? zox3nYlU0>YHOI@&6!-Oy!>aE5E?F`4Ecfn* zKEiICRzCJ3osoSvxde?5ga~Ca=L)Y2e&7=~b?RY-mZ0^2{C_?p@E+uDYKpqSiG` zB?{_PFm&;NS~GJA`|BN*m6Kd<)BmqfNPv!rUGcVg^|4U6t-IE?esXktqBdH0G~n5g z!6mYdp}-p%c6qvR={G9>w!GZCu+~eLfJb^4lQpEkQJE@wy7EG0N=E4(p%DMMZLX>! zRV*A>y|z3=2lT!#;=2Tl``8SSV>_E0ufEg@Z{Vz zR5+Jw$OaMS(amI1luErC7S=$wK(nIvC3QCgP3Po>$~gJ4VZH|R_)-I5r9VIvlj3I} zXOLP##ZJj$_^P<&+ECAFlM^H^`QTJXw1?8J>3@!n)#(JR&YpQ~vg~s4Y2K`@au|WL zyGn$Pj8TUJItA<)7@K76BK;g6+d#LC`2YL=o}@u{|EqKIoC#+`XcTLVdYJzA zXeE%$8?A1vSB9wf>Bsbc)8njqSqHDs&2|}Tdb+hJW;)VSdTpX7G&37Yv37%0(VzX( zrOs8=x+|XBGul0}axkxSZrg%WUqJO3UO?rqH*PFbG-YyZlxok`Hky29?<#i)f zqAPVtyXV*Iqf4n)Kwit)tWy=~&{NCS)tfz_>6>;f8#U4&TKLkTr26Lr)yc_1#dUdq zTe8+;Z4Wk7fNff@nHoBabA$+EdN@dXM zNNP*#1f>b^y4hbJo@C3fTs$(>mLE%&4sKxbA6m4uIBs^m9p9 zsvn?gewN`p>K(E%wAv>uvxm*mEa5~xSBNw5w`iHfqxkrvUPMyTc5-sZC#nPF2aJuf zYsWrNm5Q#hb>{ySz6vlxW#d;8)?OkhYgBcHDzmwA=4Be^RaLf}(G_-TH`jNS9m^USd~jU7WqLN>%5KCp%~c`=8RS7*(i?)zIl) z$^B5KWAuiSJAP_+halnNlvRv4#(OZdto0|y$3x59HI77Ur4K@O63q{&v&5t_l|uJ+ zat)&HuoF}d``^~L`TfR3ovOnOjBoO4HbC16`d>@?*`C!Bk_E1weg7u<{=)>Ra zHEYP*#@*>TL9CX+rX^oXg#Vs&HYvZH-w%w=g#gKehOh3Y+$+Qk9~dHNF> zfG|&LlqT-Fm)#!p;f%M?V-z=Ss7EB8Ag7Z=rNgktmb@r2GOdZ5>^ruonW;5w@$HG! z;!zmAB)yx8Tl9hC4a$4PbL#xmoy4`*4sZ)BQYlbFkpb#~O_TD;`r^nbzcti@G*dc6 zw_A(t&VgKxX=2gLyF~ex5Fx@{LA4Qs7v~m6A^Kwy zzvd3YmIVjB+Yv~c3W|f?N0w`*A5zxsvj-^7QLnl{sdwV#du;J*U>IA`7t9J2RY*1r zw9fh01!Qdecqhn-6rmrkrerP*FTjH7Q#?y9Y=bcs%a;mD82^uNCxpv=9 zs85V-it0UmSAl;EwBbXC0=nJ6^ruCpCU6FLTgy`DrIuEt?4nYPgT0=}IOQxQ2??ws z6nLBV54uromapjc&F?BN9`MQ2lvWlheN!}qK54L_GBHR*aaH4ss(rkwkO6WBm`SN| zcbC8)OOmZYR_YQexGC@&M>7<%s z->!t4p@pO*tp=+&ji2?l!ym@e>N!GL89k@*UN~qRov64?d^lTQoF*+(+5^u-4#_*Q zx=Bm?(Y=GFQ%Vw(W*BWF!z?(?9rU;v6?(;;2*;edZQCnUj&EjrWfpV`g?8zBaqwwk zeUaG=w(Rndp$=yQE#@X{;cKlW3_mWJzmBCVY{SN=Xx$arb||ez8SH&x`Vt5^PFWPc zU?%&-?5d0c<#&qTnk8&dUyw<6&ib6C%Pi8=ItOsrp1kX9uRTN`{3d%PrW+E8!LPRwTFH5{1tLwpR8jLt)7BEXaF}Xq3FKm2q(*U>t){%D0QWck(_xSuQa{;g}naXW_??(g8n` zu2Y2+u}}#Y`CN%QbEhwCS)|FCujV0#@!8~<`ID6TqN3JjHjJ{%VzDKLEU%U&i7v8D z4nV}vN5lJ;^mEc?`svjqqIg2J9)3J^Y+~)?INhx8_Zc}nZj@rf-rR zF;N?M8)(J_}Y%Fv*<`=g|wY_li%v$zZ=`rW{jCFh~a zA>q;srr7dBc;(!|+DQL6W!1aBNM0NEF4-6LyPFy41nWt{UE4`-m&uZC*N98H2FY`W z$MML{gFXt4Z>XJc(kXE!{ZWH*$8{)5H;6| z3Ck8Nm`EFJchYt?yvcv`k(o;WLzJ|Q?*=7sMiBx^XGimghx@97>jva-b!^-ZwpYmQ zqfRdy${bsBa)ntsMy-3Tqsr8|#ISJrS{>< zI!Xu+X(dp>SSOCeQ%(A?Z~98@JskDwtDdut^u0)`iBlOBwTxC6O~QKav+POiu+XE4 zYZj+A1D;VqXs9Icl#LSy>hyP|2kC3{4;II00*WYL+7=ZUk3NNQ%dFT@a^oda5;Sdh zDzXb~xA5k)Q(NJYJ|#V#`^r?%V&rCTL~2T=V(ZZY;+#POs#eD)@@U%9R4zXx5#p4D zMNj)Q#mz@k;-p6-FG!`yrLwx4)(~I7+UOXu6=b-D((dSl#9&rfh_q3vpPIZO?XqUi zrIgZb4XjfkY?9ET(k!=*Eul*U!}M4yDTMUe_5RD|!31Vqr2tYsW3s`e&W+B*$euk) zCxb&>kbc536(jczhkX|pPMxzLK>GnLTE{p;K`6R=nN$! zxoZ}(pP8uljMn^>1P+R1%h*Pms7bh)5ML>A_5lQ06u(9RgZe~0ZR=e^?EUzVlgG`V z{Z&>uEe;a0NR=wc5q_X39c66kud~rsH1atFI@94?F?z8)1tZ2M8@bP*9~%4SgQ!s= z7;lh=dgl4*Oy7Dj-ayes5=##AlNug9ecR%j8^%W#snKAcR4M*G;>V{XvCHXr;* zO|{czSF%4@nWWi4j9?=}O;YDQTcmLKM7yqmvGGkky^CvDsXubfwzt+GWjED4q8u>y z1p{BZvw?)9lGT7pKv65W0F(6A$Krb09JSP+FhBefn*xBnoUP(q`d?5^#$rHz8g1XKIAKqX&W0TudY#$ zQ@V{z16|J-|EJ|tKgHB*4J4%2@)8kgkDTVB(#W*hGcZU~czW(Yg^pCHIE?JoT75ko zHuC|zb1&RY0h$n;r`y9xyR9;#&(hFSkyKwY!M>ByREt)5cU8ve?&l!Srf{RRW$|b( zE}QbhNk@gGi*%-^2~RT9U6GmW?lTZk=^0n3p}ahNtgk|4$^6*Cm`K05U663=C9(^PExR5hp5aoyk#wN5OJxFi-En`{R zDJg}bK;(|psN%!9RHT|kE?0Wc(2yRC($QgMxLl)JQQRG5izDBe?Cu>IL2KJbox<6- zZecK*m6`-uD?$^(4}8ZpMEa=c!btaeGEd~UF=v|Ap0hj|0--5O)!pHDyC*X*8>TuP zk(b4J!aG~L`34WRTK%{*ixf578FKoGb^)q}6`1j1y1KGe!A66wtIw>^7Yho|qz#kd zNhcv2pj783T=_wF%IpiAO~$jKcJ&Ty znV+;ItrbWrM&&u#;P@KF_VlA9l*Elv#z+n>n*8p%W3cMhFS^Uw-Mc0nq8aQOeDtF4 zF;~)zp)AQ6APhiWu*_tHaCbvY^GLGJbXs%cx&HqkdfSh3h;w`Y^`&Or9)*{11se(eK&akMV%WX^v@==NeWw>fZSJ)`U?ixR~Q+56Gak*wTW=(Rw>+ z4{L3wu`9A3OZ2WvM`@Jv>53mvtz%Y37bNC#20r~veI|%jTK1rb@oe!S!sV7skpvz{ z+M~s5B~V7`Ecd;h;2Ek{wA}Y(3zM(|yWuUVnmqux4UTO2jlS)s;F%Q)@XHjeX$B9O z1`vE|wOQ0B99Qsh_2O^L?4Sy&KJ2MzgEZx!g4Gv^yFpsc9mAQNWOn(#=zsbE?}9Q_ zR#G8{*7o9FhvYntD)+%>$K1Jc?*6TtDuXX6R|W@ZPd?-$KWTCXUj8byuX4w0ZscHt zS70zm@mdM-72xO=r%q~`s}11Gm%AVx4b~j`PC<3m1nsL_(en!3zFz4n5IFM?qLgY- zGf9zC-<=|x{BcRU(l5#_kA>8j_Y(;mtk}*(wN^Ilsmnt1kyY3Yl;5)ilQ5l(aW|7r zp7;tamd5EX#rGzwS% zzjnxePh(tlM!G%YBW$ggf|CJ(+nH#ejJIFUECL#f#FVkun=2j(v`+9n&73UTrW1c zy_h{6vhD^-FM@)q)i`=5BJF*PvCR8c8Ow92b&3kWhSz22`N5XGn5aVTGnNVn6J@A! zc~Sqt4a`KjbDzmomxiSvmsN|sm6Es8#)8%~E+`EWo5PN5(21frc+)gdCtF+M^rQya zAEbpv83}!=ZW(^8W@N%yzPQgsKayJ|rP%Gef{j|Yyi=z7L6CJpZ;gDuDP)-6;z&+#*mCMi|sq_{er(@%~KQS>TY zR3$Eb=@K2kTAF_8lw-ef8d#(v_gFhOP<|)LVmGRJ%Ukk5x^YR+@~6k%IU++I16DrrGW#4Tt<>{SisAlClY z1S)1IGVdui+O##yyQ`ZFDQy*n(zXr>iS#-JBZ3~V*>TBg!qK+h376Cg;M>!AmTILl z-pO0QFLpL%xsWBwqta1(MF_i^WQiXA& zYbMw!>5V9zmCBdt-2XmjXqODfUEV?eIVYrsxpBHU~_)pXJJ2h;6C$7sULL}Oi z-#3lvGNc0e=`yH(dW>PIjSVgQ+-yW6q^%BQgoUL)5zh3h}K3td#KDYf0XO^kVoa7R9tKM03de$T^L4gY}H1u=2cGgDp1U)$&x+x6Y+8;16%Yy|9ecEcn{u zeUnlaWL?mQYUmIeh{m%V(Da4}!r*1EVYu<+$c{Ia@O*F;vgGsywM-Z&Qg{NY6Hsvv z`hHQVyuR1HMiP^w^(t^L;Gdht&+(kke=buo%z9F)Q?x%as#A>nJy6*fbD+IHT&nNW z1D|)v9zXbmJaTX5ljbEi^YI~Fd?N3~l%E>o!OLy(A*=F+2{yG8x=2bMU!y?=UszzA znpq#zGPuSo+U1F7G`mmN@yO#rxzokF4rk*Om~N9=E_rgE2#x1-GCvXc<3R4KkG0YD zV{1ucwu#)+P=ki|J>|LOp3n@6n!B%L^(>;c&K|4|R5tY_B_9@yFYTui%DjW|`oKH0 zW&O+ff=2(T;`xRSqV#3oLz8vd^7R;R>D?Bkd*XAGjvBhTYR~*~|6(fKLj`->zYFN! zh2>hU)=!$?{?OyILP<=%S4b`6qjJ6>OTSQo@S)1O;W{0q4p*t&{e8Hei1x6fM*Tvc z%kaI$Pqx65T_Wna^-?k+`>5z>o>a_10yDP9R!f}3)I|%cba0D;dztKaWDRdE3%d}{ zye$Flieel7$fn$aB+c=+($CSVLM&&O2&4%_KPO_SE zm7x-U)s2`zx2 zq0II8a65iZutuZ|0~vWq3*p(~mHsXzY(gjkKp~^7b zec(gca0b#90$XDfnKCck9dmGGM(dT@C~25eC3g>IV%-4O!sYRnd8Ax;$<0W@h3;_5 z6{2eL1aBX`vqo(mo2aZxE-uGkLDPX_E1oOpU2HbRin{peBr|4iW3ia#vA)P{dny%! zD(C#pG!#ca`eVQ7S(KbA7wm$e5Qir4!k2;fNNLqq}-jZz(DZc|_^wF}O{A4lK z&dCjxad-JPaF|L(7A~P>A_~%&l_2EBR3J_*GU@ZlD0Qw1$PV|RKs~MkfxM^+)JN(~ zRX{3#?6n?kMxs zU>=;zJhC2VXU^sOL5$C$&G{HT#?E-2wAO@qiEGXs8|qxF=d!~Thvr%r;&_WAVMquA z!{zr}H`18^m*|x08S9%ZQX~LJCA!lR8_7ny$y@14)xg@2Y@}*t&@$(ekYIsS77Xvq zqt0exz{MX#Kk};CPIrK)mHFb7d#~L<2WFN1JU9Dx3DWS+#}Tn zncm5WXH2j+lT+kW*nBn-DVS9+1&mnCGvdud%zC9EAesx@R+?w-8+7EfxEm1k+?4ix z&EH2(Grf;z0>!=fUmu2Wr97N*N$+t?!4ob}T*8MFxMz~F4H;GOt`ein-=36xf>eiv zel$8dJiUdQqw~s0biZ-%#SQzCL7NS^Lhe2ccXnJfx^+MqMDnE7>o}?N_cDdM-u1lf zk~r#G{2kPdnT^iWxMYfZP~r&hzeDhauR6-#Gb-1HdQO|1ph2cyr8Fn9_545S zBS-GrgVfix$+8QvQ#sSMRkmep=t_jT@M5z1)QDLRCnHhQ=gO}>|BT<`e(LU?y}mNM zj_MJHfV`_be}Z!)o6K@!fI&HNl0(%b;H#Dms_W6zGm6w%woCadg>yf)cZG+L_ZZzS6ZE_aa{MGjBG zO0SLORujQ?qPLa^-gt-ad~eOc*k&(oOc9=M+!B}x%@I3=>Ty_JnygE8LHo>2S&EKl< zR<2p#Bw7QbfL9!UPL?t<191Ys)JrD4#G|6LY)UxigXV8Gnk;#(YGCu3WQN=H_gdj$ zIczQP@Vzmm_7qmeTycVmPDuhC(Wb$Q$Fy@LC>rUGG8th7q0M~{^vPV%7AQl;TJCc8iz#T4XD!7n@N9N^stFc-zV#;>oo@4LF8054lyqy{ zTh7R6y17}Z7W`9+>V;^(l}J#Kb=aEgL^*64r@6~+Dl*O0OqDTeEiq`gmp&in=OWpcP$exS1~odXh06AHr;aL5XUapS zyu)FNJCBSF_Kch~(%C+*b6(e+_IX|1CykWo?<|T|(y}iWDDp^GduNp_PJf1vzUa@&ROm4b2?_tnawOqbKUFh7i}6ynz|UdXEosp1?DACP$XQV5WQ%9H~x{5})`QQ={8aklYT~Cp_E3wwNIuYIURaJ5>IEzSVV+k68)_wk3NU=8x zS-@oGQMOU6h66{zB)gTCR#EIIZ8Ar7uu6OsZSP_)OqgV10bYtJ@hKc6%54(~kkqiC zl{^0|JIZOq?caD*vvHcpr;4M`CmoZYoRa(O7ImXuhiEjji7Dd7#}ZSbVuiGnhV&+_ zxwzU;6P$}e-H};SX4dXwaK$BXl3m2;>s9#x-}gH2Cb+bJeD-NiN-}9 z$yc<<0~BD{GW`Oc8MPumTZ~q+R!hU?CJFqkQcb6I@_b%@nlE$H#ed<(ec;o&6Og<` zq-=~Efgz4j_w#3?UBxEKx}$Z(?Tw|DCb$b~nl+RIkg3+fBZ;#asWg(+3F#;OR&I4` zCmtO=qeJwC(}98EE7~fNpsKhD3YsC# zkrZo56nSqdsw}6mE-k5gD$Rt=l5e57S!Rs_OErY;sZIeYH)Y`V)5XOTDW`NU{TKgB z84>i=8Lm^%P3OEcKZb5$!>Aojk#B~k&_{%i4{*%k}e|=2pKEsC>3}}^DzHXco zSt*yVhr@%5cEHAS+ZQ!${CZb|1@98Flq zwWNz`EM}WyG4-zJyRFpk^Ss%cU0_Od8m2Cz5Xm6qPLdf`F6!>$%BA|ct9w!LwPO~! zCtN$4{?D0(rryYGs}HjSS*#bQS1o*}6;TV-V{_=@uHnD+)tmw5bR(LSO?#({>h9KD zIA%e0eT{3+NJsEaHfK^79yB(L1qMyEa4xqn>Y(}t7Hi#{?po3uI_5Sll-Zvoe6#*k zg3lwtExC&cJ-49HRu_rP*4vc8CEj?JvVkGDuXWEupJko=c?G33`ojrnNtk^T38>{u z1~q@-F`C92nO(37zA*bBM0Pm7>rcjG6&|wHI@uJts?1O261kr9}z` za%n~%&3NP#p`?)Pvi@FiY^GNKk;^S~l*trxvE<+)O^p*o#p%qBYh#w_j<)o<=pZeQ zW|d6VdoI04k0r@QQOX1rME1T$5=l)C23#QmAy3ez7jfi?!lwh_fp8e$ZFlf!9b1{k;$dPp-7*~>RTJkU=cQ@}T>Fi<-a}JZzN{?%yl1Uh@sTJsm ziOH>Z86!~1lzO?PYx1#Tn@RmnKPbA^X3MYO+=8U0LbjG?`>dg{i4B#B!Nw|ql^$1= zF})wBGEOA#J3HN6j4zCuYzJt2&Y5 zcU);RjOg*a0S~IzUZ$II^@*`ft`w?!8#QInV?F!~r>ptI3pZln@ZA|D$%)SEC8`0D z_LAd}`@-L;W zoz{^*52*|0CKqCPOX161MQ56;1c~|H+{>UE>L7hhHh0h$JmiD7CZj^t5;LO`wI4$y zcapY{W@fp1h3Sac!cnfYBTHO|WR%_5OAWvTbqkVvBI(cMx6Bib2^{nypQ@uz&E+LJ z#n7FKTCHwODs~kFX)W`<3oR_Q%5d#mQmCXtAlC=w;A;3a8H9SPgFtUe2f@~5no_ZfSBaa}`jeUV z^^$>1Z*ykt8=Vai_r#!DlDKC0FQ&9E43@N`7(PiuS3cJFVmZ4^c2l%7I3I>1c}DtC zD%D7>r`Kpk>ei7A%+N`~Um&2gEvk_;F;b!L=FsJj3cG&GorBg`ftZXH#fV+q2yYB= z&9UAr(#Nz^7J%l@W-_8ESC&rJDP~R0U_-t*?SoizTI&>YoQLyQ2dscqKvOg1NqSXQ z@iR1h=t{jtK5}c-__7t)kY)Y_`>DN;B+Rx!N^*sSWAUGq17|!0N)B0s-V_I)Bj<(A zV*3|1Q3c*j#w<>Jg829QeJEK-%M&D1(y}MmTn_TEsLAf5h_tokf4$%D3U7EXvEcm4 zbfaF%EkWqrY<0HeSt}*f72+SP)GK^6pHwi+UZf~m50Au;U#=+m%N6wd)aQjL&=_)busXmFUD?9tXVnA2!`LDN;u=-Ao%!Wdzg%q3ytJ6Y8WAN8IAI3*o?Q6MfBx|nhbT6og03adGDRW zh6zU=5sg-|L)SPb+B=fEmvkX)sV39iv-hmEoQtszAc2a@$976N+E=(Sc>YQq`1Fxq2IZqmYeTDl1M| zOhp2%(bQk{%CPptUaEYxI5#JVH!ZeTkjYe3!CEZ6mu<9ufA+2@9fUP|ZIX{uAYhl$ zPF22?L%dHSF<1JmWg3IZp1mhdPSn1T0z~agav<&pHF2vPheOJZg{*yW0wI;cRLxh=IIN3%g_g?bJl8YybEL=ng3Fl<5IF)Vr3ewr^FX3a* zSun{HukWyh2{=S|l2{gX$7hGEEdlM>VF!h8kI0||joaRD2vMq-znR~^C{#P)f9YaA zWrA_llM`sB^R9JS|FWQMkEn+6h^dBA9fe2c-6kHyULHg=laJKa2D@LH+HEIa86|s% zYYVPm(RvfEFZI{CSYN3B)3{iX$E9yWInSZHkDaj&|AEWuaCM|QTJJfHK62TgtY8bq zDT+mNwX?oyyh;&t7k1;Pwm|NX)<0aO+);_K#64+_Q%2L?=~7mab9NyEXBm@)LdY~0 z7jTh?CcD3y+x59yMeK_=%P5V+GU>z5lhvdP$bWWOZZ3_DZ>RD>3(1i4S+j16>iB1R zsISP$?p@4L8}eU#aC-_ejSu@=R(i3OyBG<1HVfp&Ldo3}W(@~daEoP2FAc$+B=gP? zop23O-TKPtK-JwcatbY}%;ow-rBj>VpBf5Rz8F< z(DLUQ@>PXZ6e4TX&Genx(RIVabm68(H*zL=7E{-)Bpo73n&3q1YCM;R$5zz_%GHU9 zv5BO5Wvhe4k*opwu6mtphM zTYc8|dCQMMLVl}nTrev!_tDr9Ur=jY@$v>EJ;)pBT^w|>70d0fxrKqC=D5(%;Bc>F zh3kh>gCP*q(b_`#FLm&7TIld>$W1PWMWS?a^V*2C@wK>yJPB0A2AQ*`v+gb#S(3cU z-X*$1%nKf^v_4X&E}HRyW0FCY(_i)C{Y4kWTv*EJFS*0jdVlc^wpRM9u1@>iwXZGd zFS-+k{Y6tY6Evj%j$O21$0pp{!EbF?)O=Zr6^)F$m>h$Y1H>HoN~c_yKyi1};F9gs zchmO`bZj5*D$zTe-BpMh5>wVWlFa8?zMlIv^p`z>EoLb%z8fa}PSjKac#>8( zx|Qj!`}Q=sOjM%ElS`^2w&^N zk)NqFMTTd-03GZZR7Wc-sgi1tduU>8gd8C9kv7!ot2u4ag(pe*@&3kWeR5(c^&oGD zR#zs=)ybi{uV6{C(ME6f`pWP+^6Bt;@1t- z>54|wx-6nZ4>khXJC@Tb&-i$<>SvceLnN_Fei;b+4f&EE^=-T%6g(&OEOkxLOj=*1 z!@{BJM3uKVw3G4s_hqy*q~0c{E!0@SX;Gyh%`S3NUW4gM5UwdX z%l}MVTnNsv|C}@Rz8>G{zo5X%VcpRQ{*vA6GtvcFr;6_T%Kd3$Wt#pz!6lrgv%f&I zm3txetO_{dk?WsqTBpBfdO7MIynCvT>rOgJO_kd<{qOZYn;VF%&o%?E62~LQ`(9Jr zFu5X2Bj z=#m(J657sV7F@$--ah+B(TH30CEgsn8i8TW&Msfkhs4CfHYC{?(Qnabifpl(Zjv(U zSA1MBUAEA0-~-b=3c$N*=L2oCJ_pj`T}?J4s*mXujV)u%T>?`F6UM6PwgpI@fN=R&Z0>H$&Y*f+ zc|EBgNRhW_#;uR?CspOjb|NlEwkRs;yuUuk#NF}p<@kt|KR{239xmZ5wTBH~!I{p_ zo|N>LhZxA_59DseRs+)7JvCDvJjF{4QghI$P_!HSI<87L;PLk zz0(}FNF{MvbmJLue7-`*Z=Geg9Jno#U!eJN;Ogq|IOWN>kCV|>E}Q-BzV#Ak+DEH* zAF;^{6zAD3qD5rEIIZWQvUSMfBV*Kx%u~(c%W)uR$C6|rGzm7i`shh;pt)CNo~2Vd zT^axzsgCdw1?qA{OJt%h&&d!};VIf`>a?#k>x*={<;0Z@l%IJhPV%$xQSP4??@;eC5Kskq~6GkEMCaqqL#(- zGkrpA0~0LOx1N=C{J9yiT_`ZX!clUbT)&=2r(@ z(m)HZrWL8ltNDXiv{{lD4f4l8ScE=53;774;b(5t^a{ZXW{6`_#0GLmLI#!8jhec`KbrhY0y2sXyeh6&jM{PqwBHO`Fm!j9a%fKFrY= zx{IzqhPeX!2q`dWjz4_si6(5^gy!u@=wp9Ajl)gz#PAP~Q2D3ghMyLn@k(uiW@6NE zPQeG@t6VN$(I-LVxRo>V?yRTE&C2BbOn+-RVTcnj%!DE37tq*V&l$sEOn;r6EiIm4ZONRN|u>l;q_XqNjGTgug?-cq#C+4gaKMJ^8) zkQpscR7MB=iiBZ)_~zDJ`oPT~eYIk2w7kADQRAFLFKU2{ACc&Z+J(J;v{?PMvpo>xx*G*^ClxeZX(u&N;#6Ll#LhpfT=54E)|Zv+tZ{W ze4aP+HF11SI)k>od>I`=cqi^akyMClhXg1|rc^;0ssT8i-#71?u{gHX`_n4^lt0f- z(Yysan0kc3v=wgOF3$%zYMKbV9Yv^VSDJ}XCzTh-xl}<@vTpI1Dc_Y=fmq&CWp83uAEBX z%Zi(317E~lEve!5vyVBb-(uJOTB#gKZ;KUgTs)FSUUL$okN=6n~#-sB5phKQSg359WeERMrc9F4<}BjwwM`z5uZp3^2LoGNoD zc9Lq(Rw!~p5f|EPu6Nf(D2ze{dEJeB%5x)cm|EkurZ@<`tt>%9ThZ|ealD9#9OK&> zBC-S^BF|zYT_d_7RjJ)Rl$h%~#@XoYhgN42D;CpaO_9Ej_>0kJC`p4uT~_dMx}zR! zDmhn_I$5HCIIe0p+Dh<$Tj8Z!xpOnQKGIt5V`?p}4fQc8ayQN4=U%c6;o`oI zKD!(8&C8jl*Z#siA|F#BxoOcEFIQKNC5O(nS=}_fGcj&=Mh$&k{F&rlb#Wzc8hxeR z!pQh=D$ew?o(ZYRc*er{t={Vo_K>R(Ok=xAg=x8*MuvWj$R>E#$XtsvYSxXPq?}cD z-9`6?>9ZL4QA#k_$fPgy3aO?Od8nTV--{-MNrlB7if&&1h5P8QaJ%!!Bg*h1VqqA46x z-x`|25`d;ODswdU&MsxR(@UXQ*fd0`T~u&d8Z$R+CBoI2v^)J)BKmgUK=NlrU|LTz zTd5M>(ZQC*_M*{Xzu#n~(SDXI^9pAX@^@?QWFWg+TN;gYCw^>&B^q1Xq`*B*E<=6bpfG7my_b?r;7#irYb71mNK$&d?CqZV; zfyg?{RKj-hAwJc+TO@A#)WoGpwTw=Fy;ruq4tEG-hv7~ZZ!*2@CvT{ZcAQ$Rty*25 z%u6(&3l2%5LE{{;idnmWZBAbD_AnpFb~ZbjC=jxBrE4XSoHZbn8BRT>g4|roF^!ls z7>ZVd*om!I#>;*;SSuetYn=Z6fTK|3fty6eGHwh9r>DhX$;3W0nU~SDoJd+|nta20 zOdEzI((95!c>_)RtDN*%D6mQC$-E$D3TzV7Xq-Cck8H63be*_C0B%;xd)34hAWsFi zUfH;~;U~l+o^(`0UIRxwmDhlZFf)lF6KwCVC%T0eQrX6B2g#&mzSW-{1 zU^_5~Qwl*G8UxN@ZHQAg3E=Poz@H6q+a%IXWkwpdgtB*2blLh2xOa@&)C^nyS<~3o z(Hk_srOiCzeSf&dp~Xq}na=46h2V50Ib6tzgF%&_a~pPN;Ee~5p}-BsWC1r0(S(zZ z+PN(qeb-CJ?RWtR>4{?#Qu~;Vxp*<%-D#OWw=E1Tf*YLme+=gkT!uvc?ZBaDvBG^< z=hnl3X!pL;-S29;5t_N1t4W11)o%p_<4dVixKiu3JnmkNcc3ym+-tWyyz)VS4e5|E!=5jm#yIRlYCV62)?pD#id7_|0FVy8v2|_JUMh} zNWk_uQeMj4pjC&i2^7tkPuZkP4>REBqsvb2hq6Gs#EqMxb)TXUg=HpOQ8c4)v(05@ z3dXyDH<6euwW>_39BvVDrw$SOG_&g$jJ>sr&8z@!VMtz`HLESzNYm05|IQn9hUJA# z>$rp7XUnybxmEhs)xBuCN_@zt6sGzLr>$G9e?qv{njO-X-Z-ZxM zeNHi?1g7^>Rm+Pu!l_I$8L}#4yc@J5Da@|kZuL88Lx>re3*}tahG|s>gJZ5P$+0I{XGRMh!fg2ZtwJ|PoSv-ALTp}1hxjwB;snJy}4BtE&zG)r3IK$k* z%V|aa*d-qs%^y31f8Nj~T2absMRug*{RKuU-YqQ*REbZgO6nASOKfhOVeUZLjmRIT z%E*I;HP)SGMc)JtzMuO3rVb$G2|Ewb7w5 z8rS7Lq+=nTPD|h;^82mXFIhiQTba(ONBD52^ZXIsP);JBSeFzre}uO-y;Tt!N;w~O z`kH3Ic}Z(I4Wx~e&MW2dbraR{`r1T&U1gYU-ioDUcR3xsva+_m#9tDetd$3bE0dG% zMxwKgY&4elPph$a9RD@UEiNlz9Hm32Fo?8D(cn}0ktF4~$=BAXr>URk*2X4X;&NvX zotTr?M)_>z!AZJrLTRsmJ<&8S$*G!8BU69pWiHtZcg!uGMEukn z4|H-4zFnWs<11Ptm9@z?B9AWi|4sa%q7)&MWg3%-yJ^8z}|M@ zBQnxzz6?(ZMLtAf9~fo&+dAxdBy}f$3AyI9qGu7;FF0wWM1R}qZwLMDq`$N2Zx{WY zLx1Pe-+AGX-++KgBUn=G9fMu{<8#~v$ zPP7TT_GSIqHdDA_o0Q8oj0kP4=~SBS$cZ$bbLqeMU#gErA0A^P(#;!hwjNv1WQwHy zt2LPdzyE4YrZb6uwIA3x0t;w8_wjHg+Tf<`WWuJesCUe#Pf3YV2D|ug3 zRXh0Y+hmj%ZGrp6-WS~<@#53v-^q2ub*k09DEOKFaA;Zo621)3zck}u?|>uR`sXl$ zf1MnzE!!p|oRs zzs{Kc?SO90baY$vl6+q0;dSa{^6r3c@HBK@eDY#+Gk5kz)o=&gL(N=IX0ycev@+V4 zJKz?48v0U)(DS+tlWCI4yaNVnrlH$vdHsZGt4TlYfN{Jj_fzmKcED)umNwZXhV5WG z;VHKnHVryUGAAW3>k@DciLg0ywS(4gll{k*>@b&8Is#Xp4y_ogudYt;y>h-?4n>$V z1+ppca9b$JmNYY`Q1?lhLfw2NKGRrt)hBWWi~E*E+9Qu22Gd1uRefoS*V1xYb3%S= z8gF@TLG7q1eq|s>`UUyYTbBh5&B%DCbf7(@d(=77&d-;&(DklGPARg^pZ02Ei2-?~ z=jZ0jRq(>lqAeApD_@Mp%MS5dLOBD_k^ctD$&X$&$>G?Z&ymu{W49ls&X5wTBT4#G z=NCz&&i_mSHh3w2+bm-0vS};b15}EFPPttfE$a!Wa9D&MjVZGyA2g3WlC!s#OPsL+ z&8f3fF`-x4b^Lu$BejmoG)A+FIPI>rY%lsU~?u=byInO{#U))CXlaFxo&iL zY+$Vd9n{v=wio^7KZn!b-#($O?Qj3|f7%_|*0vMivE3pzT2QbNzvaD z=$9MxFH!V&1^NMlevP8v9Ox$u`b~=d{y=}BLH|a;aa=A7IFIkkVZ5#d`ttyPKj7yB z{$aqc2mC7m=YLSIz7=q-SKkLds1FYUj`II8;QSBs7rpvB&?BE+*cj0tq5qlx&$hh+ z&U%aV`zm_TtAh;sXDWKptHTZYlA;&AnrF}-tLR0qPBiG3Df-Q!y{8)VFIDvS2l^p{ zeiCpTmzM`z>(y(39`)+2fTLdhJK*mHz50TIKb{AW{_sCYe;DAW1D_KCKMn8+z;6Kj z?SS71_)h}Pdc~s2diMvQe;?5Q1L#?AcM0}%_g(2f_`m7zzXScV11@@dWS}n@^otbz zf^W4f-1u{SCgm+HNuEKceWxp5JNE-=*lk8|3+#LH}(x5ozx(cC%RT|u_WY?pkM{i8fTKM>0&w&b76x4F)slc~y*dT>pgyby z9OWNV{6()e06p?KPtpG@^xMS-{Tmeh4T1hrgZ>IdFM4&IL4T8?7rnaGp#PYn7rpw7 zLH{L1FM4&iLH}JvFM4&KLH_{YI4*weXI=qh1{iIIjEa4E!1c|F(hu z-N1L>Bb5{Vil+mP@*EC0`Wq($j`VK=9PRUa0Z04%setQw^o4*QPU7M|?(4t@_5Z&C z$MO1Az=glequ&BO^7)IR|3k2cZM63BAGR|{|LA}Vf0;*n8}$1t`u91mv==bw4^#9q zkDhPPcPe_BM@Jj<$1D2HL7v41eV?M2d31(B&%P7?(f0YAfFDl3;XaQ0mj5t4&ZCO~ z$9eR6z|lVQ|K>k3F0$_59B_Uw>eX$)2gmmdfTLc0P4O4Kx)&{i0mpsH;~wh=g7pFSDGwX?v5)igx*t{qocZ9s<^sS!2=ZSC z_z{5rC-CR~yo6swe?ATLw*dXUivI0^{wF~HA)tTM-ku-J`C-8K4>&*0@4X`Mc?QrU zpZS2_3Vc>5K34@k6`)5x=K{_)N5=QHfU~U={4xW-!N6}f@Xs6gJqCWif&V?=(r+?e zyFNbcXB@952VCf-pAQ3i?4NGHv7gUSe59YNK#zRR104JL4S-`mUt!=k8TiKx{7Zo2 zzW8pyabN#)z;R#yI69l)KdiU7uRj!U+}AGv9QXC70*?Foivhm^^zEI1{|De70sMBr z?*jZIfZqo=_QL~!V|?JBfPWPD?6Z&S2GqT(aZXN zqe1^cMZejzZ~KTr|4BtJ>-${>{kH(eart4u_4@uxpvU$7cYx#i{%63^Z+ybOz9Cw# z4hXo`tAhj1^Az==18|i8D8*m&>NucBK0S(F{Kk_F`qLD>=+)T<{VGK-dNpRyuUGV< zSLYe@uTu1)S8p)r->&FIudXoYuT}J-S2r2-9|au8!2c8IR{;Kgz$XF6{rJ@eey4%&{KQmF=2qzZdB50Q`Qy-vampfPWnDhXXG5;gDcg zcYabTKhEC+0RIH=d3wNw&kF*dX9GR*nXTxL3G^>A=uZXwlknb`D?V}#_bQ;r_g)3~ zoxtZ-z&{1}X2pL|kn>wWkNkh4=uZmtzcT3m0XV*Q*8_b2OMl9H9~*FOpAS;>rtSqA+C;P~E6ijTbaLZHX@zDd!a7vB3$gZ^f~ z@x8YyKJwl>fF9rbB}IR6c<ZqX5Tz zn6nfgImfL6J-&A=;B0^9fu5WT^k{!x4ftDu{w;ui2Jou@|197)0{%I`KL|iq=*JYj>^n9X^yex1{kh@v=VF8Y4T@g& z9hVyPS19_;o_*VO2K`Nn{{BFJt3m%Mz;Rr@9B>|AJO}wU(Br=2$AII$<3YgrOhN3@ z-vR$Q;Cpe$hJ3elXx?1HJ(8F~H9P9Qh0bj(pw>_&D%+AK=L6qktoyKLCCX@OkXP@pe50 zaO86u;1j^-9KcbY3jjwx9|wFA_fFqy30FHb*pP9C61Mpb_IPy6YaO86t;2VL@hX6-Dp8_2D z{08t}0-ugUQaLMt_XEBb@JYaj0lx(BO@Lny_=SLf2Jl}2{$m6Gr-48H&{Tfp^I`)p z8~FJKezk$$Y2ZIL@P8Wk?88#|zYqPe6!0GdeiqRl7#Dq6z;~k8uxS+kf&8{+o(k{FDDO=zpT< z#XtF#LH`>?FaF724Eld6`pu!ekA9Z_AG^>$*gua8xQy5Rfqs93{t&=%T%I3r?VrpA zdh}0@104O6lL5aM^zCfGIY&bLmN5fA&%obc;8z&r<@Livs zjwAL%C*Wv@ml^mP1Amo)Ut{2R82G&g{-A;H_MEg`DF4$8yvx9s0*?AK1UUBFIs<>D zfxpeb-(%n(GVo6s_}2~mM+W|Yf&b0G_k3>J&)5(91CIOFg8;{U>oUO6{;vf5e_@`k zGw|0K_~i!v0R#U!;G7dB_TguMKLGfyY^dlD+rtL|KLGHD0DlhP*seK%{|4w!0{pjt zj{%Nyz6x-Z^LoHh&W{3)a()qTl=C})qnr-{j&eTXd9Hh0|Dc@D1{~!)0dSP_48T#& zHGrd>n*c{S-wZg)c`M*3=a&FSIe!H>%K10IQO>=NNc#chJP2@j{}Z!ejRX>^C7@d&Yg}-`vK+L2XK`08Gxgla{)&=PXQd|Tn#wN z`5M4c&dUKuId1|S<@^}nDCgaPqntkl{C9h%>(OIorTz1Jz)OI0?yjs~rvZK>;LF3t zir25_?v>Jy0zM1qUk`W*@EZ;M69JcfK_|b6{(K(j4~O^um!dy4(EkkRKLqdHwcYb$ zvJV6PaN`40hloOg2-{ocDewr$4)J?i-Y;5c8_0gn3rO2AS7 z-)7+NG4Kx=_@@l~3jx>TayQ^OE;I zIGznSj^h~MIF9E5j^p@Nz;PU}47eW0_bFV)@fL;GLqGoq@W*ldJm5Hv_X3XNc%R~b zbKw6=pvQ51DByY={|Y#c<1R;~+2X9Oe8i;MkwXgaZy7t~4*`yHehP4u^FF{) z&gUPK_A|ijnM zhxHcad@(HyTq-}xISX)<^TmLpoJ#>m zIY$9UIj;d6<@^@lDCb`QM>!vTd@4W6xj*12=V5@OoSlHr2EV8m@Gij1fb(5l*=N-Q zz7vV{H2QmYSWnLf`iH>}yiC!5C(vJO(0^Rf|1i*h!Jz+dMSp*we*ozJ0rKp=!1t%f z|6risH{h)2NdE#w|JOi259qnbp!CmjMZeRd{rk#5zZ1}(tLPsS=wAi&*q_%Z`n?1F zO+b(G+@{;B!IFBf5P%cA{`_@AKvF97_7K>rTFj|Tj9 zz+VLTR{=i;@cRMZ1@ONDz8m0sFHGBo{^X&6qd$3cz(s!!3H`qS=x>4cp04PR4)lXS zkNf)dfaCuCReb(2u+dcdV0DuMrVfgbrE4LI^&066yZl7I{U4Uh5ioeK2G z|D}rl;y}OJp#L|(QJ#wcM|rLZxX816IKO$nfqzWl|L-wg{?7$m+pDiC`kMp&cMSUb z6#bon{(lVm-z)mN1O49(`rQ_Jd8NHS3G|N-xR(Dgz_CA{4>t|-|4zWAKTqUEgZ|v7@IChM z_+Np~oxuMwi~aAyXU{;tZ@~4q91b}4+e;Lm{Q{pV&||-C1RVSA0>E*6uLS%*_DubO zPXpc#_zwZU9q?ZRj`xEOILY?|x3?SY-jRT#oC^R)JGKIFtUFNy9P3V86maSPa17e^ zx_}=}@>~n@zZLjke&+RnUkCIzEB+xY({>xsBcD4I{em!#pEc;etmtp>-Pra`gZ_Jp z{(XV|CkFj575%q^JijsM|E%cm3jF_R(C@ay6G;DW4)l);IO_xU&l45>{ek|e2L0iH zEp z{r@R?(W{-7`i|0ixHsT9E(Zo&>(!w^k9ze2z)`Pe2VC^$*J0kh80gWSpAIPndbo?v=7Gsj{4lK_=}xc0`$mdxuPGWHt?S_4f+8^FLq|F zK|i7B#m<~-&|j$N#m>CepntQX7dvyAL4UQP-|U;)c7s9x0Yxu%=5~Yrvw-8cd^O(a!ug;Am%l0XVK-k2={mMC;Y=0oQu<_<(CWc?jSr|KWH zE*}oK)~indJ?hmL07t$0TEInr{wLUh?*cvA^WOoE_Vd6#uZOHx?C}Jr|nt<_%{H@bETb^r}S)7 zr62YId@=Bu1^5uq_W*t$;HLq8HsGrO$9`B3IG#^k4mh5deG_nO*RH3;+jThLcwTlq z;3$6|;3$6uaFl->aFqXi!0~+TD!{RyZvhLp!-Y;5e_J8gSt+_VzHKM?Oa=`n^b2{?lpDzfjSOy*=KbU!>^8-u4;v zrz?7~w-tkawW1e$J8sZ#Q1qL9bKA~0=r30EVsGDQ&|eNXj?47{*Y@@nphtWAF~HH@ z&N|f#EaUsiFkVLmob4yt!v%oj`PWj#=fi=|3ZTb)sKUFZS&|phrHxQ1lBze?DZ;|4Gq{ zefx(&zv~&kpQXKG-yR!q?theLKSeM0?I{NRA&Oq?+w%cV)??7G z0365V?0{?gwif8ozO4fs?c19HUqW+K?8CnYT>4q`>O%q7di9^c2mAjkfTR51Qv5}) zz7O=s=f4&GSAt&sk3s)iMK5~wSA)LoC2_slJ>Xie_Ez+wS5GqNpQh+VubySlKVQ*{ zUd=Y>k5=@eR|^dKlL5zZc}c*vUJU^~>eV>ls8<^T&iaGv&f5Uzb>|OZ94`Ytcy4Wqw@>^vLH5Mc?^2-!Io0^fxJbnP0aW^dD37Vz)nI(0@tM z%lx|Ap#QF-m-%&{LH`RyFZ1gmgZ@v7Ugp<74EjCKO7#lI<%t2;^Xp)s$N6q}ET zM7`P@aMY{)11|b=Td)HM13mh~M*)uZvjRAtkBlq+az1h{&?Emh0*>*eOBJ7cf}B?a zJ@R?4qTgk2-!C@-J+3<+R`jJn|8by4{`Ufo{qU%A+W)vtJr8i~pQ8dU{V)Bq0O;|( zF996;XAR(ZUUMGccwX}!z){Xm0)Ew=sekxYz^?}U2Y_R|@%Mn^dfIk&+CR9S?gKcU z*Bk~ouBR^!xb&O&s|y3J?cP%0gZ8QdIL@!tivNq+{GV~4M?M=A{eqxB=Nt4FD|)ee zZ#3vHRrF%_t~BVcQ}kl@ZZ_y|RrF%_?l9;-qv*x%ec7PDThWW%`<_AnbHH(2eiLwQ z_Z|j%w0pZ&+T0)Z_t5U`0XVLwhX!2h)$;7YTsR?&-IO&av)DtghYR~YoKRrI1)Z!zdEQ}m)&*BJCSD0-3RojSN{Vz>eX)pF8U+u>BB&e_WX$}Q~gByc{Jd7Ub8sh!e7p7P62x4 zUjrQ1i;aNedCmC&--*Q9nf^+A;0B<_^N1e-j^`0i7)aYS2w^MzLcj_b~w0LOLbeSqWo^4WlMKRk{8%KGxr;Ru{nr6U`X2#)HuV3KR;B&6 z6ynlP2ORDBvjJZQ^z#5m`uTt(eLvtx{}RBF{vyD8;k{P|T*gt>i|Yce*NdBh58CZJ z0lyCV=Pt!x){CzJJ@WasqCYy!iys>FKUMTM_~B^#wL$-TMK9~c!v_6Mt35;MA6YN< z2>34aGs^!2MSoY2=Rkx0U`4+<&_BnZKT^@ldNIeKUjR6c%hG`B_2P7($Ms?Wa9l55 z2KdRKSMLb8^qc6_l>yg!bv^LG{`@H5DF2;`zv$H$fFAjLP0@cP=+(Uj{SOtr=+(~* z`d=%0(W^fg^bad~(W{+nUIDaT?GbS4PtmJ=4Eh5Vz3A054EpCNdeN&{2K|cw$8k9! z;99TxfFAYgOu$jE)&brJdi5^A(cZoraI}Z-103z)9e^YKX8}k0Zv&3>KLi}<+t#Gx z*bnkQIp8vmGB2JHaPGHj!G1m$_@F)?4LJ7y@c|e9GA|YbJ@V;O^m{$Q%YBAHUs3cj zFKPz;xT1fbXWzEbpg&*H%e;8CLH|ZYFZ1FZ2K|+aUgpL12K~*7UgpJZ2K}c2$8p&l z@WZM7c;Dn+pvQS}AK*AI{tEbV(5rpd`i_-x7rlCN!1=wXSI-Q%wllKH>rQ)rwy9>P-gyI~4uqpyyW^ z^w%qT(X00x^dAEp$K`VY=kZ0o`Z~~~UVRU6)T`eCj`nu{Vc!te6STKa103z)vj9hX zI2UlFKNfJL?*km^&j1|hUk*6h+jjL1q95pSUYrFu&WrVcqrH7|z@;C=-o7*7TCc7KKG>fh z1{~%8PsLyK>eE1ve7>mY7wqFZ`WpuQcND$o)sGGOpDTLNs|OAGKPr0BtG^rcyNr5y zr9VZl_6oROr}kC!qE`nQ^v_iEqF09-^m72mahV@*tyd=jJ?hm8z)`Qp07rZKHo(!| zUJf|g!)pOYd-xxKBmE};NBXY=j`ZIF9O?fLaJ0Akjb+DC=EXq)*Yo0#fa~$?0v!AQ zMT)=7i*BIDd2y1We_5Cp%MJRMD0-O}0|xyXMKAMW!l2)z=w)79Xwbh#(aXGevqAq( zMKAN>YJ>j0ieBc$2Mqc<0LO9pe8Ba*_y*A9y!bD`abEl$aJ09N9gpkPegW5d^^|~X zy*d(bl)p>y7riea6SM|=CibG%-$o($}n zo*Rw=J_z_FfSxm->j201-Uv9p_alJgd+!1q-}@`TaXtMl;JDuX z6>wbd_N=GvT><01FW{#Fei-0L|9rra{uID5K0g+4@rz1cjJEXw*LLqb;Dh$+Er4GK z`gxh+FLv)5phrG8DEiKQy+9u{=x(f)wrxEvgCZTFrB^l0~H1CDm@48T$UR|1au zzXovB|Ca%d`us}3QJ>!dIMQDQIMUw*IO_j>0hj(1{r^S4wf;W@d{A$9TJOh&`yc!H zF##9;qW@0-dgOC}qW?kA|AP(sXDfQq|050hE=4c;e~dxjt>{Jnml*WR6}{;HnFjrU zq8I&NYtT<9deQ%L4fUi}jI zpguedILg1%hO8YBz1kz-+73KH(Z6y(|NMal{lSV}^y)bV{gH}Z^lFYle~h9Ry;^9{ zFH!WOSEm^CXDWKpt3iW)t)dsbnl$Jy0365Vbph9Ubt%xJUR@11>eWX9M}PABfTONf(8e)W}rqhI|1;P~Ey&rRjQ_Z|Z{zIO@W z_}()C$M=o_j(*Dqz|n7c1K>!1Dd0$d1K>#iLBNsz(}1I&^CiI1&-p&!NPj=zNWb^L zrTuvt>@S`laGCG2esu<1+moY#588ocfTP~7Q2fQ7ycFn>PgT+H@_W1svzm=K`+h(dK~bdGsydgZlq-z;V1DQ2b>c{SoMq&)*dNSHe8n<>l#kA^n~K z7x`r#?Q76KS<%ZpdZt1DTtzSQsASO3RrE5Cjy32{Q1mj7mKpRb6ur!&mm2hI0mpG! z7jQj~E&zI*N3R1M=h3x*pAPo<69JcT5&QhPfNQ<_3h=@4{W0Ju|IZcw7kTw)dl2Z6 z&mR^2+xPbm{N146yV z)rkiEGDR2sn<*Xu!2zZ324KtBU|fy}AN$^!M%r9QEyc0hjR-eY-E5m5-&!f)@xSkh-0oU_l82I44I3IBA z|BC}I{AFIe5$KW6rHX#PC$+ghR~q!!1zh;cytvt*zg5x8ytu=l|BRxSdGTd~{%%Dt z^Wu93{e6mF=EW}!`iB&~%!@x8^t)Wx=Kiq$;J7?4;Cfyh2=q8F4gnnJ#nFJ{dGv~a z%eaWWJuBc^ud2WY^zG9^y-@i{r42T=+#dQ`d=z~(W~DW^nV8&$K}zl@cO3pYM+3!9->}7 z1#r}>*?^I0{l$CUjsPuxejpTb1UG;=QDsKpC18^d>#NC`TPxV znQJ!(YQJ!l7M|nOCILh-4z)_wb z1CH`M2sp~~;@76_Lj9Ql9QEgofTR9g2{`J{&48o+d>`-u824WSj(+lQ07v=vd0i_1 zAn@+07w7-Wq>37D*;FPs{yY<|K9=l5a8bq_)gU9r_o>W_kIxY!wI`~&(uEr z6!@T@_GiGa1NwgiT=x7#IYKO>*V23+XB68ya<8uU+5^f!36ZHF54&r|g83-lcZ z{ZWcu{JrB0`W{7pSKxoLL4TT}-yG=AHt1I=`uhX@m_h$Cz;RqI4mgi5-miHJ(4)V1 zIpFB;{X5{OpT7kh=k=j)@B;5b|KPkn4sf)qCj*ZBZv-6qe;;t<|7XC_ZtV8PR35Yg zPXrw8$)SLwJ?Q`(`5Xs0@;MoBBH(yV(F-`9Q>+CX?b53ON4xayfQuavyL4l~wOzUe_@F)cEa0gBUsn9ZE`1Z| zkwB-~UhgL#&;7pt_kHf`^Qqnb?O(@vtYaN(o#%O7*WTOpwdBQ>b>#ZmAFh8T zT>laH5RUtTob%(`aM$B!aM$mTaQ!7NspRi^Tn6rXtcL5igzIkscm3`Icl{n8caqda{(P8~Qobjh7eq6uLW&D>DKd#@mGX8ss zAJ^|E8GnA_$MySt#$TBDasAf0w9+Tmr(WQ3zT)~_F5_3jotIXDr|WkkeAn-0aM$k` zxa)2@+;#Uk-1A_q(UtKYzmz>{?l;xleutT@Ay6N*q?DfS^V;H9n$@z zLE!0muo`aPt`m5SANP}X_^y|ZiC<@xavohX{+5X!_mk~2exJmT`$_+dKPd6zesWO8 zKQi&-esV&_KP~a&esWI6zbNtJeljNG-voDF?g~8JPbTBLpGL5c_S}z|($f5_sBg>%kpYha`UNx6b(X+ZKs`bm+J3GXBnqAN#Fe#vhpYvEPPf z{KFGJ_S>+Ge`?~#ejAzbFG&2@ZSK_lG&eab150w-0|w z;>Z1=)>V~xF;3mUV?L8sDSM4G{>q6T_lN3?zfR)E{h?jP@0j>;f9RU=w@m!FKWv}z z`y_tcANptfL5Uythl4WyFu3z_df@5)a6Z2K!xeD%hr8hR+cdc6qxS-j^Bd1c9|xZH z+kE2KCskKh`p0oC8F-8z`>i3qaaK(H7EQ`=n`Qj95UjLobk6v{Mc{X zW&E8JKlWR{j6X2(W4{f}_=hKc?6+YV|4g{^a$(?Uzg>lIzm136Z;!$~AFX~(r5`*m z>2 zpV%qzxL)FV+$`{PJ#I}L*HJ&X^FA<%-=&;h)lhun9G>_GhxIrtv2-XewaquKahUgGVrwDdIg^LTOZ=sCqv+l>(C^Am(Xv=;v47W#D6{X+nE`ERN}{e z8=dj5PW;$!H)j0X5HH+(SX*_rSS*atpwz&=6O8a zJSW1<^Ch@>z6&?cd2sXm6>grZj{Dbn9s)Pdv2gQz3T~b=;pX`P+&sU5o9EIuR`T@z zeOb8oxlQ2S=eB|CcYy0}1=rsJu73pF``PmXuT>gGv(leG|a<;&@)T1MWUH zDTyD?FH`W1^Hk!uShZZ==^1}!;>YvLI~o6j#E<8f&ollvi674|KV|$s5Ev&{xwk{`AC;ef3(#e<$%{UwxSIKTrJFSKns* zpAtXzRn<-Zy04ZDJk~$ zLb&_k6>#^%TjBcm!1bSk>raR4e+75HUFzoYe8qW*>!M-c>AF}c@H(acW%-j-mxsF_ zt_|033)k-k*WU)Ne-PaL_RPT3bulXNbX|-lj{STK-1^^{#E!Lfp>tZ{&>tZn6{q~f=c_QP1X`QMVnk9~C)zHuH% z{1&T~*X?5&|LMe!ef46-e?9TTr&Y5v{zr))`|8V#|Bu9vef3MmU+lK>_+tHIU)2pf z?ZZZiANy+MjK3D#d1)7TtJ2|p|NN%-_EitKeYF?d{q|V6``#&V_q~hY`eWewx54%A zh3mfoKZxg}z79OrKlc9*fwwBv=f--K_uu|b9Q(HM?d5qa=F2*-9C(Z$`@b6BIO`;S z|Iq*KGJeOzkNw{@<8PVxp)ISn&-i^3KlXqBj6W#xWB(tN@sCXW>E$w3osjWQOZ?dX z=VbiL;m*t0z>E2{|8K{)|L=#}|1ZH0rmsE=JkERUtFHr3`|3yH*bhrj{MUWeIPe%h z_SGu*#;H#H8KJLQX8d-EANy+KjNdi!V_$8R@wZR>*jIfse*eUeeKk1aAC&mhLth=0 z@lQzn*jJ}#{0rgs)s=y#eKj86zPb}`Up)(VU;P^HzV{>CeXs5vmHyFh4fp+5o5J-s zha0~ST>n_O@y~?opAR?w{cwN&@C$HXzj!;t=La6wOI(i&0#DcDLgKiN8s1r{r}MsI z;4yw&kInFnvsU8QDW0@0J=V+k9TGpT$Icmli^PxXal4GabK=MK*e~M`O#HYWhi3f4 z6F;uUVHy9_#E;mSHm;@`EdK{ionyp8i#LRO@!N5Ps82kKZLtaegSu%TnN`+ z_3lbvxevF3>u&%zeh;{QFS!0;aG%E*0e9aX6L{?NxQ?z5JY7e(5Xb(X40pbsNaDwJ z^gOjxg|8Q0PB8GqHp zkLzg7jNdBp<2u?P<8KCcUU~+euA@HquA{x-uA}4NKEH8s;IR(tm&;W(Ch)Yc#uCSR z-Um1T$$`iCv9F%QH_r2kKVtQALa%20S&1L}YEH)gH1T6!eVy@tNc`AWzi0f#?=9yW z#~b^qLE!2070V}n?5n03f6c^?ebqYScY-@Fn+Klu)%N)I)oyV6>TtOG^BwSw>Q|m0 zm;*lqzT$nAb?W;)dcp_StHe1NekgHHg!{hGGvK~2^a1#q#Q6liDtxI)l{`J}N^p<6 zI^5%K20x7Wec%n@gW$f8_mjKfN6=R< z1Rm!x_SLI_r+xJ{aqNeAaP$8zi68swSA65tdZ65IF`t90%bzbDc)HIwPW;$ct7QD@ z#E*T|GUK;P{Mc6;XZ)^-ANy*njK6*2$G+;D@%tx!?5n{U{|LDAa$?|VU!8?-UtI*Z zuWp5Vo_Y=L`y{`EA4wn9da$z3`2N7g@T2fofp5n7t4H8P|I{x1S@=i!aRB}q_$T8R zm)a{gD(7(~{s{c5;KsQtiE~?ZInIOl#(5ilC-Fanp9TLBJ`%pfL*?TX>q378xc>TZ zUyt4d?tRu)@Ua|scewt3aQ)-p`Xk`_li)`)?=u6B`$OEX-VQw7uRb7-`^*Bk>*CiW ze%!BWKU}G^ah47|<}@Az|S4sT1Up3G8EfYWPSM4+Y#)%*ItIab0R*4_?s~s|a z-^8C@9&XiM8Gmr%$NlQyjDI}bc{wBSbicX)-~DO~-2LkMz|;NVDSX%UEPVHeIry&Y z@8QN-dUCl?u}`*Z9{MEkV*MJY1N?c`-4<}y`_6FJ`#`w<;c)#6;qDJtz+J!N;jZ5Y z;QCL&_1}Q&e+bua^vFNgvrUKc=gk5y>Uj+NM(4zj`^FZ5r~Afs#Bsgv4Yxn{OXA0U z;}Cq~9G&?6LqD99@y|&7xNn@7@h?gIxNlsQ@oz}{xNqE=@$XLjxNkg^@gGn8xNkh0 z@n1^(xNp3f@jrq)FY^OW_l=+M-8X7YDfdy)KkggN;jW8y;r4AuxP7|~T)!_|e+XRv zXt@5kz~ep>*TvoVu8WrvKdy^60#Db)yTq}dzkyr-ACvfTUHpM>oFyJD*FWYzFRY7Y z0x#x8e}%-4>teNx-y-qjx@eQ}H%$DvF8-GBH&6VyF1F41J0*Tx7kgy<0f`^i#Q_=r z7`XFtYT)U*I2Ye_F&gf=xDW2Scm!_WJ_omN-+}9Y3fKP`uD`@%mG!INA@JB&ab0YV z@46U}_;Fnf2|QgFhY`pAJRNTR&q?BUDW_d^F}`ueB>qWjl>funjDJhw$8~X6#(yC3 z(C0{{5OLer%w{6NvK1AeB+z~w+`pS z9oIE*$8|ege==PEdARw%7kFBSPw>tESGaK+JyBKjh;dqlI;W7u@svopAf`5x9Lg9d4h` zgB#~(xN+(}U8%ouHif%hHisL(H{AH=!d)*D0*~_>*UMdjr|acG;@BrIz@3*@llXDH zyp3<1If=h;&2n8n%lKa>eq1jt*wd-y80{>=AgnUIyd4UJix3UQQ3Zs88+EpLjidGrs4|NAQdN zc|?oSU!})W_#@bF-%k8a;r#wV!aD~(k2sz;e}#MAT>9Da@r!x)ytxY8{M!Z|`{9z1 ze@A@t@0Iv5|2_$i`R_v<^FI`B{wKrDe^e4b9(Q!Y<8iMc&P3{ed*a79_a!{Wc_b6( z1-SKmKZz6X3qHlSo(ti|Y5ZJy9^<^kzFH;lVt$SDH@Mfs+rq8$9&qb?09^k#xc*si z^S?Uqn18IpjrivO5ZpK~CUIhazJYI?AK}(viRUZp!f~wtcU&#t`WwRakAr*u8UgqG zH5%^u@nN`eo`D-@HeCO6xc;wj{iR;0)JK0S__4fBF*xuzzwtbKP~fdf;cjHz9Yq|^ zM`yvu$?`si;9cfA}CcaM#P7aQpBNxW7+u zgO@6K`uha8gZujg`@#KvfOd<8e3-{9s`|CP$P%x5pS`CJ4ypIhPP^AOy8o`svwn{e~_2<|?%!mE}1 z{WzaZ?NA7}h|i68gR1sVU>#E<)D?U|K%F`uOakM)WB=du}pmBf$xXY-8TGV$a7 z**@cMocMA7+$`hwf;%s}2A=Mp1M%HI4}!aYo&dK$N5bvT$#DDsCAj_nCfxYHz>UAc zYn6Gl|J%du|7`=0^SgfezgO)Pc-sGa5XU|_7;c@9O5(@UiTfp47G5`RJH=f7wC4--H3^A{Qa+r*Fk{By>ydZRqQu|Bb%mkPYN?=b&{i68rU zrHtP!@nb))o$=R8{MgSMW&F+I&P%Vr(|+C!-+mqlx1TS7592!F^1x%=dUPn)@ojwf z^Z9W1^Y7vIf2}tw>(2gP9B%zZ+bCU%2b*a=7mcniP1fLp)zR5_r14o+6HY z^(NeTd^d?7*Vo7R#+jG+z1At$aY4raHSy#6s{K}_{>EQA@HpPMzLw4St0aD0U(GXq z%fyfCt9{1bIPv59+AQO5mH2Ue?U3>JggY<$1)i?2!|`2TC%|1_=fLgH%i;Ft1i1eF zaQ(O7_VYsc@vQ3=XO-tY&Ts7J)dElZd9A?HeXIz85#fF#E<=aNyfi2@nb*VkntxZe(dMFGybD+=jGYJ z(|(?bZ$Hn5+t2miuIwA`|J88!|CVt3c_X;}yfs|E4_tqLxc*Ua{mbC4m-_>c>mt5i z^m%;G3!f%_Tt{C8p01-Gh-3fPeWy}S=c`fRF`u}OR>n8Z>WP1J%W~XxGJe~{kL##o z#_y8&aUE@$@p~nHTt|H}{+@{+*U_Mie_-Orb#!FLKR)r}Iyx=mUjTPr#sr?OqZ{#E zM|Z$oM^oVT?elQ^_U~~0&*1vC|6b`U`@9+4`-o0~$9{|byjkFBKW|MO`>-F}x(!U? z$9^7)Z=Ayuzv;T=@eIrOrzU>v=aCuzg2a#ge0j#dHt}OW-<0tuCVuSaNg4l<#E<>_ zRK|ZH@nb*F%=quaotMu8Py2ZRzWw|=+Q$a^TsOlzW%$+^J~;5$pK*V_ z65sRA4RF`Xop9I7RJi_&aQ*k;`t#uW&E757C(XZ2;KlWo=a+30Kd#4}0#DcD9>j4S z9SnEgk4oZqDW_L;BEE4>PyCa@dOSDdU!3@HJzkmd$0mMUj}tQfU5Ovpk7_0NLq z-vvK``->L?k9{8d_VvKizMV}R`|WGE_4y%*-zD_z@A$@9{Qa^J^J!WnP0f3w7oeYJJQ-y!j1U+tRl z2f>|}Ljq6xY8bwKbvoR>Iw$a={qW!1w)fPvSQ#`SE>sFW`^hI_G`3aTX+T z&MyX1dMun%^P#GkcjGkspu&CM-Fk4}N7otd`{=fV>-U4}9}o9B>NL3Lg$v-G7skT% z?||z+4%eRn*Ix+t{pw9VEY~frued)n5457QD#dYss236|`owiz9eClpKeWVmU3Y;S zXXhkNtJbA}rAI$}e~BdcO+pdcPH}{}5dNEx7x``*7FqJhCYyGQW;@$-v`$#eJh;#$Pe< zLpxP9%lNI~&dWxDr~5`XeD{rQ;qDtl;jW7#;P&k)aQpTmxc;?p{k!1$Q{ehv1Rm=g z*Tqlxu8S2vDbH8**9=ze{SMGQ!a1SB^m$9#Eo5o3{J)1AXX($%WG5TfiOH&Tz*y5UzhXT>n(K``9SB``C4G{oCRCkHYm|fb0JPKauAjR{s2- z>lydM)dNrW!*v2r_rp!$v{~_b$tt5 z{{guEvw_F)#_N-X_+DSt`La?s*L7pK>v~PNetWq7=5YPqaQ)K)PxHSJ-|MS85Z8E)Uc8F~{ym8w*UQ5he`?~#_3~WC ze>w5vdU-43znA!Ny?m1K=O=z#FW+bUg^3^6OP#MO{a`&C!=0B^15ek>+W4-Q_HftB z32^&!B;5YI60UzUT>oXbuZPYLJob6)+XaEAeY=o2)~DgumAYA<6$6j)W8XHzH_lp# zA72k$FXMMe{MfghGyWEdANzK@jK6c@$G+{C@dqY;?AxIk|M0|*eLF1UpPKlwZ%1bQ z(QxPGy1>)Ey%pcSy$^2RJ_+~r(Aj~S&?5iOe|Ioyb zeRXWcKLhT(oF91FS7Y$)t8sAq>ej%E`K?|06Q7@X5#RfW5Al7y;|u%|^(t}xfE#D| zZ~wWk?p?3^dDFm)zA{c%xc4L5!@a-g5BL7!Ah`YsaQz?Q?*DcEQO-BLo>>vDzcJkN zMNhc?esKNE;J)s8U*K__u3t{CYI5M|KJX-Q+y~x(+i$az_;DYYi*KAS6MtHm-|sU1 zFNq)bfm+{H>TLYFfyeR2eV}p1UpevPK2V+U*Gc@i546kp9TPw916?!zmWdztf$cMX zKe+QUIPi2II1Jx?;CQ(Ez!m4ue&Yq@d5`lO?`u{MyjXYM z7j#Md*#A8OPy4?YaqO#o;MRHnB!2AwL-CDsOyUo2TOQBJ8Gl6L$NnFc@h?sM*#B2& z{BemN`~S9#e^27a{(m^*Pfh&T|IcOomlHqs|63VmGcfJ<;u{^FAKdz&?fv4A#jS_!BSVt>o{M8dbuA_A_e%r*4 z>!@SK?~?d&9c`KMdnJBcM}0E>o{1mV(V&cfVB*JhbY#Xq74E#86L`9gM&rAVu7kUd zo`HLxGAHmj@3F5w4Lt3uuZd${)%vNjF3i7f;4yyetH$`oSvm3JeM)u4UnlWnU$x8l z9TPwHRo9HaW#Y%a+CJm=N&MJX{WJcc#E*S-P{uzp@nc_|knu;totKLPPy6Z`eEaHV zxP3Jh?*8^7-2Lr+_;B{Of56AFpV$BSpZBHMpUVZF_UEdBr~6e~xOLbli68s33%>o? zBk{+#E9ciMY)?9T%;{t<~E`}6pWKRoecf1aK3FHHQ{pI2o3>k>cq z=gk@aKDhJpXy9poK8J6A&V<{af51=OYZW%&N^ zbBKRThL6bbOEY|2hToInQ{n!8=a)15y$qk9;R`do-fxw8@wiRk)_Gm{xzw!_{5<&9 zaDU(Qu5jb;m*Gce_!${~NrvB$;df{F;~D-^hQFKP^D_L`3~%*&WnO%qaPtiBm*Gcd z__-NAF2g5h_{_Zlw`H~yM%^XU%vxO>1o?g4O*dlKCE zSHX>cFWmS~z>Pl}Zkz>hec$!^;sKkoUU-=Y!5e|1K`FP4mZw)aO2ztH_nrAeQ~( z!8psojnfvsG5cFj_@?k9;XkmhFM+#{-In2xz+GQ&!yhF6xA5oSb?a2ddp*24{C4;z zaO>6w-jnej2G<`6x6b3>`jg@M&%>RUcj4COSGe_Aaq&tWtk3#z>(dKveTKlz|5Uj7 zUjR4%TjBP@6Y!DD%PVmG&*1vCm#Bz@qQ9|PCF8?HYcuKy8S|95zQ*6AusRmLmt zkm0>E{J;!9GsDNi{W;hV!TmYdufg@dgzGO+w=ypMW^nyY;MQ{|xb-|3u0H~n|oQL81ufp}`!S(CZ ztJFb%IkraF0zXrF?U&4)3yZ*nf!>VxOYzX)Hf$cK9 zUxpu<;pf6V?zjw}oZ-)B_^b^7G{b+$@WmTc>SmtHXLvif_1_xqJnjnjynh7TIOo8P zGX`#)`{2fz4mZv_aN~Rrw+;=K`PcPnnc-W&?W;ZE=6N98=aEi_8|NyxaqfjX-Y4M3 ze;e-g`~tZ3Y1pt*2d`V!fsf=mq$6B^2e|%Fxc+Hy{fprG6X5z!!S(+R*Z%^pU#n52 z4*ILX^*4g+_kip7hwC2)*S`?1e;r)^0l5CFaQ)BW`gIyt>Y(2guD>x{zc*a}K)C+t zaQ$oH`uD^2r@{4S!}WiJ>o;7sQV0EJaQ%*O{hi=m{|tqD{WAite+^v!LAd_QaQ#o= z`hURnS6Z%82lvmlGkl{A?~&pAz^&(TaO-&%T>m<_{=;znSK<0|;QBwo^_N?|QXl=+ zaQ#i;`a8q*4}t5S3D+MD*PjU2e-^I)K3xARxc(9=RO+DL46ff1uD>N*e=oTHF>qfm zJ|FJuaW}#3=SdmO#tXQe1@mI_6jo@Ba^vLl38GcZPpOE3_ zWcUqm>wGud*W;dp`+D3QxYs#f!;SxYhBsKLQcwM+8Qwa>H_h;#8NPdl56SRjGyH-K zzd6J2f&0421MqfS*ZmAXpXXIJTBTCA3*mdi^$&+%gnuqv|62IP_z%MMpMqb4|1$hi z_~&r*Z`PzTF5|a{oBy_O{gdJ5e-7MyZqD#~;Lg`HxN%;F>(7Ndzt^u?8Sh8T`vY*t z`y$+Ny$iSgpTo^_A>8=4Hm&5-lKGkj_qcDuJ?^J)kNX?^3&y*~YL(;auLrlD-QfD$ z!S#29TmOsU_Rqe}D*0SP9gcyIho1|d2)`aa34SNs{{H~}0RDXVgYdemSMq!azBc?} z_y+LF@UHMj;M>Bd!1sqg3O@t>82m!`61eth3!tBk*X;>Y(r zZj$l4CH^iUpKUUJ@5GPqd)zJK@16M5Lw)wo_=hBZeBa|S8UJ*+^D-*%^nH(4;QPME z8{oe0@ecS?^wmuGH28b)r{Pu2E9>zY_zLi6;cem1!M6!K_FL@d-hrq6yc==s=L6x^ z?T93P?C0b0jWay)2Z#PSJL6xN__3d_$oSVKe(dL)GyWZkAN%?Kj6WsuV?R&J_|p?V z_Va5Q|DD8-{rq9Z{~GSR{2X}N&vn+Q^p*X*4BUS10e_yp+7JE${A~Di_#JT1kN3l8 z;J*WZ5k4RO61>xz<#83`9S!dmc(Jd(4Brp_3j754_FPw-1Mdlc4eq!;hI?I9yG5n` z`WwQ%F4_$4^~oOaSIPfOxaZ%i;GQ4vg1i4rgL~YM;U2g4T9t9lB%hVxp08WO|LOVx z{u*(5!o42b0q*tKmGC!-b1(cY_+#){@E723!@q&Q1FyYyr4E0GH-^uKuLpk}-UV(Q zw$JdR;pTrT+`3%~f0yIl4)?fIGkiMS_^-g7m%8gz>hl4|?E-gQ^oF~RhQjqvg1e3` zhwG1nd*5*<-2R^nci)%?|B!LjX;~TX`|!qa>%TVK z4Cmid@w<`d*@^#2@XycqFTw9@Q29R4W!A0KLB4Z_pBi|PXR&w3pTjT$|E|Urejfha z@JryvxiawLd~px{4aC0}-n--#k0Qoe)N|b$TITD*1FkAY{cR~ue3VL3^C-qu{E6dQ zql|0ff5&kb#s3F>qYx*JUyQSO6ud|$9yeaE6?_HwqQqGa-#9D5S1kRzcogxQmHsT; zu=uR1D&_S%lzIKoUoro0dX{;kgg4o?%$HAiv%ptQ_!@yXP55$wS0}tx;4Ko~Ht?1S ze=X$MCgHyY-ag@*hBzG(-Zk(}3Ew>M&I#Wt@NNm;Ht=}A80%Rp@bK+(m3jH2{Pzu|}v&Js_ zI}m>zxcS)cz43R#cRj8G-wEISWEZ%8-wfXkzG8@9%x`i2jpy0wgf|Th=yCTXj`iFN z?mYH~7kfhSDC#)~ep(GJ^TF`A{tDmsLB;h{@cr>0A!3KUet4Tzjqf}Y| z#%rFd!aeThaO=|>?s12}jdLX2`8^YEoNM96nFzNI)8NK=8SZgEh8zDIxPI;0HD&v6 z>vS2o{u*%m*7KV4^8eR;>3`dJ-z&FCO`mjQy)T(rkAvW@$HU!HxepT>m$?`_+H?x%^DYfBLyxhyV0*xeou? zpUZXl&;DGl!~a|R^K+T&jc?(ePX}^6;Q90rxaXHiaL=bt!M)!22i)T};e73J+rz!y z=nVJ#wH@3%`@ucW4uu=%WVm_GhMVV?aF5%L^QQIb4EMNw;Ktbp?s@r0xN**f8|O;6 zb+`v^oGEaRI}>jFcj5Zqz>WVaT)$y)16q0%AH3f9-*&(7Kus0Ob?C2@#qHjHJ6S#5Kg4^ev;KtbzZXEXk>u?;taZZPO+%a(D-w4-#0B-yz;QDXCUBCb7 z=koQyfBLyxhyV0*xemGg`C|54_n)zF_ve@4?t4GN-S-->zqvoJ0(ZT)gu8ES40m7J z3T_;)$ISC4eDjUwOa6Z&TdIox%LZ_Zt+;qgdDL zmi`ov&B`BY?)ygI?+iZ^zBl|V_<`_|@C)JkLw3A2Ijx<@NH-Rjko8#X`H?scMkPGJj2h+ z@S8LIi430&|A6EEnBn2u+-t_AzcJ&I@0Q`mW%v~t{$Pf`mEk{T_)47L&2td^L+apq zbpLUmbRHLFJ^o01-{-Rs?!Nj9d>7`${dphwZ}>;U-JgfU-9K;2@O$8ncM9D7_6fNA z?Z0EZfBY{S?_%{U?+g9kFu%3`4<4`A35%1be8~*23-|r5OD_U!;NB-b#y)1BEWkgCI6i-JD*Su=%iur2y?^`>en0+C z@DJf0*Y`i$SKcrDOq}pzdTXwq^}UbC^`Yb1hH<-*?j(?)S?q3*VkNEm?O*z?Z{63;r4Xa53ERIxou;$9`B5zAbSs zA|KjXC+k9kuM*KGT4+UPtFRrmi z75^wbrq=vje*W-o{5JxR{^;P(!M_LJ>xz5f;m4rXd_QS%+-pJ{pNBL~x5df_D6aSJ zBTnDIV?Nh~I0NvFb3A+!an6Rjj(mPk?(0YL`!jK-W%z3u{$YkM3_OlE&R3lmYr?ODZvnpw?(6th!_UOO27WpGTDb3zx(@y{{`K&W;MT$S zNsh&@RSdNBC_dZ(Ujgp#gKigiQPyJp#`~_0fwwBL8;R4EIKFPa6MQ`W?n(SErFg|- z0KRdCBz~Qc&|w+>*u=l39KPz5jDKe0-x~b$Gydqr-#6rQO~$`5@t+CtZ_oJmCjRu` zPtN#HB>qRie?H^C33p!R1YXSV$;{W6_`c8Kd${j&_yv9wah6}aoLJgds|KF-)tZ49 z;p~TwaP#ky#E*ToCBAWbCH}{uZhbQTo{1m(YEZ^MF!5tw9hvcuPyEIV zMH&B!#E*S-ea62z@nc`zneiWiJ1%KKp0<6X72B z^dwF^?z#9LcTD0R9nOPeGyWZLk2@)eGcv@Pg70x(Nc;Ia^lrJ#N<|PMnvm@U2hro=NjOB^q0-glJS?TUx~9R z_v1|hPmj9>zT<75_>;nMJ7xUs;GQ3MPU5^6;`GD!xQD=vb955tjS%M~eB)dQH%{>y zW${SsP`u_?Jkt6XKTEcFM1O8L?u!}!BY5l59L487is#Ub)4oIb^Wr&w`SRfj(c-!E z;*sW4Jm*|IVm`e>oZ`97;!*g{d-2?0@reG5p>Nm2w{Ba&&41e@PRxJLjK6QjKRw|+ z!g$Xmj`=*F_Oa{_=sRbzUR!zYBh=jK5vtFAV-J8UJ+n z2K329xYs#Pz}-)t3p~zm+{a$Qx1OIR{<7icSbmf7*KSy@TT!24EysS|3vQlA!p-ye zz|%ZW!#B^z;l_U*Zv4Mz;(vs1{DtsmsQ>bfD&xHszIKLh7l)6>WyI-?KOVj#{C@aO@DJf0cRt+u__~Jk z{%7AW;rrZnA)n?P*Z#3j`rx<2?+dp-cZKhP@BDh)-S9o`?(lVpQ;)v&b^JZ>?UTJ( zcY6}Yd|WU6GI8|7bu_ceI2~G28;WQ^>S9SC90WI?gW=~8=Mea4xcN_j zJHO_C7;)Us4~JW4k82+ufp0&1T>DUeQTp&m;@b~$$8~hZKNjw|j)U7j!{GLh?@MvM ztiS8xc;cAn2^oG;hMxkr4#VM&>omCID(;oydU0H5z;oj=|1&W?Cl=XE9ajeg!aO>l7n{(XJ_|E%faO+kZ z-hwz+z`YI}19v^HSie$#l?ZtyL*4sqV^ z&cwe5u75AwewYNeAFiMujPoGA>*yi4>*zkXaVEp9&m(aCDe&IZ=P|g~Y1e1+e*)k9 zpM;zLqj2+?1~>nw;rh?O&Hp*L`QMVse>%ST&w!i%vvBiy32y!`!}VW*+dr?t?Vr0c z`M-&8{%^s}e_JDz{TuSKo7k9ZG%6XLuN-v&MhzBl{>_@Qv~JP+=Ap9;4QKful3 zpSxoGrnM?{_>g>B!{4S3@#kX{b#uJ_91O?n&&B$P`1=umE_?*s{BMANjDH8*alHmN zjz2f=6XLk;K83d_m#1dli|4>&KXiwG0q+C<67F@T`zJ0I@)^1j1y zeTDD1zJVM6Te#!;2i$SF?=8x>z9WvuT>#IG%j<#f@x#-^HS;C!$-Ml4zZ+ct4ET@u z*7N5K{}tYzIKRQSg#Qlj5B~#x1iY&BnnLWuli>D`dDg;r|E~?V{>EQ}_;v8*i^JRE zF9FZp7wa$iC;g=s!M~1tw&c3*rofBy+tS2&1iv2qRe155SFHa>8UB5S`*YVD5GVYa z`kKC72HtwHa-d?rZ3yoUZv@{B-WcxBQ(YE*2L5vJN8!b5d2zh{yn+?r{=9+};Y+X& zuLN(K;hVtiXMf&%Zhu>u`0j5_;N6**RpIV$P2uiuURUS#x3w zzOg!f2jY1B<8hng`#ikoN1uncK0XgWi|YZOhxhz@8TojfFdpu8!u{~#HR9NBAHqHE ze7Jq$^YG5spM4&FUE()pAND$YS$JRk7UX{*-0wR(25$bX;cF4c=MUG0JHM`rmc&^H z|8MY?aGz)Jc@OW8%ySds6tCUK`PvC?K0g1iZy)M={nMKGK5x(l?s(UOJKnZ%$LsSj zx$%13_QZ+nx7fB;DgB9ch}RV<_jwZMrTD+aquA%`(pQeF1IP9Hrz70DZ45WgM)1xx zM=c-M@6X)<-U;99u}$D-;TvZxd{ca{PkOMQw8dWh~v8G26w)?!@WM-622M7 z-3o4h`ucc$UA36sW~D!aL*RXhvo&!}fZONRVfD^J+lCp_2B?ql2I zd;e}6`)wWa+=y}YCXV&r5pMm9_wdC0t$!c5^?#7-w|__d+c7TJ%WmXj{db33|6Sp^ z`gfsj{qXH`<5>TV$s3vR#d4R;;w3pdUnxc*>xFUGYW{7|_5$?*Bi?}jt#&^7jz!&BGdnmsBa~RzI@%<;c{_*^K1aTJS{Hwn+ z<8@#1JaQ!QJ>RWb`oD4gdfjpqzJ1;q?s$DaK!5V_{QzEXo2UE9(ZqLNj)6Pg%~(hK zaol6^hh_K(xbHu?G{f!NCH^LQ%pJueK0JC7T4ov>G^+h08|yWi%1-pFa>+|8pIS<~MJbk~!4)6=`y*|7Uei6QL?Ej1KJ?~r$ zw@*gH?UT#k_DRowL!Vf;%Q>$5{1tH5qdzCadG|icd0d6Q^1Oc~@$JK_;MVP0xOKY@ zZr%F-8|r4BCo?YVc?0=a&v9_;IToI)r~B1-e6M?rWBt!0PwRg(aqQb$;Ewkk;Ka=78$-e>ppF2NpfIo#e^2gxLyU$lRkG`+N`cEZ}^Y|p(d3*|P zoG0MUqvyLF7}r4N*Lj>q9G?d;&RF6+jqkd62JX0?g&$6w=itNP`d-h!fWJBZboegt z8SugI7vbk*_&B)NZ)?DP|4F9|-x}VX>ooi774mUiyb9NU9j^ZdTz?i^|82PQ_zv7S z@4=1pKHU1ufg9%o_}0|NI&T915Z~)heV?Dy-B;WK>a#n(GLpISHP{d3}&&lhm>k(rYq<0N2>bv4)cu9~@HZUS>zr@lo~L}DmDdB#ulr~2{a4=+-})?o zdmsHhJXat4`3L-A^p)KEX!E=jU+#VMkHq&n!JqH?IC1Pl$NMvJ?1x|Aj(6(6xgUPz zxc0+Bxc%V!))u87a_?vRo%q)054inM)v)sXPp&@p!(#aMgWP^FPy0b`Khz?={qQ*R zmFoxBqwfRRmHw$s`~x!la=7oiyBl7II5PsTRr=Vh^yjir__y#E$Nvhx1pIfn$6a%= z@)wGI-s84|>u(QVlK7{?egEMsxaY60;rjmGM$cb0-$-5LSo-7nt6ou<(xc$+Z$7_d z{+{38;(FD7UYc>a-s{2b=i+Qy9@P_bl@J8@^;f>)>!+oF6EckNxz8`6I>T@o>^%)CyyyhdH zmhoSMTb~c%*5`Y;^;wF2#`>%Zw?1v*)@KO3IdwP{Zrv_{uYrFH-0O{8ziq>~tlMrG zzJG=v3Adi7!>#B4RKLAJ{1((5db||he#@=LXYkiy zT<^i{!_VO@@&9Zc`VhzQ4$g3&zgm*%#sYgKxjx0=GU7WVqK|`m^w@!xCI4R8zOr@i&KC=h@tk*jHQNThAWwRfxY8 zd?WbQaQkE%_|Et};TyuYgL|LV3vL|W-}^sxpVg9lw&%FsPwfEre#-C1%-v7f=RV(S zJ$EGjn&h)Hd;|C{@a`Gj8{P+hAiOW!_`Aa0&%Iu4N1Wa8-QRYH+lPC=dlSd_)}bH1 zyg$4hajxKgS$`AmYxMUfj{ZIwJ`ipkzrWi!zWycOmpHBq$LqQnjPJVe_x1cwt&3LF zb3cyjc=v}p??d2@_W-#4KNQ}J;nn0h zn04xP@+tW3f?uojv3OqeuCRbQ6X#U?LGXF?D*2xPAC7-9{51HjaPxc-z90F20QY@v z-@x}|zw*3)I>+_-3gfimI>9*I;KuQJjP4xw4t&>_^X~g^t@9b=X+M8Wod+;p<9l7| z`)s|x7)gBl=j;qW7v7HJc4556aennj5y$iRcEo8*{9Ha45XW^ffIjhg*9-A|zFO}6 ztlaBc>+HI|i1=HRzrN#Il<%{!o|h2carypF>u_ntmwW#!_x|1d-ET(|-~Rlw=a754SZ+fZ(6^yzx_L|2d?F~?KtjraO-eA+<6%b z|10w{jyN9o26#L2?83U!ACK?+-UK)Qo8f;Y{|Uq~pIhMV$p1FD{_Sx4Y9icm-2pew zof&>NydC-61NXT1!tLk#;P&$*xcxlv-`vmlbKG`}>jAj+c@X|r_Q{8dvnc!IWPJC} zN8nz!Pl2~%T#v$6hCi0!kHc5Re*$hkPr|L+Q*ixhaQ&y@`p>}ipM~o`57&PI?)8TE z>7IZ6`6||V262p^yMKQXzdQAm_l7UZ{kz9~iTLuD;jUA;`N+-Z72*se&sX8?;4|Uw zpRd9H%KrHVaUAdK@OI?$2-k1=Z{fR7&VqX#_%_^q;H%8}rTAId@xICaoQdx~FdN@K zlzSa0UzF>>cZt6>^~qgFy@&r-=4%ddtpEG)cI4yF@6!Je-}(9o?tIOKJ72!P%lMz* zJ6~^R=Ic{@=S%K<$rok5J|li^zTD40$FHWIU%;)~mvCPfs15gg?faniA^z6X=ODPg z_ucY~@%_2?x4`uuhMVW~4F4EzpSz#W%ZztEe3wid_qVU`|H^v#hB)@;*YI}a^AEWG zcX0Rr1#shh4>!&aaQp4Y4F4&^e}=nGeSJsn>pOB^-;w+Jj@;LGfNz1nId$FvUJu`V8p565Mj751ZvQU} z|114p{BDSN-ZanU;O!WfKW|gt_u;$mtwZ-VdZ zCac1G6W@Kzd{)CZelvJ`;;atu8+d$vbw;>>>xW;Be+axO{8D&x_)YLN;19vqginLF zfWHG@3qBuS{0@=WKmME_-;cTy&ws9i-xTiqQCq@&Kk9~X-;de}?tIw~Ejg~&8|%Uw zhH(|^y*1pvZ3ADII=oxIQU}**Tl_z}PJR8f9dYcp_V8-*Ss!lx8^Ap;Zv?kK9pSF) zjp4@c1ULTQ;KuI^H-7QEW@2CE#{9@m_|8IxCKXHoRwG`Kd`5el9 zf#<0m@LhMk;ocX^eO*9)6!}-iw{PXSzOAO8J#Xf?=jDUR$NUe0$LDq9^@Y@29aboEa-eE0v8@ICIyaF2T` z+~W?1dtAR?(&L_v@4j&c{Lk(i<}-pg&exgn_Vn#paO*G}?K1d!#M!E0rGKo?<@oX|GJFi&`ML_e zEyukY?s)xv_5t`_Z#Xa45XU~e4(@zi4{uNWvGBfu$9?tB+^1hloc|~9({JFo^BLFw z$@l55mmA5yJ^9PGgpbF!KW~CpwUoYWqI6*#5d17;O#l?op9Hw$L)@PH@??(_rM3> z-wQtuejj`!-21_+;XV&A5k84HQ{X+h4}B4SKmJ_!1MnZ=55j9NR{lcKClA5bfv>}S zb%dMG=5XWt^XiSiFTT&S`Tj=x$Mef%@@&Jn7Jf0;dndwu9@RegdDKhronPzl z19f&?PbJST#CZa~6Wsb6=Slo(eCNgaeG31t%EIOz728Sf$s?Sb#U*S{JwbeyqGx7`+FJw zeumG1yMKNF&+VVq$Da?@o^i>Y-w%oJ{rgAo4Tv)r-aW&6!#~C!2>%5BPx-^!<9`Y_ z|Igsn_#42DvncDuIG+>8_3{PWzWNewoKEoeah)!JyH0Lr8t&^;_M7`vH+W(e7Rnhs$bbp+AzQJ z&Kce-!~19W!SI>&D#tx3!!OA2u^E1EhEL1zH#7XR4F4s=>n~RRLh<~JGXouuMTg?ahto3Z9*C4+6_v~6%?mVs!_dL4+T)#txZv?lVet(Vi?1=BU{QerR-;Cq^ zk>5|$ojS<9KgxZdk;iqs8E{<1vie@!}itI_>OlQxbb_!?W=9!)?aS@ zx5FP8=Dpb8dci%e*TWundwh>8_qdzjdt9%dtwV2o>$xM`dhP_@oN>87d%e9gzWd=W z@aDwn19v~{3wJ-<74E*V8{GZS@0+l1TbCARd|!uox_?@?e#DpW1@}Db>*+ae{TuMQ zmG$@MY*>HSqxIi{^=tgS$+JE6_xi-T<<5`m6W@Mb2w#i%uG7E4y}sH3K7c&EZ}NWX z5Pa{a+QA1B$LBr#ISU!0qz`;P&}YxP5*g-1+kRfb8?O z^!XshYk!)jb;#ZSb|X$N?jx*&uRF*OA%Bm1DBSuS2Dd(k!>!K|aO>muLs_5osgHRc zMI7@y8g8D)z|HemxOpB2H_u^k^E@7Io+rZ1^CY-=o(wn7Q{d)#D%?C>*XB7K-#i_c z{0w~a9|1T2GvVfc7To+t!WSj~v+>Q-@ygG`H~&#^^FJSM{ujW_|3dhpqVSfOL1adJnn78k6$l*^I4QQw`bx^gfB{* zJMhitPPp?@r^I8OeZ77$_@?2&#k}~uy7N`^ZH$wfuT_ZS^ZY)qGKo0e$K4P2K1;s~ z@$W7TKIYR0?&~iD;O6i1ioUOD82&@VKOdfZ9l`N>J=vA(jd-0_td9p6Z+GIaiSNGj zAineR5Zvp%hvE8@;mtYjBXH+;3cORs?*@Mqzc<|YzV7fC{{9*NQ268cC&P_@CVVRX zXt@40@F(ym!u2P?pTwUA*Pj7@3V$|Se=d9){sOrELip48KF_Sbe6fL*9>scp2EPmZ zS-Ac49K5;+QhF3|o`>td0Ix3O(xdRF!}Vvts|&gGDEt@U`Y*w&3%T?t{FmYSufVGd zx%4RfSK<0I;njs)dKCU^aQ)Ze)rDMo6#g4<{Wsy&g$hDp-1}vH*Spu@u6OT`+@HO#yq4qoyviN$wv6jhxbO3M z1%5H-1=rnd;y7RLX83y<{(gqf$?y*{{KE|YD8uJw_{SOkNrrz4w@<8_ePaK3Ua%hy zp#JjXGW-m<{p0>`|G58uM*i;q_Lckp=lIh&uKo5J+xL6-2v~eJv~;}}AJyR>g}Zo!rH|dhzYD%Z_;>MPK=^mT zmkj?dJ`4~4E_fVEfyNepl^z8z{=~=B@Q;GkkFfM{PWX4hmkIx_`4EOt@L02gSHqXB z5zBvdgD(#s0ACS493JPo`0`kI6a1<0rtmrN)!=dPMgFV9s|!$i6d$UKe}qRjc=HHL z9|yqKfDebqYx3gDW8p3Er^4gCL-FM~@U`))YL`DLz8s$`E=V=JWm&2DyBj=S+ZSIL z0FU?V1s@J?Q-JUo3tun7(#NUr_V79IcrK{b^6XCt!_rte` z&w}p&{{SBEJ&P~b37cav-W~B9z;}XofbR_N0*}|U#g_-b``{0S_l1vw?+U*R9Hf_y_R);7iple^QKhe|Tf~5O@dp0q`#Hq3{9l1K~sA2f@#V9}K@3 zeh7RbJl-o6U!DO!41XT{aCrTt%byf=I0D`VUVJT{BYP!0?!|rKM-k^}coFCQupgZa zFZ_5ezXo3TO&gShO-cA-;bP=zco8R_1LneuI0uD`f;wT3D%jEG&^3RzfFA?z0Y4T# z7#{CIiZ73V55pf1KOQ~}egb?h{6u)2aB)`Te-gX}y!hJKFy8gxMSUiQ|I-Uz`0K7) z{=&iVlY?7)7zIBCJ`sK@dZ+Qf}aVW2tNxx13nTy4}Lbh_?ec) zqsZr+ntzr5Y6CwP-V1&n{9yPf_$c`K@QLsX;4|PC!so#+g4b_Y{-ntNVt5<)CGcMG zOW_B@N5eV4kr!WkA^f1h;(LMRbyqI`hvIuY zBQN^u(QsZVzPBjyBF+y@%KlQ}9BZn$88jDwu@s<#BUG31HL2tPWa*Q zyWpeYcf%*a?}0xGFY-Jo;Sa(mz#oD?2Y(p;D!j<^nvmx=@X7d%RxE!~jQ0_Ed-xRiCh#Kuts(yI@JI27 z!5@R41~1}F4sou5KaPJByzn0j{#5u>{5kL^;8iP?KPkrhB)l5_6ucXJ8hil!Y4~vX zGw`wSXW>)f&%vLE7vp_0jQ2D6^Z4Jw3x8(t>#tmvit)aH-v&M%-U~hhelYw+_$c^G z@XO#uK5vG6Cc$6Ee;xh`d;$Db`0}fiKPkpF6W#&-8oUd<$mg?=Pk;F9`1`>N|EJ&& zhrfY87XBuDD*P?@9QZ7FykRNE^)|e2lkz78FY;L!@@WZw2Y(y*-{A+qXTy(x7x4#% z`I7366pWt)w7hARbNvy-M!EXWo0KW(PL-=6$ zNAMBwx$yDukKxncpTOtBKZVx`_fJLspTS$eKZo~#e*qs1{}Mg|J`X+~J|8{}{uO*K z{A+lf)ykg~`F{g%059h8f-o;_;NRl+g8u`4F#J3CDEI>SMELjc8So$A^WZo+TZ zQjF^-_zLi1T-Swhb%6hj-vwUyw+DZB_%Haw;J?Dhz!$zQn8no#2bXcZb)44}-^_4OIjh1FwVs5PWg?EO`8#m_?iq;6**}3iYh2F1}j6 z)?Si0)zvk)<{Eq{csF=m_yG9Q@Zs?IGvtfNW8v{VFa@6qZ$O;q;YI$Bhx|W-FN6Or zyzpNRe*Na9PfCwsz8d1Ufj5Hpf;Wcug%|N(3-OPJFN=RNyzu7)e++y%{DTkArv>JMLq`1`?&IG=@a4TpzY z?wUB~zzcs~@F&3I?`$pp!*lTXI|T~<89dxl7PxAuHETZ9%y0Z&-Ga1)uTK8kz^ma0 zz?;L*hOYsi0ACaS9J~emRd`W{pFsu4xV0iplbHx`%!9$1||1x-y=Wij;yWzfM{xH1o7pqqjrs{Qg{N207R~EoqQ=jG6 zD*wM?tZm?{!Hf8d|6k;t4}4ws_{Xz}>8zt8D2gKJsLD-l($qTAEq80$tj&^at0}j) z$xU;+n?Fl#+jNSIpy=o*ilPXLqN6BkI*Onuf{r4i=qQSyev04ooWGxQ&U4N^&uzLt z*4MVI@0`!~`+WbL@AvusJy!Vjz@H7zF9OcbpKm?C68JoLeiQKL0N(-JKI_dEegHU! zf1wrr0pJM;=NrKJ`HQUQj{&cS=gXhYv6%0#{f!L%dnNEBJpXLq9R8!M@D~HGh3BsZ z&d)#AdOi=_zI%}4-2!|*gtHsC{ml-3ejNA$czzQ2bAkUGINS3C%br=!;Lyx`xCEYG z2z(*%mB8&ildR7I=kT9xg+BuPc@WMh@aF?R2)qIKG2n}Um(StYP5bRT8~N`S0%v=P27VHF6Yv=qbL!0WE&*Nz zd@1l2;P#pW>-&LU1h$C-J$&V*}zx8^9{gX0K5q}+w-mqI24-i1a7YpG2aaQg%Hjz;46WT0e=zjcY`+=_oz76;d!1n^r10M(81AG#AFYtuu_Sz#`Uhy2){tfVaJ@6smD}ZkVo&nDF@Y7a3 z90q<9JpWqY{QMo(^P|9D0?!`=z6tm-;4cMUUgf6iWxy+ezZ`fY@L}Mcz;6b=8Tc!J z?*jfx;I{$i^nTS!?_uC0@cf;?`T6fz&rbq>6+B;&aMSy0;7Q=G0p1RLGw=fN*8<-L z{B^+h0)IX5ao}5kPXfOMcty2q{~LfOfxi*>3gBCT4*|aw_;%oL0=^IUn}Htzz76;( z;BNswv&Oamt-zlQoXhctRyj5Re;Yi%3ix*5!@%DTd?)aC06zfyoxmr6?*LwwbnSl^ z@biHW(j#|S_Rj(SZg_q%@b>`k2EG&cX5jAyz6<#KfR6!xKk)AXXFGpw*?9u^C_F#C z)=lpRfX@Z~LEtUGKLorV_%7gEfqxkI9^fAVei-;ifu97v8~8tgb9(Q#(pypIrt4$y zd=mJ_fwu$y1n>gzJ;1jEzYX|4;GYD31o)?bp8~!Y_{{mP{htP25BxL0R|5Ym@GNl7 zpFddnGXi`cJUzyxz5cKk!Q6Uj)7w`0c>Efqx13jlelw|FqJz z75D*o{_Vi|`3Fxo?b{3d4tRbX_?LlC0{;r|iUn?Z4+2jD|0?iy;9mn?0RDC0+klS& ze;060*ZEeu_5uF}JbwiEH-Vo5{w?4$pX;XU5b%Y-zYTmP@b3WM1pK?ecK|;Od_VB- z0Y3`-`@l~F{{iq>m$>$i178UIhrm|?{}J#_z<&&U2k;}n_XEEZ_yq8u054nU+WAx9 zvwfiM0RC&>tAPIo z_%QI_0^bSzIPe3&?*={r{CB|1p6A;Cd*HKyp8(zf{13oa0skZLVc>rPz7zOK;0J*J z8TbV7zW}!nMez9XufS(N-%Zyf@CM+21HKCQ-+>PU{|E4$z)t}`0Q{f8CxHJS@UjNi z&VK=)4g56l2H^h&-UOW6!x>h4*bV$Y@cb6w{{_AqxG{-BVey|s4aN@xox_=J{W;lS z{tu&bfzMdv@&^I00$vWh9r%NRzX&*|tIA5(CgA76^E-e)1o(d7(|~^!IETN`3jY}J z^WpjO%iMH56!;wA7XVKH=kOb>@LPaShv)l&UkH3F@P`3^J8%xa*$RIj@Q1_mM}W@& zeiv{Kr^O1V>~c4~7s2!A1Lx<5tmi9%KLVa_1pY|ioxm%AZwCG-;Jbi78u)F%*`7_7 zJ%@om2A;nYI6uGHdVUi4OnAPc(and)0#5>e9PoDFj|W}=J`4CZ;76ehBzYz)t{w3GnGxyY_DaJ{R~)fj0wx8Sp&tmjmAdd>Hs{ z;5P$51pF1iPXK=<@ae6t{UgBV0)G|oX5g;|-VgjWz_$Y5415pp*8)Eb{B^)j0)IX5 z8EvlpTYy&qzXf;;@HYVO2mVIjTY+x{z6bcNzz+j|6Y!J3-wb?4yKDb8;8noi0=xzI zTY>ije;e?vz_$b61N`m44+DP(@RPvb34F#iuKhcJCxO2UcsuZS11|u75AbckcLLuF z{Jp@(fxi#Tu1OFiK0`Lz3-v)da@V&r44165;M}SWP|0wW^ z>smq_~(KD4EP|KynVW9_bK3Cfahnvz)jbF;Pt@22z&+b+kp=O z{}S*Wzz+c55Bv_`M}dDCc-i%?onHYy8~8!s4Zyz&d=>Do0Urkbb>KUJj{!db{2Rb0 zfPWKs*$Z9!zXg0Y@I$~GfPWkKD&XG%J`DW3z;^;a4EzA_?*X3x{(a!~Wg?7z0DSgJ zH(lev8-V{1_$uH(0zM4<$G~?2KLY$9@H>GY1O5}>>J3Z6d%{GY&20RKPW(=!w({bBSk;B$eW z2Hp((-@x<0{{wsr@c#ne4ZQ3@h6%=h4go(0_zB?W0-wIx)N;@d0zMabIq+uS4+fqG zeje~Gz#js9H}Gk|4*@?P_z!_|IqtN|aT54L;rWUVH(eJ1PXeC~ydC(3zze`127DXv zhXdaWda(1pWx%6`ijAj|83sUIDxv_@jUqfIk}eHsFr|z8CmR;N!p_3w#pz z+|5V^f;FZAJ zfj~mN0j~u<47?8bPT=!_9{^qtd;<6a;AJ`2{^tUp z4g3<|4Zs%yUj_V9;KRV52Ye^+=L0_gyaD(G@I}DO*1Gmz27ET~%YiolZv?&ycnbJ1 z@GF4t1b!v(1Hcypp8(zjyzB%V^}s8EzYusM z@Rh(jfxig&2=EsJ9|fKUeh~O7;KzVxfS32W_OAwB3A_V%Bk)e(oxroeM}T($9|gV! z_(9;^z>fjX0Wa@!?OzML68H_k8-eG6cLMJLJ_5WK_$cr`;0J;C13w1*M&RZBuKfeR zD}fIJZv;0Q?~EA>hY=Zv%Yk(gFz8Uy& z;I9QfjRynt=Un>tb-?EUe?9QUz_$SJ27U|h&A{IPd>8OH0v`ju75H)Bw*sG5Fw8gX ze-rRIz~2mfG4O4`uLjOhjZHVh?FarAcz!GJw*ub-{B6Jw1K$q(B=ENbuUO}%>m9(8 zz~2eH9rzC51>o-jz76=hf$s(W9^m7^cLJXT{$Ah}>s|Zb2RsS<{lMFSj{+|M{{Zl9 zz&{9lFYpfm9|yh*_$2TT1FzWN+W!&YN#Gv^ekpLSKPRpFvkLfbczzi8$AIqy{&C<3 zfPVt`1n@n;%Z6P0Zv#FX_$Prk0RI&5RlxTG9|rzu;5&hT2KWKsp9MYvd>`<#jjsKl z1Aabm&gb$A&FAI-|2#au82A@}cLU!Kd^7Mb0^bGvcHm>czXbd^@B_f7-Q=e04&ZZu ze;N2<;9mjW4g4VR&A`73d>8Pq0UrbYb>PQ=j{%?d64(B30G|W=o4^+X{}%9W;D>;3 z2L5f}yMTWO_!#i-0zVG?Fz{)cT>HNVd=BvM178gM2f(|5j|1Nf{D;7I0sj&3G2lN2 zejNA_;L~2}+J7hTIlzAcd@=B!0`CTX6!>P~KLfrC_|JhK0{#o&CxA}?pZ+q}{=0zB z1^!Fm&A@*JJP-UB@GZc94SYB7-vB=Z{I|eQ0zVFX#>-v%?*?84{CB`xfd3wNKkyU4 zw*vnI@IAo)2z(s)pMXySKMA~I*tP%9z>~oL0=ymgUx5z+p9H=g_}_r<2mW{9M}hwX z_-Wv$fX}+Q%=p9p>OX-m1pa@(R|5YR@J+x^1K$b!-@p$5{}1p9;Qs|)_6pa|vU2nP z`M>7_=l;vSf4ma-Iq-ZV@NW_{V^=J+rO;^bqj#;Q1549|C;( zD>-%MIAR*`xxmi{-VFSq!1KT_0KNtIbl`6R&h|gYvVSk|3*q@M0_W#1v7Vm*{xEpH zY=l!~rt9IrX9J%ByaD({z*hl(1n^=f8~91! z^MF@ucI|%-@Fegm;O)Q@zze{ufo}s|1AH&=B=B+IwZJEV*8#71t!w{$;7Q>1z*hiY z0DK7ebAfLMehKh>z!w5P0{l|or+_~X_{`V2_CFtZJ@5wLD}XNoJ_P(S;5&d{4tziG zM&L(*r+}XZeg*JZuXpXg68J*ki-E5M-UNIT@Fl=^0AC7xKk#PYM}c1j{50@oz-Mi7 z?OzUjA@CO9D}i4Pd=v0i;5&e~0pAb29r#h;*8o2a{953%Z*lFv4tN9b6~H@zzX13M z@aut(0)HX!gTPk;KL-3oz{}s@+WBJOmB7=$8-cF^-U&Pdd<6Jv;G@7hfR6$11b!TN z7WlL`y7qSgp96di@WsHpfp-JX0pARKE%05yZvZ|9JP-Uh@E+jPwz~HB0-poC5BOr> z{lL3{-w1p&@B!evfDZz{4LO;1<%Onw5>{3P&oz$aH@Wt|40saw%YnB89|m3kelzfGz+VA;AMjTKKLUIN z_$lD80zUK2uKlkDUJv{=z*hj@415UqYk_YE{yO0MfWIF25#U>Zp8|dh@R{3O``-Y3 zA@DZ>UkQ9G@J+yP1-=9Ln}F{J{$}74z_$S}dy8x5TY%37{#M`(z~2UZ74YrAhk?Hx z_$csq06z%)oxqO)-vPY*t*-s=0$vIH-M|}xzXy0X@SVUn1Ai~@UBKT5d<^*efgcAx z3VhnzT>C!&d@k@00&fQXA>eu7yMS*2{$b#|fqw+}A>bbceggPz;M2Fe_J0icT;LxE z-VFQ`!1KWO0N(=qHsHH~?*)Dc_@{xN0Ddl+{KAXRF$|yncGv#f;rY41zXZG)_yOQ~ z;CBGu0{qLscLVr-1(u_{<%y z{XYWU2>i#uJAoeoJ_7ts;G@8Q0{kHGp8`Jy{3!79ce!@{40t8*p960M{tMupz$bu@ z0KW_PDDYnbKM4F+z>fhx2E6>;uKm9TUJ3j+z#D=87I-J{;KI1*E{eJ*n1^kb|TY&!wct7xyz_$YbGw?mY{{s9l@V^2-349XxjGeCie*<0x z{O`b9fd2z{Kk!q)w*vns@IAo)5BOo={{nsz_-WuX-dkq;;kKXu-DS1^W%C9%_7pO! z>1ScU{Ml{)-#|8>nODdT70TvyW(t|Id8-Ep%jP+BUi$iFOWO(y3LE>g>0EDDU-rgS z>(aJ#nW3TI_HO8DS?~O{f3b#rx9I!(_GWHX)Vo^@ntI^WlELrbEz-vh~hjI_DB+Amwz8z$uBb3zx8 zq<3AqV`!*3QJ+p{hsf&QHQ98gw=`b*>o69*#^*l|HWeR;gxsK@Ps%wLxrU!Gq zYx3DtOIX-sl^eF_X-qJhQ$1gZrZYF#pDA=urgKE0(u60ra8X;KCbN35(;JcryDT`+ zXVxfBTx|t1j&eAWSWxIoQ!S*_r~C$q53TjawM=}d$0;AOr66E#|G@7u zy;Lh}(oKDX=_`5)UHQzKL8#Pig}#C4-*@x{e(yDpf@p0GJmY^i+6tLv#h;?2*5}C5 z)tSL;axh!yN(BS1^1XGG{O;@!h3zQ)_9dzYJ963HLar;<5&m4Fs;j0tO-0p}sIv>! zdS4O=>sPC--x=jiFVgj7aAHBatv}t9xgnb#q<25tnJ#43)TZ0gg&f=?4QxzN1B8{VwsSCdQ#MgU?NzQf*J!j|cKcvfN-{F)A-h3n zE7TfIPLf*=e=KCL`P(|PW#Qku!aB2cs?K#@UM(oCUpGelj&#fY^@ZZJl$o}u-zEn) zWcpLB0Xa40mcD*%U#>T^ni?M0?6yMoMmGjBJ=G>utB$@wUqzt0*59|GD@dtn3&7;y zjRSOW#ZzFElR18tu$6-b>CFV?3wo)s&U6*B1OE89{!lw?)wce^KvP|Nws){^pc&dm zdZL-iGTq;o&vk4}Ri{#75zxC~R$2%~cruY(lPz4mA=_Ki)SK#A-Py8qf$_JivV}~` z(nPiKm+P{*HQfd3RM&KLaG#RxFHq&DoTo~t3jphQqRx(bQK8VgtZ$$vDspNTrP#G$ zcnqku4OqUeP%M5qdR+*ik(ZY97n@t zTT=SbfLMWeOmJRV%7JV8OEN$igeQ75zOipZZ%GEI1M!&PbQW#7-gfT4&<44IFCvkH zzr1fy8@w#R;MqB-8K8k&Ur)L-yFS;EHQh3yR}6bJIylpfh!vddad@rk$O}#*IWN~{ zJIM7lMgmQ1+cHgSQz@6`diwKb;EorSsHX8k|3J2b{HwlRcl6p;;DN*j0)CUdo|)-X zVXTyql5<41&o1Hf#zLme4{)L5laKZOPP{{Uy@kRXuJT1bK066>{~a`krnEs_*}@aBiWzH4TO~j z8EKFC6V)_^rCAbl+~3qe=KH3Gs2+8s=#O9!JzRo+A1~vqp+Z=m(d#nhDZD(<)Jm^R z>!P~!(vIfU2}<Ex^u583VTt>gVVp22 zpDs_xotb*9F7^c_H427B?rcpW=}a7S^sS=;Dbg6nNKm56oPB7_4t7L`CO37Mkr-C2 zmJtQfk;iKuF}(0L()C1mrn0Gr+VY-KJOWE7mUf-h^G7}7vm;M-T{lg`w$-OwXl*IIK9kPp26?h~ zO{%Buvh+=Xp_SGxrlMVxr?rT5cdj!#V73&_P_L#ZGZePmHNUHGU_)l0lbQnZFfssIp0-_?<&-6D3j%J;nFdQ)U7O~x zz4nqt!Cr#=sK7`l@SNE8+F@BOy(iasZQTC7)gvW&?Md=?bS_5F6dtnEs!kIp&skhb zrRc8)^!weG`Z8%R>v-BqTc>(W^gdK#)!xGRVr1xK9gTZDYYF{60}_zC<=A~ zq~?NoFpXl-ab=}-yh1LMr>q??UMi!a2Ue%JIB4M!{K|Buqa!;w7;7o0d39Ba|4sE5 zr1C?D%pL(y%9cy@O8a78Ufo3PT8cMBh@*U-=n-?v*VtQALgJ%#J@q_ew}+f@B@C`R z)~^<_%8to;BWhP?*W`MOKUb6O?F>BSHq^b@rk0+->y}Eb<`GpoL0R5Qt6-UYI!B9x zv}0m`DmXPjRL2)lt+W&Bh9@bc2>SckC|kc&A3&$j9AuihLSBn5xuESl?V@hHE5+I# zswu);f$zb{s$d0{?QCvronz|M*;d1r;u4S9cu^M7aFnHa4;rQP4`pLxjK zl^CIW`;(~&Sg?2*m1QBM8Wf_dLBWv5s0P-9k=4L@CZZZxPk5?9u&ws|nkM~tt4c0U zz5{h6 z@lr^H>^ujjx8;s4BnY`cb1U!{(7$f5n^qlY7tC_X2mkw#NWey)CY*#` zLSs*#T~z*!{o*vI-E{c`c4uoRgmRrKM$%^mRY&bCjuDPOH5`uA)f+S4{(1}VK#0sX z+K@{&V_cYXsjvu3>8EQ4vh*?rI#;)yFJgVsF1(07M%Nw|g5zJ4gT$H1Zc?01pJz?& zPP<5-t%_Rb4yb~b90fP|a<#PUS8Yc=n;F1PJm}J4^@^T^68yTh0y$vzt4KbVFJ=5z z_j=m}?%wht-CJI64eb`domI2F9IGg)SFwk(v}aL#QKLq<$+R_Eaq0#JXspCU$*Jzy z17yLTMaEjn8R&U6ol2*9Nj`0_(Wa>NO&e2K2Aa~AjE@j-k7p(JGGkuB=|T*T z(PIs*<)2+^s=ypN-9R@MPkCiaZ!wca)*TQbb?6IRg>uHw|<(B=3 zJ*s~!fm}aLY>D|5_H1NdAO_Fr)#H7Dy%JMgAK%CLZ^=ARg`gkov9 zp&1MRmiO+`xN&%vSVCT7cW6Ml5p=xrT*85A(a`tQUP~|i9W7MX^kfGHv%S4Ef3<bjb`1=aO+Nm`TWF`i3SRaeiishVHQK_=>r|0JrDH8s_>HRj9eAAS|8YuEMW zXuQ-zt9&#uPN(_kyxr<`1=SvCp@#3 zxS(h4Zg3X#g70raSd*7u+XNNFjq9?;*;UFx^WYdaRIS0vO7&WPTz>cDXW>wq??rN7 zOXFn~H;70L0`DV{E%45uS^HdMf{>&QM3Y6~arQt{a#|7)zd)`rQQeEB$#ox$FbM%5j(2dT+DIm7UlNE7G)&i07{5l|gQA zBik=4E3K9);QqUrYCR~=Q*u?L7QCrtX%6=x3Oh?$a)BUeQHr;_rD^6ju+o*T$T~`0!H_}J*l1)?E`8{57J4Dj)L8X+dV6k zjq9x#-i$qrUO*|P$+2{Mzv&|IzN6S<>9FbwRUvu{Rq#Q34&1h!^2_U;YYeB69!(a~ z$i>-w*J-&pO~u}irP5BV7V0D69ciU1%_m*FJu7b&p{cIIKp(w`B;X1}y&gYH(>LB{ zxjJ~0=gtlUCky3AY5uNw5qgZulXq6cC~EThDXZut30-_*RG7#Kh7`?aUqxedTEZPP zHj4>w3-0+W{vPm71G^5uD^Sy$-QZSKXKQAhCq)m!m!($aqIyjY3b(>`z&!sFoU7?c z$?dQ)O)Wic*KTb3sAzazBf~qGTbB0le%M5&|a*A4rL~UoT%i8LkHjY;Dh6170 zvwaGDQ|RO{-2>&m2S!XrEg;9eP+euMNjqB)h1+d73#cT5Q^nS#VPpkyW{K;}$_k?k zr$wdkrMbu$jb<&C6lSUUT7jwdx__jN+(R$-`DecRVNKIhi)&3u^2jgE?jvbolVx~~ zRc(W%P54*QW-?>!MkBOT%UX96W3cJ4jZ?P=Yu5(a1H0ZsLL@IjXe7;dhsRoW7$xjD zO>&bHFGKBZsCH~%v9Ga+w=vb|O*^`?bjJ?AE@`^y%~`hz)-jx!ww8(9D#Mi{POD=H zDd+4qKUPFy(Duu3g;||2YEV~ZUA_RXKWv5#)+0L|>Uz{gJ05-gid`zu(c8frmeHG& z<|~S5M{o;CwK zMQ#W0De{6_twbTWF(%+1QXEQ{;_= z4ezk2WYosPSVOA3S@fewI z>}D=Lob#V`p$v!-xY24SDcQ#HK5JcKIcGA>*5znyx@Le@g^fK^8M^aewQ&iF<9I?s zJT0)Nd6|;Y1%?I{3->2;=2u&14YR!kZWsDgKVd|y_? ztrz?{&^ft5y56`z8y4txW9m_y!pEHOjRWj0-NoITx?xUSHYnCU5s|GplO9oZ% zDZkRk>TQeX%vp&%*pEwH(A1}z8*ttW?e?K@M{Z{$8*FRL*H74h?V4$=OP{?QF-yd* z4(w(GXSL2o9_Z%P&WZ)HCxKISXC+#$@iAxrQt7r-zOp>3C(&hZ&?6KTJV>!Ui4ru7*uJtTrj$2t^<4HVtyk*FDDDl&$EyeYG{Xt2@Mxu^x=k*TXo;p3C*jn9yCBEmV8h zBlEA_m{lEB&;a?n;(KcMRc%E zBNf*vhzyS2E}rB)L!(O1r5M#P&@e|PNWc6j2S-K7O=AsAHx+-$J=@)5*F3LE%uC

n|bpgx`@F)JNea!HJ4N6L*2#XJ!;+j3 zE*q{&G{F+OvT0YXKwS@yO65IxgCUF9Ai67+j5l`ySO!qnX>KPWLvZmETkbg9MvAY1 zVu+D_fq4)EClvyjWUnlTO{jZ%r%<(j;+rqs4=opK#7vF6yz&|yrLFcynGfuW^gE^Q zevDW<0`j`|E1xRCVaZNmMV%IN^*Gjd@_FA1rj}Mp9R5oo*9_;IK zkCV`It{1J_ zei&T5+!f9Y1gd4ROqIgW=OuF*2`g6SI!`+>E30cT z0Cl#yK?r$jS16tlaK^31EOD{M#!~VnIppZ@es!rKwg=)?XzraoSeO(yUDatGII_+Z zH5nn&eQ3rt1R?9{R-pAAnf^>it`PVt9`-c00rJg^Fp*+CD7jhm^7_(4pX#I~Rz4}k z=b`utQe!(kPJu{sb4zo^1O)YSxgKhlgZIRmyYgcM<2Q(JM@^)gXtzFHW@_)xr@1=5 z5E8RIm=r0&*ohP~Mmis0c007$o=P*gmzjn#=HY`JkB-eHaQbSce_$K2C!3O4mCm8m zT|KvJc0^Mto#C?Dl90TTn|phB&)e#Sje!0H34MTMiQQya6F@>Jc4dsTs`&g33@M9- z$n$0~R>14Rt6QY8bf2An?Vf5xoV9zbp1z4)PnSr!8QcZoBpZ96GcQ#kv>1gcs7fnC zUqDk(biNSjtxcUli9w_5_EA=@%X^~S_BS9P&y;sQoBLMz_#Cb82gmAmK?mi+<^bHz zAKr=tPKKE^ta$Si5Q_7Pw#ELZxpg?xNo?<7PA5di4~Rd1Er(=P_Bd15asZ{}NIlgA zzJ=AZeVKd|^dbW8PH6tB{jESq5fjz$3It?!0ei;AMw#>Im|Ix8jzK7v_d>&EdK>J$ z(sU!KJ)rjWYtj)$I7&&KA$_qakzU_o#7n0Y173s_g0I1)AgVYBZZ2X6;W?0L+Tjhv z@ju<`5$}z2pi$QkF1AnUIYk9`FLJLY9Qk$6u^O2b_Td8BGZ=9%qyGqOl#!7k$!BMy zf}Q?LF zCF!=M^cUJ5k!!m;P5-p=ljf)8_+YirUM14)#)rtbb2C15c7MzV&W0sL^C9YhIo-J9 z&4@(Uk#_74{0)oTK2S^VyJ>)d_f7P_PtqBCE^%7lr9w9k)YCb0-UwqHqhD&4e34`H zfmBwfmc&eh*gSihDyoB}ri#xpENM#`6@CFFX%Sc8#a#KhGDqGF5J*xDB}pDs#YjSR z$w{DB)<{7enM@=692E^nMRlnqao-{Kzm`x;^cHbxQTL=Oq4{*|MYsXOaLRH<9pEEB?ByCR zi+rI)pxV}RsDTd64OAcE0tzuKKz(S5Dt@oL;}W9}GJ@isH|Psjr59&2_VM=~3c(+M z@v$Atsg|8#>(Dz5tJB6826DaY`);7SVD7J%obZHXx<1`xC4(-4v%C+Q1*`4Mc4Rj4 zSD8$or?G!o8(lHUGleGLBQnd|j5|`=t_$Y3omX@O!OSc=6ko{F=c2?2?@R6uIo*6i z40UKr&C6kopN&MXFx(v)IT(M)TdK0|(c?p}lrr9{JimyI|gTL3ilz)iVCs4lXFZrkUwpGbj?m!6% zCkFZcHlqyWi(hbz5HaIkEazxtBe*2dLhH}w^uBRZkUK|@ky3j=MbR-dlt{$=W;PXo z=k^`z@}2p?mO>peZ!vMc4wufX8^XJN%%bngVVq2ts+9DeNO#_t1_*CKeehbHQZLoP+T_iAT(N4guhQ{85?p?L5JmP7{K{xtpGpGpp`Qp7Azt=js?<#9Y%iz}JCA+$u9a z@H|&n*#1Y6(^d6nbVUudKy|cJv$LYo9#*j45G;ml1zx!%F-K*7EUWPLFwlmej4q5+IJ102%;Iypw){)5|N!fOzo z=BEja5=7!cr>9!_;%KsVhwz0YG!~?NNaQ(~nPt^w>L+<*9a0L)f>!vg>g>m>|sML##pWM`~HRh~=4xZJ|@fc*z7D2S3g;Nnq}qd`$~g{O61{%MZ2 zjN3$@#c1$QJZH4)ZQNB7 z^a+dlvH6~S5eG|Z$j`zNb9z^TS8JFr^u2qvMQ~I6B0OmzkGZMNP5qrK@rUteC3>z1 z;*AsX%m80G;WTc(nSp!OzTkbh#*G;CZj>_Fa5ik`p2Dd8Au|~F@*7a{w8+`Ws}^29 zq0BP99oiO>qE*PE_iLXGON_%+!iUyvZ^gZvq4=2IxP>Cveg_+0J5a~N;tefkI~kmk z@jGI+SoVs&oo0~jtzN-@NCK`$fAk_BR;O9LfhfSLIs)=mp}n zE85-S&1?UJG=G>^41ou5l>fqZd!I8as`&*r;une~XoA03qJu16SixEn9io3)hD;Ak zP82DAF@he>1Ddg@;kBw5FQpiW@m3Z0;S>0IMb3U>SGth>#Ys{aley-YoeAwkiZNG@)GIFj=<s9#r#{ z4cHw8FQm&W-q4846=e69TQH|a_}&0 zEpG)o{8)E8q(VCTdRj=#(lU9TCRKX+I{5>m&WA(m_=BMIZxzjZQD-AGAcETbGO15y zR}WsQvEr*wFY3(jr-y0k=Du;Ogz?=1-n|!i8Sc|YcNkyxvvZiLNBpZW60314T_?@3 zG*+WQGw8uagYRKLwPz8}zD>dV2=~_FVyw71HxOxp)`kqvA~ae#9P^%_xTe4fzEUhe zxgD`xwBhR`&S^VqaRgHsnrp7eUp&Qg%GKIN`l7VQ5cYAY;OV~{w%YuH^;yd+AqUi{ z@Z}4psdaLKDzNv?lD|^ zE~LqfhhkMw5XP^UAaY`J%Fdk+3%(Kli5EO;;)GOuzNL6A1Pe>7d=P?TS^PRDsAE?( zK|7T-41y1X={wwxdxd zP@3&y&e*GS`3Q&NJ)ycbu;U`x;60HSu@6pZg*RN$4o-mg5Ez~GxHpbjTL;Z}CF9Cq zbKtWv5%MK(Jo#&%T%$%tKDg!!vzku0B&x68knOFxE}L7^T^O`OFb*KbIX_oJ$LAzB z(GJLHxni8dONJ!RQ>~mpdGle6Kx@MS_0|$?uyu5yj`!o^b_l$MJCUSeG~!9VHzJ#8i6ODC?R7mf!b%iETR zrjZw#xm?r6eG>mWM^BApw*-^zJCAC=;8yMCMlono&Nm5p-5u^TX$Z>4 zT=~ncc1FexnNRAGOn9qGG&PanU8v@nS|eDeVYAlrTHJQcPIioSI^HO1g=EYz__Ij} zl%~$b?6SRicgoMZ*C9Bg!`E|~-nx+aUXR!g)S<3#9t_v-PTS(*u6WT=TCT^MgY!() zIAuY8e44(tT>S8@?>S*QusvO54w1SCGJ@xE(OIQxTjM#{YoF`o$wPB8&{!i2s2qzn z3}pI?9)T1LPSFGw4E9JWfB)5$P-X(;`awjp-NMjMe>c#IFkpkKHI^|Xjavr z7Qk~DIAA_a>*)>w3rBUGgTQ1r<#yOfT^eY?DaDR6zcycOyYB$=Qp<3<8X)vT6j%j$ zYPTo!#im0czTv5H-Kudr5?5Vw=o<6JTH(X(DLPXWnbfg5VhRp)OTtAuXHuJGPw9!Ei_{b8y2)-V&Hyey@`1C zlkR9U=7m!0TSCr#dS>#)H1kw6&%MygNH4I|y#H_$^Lg0^ompo5a^;3Wc^uBcg}4A&qWKTC4GBq;vSf!^vL^d60qNf4ttm%OKZW3 z6l1U+dUVhe4zGOkf6F^>c3EjSZuFyYwLj#qlo-~g=sn>zx7w#+^jx%dAfA1pH zJ2@Ni@~df`+_GL8o;NX03ClibNb`Iaa>VZsT}C?%jiZa+@^BmJvrsHf_V_o0NGICv z!K}dn)@6P#i`p1z)}w`-nW--(E|X#nx_5LOaQLR)DAS>p{2#H5Xb4zB(_$Za^EWM~ zlZ#!DeQYhtohs5GtqpIGLR=>KaFREn>>Sw5jq9E!d^gR=A;su=MIpaZJhd)#ymE`< z%;;J>AL^TWyK=p`zJ97!9RoSu?CyjamBU;vV)2V+28LtR)ci=BDWbX&F=V48jU#4m z=96hfCXvLT%zEzkPK=$Cc)ZM-YQvcdBSNgBo-;w5pFDFMyNtwq^#mEnPjNdmm61KD z@TeVw{Oz#x^2IR*8&IKCLmE(zO~#iYOjp)8c%>;#Pe5iVdXB-l{-1kf>CF5;2S%0pH%uDu2^l@>nv$9la3sDCo}DKmbXb+hjr+yQK4spc}SMraLmdeZ4tKdQ(?(Pc?6b zxo;RwbDXWTP8PfB(e<@>r6GMfPt#D)`j>KUPTYIsvmw}#nNCZjqb3JznCBmN;sTq` znegmh^5#rwR(bd1(!7Gg80u}zjXqP6KW<=ho>1);n;0tMVoi**^lG|5()wia)!_X_j>ZWq`h)zuUv$$DpJCDO4sb$ljd@tZ_p_hWymYy8C-DFVV6L7hb4z( z`rN{lo^M`PRD%x{d5kv7p%nu*99rXSikmVTYpQID^SLs!jq*6H(mqQ&on~fXb}MZP zt^!e#_BM%1@ z8R%mF5VK-Q7L-_@grMS7BS9t%ozAHCKoq4_8hc#jbW>1T-j#lQWrFWD9FTJI>my=Y)e2O~>I;^Y9e@qAL2B9xA$Ou@SuI1W+TJdAjdC35w{A2BESjAmcF zCOE;BIPu~QEigwAHs``027x^Y=JA!Lz@ZcGeLj40#qFSszQF76(h>}H$erd^o?vLE zX9N2oDToAp|GTquAUinN*VTmQ4roEyyz{{IRy@yQGsj=K^c_v{2JHI=ne({+1)<%0 z_xo?e$vd;pWT!1P;KA*vZ3RAzXnjbJ+XizSU~i12qV;cdHC~&PaW{31@hzY^)2N-dlcc zu&Nbl%X3B>ZtNs=U$`5% z?I`B=jTS>_e+3nHV6<3*rTuwfOMqs;QEZ?k4fuFNsx?HhwlR9k2=<7owusUwOBi6B z=@wDg+FD)J)i!m8+#CuIwQvo_viz)=Q z+V++hnl^Gi1`?5M{hdb2zPd)r)6pI6p`jdF*k`?381$ZdgEpI;`(gp z0_%eY$VujXo>4Qu5lQayO(*&HQBpQXT?3Jz9r@&o@Icbq>+IcR=^Vy&>rh163niqT zw^){#_JS|%{w<7Wl6KnP7)-m_A-G-`{axIfXn0M*;kx?~i!dQ?4)nby@wWxKZ%y2K z7iNa1M5LXS?f_5B0Cv@ljf%m+4sOV!=;KaAgwI}urNP~;$J-i(j3uG&D?@ZZ3->m0 zBI!`1v6rru%8g#MkWWmIkH+7a2n-2I+s0SJs=BwVq;J`BKEZ8ON&J3orgIxvlyL{A0)48+S^_l!S@`5aP(x{3; zN9k+t!t&po%LuqUZb9~D=_6VpYI2g-n`dIUbEAwlx2CRcT^hMNl}{zQ4V4=eO^^ik zvnu26?`QYW&JJt?&phnTfa}TGlD@pJmyU<5>nMb6yIex#!0V;$Z$hhsExzxCm$ysly)ii)PeKDX| z4!g1g+1`#Uf4tDVBWyX1M4j-WW~Z?7L9-UO-}T7PV%9;$FPdZUY}|~(cR0dxwl{1+ zMoI1e8PpNyW%Q)dV_x_x>4>2qf^9f86||*`f;5(!`Fp&Mj_Ny)t$7Iv_8+x+CMqKR z9^aCx{+;wM8GGHjpmP$Q8G(>|eSa`4DmKa(adp?66LmHxrNb9}d0v91ehl}Qs|%fh;|`F>hb%-bWP?R&br%*lR$2e#FiV=-;^7_PBSa;e>@a#*yxR;?kL_FF{D* z^Xj^JiJG;%ZWmRA&(;%O7wuY8`sGNp^c^voKy8ZcM|JgdapoAtTV4~^U(9XaV31ar z<3&;EDMUs=*4Qgd^706K#>N_^{}`ZuT{ps7uCQ=bhy7jVB54Tv>tZO@%9uOzW9(6A zHN1CEZgNU!;sr16GA@zq?ayau5b0mjq~SpHM<;1X#2hrxV2Hl|=D(YpkAzNbe9|ll zkGbudkY2Nnw)kW_jq6PmDW$%nNJ{y#P8tT-5q0u4nw;X+z>CeDN8m|iPQ z#PY|-Xu_iEDw>rT*x1t7XWxI}zZsD&OmgkZuI_XEfLv9Ia>t`KJP0v?Q&@Rx(@k{g z#MzF8-VpPSodo$PhO1?_&G$?-$+V>!d%bQEmBJKeBU~wKrwn@?v5%yrORXWa#IV-y z?8vl&Z!0*17HW3nkfH__wer9MIwjBN-Mt&QF=@UI>8z}sMfr*(*sHm;z7)aKV3zZZUmQ3QYOPZ~_a`BG$C z6H$Swk+qIKM>TBLGevE2LX!^L*xRK4XsMUi+d7&lp>Cj*HzHWA*{h;!N>WP2jrx_wErjxkmTxg@Zn-H(dHs!P%*dVCdXO3Q-A2_??}Hx|$B zD&5>%r?|BZEJi0$v;I)M)>B5Q`9K8hELfh!=`K}nFZL5NGDwMH(>5Xx}sufk2~o{ zLuxpVi{YAARB*V)ay8Af`DQjTS#))f_&ruJdc#RH^Q>6(9IhKZW-V7;nyXS>BX7O{ z&REblJxf!T;eC6cvRD(G#gbv(e_~BkVg}M2wEGb|p3219M_Y-kz4^JWrWJ#=`k zXu5omt{2MCGPR~YLyhLi z(R8VFKDscxzb~Kb*qEv|KilLc9IJzG ztUWcciBoR!?n_OCG3HENkGqxLvfvJ{dt%;}+Aog0C2+m7(HP3FK>X>t7@6SN4r80* zUJqtAH3j;DyLrWcY6ib}^fbXkpW<=evWUv&nGxn(<0CHKzLByle%2IS%`hi~DJG+v zjMX}MY)s)T|5a&Y+L!KzrzeH%~a@YLVg%N%w=%s$t2x7ZrDmkTvooEYSU zKxgJJWl>{H2uw-x>MhR$4ttr+0VYqh!AvsE+Ix-nuATe#s@rMBH;tmBj#F^dgqcj! zZS&rOYMKk{^({*cx7OcsgvJx}YtcMzqtS!#y#J?!gwFQ$LoqWda0Tid78#atJt5|g@G&XdZuir zx`CI{+!rPUU`C0j1)LHNol&ZxnE`PonbIe0tMtww)s!}C5n~2Py5Rgin%&lnkte&I zuFOAM7u5;T=H9SRaQl*eMuk>gTA@+Hvz(UC0=^M>2>4ose`VU(-XiS-^uI^uR?8TZ zKgOqEVN5|))@zKo`Oo?kGoG(j2%$CLUI-gvZex4}rAP_{49$#T8I1pVzG_N(VPX95 z8Ptl?TEez4XW=}U6K)S{Oq%h23v#6+HpIbIMzqr&dV>7I~&*VB93yRfMGb1%&Mr@XzeC*#icAlx0Cn#ludfXjmAIC*$rFGKYu$8;kI5^vz^=95)=dC9H89Ws1EK!72z0t;leV8w?&kLqkSy zg)>aBs;3T>Lzo)E9`?O*vjU-N@MC-nc;{dBZ`cOLHT}@|474 zfx7q582-K$nv1{*tfh*UTA)Pno;R55BEH=6iT1{*SlawR|D!3h6vQ+&l-{Al-(Ge0bxo z(7M0@VQRYMAxxB8V>hzc98w3#VU@I*$u-n%3pp(F{CQ~Qs!7S?U}Lxd6@Yt02;?4p zMWEe-7bT+Is|{`U&Q3FTuQwu4Q2w(pLfsxg1JEbXr z>ILbDs9trcB~x`GmX~WhLxW%kIpt}!dcMH@CBdlUCe)|{z7q5HVYOj(!23e#YC&A{ zpn$>1hytEZ_YIbK#CZyH6<%tO+0HDRzY<2xvW0a~!9ov9yiv&hfl=^nuS@yOvSyyl zsd)b}4R19G5kpSEN|eyDh1m`@06MaW8c@$Qz`r|`I_Ba-q8Jq*%1Z0yl=9iv=Z2UV zvzX?ma(21Ut<%w^Vm}MT;QR7mJk!*ruz*V8Y!yOC_CXqBl!0tu%p#~)6oLILm=|3H zvSsQLNbq*Ygm0l~ie7_o6XU-GvVAcNV1c3l>}P!iP^9mc&4VyJt&!8P2XZ`O>_Nk* zu;mkP+e)_h!?jxgW6%-Z285}7X)*GHKc&Q8&AfrWdFkuhuhonSi|!<`{3U-J)EQVK zDq4&%E;32;cdwghQ&xStsds(f4YcpW{iegcawUiC2Dc^YV-xjhT4ZWlnr@;iBvZyu z+MCgKMVj1@HoC3D_!ICH|Hl(7?q&O5Pji)X{K;R#srG31^UCjvYS&qu_p#3~mEQEW zd*Dr{H~N9_rn`Ciz({X4pPA>Y@l;Pssqt>km8NG)+?lQ^+yg1LXHyxz+jg84`)OL5 zK2($&?C%r5;xmOQF=H}6+vT)}XD3bc#b)E?B>MwQw7IabfX@n~bG==C540|q_reaW z3w*|znxH8xU2`vRJw;{g$QBA%Uhcz%%~Z;r<@T=l`8;yVU%0RPfV2sU*B}<8+m?_I zNXPY!pLC%)eQSYzz$@7Y^jxS*WfK3s4{{v1mecNd7;xaCyKf!GlhtlcyXy)6!@0~K zzswl}hNdNa$d_(5rhoZ{;~M%mN&oVN$aVBDUy96^9xtGOX?#R~)>EO-Mzh{dIxJq` zO_EJZRt)yj0VKXY&1%44*2%dF?%p^@Or{Du9eJk?9iiX>NfXt?1@t%m$)l1+qb6R# zi77q}*5aGkcJ7Pg35EMma;Zx5W(|7x7E?0o>2Lg#U%<bTSsuWtr!|fTaO#ca zj|X6vN4I8LBZzyY@I&r~=T)AfzAg_d>4r_yd^%n?xGrCyyEvE7CD1(1wm*b%74P6~ zTlPQX=dT{{y3B^E&0-r#=iqJVn1T2_rX^!TEj|XXWsgIWuSB>V`I7w#<)Vq zN=09xJ3GMlsqe&4COf;b{! zSH@0Q@C@-8Tq;xIRk#reX^fqautR_$RmbBl(}HnHr-vg9UliN$SVznFi1}2TSRdQO zm?xj&ohXaQw7S?91rOQCr=21qtBD;MogEQR(nJJa9XoJ36(JoEnUb(+!@qFyVM>A) z_R>dR=p2$G%lH#5!X6HszUnY;RtW5|3Q>CAY?jiTum22onAAJabH(<)y$~H5Yk&Tz zh{R)MT#eT}TV;5Gf4*^E&!^ zda}KRvU$CIh3vd5mtFJh!9u3v2J7!@de_bCALt`d!45pzo$eaQ^kmECtr_TB*IzaT zfBrafPT565I1Lt?ZvVWntn8^5l$DiRf46^z>6@+Rrh)$H3Vpkve}+Q8(vH{uN%qfC z=({al1?U$k^qU0z>syv_D|A3 zOQD~zb@orv&sFG8*gE?s>8lj_Q?}0jN&19BUw)1$#QI;Y&`%fiH41%&pie6Fvjlyu zLSJd=vHhr1=&J<%e1(3Yps!cx7Yq6Y3Vpkvf38BmO3+`T(B}pHLWO?F(qsL-RG}Xh z;(wk(zeR}u`3n6uOONe$gF?Sci2orEj^F<75z;?Rq2Dj)&sXTj1pQTt^p6Yrr3(Fo zpubX~KVj*y{w-JNCx!T16#BAr&GsAHuU3VAx}a}Y=w}N0YZdxRLBB$wud?)*|1VJJ z>xKBQSLhoB{X-S?zeUhrpwO?h^fRFU&nWEA3;MMR{U$+wgF?SW(B~EU9fH13q2Dd& z`xW~Ag8oK@{*a&_Q0R{e`ay;MgrF}d^rr>=I)#4vgUpJH^>@8OKTFVWQ0V6h`XPmW zp`hQW&^HVEn-%(%f_}Q9{m2XY3l;iJmLB_`EeiWL3;KsC;@@WJvHibQ5&sTB|0acg zRIvZy3j6m6`WXuSenEecLO&+xAED5X3;IVY^b>-^XDajsLH{g;ept|R znOfD_`5D&#Y!UR&R_M12`gscdsGxt2Lcd4QS1I)S1${!H9~1P|3jMgH$NuM-qW?K6 z#DB5Ee>pDb=PC4)g1&}W-1=8`p4on3`dbz0pC;(;|nFHq=51pRXr z`mKV#N0I&=g8mXk{JRAGLWO>>pube1KOpFzr_dh~^v_r5k63!F|8G&`|AY|#c7^_g zpntnUe@f85L!mE!NU;6hq0mnk^zT&YD=a;>-wlfLn=R-UDfDv%{bdS$y`aBbp>GuQ z?^Wbqi>1f<`$2_%g`oeCLcdDT?^5Wyh4k-H=nI1WHidpz(0@vy-y-OFOcb~O*)Hf) z3jL^{ze1tkBj~SG==TfyeTw`Wv-DVhKc~Ap`Rh>mn!tL1pRLn`8P+Q|N~U{euX|i2quJ{-mJ4 zPLY471^o(ze%eE=a)fq{fiX(MnV5#g}%ko zWBc2os6Q(NeO94gCD`As(B}nxT9N)CLBC3&9})B!g?_7`U#-yZ5cC}i{VqX2sL21l zg1%D`{{catRp<{1`j;u}KO*Rd75ZbA9^apv75Wo`{*4O#q@ce=p+7C8ze|z-(`c-~ ze<(`h5BuM16#5E5f2+d&*@B+;WW;TM`K%=WLHlzGeZ8PxtI#(J`WqB_KC9!mKd;cQ z6!bj`eYc?RRp@!Izu*2og??Dj_bc?gSKJ@}jS4-l)%o=U3jL^{A5`dht-~LGL80F- z=+`OqV}gFYLO(9(Hz@Qx*U5jd{XIf$P~7*Q=X(A6I~97KtK&b|etbsJe(_w7Uq7Ts zKkv2m>o+R&6@vbA3j1dZ`kNH-&lU79QRwRh{p||-d93WuzdIEA7D2yBVgE`&|5AmX zePzG>FH`6Xg8tFC z^alj}w-kDgi~nH$eN&+y7xdpy=#N@@>xVJ_&3jto)}P~o{$_=KQqaFbp)aGpg8!iX zchXbTLi{6&`0EAzBZP7UKV_ zLccMUJe@M{(U7;Tl(*HMwev1(QDTRK!p#PUbKPu>7t;qj9 zLi+!$h<~3D|9=$vgF^i05{vu(91`@eQKWxd&~H}gk6L<4ve^#O-(RcH9~a_3pIF@V zpA_PsrqG`j^tBI*AOAFZmj7V-AF7DILeO8J(9gE?*#2Lr(9aR#f1RTIl0y8iSLhcC zdOp(QgzhoHYzq2DFw-=xs*74(&g^dGSF`2Igbp&t|CpQF%^3-Mp9&>t1_ zZ&u{rF+smgp+6y{zgl7cDM9}hMf~Lt5B48#Rp@63`nM_cvjlxok^VWB9_!zHg}zFN zzh0qVD8#=&p>Gu8=QELU`=1se{*{**#LEa

    1p+6+V|1L%T9TD^` ziujKS`sE7!NkM@-F15%0<3o!0r(1gLKh`VapDE~fDdMjb z^h1jHtAzA#ROsu4_-|6^8-@5ktVn;ep#O+M-!9lctgwHTp#P{M{=A^ytRO;`a^>LlM4M2 zA^mSqa|1Apr2|<6GqW?Q3=-;S_zx)yA{1MjQH!1YfEj_mXZ&v7M3h{4K=w}P@ zzeS;+E5!dPMg2(%@o!he-yp=lR}ueWA^shT_}c~jrxo$966}AkBK~e6{+$YaLD1i! zcz=ck{reR0ZxQtGSLnA3`VT1dqm~}u-_Iz@Z?_QtE=BzNg!n(K&>s-=A64kb1p9X@ z^y5POpH<}FQ9-{?p+6?rzgJ=ZNg@8vD)gs>`1dLF<&O;ZAD>gCf4ZRmyh2}L>9PO0 zU19%hA^tlQ`Z+@UUsmXoLi}G)q<^8H->=X&3igjF>~9g`|CU0(LWrODq{JP+cMAG1 zD$?IC*nhi1KP2eCq|grw_WwYU{w+fMKT_zo3Gx3}q2DRQ{}Y9Nm!SWtLcd3_|ENO0 zU(o+dp&t|M|G7ecScrc@p+73f2ENA6AFE&5dR+(`hFq) zFDueNB*f2q3gh-)BZB@biukt*`oAdR-yx*`uL}LB5dWk?zekAwZwmcBA^yKB^alm~ zKNR}Ig8c^-<#$BTe^sHM5bXcA!u}J2{%eZ(PYL?3EA-`$3hrMTQ|M<1`u{4@Kg-f% z`+Kfp{-9Eb|3M0Ul@NcqLSHZFAFR+f2=;$Nk^jwt{+kN@3PJxZg}zfr|05OY&kOoT zDD(wEf007JNzl(w=tl(o!xj21g8pF&{Wd{=p+dhy&`($BM+N-_3jJ&DeNB> z^gmU^zeUi`QpCSqNdHkq{5u8x6BY6A66}AHLcdqgKUtwaAn2c>&>s@?l?we4LH{&` z{+OVDxJ|K3-MCG5^wS0X?-c2uDd^WL;;$6+8x;B~OONF@q|nz3@&8_t{zgH6lOp~W!TvWZ z{Pz_?{4Y_&-zn&a75aWbe?pOen*{yMiugAR`adY*-zMlcE8^d2>9PL3PNCl==(i~J zdj$P03jIDI{clj{4+{D>D)fg1J>TgO*MB=I=x9PLpQRsIH`cEqKy9E8G6#6|v`u8gI`vv`{ z75Xtj{~3jTT+n}3p`WnySbqBy`r|_Ue^u1KNkRX4Mf_!t54PX?75eFd{&t0arlrUH z<2#|^_Mfu_{g)N`IfDKx3VoH3{(}nrLP7skg?_PM|JM}y79swxEA%S`J>SU{xBR*V z`zICkzh8*|TZ;HM3Gp9N=tl(ow-x#=g8hdT`t5@LdkX!ipdVN0_XzqQD)jpW{Sk$J zOwiw{(2rYsY=8cysDDR=_TBpg*C|uN3ruQs}z{{XZ1- zry%H0DfGjF{?7{gw+Q;bD)ie0{iH%aD(L^INdF!||93_F`vv_!6#6ki|9=YoxTII4 ze?rjzOQAm@=ua!`KPBk@qtKT>A?UyVSD~LF=ua!sKTFV`qnQ7lBk0dn=#zr}K?;3? zpf6YGn+5%Y75Wu|{yc@gQ_w#|q3;*;7bx_b1pU7i^=Gr7|BphyP0&AFVgF7+f007J zThKpJq2DLyD-`;JmLC1LM=SJ)g!mt$&>s=vf2=}3A;kYUh5m#PKi>%)cl>{H+Rom7s4|=obq5YZUs$g8o{CzFp8?r_iqw^n9m%-15r{ z`dM@*YFz!0pnriP{t-bxLs9>?3i=l-;@@HEvHqqN`cXl@TA|-<>8)en#`pU!Ql$TY zkp50Z{D%a6R-r#4=(`m9V}gE-LVwcIWBGL}^rwXQa|(U=?4bYr2u1l#7xa%*=qoHe z`ro|@`)3RKK81d+pzl}c>xJ~+sL(eE@ee5U%|iSYiu`XE^p8^LR|@uTP}tus=!X>g zf}r21&<_jgze%CrEX2P_q2DIxU#igW6!b4s=ywbHmn-!91U-L~C2s$FP|)A3&>t4` zuTbcZ3i`_x{=;!W|4K#tlY;)yiubSV$z{eLY(F2P&`%fi{7sy=Yzo6f#(2rSqe1B&t%I~lc|EMDV zqk{ee3jJ|G|3QU*Qb_+36zM-L=%1+2mp>)g{(n?q{|rIDTcMvN=s%{=&$0Aae?G3z zR|)YyNs)gGh4^PH^o>INpH$f2BIx%j^eYAZrxp5cLH`+rzF_IG{64GDZxZx3DEi+K z!T!%F;@>LhKd;d55cFSA=ywVF{R;hFLH|XC{(z;&^1EH39~1OnQs@s0`d&r(9}&`j zha&!Cg8s`2{Yk<8uPF4Vg!m6C^yNv|4AYKUn=51E$Dxx&`+y0+kb37jw$prEIrO2RVvayONjrsiufyq_>U{} zRf7I^3jIRC{<{_WMj`&+EA%ab{)9rmQqccFq3;&-e^lrTmLBWh(-ie*lc0aPLO&wd z|5t_mTLt~3Lcc@M|4pIaC8YoF3jH2I{||+JpP)ab&>s-=|5WJ5EIqbA&rp>AxS*e- z&`$_@{w8+0*{nK0|2T3^nf0kX>o?|)6N3ICh5nSFe}qC`{LpO{*a)buh1V6(qFI8PYCf}qR^iZ^a~aGQ-c0dg}(gh!S?@o3jK6TkMGa(75bTi zzCod{6!eP}`YJ(xnL@u%&^Iddiv|6&&Wr#4wp)69f3Hx)zf#c8Rm9(E>9PEq6!G^9 z@i!~cP_bBvR1%01Fzr)gF{qI-kM}_!rROt5z`sXO>-+nTbbeS;AHD;4@?K|i9KFM4=xR z^dD8|w+Q;(3jKCL|1pJrRM3B1q2D9uYZUF@enFp9=*I;8Z3_Fx1%0g|{s}?y`yF%YC#Q$7H`K=Q4UsA-M7xV`d`XNDoheAIh=)bJcZx!@kQRsJAdaVCn zRp>{B_`jym?-BH0SLpW(_Kzv_2Zi_#DfEX0{UwU}dsNVWTM_?pLH`|ve$vuo`7c!1 ze_DwDQiXonvx4pa^A!3DLH}Jv`e$2uO#fkpevT0T^A+|d1^xFF@iz$ee_x?rEX4l< zg}zHmpBzg5sTDC*x1LBB|$-zDgO zp|F3ipr26a4+#3Z6#7Gg{+9~<5kY@Up+9EnvHkg#LVrSt|JMrrDIxyfDD-7>&GsAZ zpGF%+i$bu~3bqCG{?mC22ZnIW>}&R7hG96KP2-p&^W<_quoepLXwhe!tIk-}mQzd!Ehn z-S>aGuKRutrm_#dKJa(;G5=A(f6zz$qJh8A2j3C+fA_(61^gi&d~d-2;e#Il_`^Q< z!GJ&FgC7d`qdxeN!dKDDpyc)cKYj2iz+dEpPX+ujAAA~!|F{o+4)AyJv3}+OzN-&@ z5s3e^kNC3z-_3`AE#S}k@NWe9|Lud%1AKQM@#lm1Z}GwJ2Ye+T`)?88FZ98m1NmR% zgRj^!^!!!L2VVv7)qU`_03YsyuMhZ3eDF;GU&9CA0`N6`@NEEJ%Lm^P@U?yLU4^gW z+P~}g;NyY6t`9x|_#=GqLjfP>WBn!r{z@PIi6H-`KKN;XZ{~xa3HT@<{2ahv<%6FG z_~t(NMS#EB2mhY%?)h!$gUPg`U6K`QXC<-`)owF1&mGVtnxR0pGy~9|hv?=!1_2{u_Pp9RVNfgO3C8 zclN=@1AiAE{9wTM@iG5NfWO6ue=Nv9&Idmc@I8I-(*WPY2R{?!-^&M|0sOsv@R@+` z5Rq!S4tB?LPR!Apbjj@Wp_? z)kps;UK4u$iTA-*0sP%Q;;#kxdwlTqg?FF7?(@M%0ss9z_-NoC>VxkH_y>IOaUlL- zKKOXRKj?!`0DO`UJ`v>qkPm(&;QRTQ-xLu4!#@0}fFI?9PXqiTKKOKy|7ailJizz& zQNJt@|D!(q*?=GGgU9cdwruAN&fy zulB*O72Z8RYkcq-(B>4V<`_%D3$g@E7SgZ~TgpZnkgtwZiH;Q!VKpAGzZKKQl3|D6wh1MqM2!RG=0_dfW1 z;NR|p-w*sh_}~u%{|+B~G4Sv7!3VAj-T(7_@Kpf6%Lg9;_#b`nO@w#vKf8VKEr5TI z4?YI)Kl$Lh0)DR#J|6Ht```xyzQ6~c1o&Tk@M8hL&j+6h_+NeSGXcNf2cH4>-+b_k z0Dr&-pAGnfKKQkOKjeeo2>9Q9@Y?}j=z}j1-o5_+@WB@X{;&_e2=GUI@WmkiRX+HN zZS?-_&i_vzd==r{&%c)X;3I&)$Oj(@{6j7def=s@rCCn>U%fO@=4c`OXID9YT6&Z( zy<2kcGWB2Q@5IEGd{z3NqB}1v`6B*9%~z=!2z;R)7_4W6JhZ90bskYcz7F0Txm$B( z?fS97$2hkw)c>8IY4S%s^6eDsI4T*51^EZmpScg?VS#JaAC}~A{6Xx+7nNYaJe~YE zf&2%kM8-46!*c*0K<-ignBnEe>7{t)G#0P_C^+q*T`l}h#iRZgO8w)6ch>(dQ2)AW-*93w z_n(s<{WTPi{LxkP{^R!V2L9f}e|@5^RHP;UTZlhJ{K5V^`Tqp`_Ywb{;9A^An z`Emb!cT&)ABoJOr_dih{>D2#M;LjrdfiHXWze)Vb!2cWYe?$Ch#h>(FruvSV|3>1^ z1O9`+e}wp(yrNCE{r4hu;&L8o_U|^~ZRhW+2tV8---Gazg|A_$-$=s0D7;IE>-hvxSgkNyi4ul^U(|9s#-3H)t{f0_96 z-N~!_X8zX@e-ZGX2L4-#e@2EjC0hJFi9aknbpFo(|1j~FHwLRn{O0+_y#9Ti#2*6<`f0+I>!siGd z^|0Los0RDQw|0?lM5x?>DGvxYb4)GU@-@X4| z2K?_5|0?k}w#@H(!tWB^Hb0+`_@iotuKxxg{=LLsX}+FcGk#;41a=XBEbvDHe>3$m zH(Ec<#c$59>2IWXG`|VJ-w62o5q}TyH`HbLuZzZ)f!@TQEdC-Po&CQF@INDd^ZMzb z*R{#K{xkhsNc@w9Z)5<*mw~TH{E4;o{J7(92I8-%F36mpX~zG?0&S}GU*+^2)89vU zbN^l`e7FG^Uk2h7kNO{9Cp7-%ApWT&{x34U#lM-vzf1U=)$>dDVaERji9fckjz32R z-kG14ApU@QnDaBu`bR9%Ci!ce|Ak=sZxPx+O>O zx^^OoKj|$UscnANllUhJZ;SsE5`W>Pq48e_;y+E|e|@pH_;0L?_rGPrU+NUsb(sB& zQ9P=D{$)D;JlFiR1M!a)zd1i!-q9x8{H!JMpB3H~|3@VLtOlX+Uk~CxLE^9Vfw%Z? zQ19tFk2L4Mv+%a~+sS*v=Ji`j!_fFUg80XX->m=O<=*07L*kzyd<)C`tR(TrH3^OX zW)S}|691}?yv2Wg6s({A0v#*1yGP z-r`?F;vXb@v8DbiN&HzYL*wrS;y*#+@Bf9j_;0L+>z^w8istd z;x7dL!N7lv_%GV3D-~|!Z`SW(^#JF4X8-GX*0qK3x2e9n53`@w5r67#9jR^p zS}7jQU!Yy+{M`rY=lq+1DE~~4{@}lS#8eayYh(9()uYb3HH1OX+{A~)f|C;|Y*LO^Rf8x)- zUi+hkcg}y00{=SVA1eM>i+?rohjj?u|Hc9T#p(h9^?#B0ZSz-2@u>fi!2cNVUr+p7 z#lOXpe;eYD5r2`Z{*MFyD&n8}i|)V8{~_^5-Jt95_CEpqZPXJ)RR3Z7y!o$DJgR?0 z$58*1!2cofxBk_e|2^UlyD`*18TebN7pPJG=f!Vczv-w-hvF*5qx_4-@4o(g3i!Vw z{`&j1$>v{A{NXo+u0Q9!Zu9(;s6L>E^6wyi+x5>~ibwfJ0sqs$KZE!OiN958sY~ve z``^>V9}E1?0{{2KpXSlOmG~2Ye>(8jt&8iwO8n;YS2O>+2%jgs?fPk;;!*#SLHsj8 z{7Z=cFY(*vXA$wI0sk!E&nNy`zv=Z8Ynh+#i9Zwg(}BOaTHt8?3=qHV{B@<`QT=m( ze>U*nP5hI@Z>xU-@#g{m9N>S2_+R(vf06hLf&UfY|BU$8dGxO({=m(l`)3C5|3>_K zJ^BlXKLYsY0)K;gc>d0N^w(27n!jk^p9lOsh`+%Bz5ZI8-Mzl9$Cbrg^0 zFAMk=1AlkoU+>Z1iTHDYKMVLDBL00I{SOd-KJdQ-{Huw-{6TN)Zw2ue0snizf0pH$7bqUhPptUUC4Q%Vp8@|Y;x9a;O}6W&>BOHW zesy_Sayb6afqxJ2U-*aio3B5b`4GDM!A|~Pf&6C?fB0b?X&<*t-8bW(PW-vf2g}qESMp!=*?HyP1pEhy{~qyQ zVF=Cn-$(pmU9>-5c*p-W@K;v@<~+o#|D+=t48|{MguMS!S@Ed;30*_;{|5N4BmOTv z`dblyCh&g?{QZf)+EJZbmnI}@Q;ba^Rrj{>c^Kk%=$g7cr-tW;&GiVdg)d_|wIoCAd@nAA$cD#iRVUiC=9GWe(H7NAalsG2M0l z-SyuC{2dx%f06iO%Gg8Sziq2{`~1BZ_#Y(xs3NB-fxtYAe+coXir;RMVZ6Q{{h0!6~1>FI&^*$6_4gO6V&f_ z5dZVUe?a{47JnM?=K=p8z`vaM+a1^ai#7cEyQSY;O8kL1J-_bxIRg9@)C+XZ&or+; z#*4p6sjB2&dE@IYgr6(CdV8$QVe-=npDny?{Z3On>VLH4?_R&hK>lwKf1XGGJmQZ7 z{u98zn)r*vKeB~M_@juwK>X^D%N(Zv7Q$EjOV{t(GPKUs{jK{6 z-_|2Pjqt;S|JE&2_jOdI-<+)Yit0=A^FuLGzkGS9)4%heetU_(?MZFwRjMkvXZm*$ ze|S&bKll7Nza%eAeqL~0x4#n4PpU_MYsI7d6M(-w@TU|17V+EmujhzAQ~bd^o%&Y< z{w(4@?9u-w@#g@4CE(8`{wk+*r)~S!$Hbo}e)s+r2K>Jh|4i`*=O;LZ=K9@F{BgZP z=dUvGS8a;tZ@c(y`&XFa(fp;0-`)Qzz<&+#pA)~@p359&{(T7_d0O{REzdHC$v;f^ zuEML2N0m9u^*@*JcL;wJqy2KW6?_6_5OxLqe~=Y6AcL#6Lm&w*BKC;?EPmdwyyI|BJ*wSNyi;FEfe1 z5aeGM__q`PVe#A6@3+JsaewIgtq1(!YJsBux2dc<-L|y#CHKt!U#xi4|3vY-ufHz? z{yT_&xcD;-zd8T?i9cQZS;9N>e>w2aC;rM8>r8F0e`XMW%uro__v@cm0RKsI_&+E9U&L>F|F>34JU=Hq^0yKGf^gk`oBt)k z*Y(J6Bm7k!`8w(iLbQK$6y7$!mna^spYSBT|GC${^Sj2vq@Tg>KSlh5J^CjSe-!X{ z1OCIrKT-U)_s|4oE{UU=L7 zb;otszd(3-`Qg7T%=-OC_?5z|?Y+!l^5xoK|2~iT#}U3*_@-s-q4gg@_y#p~r_|H) zGKbl}mk8foc-#IvOYvy_b7lU5`>eD6oqs>t+<)IC{$b*`oxc|of4=zL=WplVUpD>U z68{15+vfiS;VakDm5wS?{m}j$CGp2U6ng%;2gF}ny+DH2Z)@?};_pZJfx_E9Khaz9 zsQ*bI{`)}u9}xdo@!RI_9pX;~{`-M{H}TK$=-)y78NmMl@L$*t_iu&xZS!AV@u>dU z;&;#ggTUXN_}kRh^BX*W2glglKROYA)CfJlLEc&a&hIKU*Iz2}PZYnc|4$HqviRNm z&j{dOMEvtS`WFy?8t^|1{2PgXt@v%%f1eY7rueg@aZddo0sgw}@%$y%@z(zuibwMo zH!`&Uqk+FG@#l!YWpME1)||g*2!B9$+y0kJ_{z>dOrs9l>u(>E{L>|Wb$M2DIQ1U~ z@;^@e-Nmn7o+xve`5z(vT;P8U_^*z^{U0NK+xrhq6_5H~1pJQ!|3Kp3;4%Mx#2@)^ z==J{-!2bgA7kSKo2Jy!O|C7MKhWI;2=uX?#|4QOd6~BA^PX_)o#D7Zsw)6jA#2+;( z^!nu~;J@v9Jb$(7>HO8plVuKb{`)E(&0j3=JHPANJpUXe{+qA_@4#-eZ;>={844H550c(72R2wY1e%&H|+y32{ z@I}Jgp8s?we8i=?)3*9QM)>Z+cem7UjN(!MGNpd*^T#Yuzs1BqQ~b8`#~Z|-2l|%| z{M(2>NBp;1^8beT3qk&~fxkjWJU?NV>HgXBKTrG-kA$8-oZmHX?jIeAzm529*RSmr zkNO`c{s^g`bN%oN@DCvV0phovzi%b}M38?5@J}ZG#o~{%^nU{Jr;9(xJNeHA{x^yL zsK@@lPW<^G|9QZFocJ%TujkL!|0Bd7mK=Kid>!~3+=%DDgZPJ7>R(UsX#OI_@7{kh zf&X6OpC|rc|AJxYssz6Z1O^d*y!eB>)BlCQ|0?nC7r$-&&msN{kpCj!|C#v1FZb5} zABle>=>KBiFMkv6f3)~*{r`vf3qbx^z~6!RXNy1BzhKAB{;n!oVTdj5mF)BksX ze+2RWD1KZ2hY^2*_}%mW9`G+E{z?tJ^?wQRr-|Ra|1Smpt;F9{{I>mn6Y*z){Ih|- z!p(U8Qp6wZAM5{l;?D#9Uk3awh(A~S)q`TWHDAAfjPT!j>c52e zi$MKzfqy^oXNccc|DTCJX?*DZ@BCX!VF7&v_r-`#c>apTKSLfLypiP;{4!8e@o4_C z#qXZK4Zz=v_@^}Sw*I;kf7D~4>+eh8A4B{LJ^CLe{#fAO2>h=S|9bI1`(LH?9kc#( zh(AI6?)q;A{(R!^d8IZ5*KaTk)Bip3=ZQbaJL`W7@K^1O=Wn+7qk>|&HT_|VNAnkv z659W*z~7Dd*Ld{D5`Q%C=K=qN#J^AcO)UApO!!Jo%a@Sm=Re9BKQEocA1Cp(}(3ApW=~bpG!3yASvy zyW#q0h(F%Q-`qbgQ+!4Bm{=pQ6!E+F|NX%KIPn*X-}d)6#u9(7_}zbh;{fo_BmS@! zI-wXNeVFKx|2FY&5&s<)|3c!Ao)o%&oB;kj;&0MQS1Q5c|C;y{f&V1% zA1D5a;s|d>@4Bc$CJt&=KgoJ;!*$7C4Tq%DF*SsMf{(OUwu5e%whU7 zi9ZYY&jSB<#D84;!^+r0=XVS7=ZZg3cxV4O2mI%VKj9kP5L^E`^uYalNO)WQ+A1FP zFLJV;e|P`>1@R9f{tWTkt{?6r{uuGQ_pfpnhF-sXNc?r9asABsIYs#9!ke%E8OtOP z*%Q~lgGc^W!uJ&3_W6aLibwrR0{yE5>i0PDj}?Dy%lu~%eunVs?fo)`S-&?){HdV- z7lHVHApQ;Fx9wkf#GeiP7XyFeUbz2f#lNCV_M!Vn1I45M6^K7wcxV5p3j7m^f9bWl z(hV&B$A~{U^`a8ez5c5M{}05!Mf~ReWv;&{b%F1E&piL^75*+u{EZcl>X#<*yZ4V8 zApS>)ze8(Xzt$Fi7U4$f=FW4s-o%Bl+im{?!5bSLuW6ze@aO|4e^R z!fzDbcK+$6c+|f<5Pv-o|7hYrCH{tGsu#NcRuVqqI-RL4|1*SdF1)S(jr-#Icksyf zCwvd#pEmM0*IysSqxu(0{oVV|6`=kziGQp3ZLeQEOZ+KMh3-EMf&UBQKPrCP^;<6S zr;9&P>gVjgje-9p@egRD=g0Q`X%X=kiQm0{T?zb=x8nI(EdH)W|IPJ#nc`9ZBd6&4 zyZz08e*p0x62EQz%qILf;aeH;oAdiTi9cTAcb~tSgZNhyfAhAwQgsZ!S--=Cj}^YU zg|8Ej`xoz#??(80g}1H$SjD6MCrSO%C4OiAuL1R&O#J)BZ(IKph`$*0?^@u`A^wW( zbf;|Ve>w4oPu26E=*s^(;4dWphs3Y`xXfYB&#%OvEdFShzb)|BR2vMsep(}b+xoAj zc+~$);BOE7w-En%@!Qt_(}WLiuPbF+|EVPYJc-}E{yTv9-y{B;#BW>w`w2flc-#E{ zOyVyD@!tsIui78?f2{ahmzloM>z{iFzgT$N{`Dr|R|#*setli>X#N7zLeKx5K>o*x zzexNwEcwUXhU-^8#@qaLQ9O!20>s}H#6Oexqs70(692QrA1(fL*Z$ER_&FhaQ#9t(S_x{}*_$Ltm zAn`Xd{O0*%4dI^>UY%ab9A^A0N&JydhhG2P3gWMOJMP~$@!QtlD8la--gf>Tu6We{ zc!@tzvUjdu`h)ma5r5APy3)4y4?ZORY*4=ez<-+f7m7cjO!q?f@8iT@0Q>`izuiFG zzq8_RYxvFixmNM0{^8Gr?%xT(zlr#xZ_xFRGW_QJZy^3C;2#A14OD~C^;3fQCs_RT z6p!+c6@QBG&ioAq{s)MEoA}N3W7dB&;r9!_*%JR3B>o(UzsME;y&(QpYCuu_mUh&g zvc>-l;d6zz)&D8Qqxxq(tLM*M{~;j$oy6baMjgL;d%DbF&fhlTk4y{QKZgQ;!#i>P zQpMk}j6HPyPbU0A;m!J+@$V!23gL%Z{68rk)juBOp9J#1dJxV(;wBxb?fUa}!bf@J zrxCur@V4jAlNFEhPnG=L*Iy$+{_BW;t@yiI>c5)!!>5PN-zeZePW*wJz19B_^#VOw z|KT3_V}y?s-gfIp}e;K5B z6o0zJ?_PhYApYINKUn;>^WXUUaQu@z=6{J=U}*nJ7v6UMi6#63;ce%S8x)V~mj(L& zG^pPM;&0Nqe97e5oPgRruhIjw<+7An+FPXNW&q z9_sj~1OHy)-zxq(rK*y9roYbpxc~cwZ(!jsQ9SD3Mv31&|1&}SV~M|f7agf>{+1EG zhDZJ+;TsBXo8Kame}Uxh-oMg8{@sS+`Y#s0?ff6BcvS!7nR!c;p`>d?Vp)>t~4KQU4M_{tH08jKZt*%_=i~L_W<#y zgZX_E_^)^n_iwBCZP(9(2)|$W?w0s(S3K%pE{K0Ii2o(xZ{A(cukHGI7V#GWe-`i` zA^t@1+peDri9d2y==JkEz~3(k_piY%I)B^#GoSF$!rS(r48^1VWlQ|-{pWoU|2E>k zhs5vfHSVMNaGZa#@Dtn*R`+#PO22uf;!*q?CH@q7sB`|y2Jw$4{zKw#U8*X%XZlwV zzJ8p3- z{!f6v?Fc+SW5mD0l7F<~(fkyO-+lf5De%uC{-!;3{cX>mrW1ckdg%K54ET=_e^>D@ zvE*M!{ORJ46W-asJ_r71M&kM(6@MFx|4YI*>ZR+KW#KQL3rEq_h$(|LwMWyYl`Ag{@Ieh`~0;P{Fe^nHZ@(-UKdj8r2 z^1q$r z|67H(&Ce*p?-kxQKf@J|<|hH<{~O5v1LBXkRnKn?WB)bx@6&{DF1&4ij+6M4LHvav z{-{TB{S!RKKZNiP318R9-#q_JBK#ELuQGVEevd02^)C(Ne;DMynE2lne;>nd#{UNK zXNo^sc<1`{DDZD1{$lYrGyG=$=LsJfuOF#CUR35V`7Z)cjl=cdK>TL^YbhSpKW0v7{m%gZ&BR};zc$(SuMWhYB!2hx^WVTfpZGhA z-&X$&;!hR7yZ-+G{|4fpE&jHa{(nmR8$tg60)P4Oxc?i)Z##c>Bm7R`pSHvwt9Ud& z1rmR{YyFfD3;q28FBAXZ+w}aXw+G4`=KQ7;fAY(EeiMau)^A1N-%b4M#ox7zJ#_!t zLHv2*ckf@7fd9(JaQ`X|(3zUo59a=nM)(NfH(TPLs(94Dz$?0bMXvf?1mZtL{QJaj zo4+P0IR9h9+ve{I#iRI(B!2h&RRQsjApY^Udz-&u#Gm@V(D|zd{8_}mNBp+=dz1JJ z#P6QJaNz%)_#+2;o4+oPyH?z$DR>C*AL+4*t z-~9Jej4uP5Nc@3Ub^O_~V4eM^4v4?1dV|M#i23^$I^L;GL;o}DJElKO@u+^;bG6^y z{|MlZBmPYB+t%+>gwOWKe@ggIg}1H$H6;H6$=|*H>x29&KY{Dtb&#&K?fLTsibwTN zeofauLi*?QzX9+MApW7^x1IlPCH^$=C%XKR!2dGwZx(+);e$8k{G}6rq4?eN*9iE( zBmVk#X;Y7&SZ+=K7UGYX7drnh~Mr*LvhHe-hVkoA9kI{s_gR{1ZX{t_JzvOZ*iF>rUDHgNQ#J_*(-1 zLgH`d(LbN~bAkUF;NM65N#Zx>*SvmApM?82!6W|%;b(f}+fK&**FEyn2%qhdFChGS zkNi!kIR5P(`82{G@W_8o_+pQIji+$@mG9Q;$M*YAt12F?-y)gcNEtZi{MP}@Z}lnI z-~1l!Ut2L5z) zf#-y1UjJ?%qWwE9{^t~r{3#2yKi}o=3jFP!!T$H|*Z#E@|Fw!o{`fbwKiB2&4*VO5 zKWV7;FR}PPC;sq7+Mng}#{qxqXL0^D9?;k*k0Ac{#NX25A4dG~;&)%a_6PnCh<}&(8yfqs@nzs0;!hI4 zd;c8({JW>)`k#GJcY2VKzvuJ z{W~Cj+x%2lJnCQI-O%$#3h?(M{`wDlo1fmqA1!|O{xt#kpC|tD;Zuc#`(_`e|sZ;v;OrIkNicT|I>hf0P%k(e%tIi$_Wwr`ecpDq4q zSN<;o|JTGnMEthvk1vV80ObD?@LxCw&)-_{UvKo^tbcjMqxmZa{+EHj6Y-xEzwP?7 zBk_m7AG&}15BR4Mf1}ZO{>=41nfRlC|5f0BoA`T+-?si15`V1t<6QIi8u0%}{HsX* zX8(U6{xp#PeBeJx`~~8-t-m7T&lJCV{Vf3g)-U7vuQNtxdZ%UnTPhySe;&yH4d5R| z{K?|Ct-s;KUkLnf0)H0qFA=|O{k=*2fu*7A?=9f}gZOud-`xMr^>={yBY^*H;J@e< z-2cF%y8qQJ{qIZo>cZQuA9^Sr%}+Fl{~ZwjBI55Pep~$(5Pux-zX$yJ#6Lp(HH`Y1 z{cHI@T>mG9ztO@+DIV1?QQ~*6{|`X?>BPTA{5eLz=Kl8_@u!PFNqFb_=R@GXBm?I^ zW}G(JUVp8kc$9zf2YUS_O8p%Fa^UYp{Q2Uy&3`K43qA5{2wyCG7o-1X{a2Fw6S8&w z?*4xS@~`tMuKx|=waK>sU7~nY|1|MOxaz+e_@@$olK9(M@}ETfnIQjH_%{)Mr^j`sw)$0l9oO$R z;U72RH}{V)#iRPgOZZ&lZ22%l|#_A0_^F6LkG->;HG+ zFBZRh{r>>`7cIc`PZYoH{994+sQ%&0_58T~JAuD9@y`;!?fiQS@kfc@ef_Zu`1cb3 z2Jv5S^xs^+yNEvlUyW-`qdU{sk0|{L#RF0QjFE{*mIh?Vsa`KTiB{uKpJS|5D=L zLGm~Izl8YHLH>t;e-rTsp42AW{Y^B~m;#~84 z0>ppmV%)!ulXRr!`ZeRPt$5VGc=5Z>zbApeC-DyxziocI5r2~S-Sc}I_~#SyK>lZd|4-sSC4Sra{}AzK1OMN^-|21K|7#}ePTST`N5!N5Zv_5-fPXIW4;H^U zf9Cn)W#TUY{(pi0JK}#v{I>bqLj1+zch6rrwIo#tB}Z^w)X&28Un_oF|LZCq)jw*L zUjIRlbNx~Q`0peBQ{uN>|Gh%^Dye#YY}bFqguhgH+x6e0OK|;q2>-OyJ9uNx&nU&C z`o&89+}D3$pne|{|5owauK!jMe;(*xW#Ipp_%C=$*WY&icb51Ifximy_kRc1zk~Q~ z>$i{MQT+oSDVaLl^IHx0Cldcq@!OvNpCJ5r;cc&<93}BbNc`^ge+h`c>$|vq+r-~3 zIJk0ao_}vvJgQ$D@Ye+X4C1dlMVne1eslf2MEohBf3<;s2l4mt=>LxR(}BM(@K=8i z*MFk;ZTnvg;b#kP+y7<}K2!Jxmiqrg_;teD_Mc+Kqxs1K^{Wr+cisEAeka9m+kaXq z9{F>DzX9-%CjRDA_59kdzt<7|X5nr1UrplA2k|!q@t-69yFL0(6Mqr#HwOMzOL6}u z62H0s{7U$_!Z$SL*W~vq9`!#gM+K`6_x|4$#Gn2Fj(@HA>stI@6Mnn!w)KB;HufJ9 z-nM^MQas8(66D_;rc~_Zf@jn?%!RBKL+?)0RKqh?=F5@{tps=Jn**y{#@cu z@tFU|#GeHG(ZGM2_!oN2|2Xld0)K1Z@BAU2zwI9LzftjM{xX2S4e*a3{`2CmZp@## z{udIy>eKp>w)vS);?D;0w+HbbBL17jZ(BdV5&uTuzaIE&FT?$NNc=-B^$S-#>R*BQ zqlI^#|J?xmU5S6I_-h(|bN_gO@cV?f?f<(7U-22;f7|uTc9MViYL%=y+~<#*LH?DO zku^f42hvv&27C{CAh8T5`|Ke+u#EgZ%pe|1RSHO#CekzuEuo#9svbw*i0aRk;60 z#UEq%&HP&`9@RhWVDQ{$H+mRR4{j{&xZY1H@l{hOYl`i$9V0i^cEWf9?kUw}}54 z@!R&FOyZASt7PhM_y1nt-$MM^;%{fqgw|GmWD^EsWVZT}fW{F%W20Puf6{3+s(vE=^_@#ld0KM4FMi9cKXw)ra}{z8!d zaNuvg8u$N4@!R&FD;1CWANWMc)R7|fbI$)GfPWhC-#JtF-?sluCjJ!hyXXI5;9p1l zi^bnm;t$@K>u)vjX952s!2cKVZxO$3{{JNYeDO!P>OUIzJFUU}KP&#CVB&IX=HF5A zsQ*Rc5Asg`9|iu&#DC~{ZEA1vPayt+bxNiVcmCsm|5M^`I7|EQH2miIZw>Jmi{Cwe zj{*M$ALIH@5dUzCKcIM2|EN!O{_gYN;DAsk0t(UFX;N) z&VP>(f13E+^?wrh*AstN@eeZUZ_fWG#GeE5pA7uf*W&)aF8=E+{>q9+{m%pbr+|Mj z@$VFWTf=Y8{~g3%2>erleE_>+nMOYt`}){nV=)cO?1UnsnJ{x|vRibwG$g7{wq@h1^~ofoyq=D(l# zQ-J>^;D49+yL$95CjNBbe;N2s5dRpD{-ea71^oX5{8 z3f6TI@h=m1!w)*!}Jn|QS{GDIvV)_>m{|xb)*Iyz1C;npK zcm7Jf=|4;SSswi-i9dY3lBpv&=T7~dzY=Wvdw+@R|E2hC>-QDH=L>Is{?)AC_k_Q2 zj(+5DnV;ZIx!@hS-Ku!h|0u~nP9Epv@BEbrGyh6o;rx@uZ=1hzibwurP=Dtmsiwan z@y`{%?fGkc;!hL5`}wQ$mb&S`hxqfvZ(IL&5`PZJ-+4>N^v@;!<}d61o7azK|6eBl zJm7a;(lq_~#D9nQZSx=4i05ax@aFl;jQ?*Ee<6t9c}i@?KbZJuh~N7BTk&Xq1D`9I zI^6q*v!$8-*~I^;_&dw`3*MOdKTrG-;&;!l^DWbVjQGpF;%$Bo6Mr1Y{|n%c+l2ey zO8mC^cTqg5eKUDm;2MaH^x+arXe}m09|J5GzucvsFe-`k6 z4g4dC|ET!87z1bK{~+<_i$7I(=lt;v@c&HwtuwT#d1pA7up0sr`~asQtef6Y>L$vxBmF5xph^1l&&h42jxzghnR z#iRL41Nr{|^1pfu&VPsaZP!0d6_5Ox;&<R@n7(&&UCnuzghp0#GfbrIN_c7 z-v#`Oh<}v$TUq=Ih(F>>B~yoc{&xfaPsBe@{I>l!pZKGJ|0m$D`3>&>Ht{z#@;B$N zn&MIa7Pyf zS-^h?_%{=OrPp+)&HkJIir?b?)f3(}zyFf>b3yz^K>RVpA4B3d=XV(4dkb%y-$jI< zB)n~Y7bqUhZ$8NX7|8!e;$KPPH~aSk@fQLA3E+>&!~H)fe%t)kR6NQ*%(>WBhx`0< z68L)&f311m=J#pBM+t8`f25N5BSHMdApU&f?<@X}mi7NV@yCciT6pLB;Vke+eTVxu zL;SrBzqx)ID<1VPN&N2Xmvg|MM*KU(Uu5x5CH`#j2YDy|^T5BI_`jU5Gwo{l&Gqvw z@n>#SGIhA?AE+Gq{CVIuT>r|iYrk#(=%;v8|2*-#*H3xiUqt+!#Bclj%mU&s1o>A4 z{$k>vEB@w|`kx^F2n%;wPf3kUuVKj8X5B>qlD{$~9%3IDF}jSb$c-)o9T{fh_j z*97tZOZ;2KU&G>$+kx{h5dLus-$n5#{v?S%QFy2SbwK=o6aSrW=t|qJUr!PLMo_;9 z;J;=k&Ob%`HI4ku{tqGibHcYYc(ea^D<0La0K|V8h<`5e=ZL?#;Wy)dnfQx=|8n53 zn2+nXPyDSd{{Ioa)uM*JU$zm>!vyfOXn6Mr(uzX|Z)^drvyC-JupisjbyU$1zSe;V*N1^!LM zU*k>Pf7|u<2I9{Yzx(_j1^fed1N>L-!TsCsv42e!kNOt@{MQ2iqr_i+ zk*>6@e=7-JM|gApGy8YZPdNWL;cfk|sCblrG|0a#$p3cYpG@L6_m6nuj|2Ypz`um} zKNP=h{uU8`BJf`i{6)mSSNvxG&HQiLi~Dz6_~w@Nd%faO|58Bw9YOr}5r6%+^rYG5 z?|s7eCA^t`7KuL{#D6o0{|xb`ir;qr{)_mt#P2?TcLM$nKjZ%WE`HnjyRG6;|B68V zU4VZw@z-6f`)50UPays<=gnSqxUZkO0slJU?=61Y{Qpb%dxfuI%)dE5aRs>kqlJIm z!go!7~-EO{+liR zf0+35#P5FobT{zl68~C{`F~9O0q4U$>TsWb?gjp^-*El+dCb3p;wvgotRA-#XNgz* z?)>iq{yxOt=pAqK8%O+cApiS;e>U;QiNE828RqBu>iMr@see4-y9j@+h3}-F0r8I&e*=quF7X!u{}|vu zNc^+KZ_bagO$AyN;`%QZ-uC?eD#fGvh2@3rzvDps6NvvO@!Ot%k0bs_;C~GGKO_Fi z@99dJ>&L8Lo!@c&>I>i8=)cKdqIgul7!dyi5dT2p?=JpM7XKQ;KP$ZL`g0|TKOV&Y zB#6KMAzZ(A#BWGpE)G{R1p6Z5dR+Hulc^7 zG~4>wN&FeWKMnXV{R8(eR{Xa0GmP;4g>P<|-}@Ag`kxKre-^|)pZF(=-?n~=3IB=k zw)Jy@#J>^5KLfmfIq@e0|I5JtH}O~bK+lhD|BX3{`xhyE>tNyJ);#|{OZfi6+xG7% zibwrV1NpxS^507Q(@Feh|2GkTCh)%o{FncU>-VwvZP))1ibwh90RMd8A42>^;y3$m z=D(Qm6|(XEWAfVwU)v-9FX5YcSirFc~TJgI-W@Xr4ICa8b4W7xk? z{O0q2bNyVTc;rvqruSd>>mP3c{}AHeD}LMlb2ss)i{HKfzYY9*i2oe%oB8h~{w&~M z0{rcd@M#|6JgI7x+gIe~kFe^=Ib)I^lZ>-_n?0bN#(a;?D>1F9q>` zNBk+`uVMJj_%A(y>pw&I`wiZVzqaC0{fj{SAA(5k{~YoEvs~BT_WDE&A|UN@h=qrord4c|3~6462JTUZ42;67vuh& z^Vq)!314fKu9U5RLllqt7r7(!`Tw^d{!HTUEPh-6UL*b(;QtQzcN2e-_y=10w}bfO z#qaLl_rTxc4DR1L@tfCA=KKsNe4g-ESmGb5c+|gCP`@1@{sqKeFCX}S zCH^`ed7Iz8#J>^b|0D3XK8yR;Rs8KN^>3+oRR03t-vj(_5&s17o3DSG`(GyU7X$xZ z;4dcr*TtV=46NxtLHyy)FN;t|w(!pL=K|m#_cyM8oz>c8yZ*WI9Oj#QFZ`(hk&g1-RtkIQ<2nH^uJHZ(yn4 zXT+Zh{KtU5zxwp2(-ZUapFfJKh~!Cmi%`Se}VY3 zgm>zH68N7ihx7kR{I>OTn((`Xzrqs#amAzhrTrLs{a6g*f1o^$zgYai`3r_&_HQ2H z!`JFZ+OB{9r+5^9CW!xU5dROvf0Ov{2nH^<=KhgK{5j%x?;rmFf71%Mey@n%R=?4N zU+S@bBNdP8R|x7CxH$Ch|F0wdJ>s|RU#p2f;QX>Cbp-Qw)=zohuU-+?uly%^eiEc_ z!5g!Gl@*Wjj}X7Re-(j$Eb$K#zit0tPWVy6+v>NJ#2*LhcOi)X9P!T)zit0NP5g<# ze-ZF^x&ZfYnfPt{e@Dfm{-ucDz5ibf{Lc`7`E`1JY_I=3Mf|xS|Ej?MEAh7xzwPy} zy~Ljn@~;m3^(*1}KjAU|x{62jkK7ZwelG$35ybzV`2QDNxN>WrKZX&14DiApT_HA1nTDmimn#{v`3c*G~lSzd-z}J?1}y z__IO&mjeGv;?MWkzx{+iht^W%ZkLITU^zRDb zk0t*8;y35tTz^j!eyH%l^LKCzO+JtC6NR^3|9wsJk8*z5qB??p=lY>B$p7+-aQ)vA zfA3)6a%)~cL?|BBKUVzi>xV0We+u!RBK0@@+Xx@_nXYs&|6m$s{cBam`9}(GtABOH zqx_RW{jUP~ClG&M@!RTu8}X-s`dmRrn*Y7>yZP!nKllXH${aS(e zI}!h8@%IQep3HAY#iRb^f%-)Q{{rH#wO&t}ZU34}{DHlC{kqRTt%1LQ`0o<`gO>bv z6Mvle6NPv7k2b*HpbGB)dhy%V-%!H;D700^LvTn zQU8lT{x^gCj}!k}9^-#29LJv{yt)3&^G~MYQT$=fFDq6@y5#T7Z)XtyeV1T=i~e*y6K1pe+dasA&Bf3Sb7|4$Nrz3{fz|56l>`d2LRyRZNHfcU>B{(4{N zNikplFt0zh5`R=d==t+j;BQt7*KfG^>jVoYx2As};U^1U&BDJ+_~(Vc)WX-QjpKh? zc-#Ixitww2A7$|mS3K%}veZ9CcxV5*9n^mt@kf2B`)`}yi|XL~V}-ZPPesL}_zS`O zB!Kv%i9cQZw(GyEi9h<6(D@w%{KJWVr}+C?`ZtvLQ-FUk@V`X-O~2Cp>u&hX{eKqm z=K}vdz`up~M~dH;|5wBx=KQjKb-1s;6M_E}@vjm8a3g_&e3b^WS8n zuC#6bIw~H`e+KXm1^#s6A0dA8`p3-wIpQx6f1=dS**}H>|1ZS9P5idcpX?$2*k5)3 zgZq!;PXhja5xD-R#BaO)>8*HF{{-NF2>4eLf9*}W|D7!JmreZ1;&-3FM*{yX^>F?J z#ecKmH`jk>#iRT)LH?tFKbQDl5q~3#|6}6M0sds*zu{7xe~$Q@Tm0=5kMhq0{xQHm zg!uP~-!^}D6MrG_j|KiWh`;=1-D#VD9`OhEhpxZzz<=Roxc-gAuRh*d<}l~4yy8*) zBg7vm8^QEG}gb#f2MfU|48Q_yHbaH{mlXK*KCO6Unc(cmiVhF9{H2R zpCP=n|GfhIHxU2t-|0-v_y3J>0_})Da7gFxzW&Vs{t3k2YMb_(^J~_B9Pvks-(CN? zz`vUKCyBpfpCi2O`fVqPKSkmXu79V0^FjQli9cWbV=VC>C;oKtCkpTQ7XW{q zMtFXEe(!C5E>S$1-y$$SZvg)-#GfwyVEuw+Fz4qP!e)%r({>VS{{0IF`{T6}v z7ZU$h;=e8!xZIlSZ$9zIh~K^b76X4C@t+mHdHrho|0R5t?Yh$AE%Bcv@h3_A?)AF_ z#NV|sp1&#Lw_QKntavnk*}(rU@XsXvP2#s*KRrwQ#lZhQ@UJKSxbTfdWuKTiDa{cA1oFD3rx#XrjEzq$V|A^vpnr?}>C9q`w^ z3fKQ{@i(;iYbYMoKkR5||8s%=5#q1EQ_r7m|9XY+t%PrDiT_0seO-s`=jLZ#Utu5Z<xeMg~D)G1bQD<5`n7G`U`Cr%y z=il8UKY{SK3*XA{o9kzs;!*ywe}eSr z$$x_4QU1xm{|oR}ipKsT@!RHiEaA)V){|oU`*+I-Uq^V``F_}$MR4g!BN@lO)Jx&N8#Zv^q@f&705{`ZN0p7?uN@~_t#*YBY4 zw)x2-{27n@X2M^*M_0;x{leTo+Fgg^?=QUV{B<|slZ3bB{}SOR2yd%@?KU|6FFeK{ zPx##)Hgd5f0*zO3vY}6Ac;S*NS}Y) z@t+3qH*Smjw@3W8*H0TL9@Q^j{K4xN=k=2_z~7(v{}%sUmi6C<_>07!D7@qU8~9%$ z{(*bF^=|{=lZ5YOiN8}jT>s_5+t&YMgx};b{*8p+?U5g+-r#cfH1qwtBOdwoF_`~X z_(qocwN^ZuzxZRJ*IyN?g#P}5{lwq!XK(#)eLar9jqtYDA6qIO#h)bcyI+5-1mfRH z{PV=$(x{)#s`Q(ih`$)rFAVs@JK+2`i{G|?T&#GMfB5mx{#6G4yNLgg_|4CcoB7Wn ze8mE9{ohUaI>MXRPv-f52gyGQyZ(5M@DoY?X8nF6{4C*ZzdyJ@ z@u>f?ApaU5|Lz@e{>MD}J1HLd6M(-K@Gm6(0l#>gpUs3HCcN$Z_XUYRS>jKWrS6>n z>Vo(i+=%P9Ui{|Ye=+-CPw}XJdE$3J|E~x9uMmH|ectL<^(GvD3*l{le=$t)DE^2O zq5DsL5dSRV?=ODa^T+AL9}WBsfd4zoOy zKhu9V;Vb{@t^YF#U*9ADDdAfQZ+rh?jpEV##A`0p3L?fUQ9SX}?dJ?4KW z;b#hOzW!zQZ-C-a{`nyPt3dt>iGPj9_zx0(i}1GBPxg`cizI&c>nANh{2e>t`d_qP zSIS)f=J~(9;!*t~{|eoIS^@uL;=fM(w&$Pg2;a>ke^+Om{~aFr3%g)`xbTgP{mZOh zdBvmp#en)<2kN(k_-BgWcKyDH_~XUzzJ6~D{LQ=K{QnZa?e*{RgsCYuK@I~1E}9p;$P-5|C5B@Abf8lel!0f5`VG8 z@4o)I5yXE>cU-^l1KMQ!`&*qAkLnkDGIamA8Tg+i{_DhV+dsAtzMDt>wp(!icX;HF z5`MVwjg0=8^Y^>rQT-A?{knqsO^Czs&lJCH{}`uufc@zq@~Zfd44* zFZWo#&b@H`zVOJuMEIQ^`7?z7U3gpn{!%>ZUl!b2*7kZoD48k`M-q!z@ z6p!N11^vGr#Gg<6&xrqOWB<}wm45Sk;*UQaI{$Y7f9*cFejj?Q-vGjY?va0)@IQFu z&l3Ki@V5Cosd&`CBv8M>pnkD^as4X)t~+I$pBog9{Hfx1?;rO7|77AHBYxZbOd$U7 z;?Vg?1par4f4=x_=a1cl|G*=E*{!(#pLygb6aIVQZSylh@u+`Mpnea4`fVWoQ{uPH z&!@y6D}MLd<8l4&Ii%;;Hb4Ime;&wxIPkY3{^{bk&CflAf6XI5pYZQ{8_T3Ggk4)A2_4ezx(-1GVo6&{ziXz zo1aOoLIp0rB?{zioc@6aHS|ZSUXwOyW-j@s9)XH|~$;=Sh$8k0yM& z@V58=Mk*flKLx~}0^^4g605e}{p%{u$!8^}ns+k-q@ce=6`#BmO+`2lt;~8O;7q zCjQvJL-)_8f&W+Wmp2Q0M1P}iP%O8R#@C(i!1aq4-WLCjibwTJ1M#PU_}eFFzZrk3 z`1c#}n(?=|6UU!)ROfGtznS7u{COb$=Ro|U#UB#C_}f_GUnP7cqkk)e?_}`C!vY_Y z_#@7R&d)3mf1N?Pey0DF_-*U=62+tXC5qp@f20Hdc;avRr_QvCCI8XHpAPb$4g3Yf z|B(1!viNrsf4=y$gm?DuIl$lZE?ocf;vZ)5cT+s7f9gM>=Z{x_{|n+DTBPf5d;K_< z_zT1zTz^jf8Nk18FwQ?s{O0wOIsdB_kMb`D{<*+^_1)ONSp2s4-hhq>VfKHg;!*u$ zCI2V^ocUV_^1nC{=Rfth25tG5yASg(3vYgY#7HM_PWe&%NrB2GdAs+|w?O>=692~@ z{bv=A{29RiHt_#51n2*YM}M*6k$)rbF9H6)@5lZNPw4*H>VHb{$X^Wn?*jjsq1fNj zqyI0(BY#x6(Eh&<{HGtl{@XnIk1HPe6M+8%;6FJG`ycn{FH$`6rvd+mz<=UF>|fx~ ze^l|vp9B2Mf&W+%_J85ge^~LzUkLmwf&b6p*#Cz|{~^UAe?+imP_yqWWei-|Qi{EzsdTA8qpYRy}EXAYvvqAj1ApV(;VE?Ng<3B?9cRj{m zsCX3rMiBq!ApQf%IQ~r@<8M0}^SeF9AFX&4e*uX9OAvpnG1z~`WBgMIfAJ}A^FK-P zDE?v)|0WRs#7A-bSBc*?|9c64gU9%HDIUciUNLn2ZUOP{9E;<>(_{QMjKlm0kMXxt zJc>UG#Q!abzs-2;pYAdKsf2&cWBij8kK&I7@oxk1Pkao=zuIH`YYG3g$M`=|Jc>U7 z#Qy_`e@zOG|96l6m5N9HWZ>Tk{5g+f|AnW$t^ehUNB%V6-v#{3CSd=S;>g~Cb!?}Lg1L8kM{DX_V)$g$4k-zAI(EVpG@c%v$=Rd)t zf4}09Kde%yzX15xKZ*Sr9{ryv9{D4Ie;@EaHwpV!dGtS{c;t@({{6t;bu#vE_vpV_ z@yH(!{0D&lAn_mf=-)^DNx)wS{Ew&N{KL<9oBy$jNBO4${~_Q%K>V#d`hOw*4B$Tu z{9T^H`S3iykGzx-5O|G-&Y z>Dr~GF1csczb)bGc;tr@zJ*7Ak?`jG|8c^H8}Xa(|1MBGnx9yypZoRSe?a}d5x@EV zW1{$juO9@%FylEz;!p7y|1lDOGKfE*u7QI22jgltP50l7KU4hX`;U>qMC5k7;;X6q z6^!ra2_N(Y|2X({@OHG~OX?qdfZS$+`K(}7zTH%d{1*-gQ1Z42R$8S;mh3X6Q(c!{J1|KN5!MX>(y;8^K5FTa*7;@u+`!Qoo`}nsctdD}(wubuirykuXnBRKZeRsj7-a|NgV{E}uDj q4jc)Awl5X{ literal 0 HcmV?d00001 diff --git a/infini-qwen3-moe/build/.objs/infinicore_infer/linux/x86_64/debug/src/models/qw/qwen3_kv_cache.cpp.o b/infini-qwen3-moe/build/.objs/infinicore_infer/linux/x86_64/debug/src/models/qw/qwen3_kv_cache.cpp.o new file mode 100755 index 0000000000000000000000000000000000000000..0364d8a39935feaff87835dcec6224af0304e175 GIT binary patch literal 144584 zcmeFa3!GL})jxiQ2Lz2|R8+jABckCAXNJ3y4l)nS$iO~TJoc)P^!pzuE= zekX-%CH^@IKQHmSD7;(ZUy$$~3co1vbrOC_!vCc3%M!m=!uu%vip1}ia6N?&Nc=$w zze?dl68{>74@>;(6mF3CHz@q3#2=w>qr@Mj@LLl9Hierc{v8UxEAj71_!xzcOZ*86 zzc29~3ZIntQxrZe@y!(eK;q9xxP`)JCH@?R&rAG=623s;k0kzM3V$N;trGr}!kxd;RhvtD20P0 zUQFR(5O5#UTI7H&dNO-J-$5D8^#6KkA2^5|v@e&D7l5i-6Crf-7 zg{MgTR0%&U;cyC1llbWro+0rwDLhN!BPcvu;-wUpNxWRbkra-S_-F~oP&iiN<0Kp} z;W-qZEAa^wenjHuQFy+@DVz*C0;|} z#S)(>VXcIhPFoDijt%p*Q?XI zc2lLRRj>)b5J8UB46|bgddpxvhMh|Mv#n_#;)p(VV;PJR_~Ax)p7ZVbyA%P*eO*{e0pA zuAVNn>nT%H(^Q@{L~*vHS5Vxm+LFFS!Zj*foemUXOM0E8t(S0v3ZbXAtSZ)0zk_31 z(A^aSs>VMxXeImC)O8h<^n~Ko)$==~yOa`Zm~ZuJrm0$ehu_Phijt_As7VYhn)U&{XS1PF#y|IgjR`)73&{oeT!3O*qtrLRO*vvNZiiuZFyb8=E8OxP#&aLWNzP_rf zb6wZ$yQ{j>6xA|ei$lLd298V*V}WVw+*;L@Cg~}V>{ogR$Gg2Cmt9-Z8>w)wJH4@L ze0uAkn}+ly&1MF+DFg;R$OeU&(VC$+!?sDfe8 zMXlZhvePApj0MdR#QN;Muzp8PEnP3sY@)NK>1gHhp58m7=YK(G+*^2OI3-RBIwY*1 zRY|12uk2RIZ(g~4bMKPh{J%i*4;L=^mFZ1IZ~Wt3au?DYi{5zoUu>Vr>a5&6mrPY) zsMt)Av$XtKGE+T11v*`6vM{`(MnmTnvM{>_-E=+8){0J9m{(wwC=2s!*uv~wxqN5u zqTTtwK(yNn7wwkx4keP=Y>&y1D+|*xmwxng&nAh44ap7=ksj?#>Z>pcP>lc^`O5T; zqTOp&re7&~Bc~?st9q(dy_`EzyUkCq>#cvRp*|jWMbN|4!Npz6-|gz$yK?!vMV;k0 zEZ@r>Y3FHGt3Sfc@Mp2rM{-=-_wX#R<~g$rc!F7dhwX0ijG&aR<$Fnm_E5|A#L+cZ zJByG%wT5h-3ZepSz)J9=*H$sHk;2|951;W3DWp(;~& z9*5cTU~fe?R?DqQji9|%-5zHkWwy z?b*6q_Tt6amXT90nN(N5Fzc1nXv5PucErvFJYdF3VJwSDxlqNXyAupGCda z$3?L(^=W8tZ3_u}QMNwU+J4SCnduqoji#p7`b=X>N47mjRj;pWX=%-s_}x^}k!v{T z9G|?TsjhuNw!I{`u&$+~xo(M)!0^+Qhs|0>bUL z-n5yfv2H}jW|WP_flOC~|d!=JEdfMWfFq<$JhKzgGr$x1?S! z`q#nUrw;0O+ri$~59;^M0Pp7h{hm75Ti?Img9m%N`k(mH!QOuj;7xC0#J__(`d#OJ zwuo}V1p0N?DMiPgdpU3T_Io8&bYqGrHuXK}SAD(ReJRRHtuEpkFuC7!FYD(pI8EXF zezcRkv8W&I$-Y^1vdVeDIsGPjms6TmU!V^o`eDd}`H@tJ*J3nP#_hZzw`d!=2P*UK zssVmTQ~r;2x4DYNe5kxtJ-I=K*~7QXNXGW|%gW5DU)?p_Vl}g!_Ec5Z)~YA> zrnotqzWShl(PE5nX`kP`WNnigm^_lmNRR678Y9?nEq^-a#AA0)s$b5v1m^~2E9E;S zU`f}m_x7r;J<4f(iW0`{c0Npem`{&OR6Y4xpQDE!7MFtlyVQ4|F$hv9<>>ldJw2F{25s&IZV{k&{hqQuvf=k$1Rn$Xz zQT61GR8`Tl*v1gG)}g+=>QJx>QS-2tYxe;vuat77ZOY)D#$FnAt7)G|kCg7&+x_1- z0!{K7mQQr}?5qD^(R5`ml?^Re(~Ww8c9j^}bn#>%}# zgKpjJKjR$=)T`)RzfEovfHIc}+7 zKa_NeCYvo4>_2+6dNl1PiO^3RZDDz=T)xN4N!?gH({ZGP>bAeKjn*XkTesb4IS(mo zFG)?vn}C+_+%kj&L~YI`D;+NY%hLwE*pgxr}$M zX$tG+Kibj`DowK+Rp%<2#%L-tRZ?sB5Eb2Y_^m(r@QCZ`qv%P&x~WH=e2t6=vrvNa zDIh0*%422fl%l2CDw_)i*fvIt4*ma4zqJ0SfxyGS9}{6G9mfU9l=c}k66jA?=MGAu zj#JJNh5nQ>fk0#X{Mm`p)C@xkW@ZVWRn0Cu$SH|tm(nnASSh=>8;K4*7ax({jHzpQ z2%~~`k*3fPW*G!k3R^g4Ca))K^*N8#)ZW$Pq=Wt{pM5$TdH+_H_(;LL#8O*G8Lp=t zopm${)?&uK6GwMu)J_{eF5?BGfNnQ*r`Pj}K;4(-C<#s7n5MWHEa?qZt7p?+nsPV8 zkjS?B+;qz$1j|rU_l?#DxjaKdi3TQ4&@3q$Q>zE^Xrltj6%G1O_B;KQz?9%^JGy1c zql~vTyM7gvL?zSS<&eCbJg`Z{Xv{`9l|W{bBxV*@oiUp-{%O^d|4dbV7uF0vrT(Ra zpsGF(8HHk1Cg)5VMOjFi9`C`StVkM)s<6IC!;XHtOa2rCQd7yh(w4Wl<6S8iTtv)fAB{ay+*GqkdU`O;wz zwDnN4k9Pe`tEd5cX}V{z?x|Y&)BsSr&>UGnE`izAmWOIuV&)(AQF@z`fw9;ya?O+= zOr=9L^JiP&3>V2WO>c%HNm;5YG@t|f$`uX&GMm6>kZ@^v9Hg>U8BPqO9i8!IA&unK z=|4ah_&;Tsj!$nNbSK|oCkh|B@#WmW(9S<=WA}8{{g727w_rG_7y3#;lU1J_={Jic z9T=V6{x)4m@xrq0;sIclFEIVT(LesEh?9%gA!y>{GT@4@XW%iKw7H% z9$FCxWE2AnxW5yOZ|u!18GfQfd5#55Fi*bImqtPl*331Z(0r|06JZ@^scCvu!lezZbH;On|+bN>am)bFTxfktiAptd|*p_Kb`Ji zN`}>4KOp~nYgN~8zCVO!J~ivmQx4O*{yeSg59-(M53$W;x~|`7Y+BbYoq@6${(~@q z)^k1dZiZgz@trtB!S=0+s!R2tThLvzH=$sBugO%=-@OWcp7u0*itKjKb6vObdUPQD z+GHxk_h+@Z)T18oll>iXD^~MtY^Diz!HNh_K2tzEKE0#;aLPG^EIJgC-K0FsWi_Q$ z;hEZQ2goY!i^n;HjS9(*%4Ww|D+haBJA$8H70YV7edpEpsZEbihZHOQ;eJbG>shh= zx^(f{Exe7XlzjSDGXFH=M60`;b$LRex*V#Zb+R{;yNm<1K8}Jh)2`}{`sWW!5UWSJ z!M+8Emdy`Q6o&^VqeP@UOKw|eUo|w>`2thQ!@2XYCI3V7`@d%I<*Z=T?1Kcr@9Sls zBbc5Zr4?l_{Q25Vy|;mY*hA-iA3Qdhpo z3b#GlVViiTimpKs0Fk_<1yNnfA&KJS{zisLt0aJ+ZI?qFyD{a>q& z67k1O+`Ic9vg&>F?c@D_-G<9Jk1l|;*Q8uL@-biQTlJ5Z_?`m~`_%~qAIy9>TNo5Q zr)SwqL4&iMUAL>ArP0%)M+@(VtxmrfvM`wXq@PN6JjoEqUZ%o;1XRxo6Vu2|R%PXRfJw+9(?Ps8i9=iEld$mxrB= ze2jUtSZ%OyMz(4zk)}EFNt}0s!PbOkhysjnNodwl3r@Z0B!33#2z3jU5zJH3T?6X; zJ^L`nNy{y5g4%>9wqn%?NQT{Zl6Sa`Kt@k#?iq9^d%|)DmuFqx>6hLPmlDZ_C+XOR zDM+c%b{NgzIcEs9#GiAzuj1KwhpvixFw$3lo?onYA+TlAf_QzZ%>B+lqxxEettPAk zt3aluW(??A(J5~w3$!5G$5gwOuo8*iPbBH>LOqD}hi#50nX@kr8S{Q?$N|Df1E_j2 zyAYlI9ex-46Evvz-TLOsw_-rzjE%Pv9WMoKLOCk4X|{j$J7CUbtQV7wZ{C0l?QQ$r z2Lqj7-4pjjtrKN!MoM5TzdxhE(2U%b*+RrLuRb}g>or>Y`|jb&%_Ztt@?Z>bON{d! z3_zP!x{LC6iMnWJAouepx~P!%8MW4CUx8}rZ%u0**1zLh)kM0dN;dt#q_}$BUs1f@ z|A6&+R438W*=;!Op#4vGEsP+4LQ>X^G)~nPOiFq$+NDtcSg9>d30qJL7q5Z&{uFr* zyf~&`fg{2F%l@C*isO`G(0}6(s5^MV%d2T&I`bB>N9&!@}c@SNcs2Hk&KqWJ)U9 z8X2F&550fj+Y{C}8YWWfi9L|#^Km@Dh0bAOH z47W)}x!U~b*Qo-CK@%}dsolg z(NR-`_z`ixet?oG3y!oPbo;5WUgaBN)DMAaPO#>S%Zm}2TtYfZT`N<+?*i~|Uw8ed z!pO1nAO54-=rqz$#iM!Yx7tu=;}N~4Oqz5~$?)0pI$Ls`CF4ho8c}*yS*Q9{_VM!4 z5v8L>l%M9~ij!Y~;ayXDJ;0kXV>W+orH8Q;{qZ%}54@u5W_m?Sii?gL*na?jQ|NfY z`H!!o_`md_#W(et^udFs69N6?vv6N6Shp!RsA&oB?154y`(?ZOZqF4?yBFrz3}t5ZuoN5kG~s} zFa1hMr`J>h`C$L|!hd%!{Ezj*&&3v9S2Im%cK^Hh=W$xUhtrQ?jL^mJ`tX=AeK>7( z>ejw1`=yTUr#hpAl7`EGd4LIZUro6+wX$#ObA5@5`*WJm*yPQ5G)X#tugT<*UYSa5 zD>6E!Cd5usmLn^ph!W+5zM3O-Vl- z)2oAa4m^kxJa07rr|Z}hq@OA2*JFC0MT$V_zoegq=?wvYlcc-ke5<6Nfc!8YJk(46 zFZ7cCo049K`S+P)>OZhQC%E-~mg+qEZWYnk`jKT&BYC)F+$QO+vAIUlT{7N~r;c3y zR!J|Xxaq^jpbz#)`XEf79HbWy-~`Vb&;My1XbaLyCEeA&8cBDxuT9cj<*(@_eS@T{ zIbYYWE~wvDNmoxY>-1?s`W{I?2ir&XfYqGkS9~xh*vCK#;iZzUzHOuPXM6IMpp02U zW3vZbMo!XQeRR8|yJWOc(p`PDP10R`^tPnC`e^VWR7tvA>7yjwmVXKH&64y7I1`l7 zU4e|22#xyAgQm#^>1!pu2J`4MMG>(5*dcnnjO=D;Eb@HX(nry;tanI0E}POr-0%EU0INq>sV$<{*8pq`UfRg`}U3{GiY4B;BP8J(BLyr5(NG?+xUF-9CPa zq#w#ey8Y>(eHD^!lWz&J%$4+2A$}S!i-SCF6q>O}(-NfLE9tIzX|treWVBP#T{7zX zK`vO<5xR}g2g4*?ty6UR#X!=-bIeYs&JcQLluc zS(3g7)7eJ87UZ!+Xr>`epGk^<^H?kCvoW3X@J)bPe?BHOF1>k0(l5h2`qUViyesK8 znXt@`R|@6UiwTnMnrp63bslI1X0W%8+ruWvr9*%19pU zb#r=;q`P!vhorm4fTxtx^8?*RSQC^;x=Y6@B;D1Ab0yt2M%eDG7acsY$ZmI19}H~V zLz0)PKVOh^*Lw0zNq6bfK&9yPEq0VO%;jfFx@&zkP10TCrAg9V^}kipRW&Vr@sOmu z{K^+3-8L7quDmJfQ?Pxk%=DYn2g+t(8Ky4`($AE1SN*0*x~qOok{;?G8c$b?F5fEY zF8Mqp=^w`WeKe@w3zF{Y-!~=Q)xQIklF~PuF@IP;pDF1zm@X3%%Ws;b+vLMO>+Pa{ z^C=#hV|iYa`IGaxUGlk<;>PaunV}lW=^G_|PAL5%lfF&TM`JpyvJhrlH0IEXoet--vs4DthJTFa=IYs@5hAZ3e2O=Oe2R^ zB;D1{cO~669!e8t_%Pmjk)J1T_cETn-wkGYaQ2GhZ&tdR7}F};uPT(X?j zNxE&0;q)F!clGrSNq5-9BO~r7up13|^58OV zlyp}c?v-@aUAm5NA7XQ!{7WeRos!-{aZ^XXAC>jsIt)056Ftu*r{R*WveI<`IaNuz zZS0p&z6&KipPX(L8rxjUc|0WP%P4Nz!Lo+6(pI5yjfp*yUXOY7@r@kIp;#?)?ETH@ zrIPN-zedv4RHNkt_k`Lcy#mXpHnh{I<^0!3I@?RrM%atpAnD44>HJ}TXR9rLSZnT) zbeoJ-`;X&-{rz=QKGp9!sb8t2+xnULYb4#Z)@$p9e@!px8zjAgE2-NL?c3T5{~k$q zwx1h8mn*$g(p^4rjij@0XX=;nZFUJkZIXTorn3(FTj6SMxLs&mGTtcZO_+x)P&sLv zq)!i}SDJa@ZAn){LCerz7BYwR;`0*vdVhW2mYDLk+hDrK`jP{G*@3^}z<=$)e-p#W zjiwLC1vvY5e3jE*AE>AD(NqfG3*I5IAec(gK7h-xQ3NUP)qqf99g04{#pDY8Yccve z6n%gzeG{sbI{7Nh^0cD=odLu3-)V3i*{$dUyizAnm{O-|0YK5?NN{}s57X}o6VtCW z(HD6^KhjTgrMx#xTt{Xp`T*}M2_F+&|JM2N`8-9GkKD4MxPS4vH;f-7zctQh#W3DK zg7eRmIw?#NAV{f`!vK(IN}VER!_`VQmXPw^ieU_-ytf_r-yQf~2mX!&f7gM(=fI5^ zL+L56j}~JD@9V((Iq?1te1HQ##DNcV;D10U?bk96QiJMd#1IFF|ImGX{v z;3qin5(hrife&-wr#kTA4*WC+eue`-%Ylz@;AcDVG6!Dnz(+dp(GGm910V0e&vD=r z9Qb(-yuyJ`bl?{_@Ja{H);NBpyeSU6%7Irq@Tm@bnggHVz-t`%Ob5=hM*K>7vmN*) z4t$OSzs!MO;lQtQ;2(A1a~=3R2j1Yo=R5F)4*Y5d-sHer9C(`pZ+GAw4!qNWFLvNd z9QaZPew_nf=D)X?~q^7sVN_?7bZI`DTK_`44LJqJ$T>$hAf zuZVVg;#bP!lS$)O$~d3>^I{7QLz%4z&cd3@q& z{7QLz`f2=1d3+LT{7QLzGHU!vd3;K0{7QLzVru+Kd3=Iu{7QLzqH6p~d3?HR{7QLz z+G_ktc_j|~BnQqXv&OHKH_U9FxD> z-enH_atD4z4CnKJ>15|v=JzRF?_LMVe^L0s0U)t6QXDxXhEG(u2mw-WR=5ZO@b4>J z9+m@qkHQZH#E*_t8@KWh3#4AE@M1vx=p71|hjJiwqr%l44?yY6afQo+3y|_Vg{!+@ zfYLo{wPCF8l!ow}!sTW#q&%Q-+^zQ$|E6%%4j(^TZ9L0OB**aAl-pY2=^BFa}cQ?pipR{Pr*)zfzWYij&BnchIZB6iXvNKZY@o@|2Rs z(#ZdhVLrW5idY)03u2g0uR14|M(c$b=F`XLDIbr|Q$8*a7)ElV^+g2nf71HGfyd`B zCZu(TgFZfg`SkJm%g6CuqzwRx(fTEZ zF_7}&^PP{!=Q|&d&v!l^pYMG96b@on$~!fNF+l4b2R__^pXR_%ci?9@@G~9wSq^-J z13%k=mpbq=2OgjInUL054*K}K@6(TV(2sH8V;y+hfAI5<`wu=I_aA)xTu1&B9Qa2Z z_<0UI?w>Frt?wN4asR}pzraC1$$?io@VKAj=Res&KgEGpIq+%+exU=O>cFQt@aYbG zh6BIIf!8?jiyin(2VU#IXF2fM4*U`aeyIbWZ*kzQ4!q5Q zU+ciz9e9TW&pGf;2foOGFLvNd9QaZPew_pVm;+ztz~k#FCQNxBchFz&z*ji%8yxs2 z9QaBHexn0l<-l)p;PG`L6Q;ad9Q0id{F4s6+kxNez*jr)PdV^UJMc9Q{4);xHV6J$ z2Y$N)zr%t5hXcRUfvf7o&$f(fj{oRpK##cci=q^{7DD?lmma-fp2!;KXBmBIPfhF{8@st2jI`E%4@Si*IZ4Uew4*Zu6{KXjl_d%XF)O*F? zM@j5Gh5y>%M@X!HvHv|hlOT}~5kAyETSOse3qDfee>CK}g;N#&XM>+8u`et9&?23n z@$C-!(!;piJ(}F?5M8J6;5qO!gyLro`r{AheBO%XGsl7RyQo9q`F!2p2OW5i!v7x2 z{~!|VQ166Tdrwn%oW51zdt>?BukhndKDxbsa?qbLn9F@9md_;){APu}Yv@mw*z*bx zp2Hs@cz`&((DN zEjzxhApE30!F}5SKI|8g5&cl_R8xQ5kAHCFGj0ge4>$BQ zeAIQl!h`n1g}5Q{ji3e6&1Ej^lFU^kWETIsD^bzv8b6=M+Uh(&SSo?dl@jCeP0( z`u>Jq*XMqP$NRBI;o}Xx)}OyAJa{id+pEFHbN$B}`irGLqZA&uZ_NsyVCc2JeN*A_ zdFPN1asKBSdOfbsS9pcN%cb1A6+Y46g9P8D@c8`rAySN?Ua_IqJ6sDCeu2TqOFmCI z=m%28hI;Y-tx|ZU$w$lcK82?ZuI0H$;c>h7Z-q}W^txRqm2kOnyEj7NRWbUR3Xj{* zn-m_rhjO~-Uk@sLn#t!J!Cz4LRD;vBqOLs(uQB*W!OuI1>lvSCZ&dh9Lr=|A*Zm3) z-b2y-`cz+P846!)=xI7u*9?Wv zFu3*;KBMs9{h&jI{#Au9HT1gQ)8t4G^@8_Kd`>Cn^I3!M5&T|--(m1+g6~xLod%yR_<}Od=W_0UnjrXiGAu*Acs;8X9=s3LDD>A7ZtJg4EBfF)DNe{97xE&Zuf*tBTYUp+SuU7cC4X*3|b%n>r?=FRZ$Ixqi{s=8dhkD;L_(@Wq zI~4wy!S@LMONBpSaGlQwXkr@b^%%TT=s%(GrwmTjP}dI>{*1v77yLuyfDQFF8~ifC z=PUdn?id@kqng2DMaczmT59=8JzDE!BUUd#DU3g2pQYQDP8B7qO} zg7>O)|E^T{&ka3wtGfQK@Lw35+#z+c-g7&?S@|0 zfAD#B{XCBFp6C`L+ZLg;kym4`|$?~ z58m6-{dm?y&VP%c*LG})!vA7$-M>8wkL%%oDg1ASUdv(P1)P8I{#T7(O>eEjgZI4V z3;qj*?=|_%5&YmuoKNun*QJ7=ukd#ay|zDF6#kyUb$w2uflT{#2G`@RP2s_NUb>#o zD7;9#X8Fb&m3;y^z(c*bK2KJ7%FqvydbSg8v#Z^TUY|ff--QzT`xPG7ljjs3yf1cz z&>u3H>l3^$rrUL)!h`q4v_5}I;lcZ23nZUy3J=~FBezFgV~CN~-KJeye^w|wcwg*t zp?^-{@%}xiit`EHBRfgxGYSvhBh&r3UE#rdWO|&AsOEfv_sF!IKcVp8Ju=w zMd88wce?)fDm-}qPWR(a6dt^fRxkOVGK2FkG4el3@W~1f-c!^0*DE}DPfge7A%zF; zsp&2tFS>}!4cJ58f{$^;6dlg$M7o4HJAc4cMXZ-kGlFrxiZY zw2P)Qb^Vj@Vu@o=bTQ`>umAZ958hML`L9)Y@ZMa7WcHH6gZI=le)LSvKTdy{!h`qh zw4N+;;9pdD@P6FkQtobr2k*gYd`c~s8@$)1{i}x+9{7Qw#^S@8w!FzVPy+=_6hQfPxO;Vo*2mVdM zZTj#$<3@WV6AE5*(Er7Oze{*Yz#hahmtQ`uY+NR@Ah)nRTi1}OtIstq%4XUYrOQj> zXs+GMIOr*1#(3gr>CDtGS)%gHE~#Hw*Rmj+scWH(8yeCzmozp;>B=*ix?F2>W4()_ zd~_xg%_-B-*s`E0o34rHM|BADbJLZNwdC5+*wI#(t6ylzx4bNrZ-lZoH;rW2*xK0A zn2R)!85^?m>pGiqnWonIYcg%EO^x+S(@ZudlWFV7HI2_M$<}w~8e3b+7nMyL*^#|A zPik8o`4BGN@syhBY{{F$Xp=sQs@9RK%aQE4;*E7pjn~ny))v)a4cU%ddnc&_{VruW zPnmYn1rw(cOXs{yb>`Z|*_M%+YZhhd>*^O~N7T2qajj}|W#ySmvh}&v_RPGxj%;;p zZp`RY_Ibs9yX<}E9~m|H#dl1bcVjau65Y$@kja0~P15|nP*;}$R%8!%7>GSCxc zrPH&`t?f%|T3ef@!5Aj8X`ry>=`{2Pb|q$jG~r^T!}qBKtt`3%;f~N~pli*fRn4LL zjxtTIZ>6&ks|&8i0aa(y2c-0c3=HnKutfK3Y{dmtTkse#yewI0M>EtkZl6^%$WXh&WH865&iG!E*yzv@3$&jwuCOha9+PG=rC3uPl)vWqiyO*B2y)ZJd# z&|tf4de@y-psV?4p!32F4m3GXS(ap|kF|9QhDCf9v@-DAr)GqGlae)0BNH_(?VFv~ zT@=frAR8XyMtfydV4^mnb$3xg)>5g2E*YIU;NLNh{<)2>ivo)9*zLU*&6NE}o&7#|E2?zJ0F}L$ z(!*BVmx!MeNW|7XQ(YMc6IHB#M*D5dYq&@h<%$f-kMxM6d8g5$u&ISNmg7y%`>LbFzLR8^)9nofQxEXZY_(OYj|PR&{VUC4T-Gn zE}A^KHa#mdYvKjd(yR=%xslC{Ey`h=l_CA`Rb+HSN8@$bvXPA~^U1fX%q?xpX4JCR zBH98>LY~!Mxvc(Ja|_7Hcb7F&TwVoak*6|c2W%<&-&Dg%2d;)DCnK_g7YdV1>-K?H z$_=mYbdC|!)G-(`g?RP>wep{JZQaJPQj^3L%05$4R?4e0W=0jXkJZ(Uw~Va>rRPzN zytxE?rZernubJD8unJ{Bx!}ZstD;_3>!!?vj$7pcC_BgkXDzIJ8WhM)6XdfG^Cy(B zFWf=dSU>G!LFU$E=B}sB<6IVZ$Gpl*N3>_>kDx8#B}+1G+4hds7TRFUEzK+{rJ02$ zAYSB>H>35WYHi$M4P!$59(LZ51UTW7~Y+ElK;hIoBfHK=pp z`&f_-H9BD(wAKU#i?IBJ*)5A3=~f3RZ)bDXBR4olwawHoye2chuCd8$=xn2{|L8ag zkAlJ*qVzPvZ0`$gn}MT@y!>DbnC|-2XFEFdo_tlTiZgmj)yjT2O!ge zgqa@Qp6$r`f@D{mZ6f>JbxWc?H|4LJpj9rI^^BvT{2xiDE}ouQkj-UgwFP@?s>^C~ z<1&?ecc*n}bxm_~dJ*~dWto{IS++8M4+dTm<{8K{=}a3`mNqupGHMf^?gXjK^Xgl= ztd!c(o^8s~?If82;*OY>O&+Cjd-le~O0vu_O#5FM%R)`Wm2WLxK$HDFQZ4!F?S+?ozU{#e?=&4#1<|% zx}VBpdu(`Ys~yB%6^wi=-7w`Z6~GX5wbl_aMU)fuQL-CB0V_?@{>>uZ`m$v*l3Xd@ zGPo#4&&w`oY|+*uQafcR+tq^#!L6$5nyGXHk)<~}Q=2OIl?8LiX#>Ev4 zbq|cz`~nasEq^x#dtPo~1lw9PN{#N~0xm$+-KDbN-n)x~`nOjHEZChR0nnq7xjNqJ zVGoLz#iR8taBW*@52$Q(LnH4)@t#md?FW%6_@mJ`FouWsOBT@6Y5B7o8>8o1tI7k1^OL3hA9gA;P1%+--)`2^K+qAP?kOX~ zXcjpBZa_Jix6YQpGi{*7bKBBP9Sek4@3dxDZUMDcE#b@jCA?cpSRldPkKU)Cd->X5 za65T<2D>#+GYeamxCIPtYo6$7ItH?xk9HD5!**n*hTH&O$gm;e5|8hEiO-+#_A|}} z`{;@JS=8pn%1XrPh?7?{)4umg8X}vo(u!o_Lno&^Bc2qwI3!1hi(^+btrS+jCO?bS znekn^-d!2Gea%jfab*g!&pgMCQk%|vyBx$%W4U1eFi)3^!!F4;X0?5Bs3!|mxpde| zW0WhZo*`Nk>LVz~>h%@UGA=Flq>`E}{3n%AK81U$dE-n^$4|yMomv8ifNL@zl$k59J^twfUYh+4yn+AUo zHdei{KzrIsT6*mz#kFVY1((*v@{2Z%f}(jzvkyy%k+iqIUkez2x}_yCuSKeB^ct9K z_xNl}-Mpr3rg47toQ`%<6ngiBYD}wFj<&SsR7cQzF^wIW9NkpP(sR|PvjV}_#_)irg}7Q4(5?AHdmn{2}4HK(kw^qo3~ z!(+yT@<+=9Uz4<6Z9&o)vX7s@(UR2_=ionKtQ{H0!k#wvBCbGbhCC@zL)sbe?~P>Y zI+s*Wn^#85dwzvS-!JiZ1^tC148Pd3as}wEGT7RuIuLnIVO-!n$Xn7cNgeC-ZnAw6T~jP%3E!u!FVpOU)V{ZIUKvW$l0 z?MGK)BsqM_%qDastnC4H!)H~MD*1Qh(Rpy&ec4d&#be0NpU+K5Q2Bju4KB#ks)sDu zSvO`-tp@Y>c&f?!;s_*Yi+%FXFUIJMPtG>oI2lbdzHcde_f!&8u({EH(BOb(+CWpE zKLuzr)=M||J0rfn!r-#_hx_h_(VCp>=6*QTfq}hD#wGeU(lJHiApnyJf7~ z(0P&J?#e94iw8-dCIvR#pG!;+C-ga-H?OfFv`X}EG|Drv+}n0NNV7Lddlt2s4tlh! zK9`}d4Dh}-Pj`M5i7{;E964!?qUHtfbKyK^uYhI?b|DMr0iTgpGp_x46pU?nQh8Pd zhw(|us`vH$T@;#k{QZ8~SDF;s@8{g2@+v?vSd+L6wZY`>W2jT%Gjw$M=ij=|r}qU^ zSk1hF4<}}xTD9JP(~TqV;gse``&p2weD;TNC%gi$;M?8{e&vXIv6*}#`U{LW4<_t7~6-$ySGW zFYJ5Y4Y6u^d{(ygqm`AatVgilYidr`y++ws{RYBqQ))Urax9vmmRwz93w^CX-DnGT z(OO&BI%!$OMw>NjUCDqtxbeD=Iq~M2|Gou1VBXqZw;)R|Ecnh+)BxnI5;+oq!2~_u zM_Yxygj&gK7s<9m^WR__(Jf%ajq+;4b#%Jx-*?7 z3ra(iW^)`ec=xlQCv}v&Yeq`4Q61P_Gj8+tpLMvzt#_=q?0h67<;mFl#>&%+Rdah& z{$7M*ubnDjvw9BGq}Bl<@H_=7GQYqBYdciNzN_GG6XWZ9}CNm66YFl(K6HM_1Xr2;$`qqoCE+Tg1u4(AK6lNzEn~2+Lk4 z?nyP7_Pz@?PVOSPT(<&;SnO$BHQJ)j;x=cSdB)neg}85ReWg0E{TpP8<0L((6SW-HT5qH+<_XC8Q>cG-0-o#*))vb6sDrqt8$^iBf3 zbw5r;^MlDduW2ekbCq9+H{9si9W?=|c_J9i$@>&0YGNjfp6uxZ zk3P|W@s9wmPaM$s>u(zC69J<6=r8tb`r|1-zWOjgSCsx>!8JX7CR$yGIp~K(=;?FK z>e62tkLItx#IEzdQs~ci3;k;(4Pz{2K+3*ZvY(Yb3fo%pKk$<^?4R>tj~*pV|`u+9P2ZH1w$7HtQ)#r zLml`Cz$XL!IS%|v!1=4}I{!}sj`rjZ!1*idn*OVRW4oRP9NYCW;7I>Bz>$7HKP3hw z2W;112Yx2t*sigFWBn%pUJdPP1DsD8(CxhuaJ1Xs0{keT-w8O{;lBfpcKDEk{Pt27 zAv=7S;88n#jNmLjw8N(Xj^n*FlE1dYe%T~=)DAxl^k|2F z1UTB^LH&J!us-mq3R-VV0G|%{aKQOg1x-I1@QVPy5O6;AK+|6V_{D(N1CIH(0gm~v z1e{M#(D~mfI6v3_SgZ|VAy&e}M9Q5Y` zj{TArJUT9FfF8%i<$&Y3SOqwai~j%|$HhH>quzcMaMat)faAFM1>iU?ehoP0|3|v7TKpzny#>v8ch2mL1^^m<%;(n0^32)!N`pL5W!i_q(FvED)dDB##H z-xoYOE}jQ^92eUF$8j;_kib5mKQkF{^jqctj($s%;88v75d3s12G@@tgM3hbx&g=b z-X^%tU+dxLfgbbuQiNXnEe|;8H$>>Q9&U2be?LO6^>B-W{>KsepUAlQrGx(05qhnM zzjx5T8KKvD__l-oUlDq(hp7()`hfB*1|0k4Siw&xHe5fR0`#bdX9JFUI0taFpN)W{ z{k#_NOGzoTpKvqan9o|kF`spSV?I5A&xZVe1o#}l|1CJ{C)-wCpFXl-#MmTS>uLO8 z!I{1i@M6F*pQ8n5Ij4dCM96t zVl#N1;MgwCr%CX^66*XH%Wr*NpT<8SxVASMUoCivgc@Hfzflfc?w1809-)6&exDh^ zzazgv;^|m=Eu5YB^xLt_B?2l?5F0X>#B#fLB62+W@Zt{93>>fVTs#=Vxvg@4@Q&cL6=h z|8~G}|7tzpxPOK1>V*8UUlsw5^os#M5Aweb@CyMy8F0*JIpEkYytk~|i~X_`aLi`~ z;5c4xaNsKe9|85e5%3DYR{=f`@S6Z%0{G267b28|Ji_}{=53O z8R)UURs)Xxbt~Xl&rbt>lGKMF?+E(S{dK7WM>(v4e9DEM_560gvD`ZVKN;xP0)7nO zIF8WneIDphPwoc%Y{=&efMY&)0X_lf(SFjgvFhS>-2?Rc{C38_2>4Atx$+G$pLKw* z1A3I_mjFk3-Um3!^D7Sge!wxG2LRXeGS~k>z+0f6NWUELuLAvTfIkHIJ%E1=@Djiu z27Ci~ZQ@Xr9g5%329$98Ey zi{*^==g&aj2>JXQ@J9hZLfXOkd<*d5fMY&m0mpo%1CIHu0vzXMywCe>DECN;^TqYq z1USzTd@+vo;k^aE7{~dV*KT|4^8YE|`(tN*2Km%OeSQu&%4ZwkXlH%_INBL(*N-57wCAq^j((b6 z&$0Y}3Hj@F8{;nmejL~b%;$8#F`o*+F`os1qo4UTz)??dJmS2A_Ix|k6X&&G0X`Sx zfcc*SINI&;fFBO|Oac5Qz|ns@0`U1jk9J9~-&xMD06mV!UjvToi{}8p2J%OG)GNeM zuP%dp?sw#)_Zy=1{2|a|{q_C<)1#cxKgW3-{b8j45af^Zm);NI{1Hcc@*8Lu%Ja8? z>wVv7J<*=*0D81PuL6$tAMsgGf5bZg$M#~muR;D;?oPle#zW|Q< zvj^}(#v{_>ctpIA@rd*|99FKnmxuHDY0=$s%_%_f_ zg?_~PyYy4?_(z^?~-oHy?P{GUL7FW~P0{wU!8 z0{k_=vD`laj`NNO^XyW{=L3KveKFuT-qC-47xF2Aa{moD+K2Z5$8q;Q;JB`z3^?YG zb`R@=^oSRwf^`JugE-cw571XYyGB8M&_1Vt9?LBQdbC%#e}&_ypCcdCTde;-1qpUKqgDU>tT6aC?XfgZ;T>I0U02+*Gka{B<_HGuP)QP&6kDAYsr$36)8TnzbO zea-2X{mJ!yl zay|y~IT^}5)`1@nIO^vj7%!L)t`m{|1jq;N66)bts1N#$INlM*ae5-;kL3;p9P3#E zINF(EfKP<_Abt_xCjosO;3or)?L~dS_M)GO?L{2hdkW-#GL(BN;HYmO2D}33hXao7 z!g5i6P6PV+kk9FWmjHeS;HbA}0)7F|WB=m%WfSNd?uVQO`Cz#?uVJ~kFM;J^d$HWV zLAkgdJR8czc@uG*FUo-aWN2?G;F$ktz{`Lh*Ack?Fbe2V4r2gEIgAAy>p33qlcC&m z9r!tbqus`G(XL|sD#EcuY=y(8&Q&&w#E3ZTdFj`j-ue56PHnF#qvb(DL`WuH1w zBYs#!!G)Y4SBZoS^^-UU6%uFsAc=D@SK^HGpT1T|gt0?3NZ~q(Gmcg10sK&jbFf3= zOkXT<6>!DrV*GFw^ZzRWd@$g8-;C*x0(>sep8)s@z)JvM2lz>V_W(W=@Ew4k47e5+ z*Jl{uB@EET!6_={|5pL{sesP~T>CJbyaI6TLo&V&@Y5KOs|Rqs#^fZnA$)N#QsN3A z9pj53{yUL>(8WG2UyNxVo`17#G;{exlF5?pc=W;*7AYJ;sAGy`Xp5qka|2JIv zjd3kYPN@Q1uMrqu2)JIeFn%rIT&`X}+zPm!3z+^6z?uFM2I<-WxRx>FTLIU5iHz?7 zT-#E{i^ZsMyDAuv3*YzAeZomKK!1TIrmzk0Nr0~byb|yYfTsc93ixEe_W(Wx@Z$dd z53YX|;H7};H3BD(2b|@p{h?Wa>p6w#bAV4}K(5;XpQb?yHv&E#@NIz40Q_yhwa>=M zg9rFT+^!lws{S7ZxL&Jr!YshGPBESXTH$7uLb-{2ISfV_*ELDa68~11$-~y8Nl`P z?Ogx4fRB?&o$)%r_4DnF&jWlJ(ANWgH{cC`KL&Uf@E-uqdOKIrb^<;h==*-a|H1WH z0QfM#7Xn@hcq8BqfL{&xjeuVR_`QHP0lpdVX25p>-U4`EagwPx;5P#PdBE=l{4T&Z1AaH)I|2U!;C&DGe{lWp0el$XUj)1o z@O6MU0RAPwZv^~50bdO`k1s9j4S;_c=(hrXFW`FszYp-@Bm5uSuCD-I3i$ni*8sjA z@HW670DKML4+6dc@UH^C74U}u-vjv9052Zw|KR#R40tKvUkAJf@C|^s0salZ*8u)a zz&8N?2;f@*-w5~~z#j#?_(=Z;*Z*69mjeE6z-s{C1b7?Z-vN9n;H*DdrfUKJF3@iR z{Cj|J2mCR>_X7Sn;6sk`e{j2=0DK(a-v@jS;5~pZ1N=$A?*{xSzI!S(+I;AaB;G%OM*;pDz-IyeTflRG?*ROEz+VM? zBjB$Az76o-0sc1NI{_bjod1LC|9ikk0scDRvjG1C;5oo|0e(B+Zveg#@IL~+4e;H7 zzYX}CfDb<2|H1YD6X2r&|1;pT0RIc%Il%V-z83Jm0=^0Gw*cP`IM?Y{AMg|RMsTzL zJfdT1bFOZlH-e6T^TUNYMxTML8$q8E%XuRj>T-47hE^k@fOlPR*UwoiS zZN?j+SeTe&BidUH0en7iS{Z$dHM^vqK4rEbtIqwS&+nycE@^DEQ`1Kx>vFBljrIJQ zJ~v%C9TynQl@4ZWSwdT zC46EtFtMDesM_4f=EfHKM(xb0>KoMj8ETTFFMrHLruv=vY<@!Tx>Du`p>^Kkci(5p1oKX}kT4p?P_!zwo=KWu?=z&2#{HO>1jY z{-YG^jlmC`VqR%GdARU`K9O$E2+P%Hi~5uc>g#!oX4`TZGHLX=#IQPsyWD4i;a9gL zI{dU9WK$4)SXEA?vR5wIcWLc;M86_hK%cvZrA>=y{1)!UBbh>c~9dr??PSVem8l{mBtXkSZ)w7Kq zd+P$@UdVS%^9smOuC`TH-cU_L*2SjK;f6)YD_e~4_4j5SRVv^svh@AiXisUab7ZR9 zIlis4V8QQxLaaWYv%!DqHa>;gv$$oR%0fgQ zOyXooBeS)2Qa4BCg%y}hpUnr5)A&NfG}34-n5%24=TkeH>7?ac*=RnGh!%_O898Zi z7M&(Ix}kCY{A_!c2s6Ck<1+^7Uu-OD9n{u@JFQ)Fg<0 zU#Jd#NzbH!&x(j=d!P?hq2bSm7)6g&aki#lXU%2-P%H&!M5Bf$Q2#vA#+Ht3d#;*K z$WhbFBG|LC9*aX{>i7+tG!Km5HioLVz~%IydKTghiiOFq=LKans$+3o8!3{-ZaTar z<7vr~N@j+5UC`^$wl``-njS&z-W6C#@luu30D2|Bo*>mzI z(x5q1>111Xkr~q*=HzO9o}-*d_vN^yShpsV4j!{qPRA^%4Lp4`w;Z(;Euw&N&)M`g zd-XD@+omk3Uau1totnqy%r;>aWLmiTbo>}?y9BF6I5CT?dLb)LWuWqO;y6bq9tcB@ z{YuAYQH|9}j(pIf?p4dMv>nn{P*Yjjboya6mBMCUl||}K#g2_y9+X|-4oo?jCEut(pKak=SeeH8)pI)9 zNxF0oxuqdP3nGrTwC2>npnKnq9hqEb8&NhjUX!h!#bY;DUPh=oxHvfEn=(^lnOZ$g zr`C0JG%ir($S@zhsII9qn_eUn63a{}v6dMzCz~eO@QdRe-}07QE;Lr4TVg{b@P_QW zb2#_2D`cHB>G0Qx<64@b-OWH7RI#$}nW^hsLg&kt(TSVn0E!mT$*{^xkF6g93sbhw zyMm#*@d3RNXU%uxwnce;c<&bXa3 zi7(O`lC|yZbU9r+@H!(k%qM8OmlqN1$Y@0h601FHB}6(pJuJlJRm8b{B>8OqLTF!R zc56LBE8x&ye>RVrA-sc1U1Ymg80^v{n8NdCRp7jBuTJ2(LF@KmuD(E>gB|WmWf5>y z)tgKTotg8?gC_XUbWH9q?bME@>6k5#HpIP8#|Z?MbPBd+olXZ6F1y@%-4Pif2ejhN zjp?TUCYdzQEEuUejHUctSitZq*sUFBEO4-T2@TYOb27Uo1x=U*HC7p`G}?OG%%$g* z!tdLXlTn*#Xmxp|G!Vx&w$LU|dtDQq-P^+VP;%At)QlbJvwQ~G)XUwQyc7Vv=A{s? zx&obZ=dbQi(X2_QJ%{Z|K9=x=1^x*&nV%wTXrLR)9j)^VyGyFNhpNkNT(}arM^boC z`%N!MeaKk~G}q=kilNkVWXQjlIDZHPa#`bJTt44pur}w zjAJTbif}YSZq0E0d3{IsLeveo%&a!O`QO?eb7nx1_j%Qt&rYj$m@V7Asc&slH->0A5BE31v!g#^67LKtD;^s)1vOjQ zz36&7mX*8B4*dp`9G^*TJ!Bsx{@RM`=!;Cvqkey&bXjcH{Qy_((j&*BL2Sv@HMV5i zm1(5^$gOMdtj{&Jw&)vk(ZZ71N<$so&W%Y}$s~cPRQKZR>1KQ`o9UpZcIp;n`B7eN z(4oU@yYU5>C0N$_tBUB86;xx}Udn%iMMQHLA1{uMFVJ?O3qux>-Bm3^Trlptr84$5 zaFA4AXXwT%shrx2FWf?<>Aq<6UI=AJ9s#MS+DFgSWy%h$$%|K)ELbMyH4+)Qz-;CG z4d6QRi%Zk8wN}P=D`=9#tpv<>aXarqbH3CGWRAe%`ee-@EqY>&%8%~q5*>*F$}e|T zAUd2KDrB`u5ryG_T#&-pGqd^6C>O0Bc5aeFbSur5LR1;7?9><9ZR*9k(wu1b0QlBA zby&13$R0bFO0;<#T`C;_li^R?>V_dXQ5^bBnjGI=qo$B`G!kl^cQrf4+{6M+2KhSL zOJF~}hDw^LT!g4v(Yo2FqtZp*C&>R~Mw;GK^m*v@jWp-Z?lAi$;I2~($R9=f;T#c; zmfH1!7CR&co^8YJ()taPO3WTwa}@Q6>L~Pg3qOGBP%h{Yn2gf%W=_p}OEg`ZhA|0s z-v5cPtqtzv&%pU%|MRwP`*A-9FEbSM7K6y#=3qEB0F=cI&#T-cnMv9aa*cj|9zbhL z(t#2&%a1K_~+xY_)c z>?4nc*hf^J2SI2F=6aeZ*{##SsK;xTpg(CYSB`l*z3jOtOBPq2Bv*HQT{ShnvOJ_< zm4Y@?YTXmFE(KMQ_14Y~79(cm3Q#MxkVmbe=VsJ|q>QQ_CNVY6)V7h~?`+9s8+@NI z|6|i-qiLy4o3bobHDSb-*=mwVvY;@MmsQF|G!~3+W5%<;LM<>~9@oL90_U49;6*pS zL99ny;#OGSr16`gWUiiPtZQ$tTPn81d^>>M=cW;rmy(H@KO&P^uw+T5E!*DF+EUll zm|L1zRO*eWZ*6YQw$MxO^b|yP#FQDc&!QUu_1DPn1udN;+S*%*DW@~fF3iktuWP3F zffuy5cDBK16ZSd!_XCnZ0i>3a2&W`5Op^JZw@DG-d>wi=N{(FUQAn?CK zab2zsUFC1E@c%0T|6>;ZcN6eGYvKQQ0{(3l{`V5_@38Q{pMYO~BL(H}6%{W3y_WKe z67csWTE4LSJ_-2sxiH9|O2B`-rF^cF);k@#^iQ9AiRJf8z&}dLNBhC&q3LpU=qi7L zrT%<&p61h`i@(ak-#-EWEDQgD1pIR?{0ArCZ?fj%3o{Y|6l_CdoBEjCg9&7{Mi2S3EIEOQvRR>_;*?O`HWp%Zymbo|F(sHZ~}hyp;NNI*ng7}^xpuYO}O;GbjRKQe*- zHCXtMO2FSH{Mddz?_7@?9lFL3pH&xa|CI^&Z?y0qouK|}gdgiaBmw_g3;!_*_}5wL zKc6ryHyyg<_mHLh#svHuE#+UGfWOC5{;>(#|Ez_d&luM2)}gEYTP^ixpHlPb(8bSZ z*7Ak@-%gn3)1ixhr=|RB6Y%e`@ZXhypU+(63++co0)9RNFv|a-1oGpv-uOcKosfW^ z&kBn2pO}E3&jjNO+t24o>weRrtN-}yC%&-$%M7{L?J;KP5r^XIc0kNKn2$-wXSn&u5fuP3V&ULQDI|2Grx_gMHJNx=Vt zg`d|X0-MmKf7>nm-%7y0Q~0s}A5Fmjrlov7C$-S}@3oZwy#)OF&`YfUs|on^xp_GL zA5XwPxL@3V=JT8jZU6DYkL7?~;7XEDs`1QG$DF0t1;D5zZ{)-9tcL_hr|Cb5) z_4%Aw|Gy>Rf7ep~?Fsk?92B>|ze>PgEc{si(-Y|5@fLpGQ!7;ehg<6Z>jeCzmhyj- zfPaF8|F;SFt1SFG67bKmwEv6*^1sqjKJQ5uD*uI+@?THD-)1TQ4+;2}S@?G);J?Mf z|3(7-+b#TmOu)ZR_|bmw8PtW!f4!yrHxux0wD6yqK!18H{C`PM{tK4&^PYC0?cZi8 z|E~%7cUbt}O2Gf7h5zjY{Ch3@|46{!w}0Gz|2+Z!K;cLGeO7}0A7bHuCqencEd2jW zz^~7H#`^y&0sjO``TtJ9UnTrFe)*nCq54~6DgV6${8w7|-%r55&{F>*+EXmF{%w}> zQwjK&S@=gJ(7#(O{0Ai{U!UiV`qMuF|2j+gd{3-U`8{OeKR5yZCJX=B3FNog!vBE; z!)c?{1^2j-vkT)Q3?2~EcHJ+0e_8!e@FuUIhOh# zlYqa$!e5qP{I*&6Pe@SyG7JBS3HWcZ@DEMEf4lHYpq}3!mVp0mOZlfH;9qYkzdV8b zH(1L5aDwt5vy^{&0{+bw{^1GuU$E4FWPrTu3l;NNR0 z|I7sZeGiWN|7R!QA86qpoq&I^@ZV;>k&EEJOt8@X!&On#(BE< zqx8|=Ws>JXR6o3Q`EqQzNs98HPbrK=`Ja>eA7MECNR+=;_@njTA^ciiI-BV48I)j? z|3ZFPh%PN>mOr0^t@$-RmoQuXcS-qjY&B8&SG4}0BG|^?M+B$Kaq)l4!GEmqpRO4t zjF!)5SKG?36#f%?;pb-}ZT$7ZU)c-4ex|r0qCYD{{#C-y^504EsQl+R%D=zB@@t7- z^?#)NXD#KwZYlp3NBO&@{Ih#$KfmWf|GY@?UdhouV_76Yw_j)b4M+JG%X#m3hCTCe zS$I7}fUW(-{ID5ar9#Z|XI;_dYrWhi<@brybC@W;eveeLQDx~khd*(YUm@l1@T;Nh z|C>Y~E&rcVe$;;Cr2MGOYZ(E`3ihTK+;QANB8T zOZmJew6(u4PpowH*8h4<9W8&YrTu@W{G#o@+EM{>j9HSG526Ia6EvcUa2j_n)KkA4|-({=cQb_LmdCt^B=~ z^7;MeX!-LU<*zTW{dJD=i}_(Uy3qb`L!;$?-BJDv1-Ab|NBP4o<@d3azfH=I_UoHc z{wCFm;6EY{UArCizgIpy(3|~u-BJGvsXy9}ewO<8r{@e<(oy^Ub4U41ix)p;rQeCkE7*}q=g?-M)ha2 zls`}968snK{}IG*>;EBAKI$*OKO8Ne@AKQrKlKE^%&MSh6^oX?(oueul&|?&|M>mk zX!*}O%5RbK$MsVF0J_g$tN&MozqkHN5x-4-Yb^c8?|(+?U*ahLnF7l{&Qbn)OZoi% zWwiWpj`ClZ@*n9%ej^>__elBcg`fM6-@lBO-|i^?^CI|Ry_A2agZ~NP?`{0w=BWQd zepr<*)PH_|Fb!!2BN_ ze}@sjO@B97+F$bjG>h{s-XW z_MJB)pFXPiYi6R{CuIKjKtHChYhg`~-#F2;`l*Ni5#^ItdkUEm)d;152K=Wi?Yq4{gNp)VTd!I*|)7drpvik{`aht!XIe?iCJ zY4Ee)hvNUp&`*Ot6o2`5c>b<}es&zWU(xX&BYGD94#b~_-i^O|{-xtzVDLACADTa! z41SG!Syx(e{&3H4H2-0P|2g=f`TIeGp9epA{JG~Zn!nlLpM{A_&mTJfgXH;wi&UR~ z>Y>-y|2`{))D%7r7CjrkHj=+8;=kJ9F9$y~e~dNwog{zm-;~n%`;Ec>HVQvfKQ|ft zEhPSG!hgZwj{yJVzEl_1^!SZF*cdJx3+}GzXtkH{hlp)=D*mTY@{WR|0%@(y#{|f_@VqiXXroMNBu5CUvzB2erBX! z#Gkr;J~iUclK9Ud@n+N;K7MmYcwuP% zJ6iNC|K;vwd1=Yx=X}DiHTcWH56yq~8v1VNB|WL8^S9m5KM#Fq{(HxWKS$!9LE=B@ zP@eyy?|Ivy`R_(Ue>n6fq@wq%UR< z7VxL|G{m2p|FFUD0Y6!Pml1xETp+ki>H6CM{%JAUuW0_)20!CowwIRVUqSd24E|T( zhvxq((X;Wd06#hY^9g^E!5=f)$E&a3d{q8t?(-UhUqkp!gnzfe&x0R2f9^E&S3-Ys zpVh*e_Wz;bKTrH$P5jT0_hj5Yb^YA8kNG=K^lbc^34a0M|K8wl0$+ZdYPx=(HT2t| zm&@~1)A>76-t%Gc_d>7dFRj1X(3c$VvCC3)&q{7uCVG~?Hsr4jo!$I?{ga2Fb zPfx;RUFYvq|=>e~Sr!kHLQt{4s^S^nK|3{ny}E68;jxuM~rBpSpfCnD|4- z-}Q$65a>h4U$f}h_|*~rza;)&Gx!t156xdM8TiRj$ z&@YExF0WF}v6Af`L*EX4==$w0(X;%wApYd}>qZj)HwOQ8@avLhWnIVrrNQqa{F?}W zoFtg#f5eF%+^rI(`KODX#lH=Fcg^COn?G(L{J$Fftw`KlFONSpzuVx~xSP$=@`X>& z|25%%VeluP;70pTBLXW?;!r~H27OF@zne(3z!D0-IvCh%o?E;N__-x2;x z27fL1GogOn|3mg`I>8YNmk3U_%cN_X)(1*@nw~L(6RVu8}Ylo$+pbUUk&}DWM^5|`WDf%_HIAeJ@a4T{#Z&{vi|QU{%;e#>#2{wY2afTa=R)t&3`$|^(}_p zyTZ<7u~GEQe+}{9N&FWN<@`s%m)k3;rumZ${RZf98#UgCj(@!9ng2ZTzgpn>`P~Bs z|6}mS#bm!CyGXSa2EQ5nkd_@Vk+Y4Edz z|2W})WAKMy;tJK@mj=HQ{AB&DCH!gfoZMwf&p$K34_*J9BYIYU4d5s1?+LX_o3I< zAKHJJ=vn@1NdBH9{+kSbX_e2BKL6-py4>LBN&a3S{1*-WwcwZa4OL-H^GC^ZHkQ9O z=-vObBH6Tlgy>oR+TgzqdUyP9BK|Ko_+4jss)Uznnm^azZ*wm@NlVt>ON9Tp!G9fm zOhd5;&CkelGDp(&SAr8y=={Ie@Lzg{_n$m|Um^Zy8~kzL<1`%m_j`@MW-byv8^3bG ze~s|pHTYM7FQ=ze)A@VT;O7Yc?}R_$IG+Ezz}MH$ntzp{Uj==r{(dTYmcM%9e+%*7 zY4Cf%*Ylqqzb_2^cIfr{BU=BN;lGLae~b8^AQP
    _6~#`|HHU$0Y5^Q%P9^4CiE z?+|{o!Owz!c8cEr{_E8SzXSXl=-vGFF5y37@aKV#+isr6n@ZSsndFb8z>He*TVtkM0g;7$V zEPox~Ux5D6Li0z6p5?D@yxVK~`fnHEyXTaQzZ?Ae2#lrX&olTfguk2cR~!7goCk*P z->fwFUEoJ~cKP2!_|@{9kj1|Y{6@qbEnMGZ8!LL2|L4I^oUX!H z&rI}sef;V1`;XzjwAx3S^#4ENf2N#ZTu(j!)`36I+lW8?UgNKs>7r-(s|P=iu5SGH zlmGPbv)15ugCC85MAY%`H1s{t*F{1swf_$d|J@{i1Bm}Jxj=IJ)cMO!^1xY&ul<*b zp5-qy!8;Xuh34|-7&?Em4gPiDhwA?#gI@uDvVMvQ{~?3_9{8c_*9Q!KGvOad_~OX8~<$1_l)Y_jeiN@KVk6K zfv>M$WEZKn#^6_ipRB)9!Y`5Mw5zB}De|0>eUM=SrnD*86&M9=cqPWay?{QC_4dhnzAi-@v| zRNHIl--dors<)3dxoN5-n8m*n`i05PvaaKwB6=2o7vfLOUq_PoR~h^{Q#?52{}n^O z5PJD>s%ih54F9{y_#I9B7s&z6^0ylNBU1GK`8&?gKeLbVpJ(V_gMNO1Un_bxex(zA zrkbI5^WQKMf2+Ztc$OE2j-RClzZLwb{vCff;lFC|^WaDNC=IfURCCYy*!VR=AF96# z4E@c}FHF(;kN^3iXZi0y{K@0T{VjyP{#a%3GdS_f?VVK9@!xOocaZ##B>cAx{z&i_ zr|A9Tf5YHsCiSns(S*OKl8@g?@b&ed-)o<#YediT-wb{pU0wa1K=|(({QWR+m-e9w zYnuP2!QTS@-~_*d@XwIvRIGmIfFG*gOAY-(=tKFtSoAD^y(E9`-_O4Tx75u|g z{$v-aw%5@A3;JBDcmMj^W%#c<(~ng0`s*~}f37^|V)-9&jt37{eC_`kLq8SzvVeY= zJZED3JoLGM{wUG2@oPc+$@;4z@n321JHQXkKbIN&cETS=`1cz8jo_mU;w<=4jlX8v z4SpB+k=`Bu)r7y*;18PW8KL9vOnJ`2#&0zso#mj|6>OK62rgj zBGm?oLFWG&L$Cce8v5nX=Tfx(zrV~8J&V75vL_|$cM6IBA%nkRAN?OD_i0)DTc95q z@LwW&=08XL*AV|R4Sv!0JvhXlZt&{~e=6a(8~oAW*9PLh!{9f8pFI9*3I9O3&&u-O z41Rfl|1Ht8{CAW1rxE^-4SomsGXnez4So;!dFb8zQ%Cp@8~k11hvuJSb3A{=7`Q6} z{+EcpMEdNn`ZYsethPtrqvaye7xJ(ARk(kwLt66qolfGvHRt{BFF#ARmHhw?)=!9IsaK-vu4fg+55~KZjH^DT3ArvaVqdG^MX%|dS21jf^vl_R(RXIrxfj_ z@cA;h5T63q61W=SLOEZ6>vFiRfU6R&7PwmB!nzn-Q>i=+@@gvAKz=2ar$au2$}=Iq zipsT+f04?wAg`nHY{=(O`5MT-MCG}VUrXhAkY7jT`H(N5^7WL*Azw)41mugTTn~8z zm6MP!rg9VHH&FRT$eXF$26;P`mr|aBJWb^cF-cx)-jm z5OyErU#0T>kYlJ||2{zZgOGon$`4V#3G#1H`J0e`i^`iRe;D#dsQf79k3s%zxaL%M z{m1jDpt^e?{Pn8W_H+z<9!tZXeQCjtjUNs>=M$u}ja(9&p}h zr0$9xDC_w#77rj#b=QvSoe@>qSveTaL%w+OF8XHLLb{!|r$;L{x7WuLP0xETHS9?%2M7g{iojnEv7)w+})S zSxpzTL_0aK?8oSPYV?9V&{FUNH+=5ktj`VS9zfBb5#1;(JUdML-qj=KaOcZaZ9RVS{ zOH9Q2qn=yt=*f(ipR2ozs=J3k+(XR&47_yYUqYjJ#Csp32eU@@-}795M(*+O-)NMa zox4{)y-`vPbNzF$QrY&GnCC&TT&ciE|Mm*OyiE*{uGsE1t9JLjQrZ1Ca8DkUdp|++ zZiAMJhO3o||9NMHTJuSgofUfhw;{G?J17d~O(_Ov7$6W1)E~8bjeGrh@1GOIi?6!6 zYcF>PuLYiiwLrC)MPVTuh4454<{hYB(=&qUVFoKPF`UKLyadZIoXLRWB%aLPdk5BS zNZnmgw6*7&$E>xI5xiKuO%UVin>-E?&^SsA1>%9ziri_BS$>1%S1abb=9vCGjlQHVPu4>*t zDFi_qm65s2IMF6UxuTF95=*7pQqjt`hGf)>Hnm0<65iChxT&=%m6@4nZf>g|eSSLA zmP#y1j=e12l5A;9Egut2H?2rs=q>c3i&JeaQLjFgOk|Q?G?9sV=~VqinPh9aEp<^k z1M-2fq`tk~`@H9|+-b>7RkE|GJ}E^vC?uX)-kuy26)FjDYPitn*LF=wFK=}k2PS5~ zsFn-KHbqS3sbF($TU)asuevRrkp^YARW}i7e732nWKuDaTQ2l&snedyA{p&h^V1VX zJO%;=3e1>;dB%i3_N+8?Lv+?>4=lTkJoOwoQc!duyxh8{;6P;cK<~kU@Su&@SitTF zY~4WbVTrw5@bv))270#*B0A0cR>21muX}*Tydw{tQgHlv17ihk12f*81%Tqix)-*$ ztK%T}2CgFIqAtu1R&S_<8wYy~*T8&7I%M;kmH|GVb5P%8dZk9WeR&yfMQeL9!;}rd zZAnZ%e!tpv+bD(JtQk??QM)?!fi}3{CI|-7{qU05q0A-Kc$dfwRmKFKQ1?qRQrl|n!y7GYZ7ssJ$9ZX0tW%d?>Syt*O=GI#Lhi|!HK>&*b)szKb6EuhI#ao#Fz6#b3 zFOPs_gw*c-XWAxQ|{`#^V=g5164OH19#J*=U+>m^ueLJP$xowL`36oo&C_f~h$ z#0C#lFGA_nlV}mQMn?kNn?pUXMuh0A!M*gAh+LQ-)F||AmHt7k5Z|CKh)cGl+sj5N z$c^!eeeQRCw;F;@X?~M@#|K)uzx8MXn^70G_Z)I&+kBwZuk zpNhWprY6^#5 z6VspSK?s^67#IhAg$l)h*@&*mWd<bu8~dS}I8HDjpaP%L6U z_BRP%>2XJ|(G3g9Raq9gA_gOXT?P^Jmq~S*vhne{_IQ0;M{6e85O2?$F;aA8ufL?AC%G0Q7hR#b52&|?N~ zgy203FuX8*2LDVsq3F)S%Hs#k0|NMnb(bOz312$;^>c#y`G`YOU$-%TT~NPLS9_nu zoa>W8{U)k+>3@Oht0e!e0sk-hs6R~gqcD$!+jkZYVn|9q*IyIJzm)1-{PX*$U)hIz z)LpQmc82PiWAK|+?@aEWbKwJCn6^WwVxRhLh^>i4wiPNvkp><=RC?7$_nf9X3n+%t zLvAHfb9+|=?Hxn)V`Y7Vu17u3AP(t#2(FMlBO<;%+%Hj^Jtt{IWDl3$F$9;%-#>!JT!Kn3dGNA=G3tB&%#TP1&Cr*D6h>T&Nw(7wff zeFfFK{Iihiaq-~u$3^@5;NML3CrSRAfWL?8SIGLprM`Xps9w!c+<#vT_|?P+yk(NV zF{qEK#hIt)DCQUOQ9<=5Oa3bZ{)JSpUd1v0jG%rs)lZcCtwH_fKKOg6-sPWtRDYe6 zU#J5R z_BB$yYn-m5`tv3Kg24W*RDYVRze?9f0oz6Or_1_UU5}6}eO}G2e>h^Rk9ZSouov+r zIq*vyIK0R-5>D*}_=1xXt|>5;!&NBss(D@2;F{%9Dg_ztCo2@V8s|Q}dKn%NLkah4 z6{=lq8dhr)qKK$nZ}fq(uR>lZaCM*X=`YVFhEF>beSl}%UlF2zU(pA6m7s{IU3LP1 z;+f#MLV+`V4DbRk@IU&OQ^aE`Mm%q}q7U!^RCP1q{F&uENH|R=^cK?dzE3y}FCWM6 zrV2nWg{6<9nuwll0%NEY@ur0^1R@^pAPlF7hrJ(85l@Z6P({SMI)ot*@oF9TEC)W@ zf#VLvaEf?y!yq6LZ=M66@4&Bj;Bg0@aNzY0Jn6ufIB?u48BP&Tt%gDs5w%-11cM*c z>LOGT@o<-AI7Pfv7z8BZWgK{i1IHbj;S}+fg+V|fp1S*l>LT8X5Qad+yV-%Sbl|HT z_^l57HV3}OfvzuST1j?-|8cpJkYAQ812HCz?(?hAv89(SpRQ$+1p zwc!tjK|mtjLm~W6FpuH==(ES`Dmk-6yiFl`1R~xy9r$Jk{;&goG=zT+?$>C~O=QpM zf+FH=5ubu$#Ct6GAW9MMJK|GNjChX+A4DnQJt02n;xzSq6$#<%70!2&p!!LL^Iass z@fp@%j0UjxUkc~DNKkRMQaL05s5C|4$A$0}3LhH6A5r*;AzbeuR;&7~x*@7b!?Q5_ z@I{5o-q0lL6@F%j{(Xgy4B=(UHKRg!v%=2};ZH04f)M^kg{!BothzFFqn4|fpgQ7h z4L*ocBh#%VC(2azHkZdwheh&L<@0uoVg>cUkKFC3pp74g*5Qivhqh2vJ!hvQb` z>gm)X=hP6U>D5z6s4k-4^jYxmxX^OsYCT|!c;WG)>CXy?L@DCo2ETBM=(jeZiim!r zXThT(41tIj9;X^FcF>2%uci-=UyYyVkaNBR507^(Cp_LYexXB-YaaNVgMN$yAM3!? zTc=Q6#8Xev7Q8fsArR4TXF_!m^|Tm*f&St!sN{@u;NwGh2Bsc-Z*HdbfESDB7w+jd z6dtVqx#ff4)@7rwQTQ}pJU=J4IPfnkyvCsVwvV%%R~+;|RrqwDp5+XOsk2x- zPcdGn@L<)$^3x9dD}bK_>ys#H(D!xE+Y0!|!pD5Q@KL3gpwp)F=O&KS4-_6e zH>vlIAZZFr@isZN4t$;izt@3pa^OFA;LR|F+3Y#sp#Q{y9}BlAo1CW{emeu`(f{w# ztr_3>W18rYGac|^@%&q<$-R3O9z2KMNcfuy51xy8bS ztvs`LDm<=<;4k;%Yk;F1?0-yAvj5kB+x+lHB`2YY;V;MSnQ*Hv_Ue6nGL^0X9OeHO z?8Q_;@@D~VlV9(^+Z}RN177Tf$Kw`-C$&=e%k#tkR(N>ch{BShSe?AoEVmM8QsKe- zgwGPbLE%mG$@8D6-&XjIKF)p~4NH?^@jO17C{hX!p2wN~b%lrh{C^c5Jg4)Dc^NJ-0UwomDKJ(|N@+5SB{CUf2&a6u#V-!z-S9 z6n=9Ee_P?fb2_)H0t&@m@SJ`fao(wL?j?Gk4s~xTJlx*FU~sV)yq72=IWrX=Zts%{ zU+dd>FO@!2IC~eyIpO1n+wHtr;p==k?2%F6kYevPA0J7j>j57rp1UUze!qkMUlhK^ zmva%7hMsC4U&8?}_QLJDR^jV?Ia8?gO^2Lq4t%=<|A~@+moFc>Nu{B%v@Q0+dcUaf zyM220^WA`t^zQL-_Rnqy{V^ku{(hexKB_bg3dQ1m2*<;Eg>!Fv{>teMg$M5+cz!zr zmfFSQ{lobre~H33(v!!}4IuP9V=sS3c1qwV^59p2*ZU zTI80OWyv)t>oPOIhfQs;r#@pe5E&biixVBqnK*2<_jeE>*@Sq!J)LR3B)KeE-;rr* zYc20An{jbExisdkZBL}!wUO3FH0xabQ|HlgNPY(eS23bJGg=)YhsxtRV^e z>pLJE;BzVZdD@Jtub46eSUMKPYk*Pg{D-!{P9f;ZbYmh#+o;M)XR4zFxDUVg9Zyx` zW1ZMw5K-z75MvtNn^hU`WUFO@eyL<;UtMn|2@}ToEms@K`q<&cZ83!V$}7j$)Ro3N zW3k%V@xJ^5MpU&DQMMy;_rGXJwg|SUcVAk}&Cee~f4+>tX1!Q!b{rf!u?hC^ElDMs zVM|{t1b1tuX3>mArLbM^=wcLXyt?jU?v8x-^obrnx)xF^`nzWlT2WWl-$saT>aTw7 z2bxH(o;tM-I&JopD`v#lrtJ93J%p>u{V60r2UnH(jQ!Rfh;UC|fqiZolEb5}G6$YY zlP9nxyE!bh_n|D@*e~`#yKE)vjtTMhdZRTH^pUJsi&yofl2meW39Lw#EsM7&Q|Y!= zxLakG$2&`5NVMM6LIcApwz4dHKAjHp1S}JRTMevw>XYeo4IA34lW;asGfH~~LWd8$ z=>uzEMO|iGeKLjXq`FLVx&~C0!73;f&!iGfnRHEkEH(?Ui(xk{+;-ut86IJ)sh=H> z)s2tWPOn$(XlqW#aG?aUlC2HRHTAhkfw5Eyawh1(ghFbk&Z=vxi^qI3{d0LN^+d`NkQ8 zIm^n9L)l%W!AV9KI@uAbjyLMAayAks6@N|PBm!SvP4^OZr5fI>j*m|I`Vo*(C1zfy z5-|n#*W*G&enBNBt9))onMf!7+ly=@dn{I2o;@{%LW?PT4=YoQGrDpT z=SS8AXvUHxPIEe9^7g}c5C9K;;ETnu9&2CzsmH!)p>(nwN#JN5(=esBv~?zHa1wSB z$F9a_kqIeyet~CUFg}?|w$>+WV3r<-rB z+^X}{eUxkNoI)^mF&sOICtDMXnv;B^#)IhbOk-0zep6GX5yuVui^ZU;E+N9E#WizV zGwIY-a3g_pJ@LjwI-X1~&gio%aC3xbs_~tP<_;K@)RVy@R(40`W~FUq(H}V%`=jh= zZZ5tQ8RXgKDbWtAs{Rh9V40ow{D$+qdQZ|uQyYxP zJcm$+VQ=9 zn^?3ToiKcZF76Ed7B&v1Usxf)t)@AdNv7~gIdcgvMB%;t;eawfb_%@J<;wy5le&o)&nOz#kWQ zoxtV%92dA8FC7B!5puc+N3-PoypeD-{vQx>Wc+Uxxb*W=hJ23y?+bb>Y)FHLC^8O-$DPrLC^P_4;}QM81x+fh1ju>Fq7>)mT>kz$Nz~A`Vj^_ z$A8p8KUUzfUnUZ6#{U(9UdI0nfy?-xFK`)e%?^B}1Me2NoS(l%IQxe_dXG`TjL#>9 z9O>un0+)Vxg>aV7@%e_JmvY`V=y`trg@gXr27Q%^0?*UutHSZ{fkDsl`H4e5{^BAg zwx8qkSi+G_+FxYQb9|0)(0|6D=lHz9L0@9fb9_#4&{qmv_RCDd&G?)v=w*B^61a@d z4uQ-0d9}dh{QOmc%lY{&fy?=M$ROo@HLf3rBG3P)3tZ06=LuYnmq`NOA>>pE{1t)E z7Wk_I?-00@-{rs`5;)GaZ2wb)b6ipQdCwAVjyv_gcmTf{Z$A<8W&i%vkk8}p-vqsk z+x-STuiJm;p#RXIuL9dJ{mDUJ2N2 zn7F?8YPF7qAT>9aCLq5mnhk{lLapC3)2KH$IG<>-zDU< z5YF+#@t+~wjQ^EFj*Q#81up&k6+=G9|3iXa%6Y_~=if$q+(G}eLC^930|$MNLC^8O z%R&F9L0?MG;dmYn6Z%Wq|E@vL@&6kK{qGHWj{lDx^oI?4j{gBelsO@LP87K8moo{+ z{*~`l&KLAD{!0ZeUgMO&sTU30%tmy1-u-^p6Sr+XDY*fy;6F zGl9!-`i{WmI2{D%@GxP&%XQQl0)G|i*#2t-zFXio5N?jUrG$SLXyyCN0>?IC z0!V#7r9QwPjz_c;FSZBeVnPgWp^Se4K}d+BD#mk>gHr<2OF3|q*T?z!65F~(;3J@n z3GqkaVmXrjI}Y3>=Ltc736wFR{FmTj`J)hog!qeaF)r=dBk&8Lj0x%SJ2j@4a^4oW zl=D-8Plhrkl=CyVSk5&FLNet`IX{Pbrg!;gq>`fL{HvfJBk=zvaM`YR1kQ3W9!?ba zm!OQv#HVOj+5c})W;=PFz;?YW@Fqbo*A=q8`vkqT=VXCP`hOSr=b(%U_5KPjwuj>r z@m~wP7Rs0q-wzkl&qojv;{O9K#xn>)LR{J}c$y_EC5 zz%O&q<5R!YiU?fV{~MygOxiE$nFK}P zvoiZp>K!QLb3H=Ro?PS%a>(Jm4Jf0S1ex)Fk-!f^nf>+|fn)zNPO$1K{7)v>ZrQHm zg*}oU=V<=jK#6bX50FQxjI?-hs-Oa5tP7!q_y-c=LsUusSxsfcX$hu~%>qAx%9!;~ z8R@xA2<^i^kPtsfmFN$(5%Dtx9u+u`O{}gE_}PMfp}@}(_-cW3UtskU0^cU+djyU) zMXByJfis3xyw3q?@#5d>`GY|}9Elr3gU6tXP%wpIU)qk$;y*?gxkh z&f{vCLC?RxbDO}C{!+LwZ5B8d`S*0bBk(BTwQym2S>QZ3BEDbX=ORF9Ktva?`aA|9 zKUd&4TE@tpy#$Ceuuz$jzWBkzfHhSD=Q#oK27#9%K3moP9_m`6N@|H|ukyip+ zc6Hv!B{;~JD1qIZ8Ly-vkx6(Zi_&Qijw1t#pRrVuYV#CH3G5UN{?njdeg(x!4M!@1 z-!9R=mSUsp>laaMJb2(akS%`~MYjaoXBgl2Po%h6d-=5#@RfF>?^o0d8-e?KphRt8 z&Nc<4IS=<)$oa6BIx8BIlx-K&_I|w1qQtO?<>*84B}?k-@mFq=?U^`)AM7UX!wh+3 zrgu~Hp=UT86_V(5k$+A)cNV+r@;gM5htskSyk=YZDV=>)r~9)dLl5THAnivuG3AfQ zqdDHE&yHuCaFelBf5$wntpPrvZ$k z+I?8ZgSIcr#G&qNiO&AYNwVKp!LjFT@z4QzJHT>AJ6ovJ=A8S=W;fk%`~zFP#ZC@q zQQx-DT1v02EL9>!Jc=oWS=YzH7#FrvQOthd)&X1BwL;B!JK0B zuX=LZw>3Pk=4h-wcU1$ZF{1yE>cFV> zzs0hCv{607vPJ#i=XVP9?AzbIGY&0(Y7wU8`sbJVUUW2r6P8fDo?>?PPCln#jZn#t z75fB#2sNHar4q}{K!zW0O*EJE_Cs_f^=&OJ$yRtz1V2%cEP?;Oc&z$SHTn$ylXOXY zstta18f0sI}qKf>p?;G);5e(vI* zZ{h!09{kIQpTAo{{nzHfe}{#CULO1#E&SKz!N1AEk3PxOe_Jj53-aK9#=`$%9{fEP z{vCPn@3!z?p9lY53x7Nh{{0sIg?aEFwD5l~5B|dz{+IIL$30b;_`67q|5xD3HGYbS zUyh$w^WZN< z+5fn|DOdhxi+$W#QkS2mjl|&-(*#{#%j<{{f5s#yt2x zB7PaaxL+#Q{yS{ZkNetk^KkmcIl|N(A|K~i~zuLlo zBoF?(E%tkP*ngizKkhfnwf&nd`f=Y_uKZgq`nOry9|a7=dV}DS=^5hRc)UdQeiMn{ z-Y&GMLV{3Uhz}c(HlLw$od0k?3dbYwM?4=e)KA-Vprlh(5tp)aq?Kt5d_b8lev{t( zj5gWySCf9Wmvx$-ac_^U{a&Z$;C5pBalZ`nGhPW8>K{b@hN4Qfq@Uc!u%h)pKw&9d zHvSukpY<@0+5QI|{I?Uo++&3@O#U|<{NE*hxi`|1Me^HD5wIIl}|2>EPF|_a- zE0sgp)Q|gPZTcr!^v|;BKPzI7pLwLeul_p?_-*=YE&69$^iOx_UqSli9$0jR+5V{x z{f!p=*I4xbgG2vT(r?~>O#SN}`d3@@&$Z}($)W!((ti#%S|n5dj~x0Ru;`x$a?JQU z3p&p>e;g$J_{<@bssA+KxAp%v(qBr)=jzs;flkEFk^@q2?qe+B86^MBH!|3?n}=bozVzg1`cU$gzsIrPu9 zv>)G}nEpQ=9$akwzlHR-`I>dX)IS*bZT-KF^j8xqvXz4#((gL|kB;GsjSiAiz zh=7)X#1;1c_Z|HA5I;?WPW~ef{w>5m(dX0!(|^;3*zNyCj{3ji;Qs^h_Z7cGjmUFzp`*{I>qvO8qDAKdV5B*?-qL^e-g+F<-MTnELT|W^DTRTJ+<+#MJ+k zL;qu>e`Fu}A9v^y2mdDGr!XKZbX)xoPjFY9SG${p;W|?Z5qayZ=t5h3^#N=gjoqD&V*IubuSE{#$R+|CU4lT+-jy{PTuG z|7OxJVo9{n& zk^a8!UrC4l>Xllt^gr5S>VM9m{}a;R*ZBROL;o_1ejJOY{z1d+{x3aC+uxV}3xMC| z|E;8dp%$a&ulp?eYaRLN1T&DfI9r|lY|M`9B-|5i5i}Z8+p#E=K^nV84bKCA;our?Z z;l34S`_~bFX#OI8FfS`{`x#sV{5JpXBm1TQ9tI}n=XUKP{bv7fCjA$&BFdx2=fQA; zgefHS{nZ}gXMN1ZpE-{Jzp|eRC@;E27uZhp{}#BIkMYxiAN_CoZ*L#^wRF!L&iajx z8Ew)3m_>h#^s}E??mp5#h5^daR!p23x~z!!qfAIS z`;XyH(r+~1D<%E|C|o3ye<|?W`mcigR}IhSn6UqzfQ$LL|FAu#Zh*WIG;kYf9rz?* tkHBT}!{bjV_5ku$Uygr0@0WBR)Sm*s^JfTuEC0GXG{s%eKABAa|9^?aMM(ev literal 0 HcmV?d00001 diff --git a/infini-qwen3-moe/build/.objs/infinicore_infer/linux/x86_64/debug/src/tensor/tensor.cpp.o b/infini-qwen3-moe/build/.objs/infinicore_infer/linux/x86_64/debug/src/tensor/tensor.cpp.o new file mode 100755 index 0000000000000000000000000000000000000000..ae276f10071bfbc4d9640b4274dc176880a6a40f GIT binary patch literal 227392 zcmeFa3w)H-@i+b?3xbM*cdDo%-arwPKp-G02&;(%h=eFAxP)YZNNy$@E((ebsB1*g zdclf6wc3_iTdmcK)(Z+MwQ8xgEh??l`m-oh@lr)g^*=Lb&gI$XncWRw-@gC1&qq9G zzcbIA+nhObF3-v3(eg8Lb8;dcKRMp1UXK3Zgp0nLYkrNS*hp_~WKINmN5cO`!9R?j z&f#bX$8h{i2utDLO!zm8u`&qHf`4`J?;^&=ayTBs2^_D0@NAAxgs_t1lQ^t`@Enex z%i&}W&*SiX4ljUk3jCYOzc1u4#^E#$t0A1u@frwg;a?p7&0uUcg!LS6fUuF{%@AJ9 z@fHYMIi7&9jpK75oXhcf5YFfLB@kZ9@dXerf2f$&O> zUj^aS98Yq14TK#WzZSx!9KR02>p8v*!W%e#BZuFC@FtGm4B;&tzZJse9KQ|1?{fTh z2v>0Y4i5i^!#g3oi{p1g_&tu_1L3_K|2~92;P?+AT*>htL3kg>f6U?i9R38tpK|;G z2p{D5&meq=<3ES+VU9ln;V(G;D2J;!dC_bNsguzRL06LHHWSe-GgwIQ~Zn zH*oxQ2;boNn-Ff~_@5yBGsoZJa1(@o;rQDS{*~kJK={8Le;2~PaeOm`?{WNn2)A&2 zD};aN_y-Vv$nk9)egxsi9RCEuPdWY%2tVWa=MZk^_!k_03E@9Ez5~L4ar`R?zvj3H z<_V%+4urWJk3iUk<6R-#gX7&e><;0c9N!DVy*b_k!hJZtFN8fg-iyQiINTq?12}#l zga>i_U=9y~us6rQ1>vC_Ka9gZ5FXC)BRD*g!#ocAa+nX{Q5-)S!ecmoEQH5#ydQ`C zIXoW16F5Eq!V@_@5W~8xX$9@r@AviQ|8U@GXvSg77aKe;dNTa{L_# z|Ci(MLijh1Z-($aj=vA#7LIR)@b4V|0KyMBz74{UIQ}t*pFsF2$N$0MXApkQ@$DRb z!Qqz>{*&W7IQ$ofUqSda$35_nX?)=@m%|8$T_Ehr@jW2y#_{eD?#c1JINY1V9uV%s z@qHoe$?;wg?#J={Av}QN2SRud#}9_^5RUhT@LL=|l*7Xy?8EWHAv}WPM?#p#@xBn| zbNna{kB0CVjvou*aUAc*VSfmZ=lBT_4&eBS5Dw({Ne~uryb!`7jt_!xFvp7_9K!LT z5SDO!7=*()elmw6AUuWRr$Trd$45dq3jUoA|3))52Er)E&w%huj+b&+2H{y8AIo7m zgyT3qp2GeJGW==N@lr8)F zpg907`<2SRBV%7t#)&9nKc9>#O2#KWkN6U+p|Z4nYYAkAN|vnJP?{XK{^^lWQJb;0 zARnb2OnzxhDb)`$AYs~MZRcjz;h^;S8 zzG7jArG4Q#ujk@xDL?pj#n+0PY8Ozi^aV0xkj}dH&AG|w+V+KOy?GaQKxB1JY0dgn z_72F{8cU_@Sz$o!`2b`PIYM4e_HJ)jo3mvFq_(PbDb?x94lK8-o@FXA{R1kiwserH zjl}3JBe-^f$yR}_TT6w?;Oe$i-C(Pr`X@v6FLUdE9@PKxWZO+p|5roXFABB+z*;0! zd32HaIc!bOn^yPy#p=iMur8M#L{+-97z6!AN7Q*6%H6(jmDhGCNLdalb#=$Mn{3m) zcPNZ6P`(F^d^fG+Go?$b9#nPYwzkU+;y34j>OAPnWzwH%6ohF!TblfHY4SI!ZjzvA zzSOWogYJR8|B*SKRcYJiHoCKlbm!joh4*?rul@w9BBLm(7Ap$%DvIiE>lX5tF$eV8 zx{m0*rOS7r)ayWNA@S5|$|8N#=zDYCi-19a%r0l8zPbaIN$W9nPxZr9V%uz1P17qe zq~P?3u3D#;gHF$Lb$V3Qs7Ywk`k|?Um*C&X5-|TgmlVKwfWVlJeuTHL&eQoUORi<8 zyM|PStSfC_?Lf~nS4lLvxjgxo(&QVOJsI8DzHlR$n-@_5R3|9qu}61?ojgC$+ zmnQ$aem*q_v#3)|LPM(|(&h|TbJ+gt3T$87?}AZNMoo3B2UKz|sO0TXC-9FdInl@b z(DIS^!pIU3=)uUPEstuWX=rro%doiw>*{+2wKzqn z^RqnpHkmwCwaGQU0Y_CUlQuvtVqfb)V{zh)EvMP8X{vI|0Tx_uo4_&Owr8u(+P-!g zsE|AQU$GU}=t`4M_WKG5^$gWR2jT?W=p~PNsDwrTcU&L1`vq$t5H4 zXNL;(JIN&lp=b%BOK(CKew_Gh7s*ij@UkA5Lr zmV#9ERL#gRHzgv*xkj!(Z<(6SV?cOQC~vpz)9MA2Zi^7GtbCW^=040-<*!G?(9fXU~NS9^+Y4e+C?y;T%BxNT$;Rj9=avTrOgne z+7#C-?O381bqz%C>WD7p{E=IbLPy1Q-|(AN@WXKYS*=S5PAm9zfwS=txg z1%AWzXyrSiz?o*6au*FxE7}*X@a9!_L|3F8o-|eG*AAM_2(>~Op;qKBFDeh`SX8aheTLgLPojX`AZI9hO$)scg6cZ56s{ zzc8|ZoDEgwN>+-gsNFKvmRLw?sH5|$PkS<|qf$YvuZ~Aln{-|avEQeR@UDKK&te*orQ`An2JUALyUBjIj({kD%5Ji$2)%7Fv1v z9#+0MH_+PVR&^3rpRCbiOODuj*P6R5xM*Y6(SS3}0HxMKR<lhVYEx*A^#r>hI=|>w#73)eRkutLQ z=1OLK3ZE;v`3{a9xUkV296M;ry&3dy6OfUTm^$8(1g!haQ`0pD=$nBPXuNltK?$2RVmosagd~jSr@DAxOv6 zn9j0Zh0$5ZH9J)pTz7mXjrK!ITxu8ygM!WM`q?xxGG_fX?!EmVzbQIQXlfU(?lWk~UxFK`pZ1@kToacRoOh2+= zVYR~jU$hjsF8d;;SUO&TZC_g|&Y;uIfc3u92CxWVO{<{4j)SR^wXC$5KPCyw?{pg0$ob^2) z^?E+KTALQV3aYF|sw3AP-ITNCDrc$G&;H(0%2r73DMKxgb6(sKa=uyRybQ7iPUPeq zxdfX{36_q?1vXgV#$>_#Jiqx)T(W78ZqC_q3wREJy|xmt!2hjp(v;m(-=vIEeP6K^ zPdLKrCM~}uAHk)@)FIEmqE#xd|KXsnmK=S(3dSE^(X@)L5xb(6;{01#+G6b-tA8yu zJbN8WROwCjz@A$Bh*g?Gi0prykq;71#lIh2*rAszp z4RqYBD#A1mvyuK$4M+@}-GsfrDI5bGiDTk_LZ67c{B>b~|_ z+PCH4AiFm;)Ka1-MR>dm)uglKu3Eh7cEVM|!H|2J*gZ^2j@I~=D&D&SMWt(2Q01jq zFr%D-?C3#U>Z2Jmfo4f5zADmtc+yN~ilw)2?%KY6-=0^|p-&j44_pEn2fIMqryX5# zVH=*O;yLS2{d!Z%+iH!@?H}Yg`+5u>OXyJTle7LYl@cyA$EFMluFEpm_fhUC8N z^n9+x-ICw7Z%FRp-;mtrP*HY&yKQMMUAhmd$>XPD###H9o`TRS{2vWwc01n=t*)*+ zLuf{ZQs15%PXex3i=(Q46#@PG$>49UOc`EFmqwq~RtFZJo`&Dhm3#(4tx3T~isQYz z>rsw`6z^g9mO-WjcFym(4o=#sOD^hc6uwTU1L1H(a^a3-+xGT_J965Ju2{Goh7PZ- zf2pcn<<0w6-puf*EQoX3;VFDowZaNmVV=4Mn7oLHlzyR_rF~LEpkAGU4Tvz(}dYrgH;Z6B;bz+&Q+VpWp zlnSL}sXB+9WH+_u6sWYL7l}qs|Jl|(Aiix&7euJ$54&JcVZ*4_)_6;zuBkC^Ms;0% zymt6u7Zfz)9n+duoyeO307M7m)z>vbq;@a(j)qIq+EQ~;BHq~A)WX5Qn&xJ&byjtA zd|+$cCGmd!^ZK5W*VIxQZ{goww56q~C2w?7Z9LD*t82`grYokdaYkKZT~qVumUwj{ zK8d7`iMQ6Y)HNrXTKeVn(+sJkSR(KE{{4=P)dF+>0eLYudGoWjI#Jz^sEL+3;2dax z+Qj_kczuN8%|kkqyq+f=iv8*A$!&&Oh24&WSN91BIssUQ8lnx^LY zXF#=sWG&Z7?B{!%+fTi%e-WdRq0*b?8!_-(>kta;@8#9T=hW52PtlQB zU9D}EF+{7J-&ixNrKyor7&Kyf+sxEU=Yx$INZ1O8@}+wU;*r%1;0&PhO&ln1GO9)!5qB+}zZXh}V+t&#*z?8Hiv^E=X>l0f!acQF(je>Yk5AH$Jne z*D;XiMak%=^mF@O?VFBFMz`XRW4?~Sq2B1W_Jv!IY#Y|TaGRGn0<7vYoA%i|`YCuI zi!M*X0I@W>`IDZ#y!UPbF>vZV72k3>yzaG~kFU4kv3Td5Ej2Ghw|Bnl;e=Iwp!8Jd~5 zV`Q%o-j0z6I;cMXx*5M#;4}S@&gPD)cak_VLQ9dm0BeDEQl{WMAj4{+ALjF$TkedOsC(wr$$@P} z=r`~gM*H>ydS3Miv_dlTJdOGcE#Q)?^jrkth)Drm#6*Ssx_t8gy7W#+2x7)(>N z^Av5OqK#FwF^YDIq76~BlN7C=0{bfPPz4^asu#-ap=inzg7)0~kEwIeb?fLU4-CF& zK{J!s3+j)ZGQa8QMP8}nv!R#gNVIgPWkepE1QZkXr7!5o@UCZjXB|hsdsTIXuL`fv z?F!%Ky_9!x_;ILK9BGEXcwa!Vm{)O4yD8k-~++y zaxQ?+7jMt)3h~c!)q6*rgkNnxlY3~#zVMyi>k;^JaF2C4_wVbi?Vj92W&YHzH|Ka~ zcU_bNAM~r24va!Lqbq#({%}rL_= zqk&f5f3CNT69M8V?|el>!M@+#*S8;&bLe?B)DZ2tIa~3iTXOfhE7yAnONP@8dcMB> zpqzb<=^D*x>Y9Md7z_yn?lrdS*E#TUD<`o#bm$wAp=Mw(^HTE(GM;!E6y{^FFW&wv zSU|IKD)p62eZkD~Kw@TTPOk}+I4rl|8}Q7K@?mm|!%kT54bP3@HVN9P*7lyiz^d5> zQ_+%VVW-Y#sSChfY)BbPVIyoCYRB`@4`D$BJ~xAxOX2glrO^+aQHPcz6WvPNquX+} zyhc17<8bKaBk6kR4!1sGb)-jHs0El6F1Pvge=h>36Wk1Mz>m`jB@RY%*CoSuQLPc2=;ichHAC zQ!bcvs%m7T1(n<}4swReb!P9y`rJNUT58wysPgC~(;Q)A4w(=fMnJm5=;TakY9_vH zTaM)D^-`}?=p9%a@Og>-A{nO1P@wFa%*lVaES`sDs{)QZSpzz+Sz4_6B#&}xyB>Bh zr2}eM+N6K#Ss+vW+f*yso?2W8EovCm3{o&hKGZe9%%dc!o*-z%xz zV3&iNPI+S~y;AYGZ>H;qyU<^Dwn)0Vsjuy9iEMx1`VZ-gC_6nhlV|Ek$ijl_f@uTp z^Zf8@{-x2s=?p<%-vH(hew+TDLUumvI>wrCSA_L!)R>AQ)#8KBNx-Bn1zS$$$)pN< z8-o>OE!|RBSEGcCrBg=L*5{T-&^mUy&cpn38lTgTs41F;qoF-)NKP}mnb@yV`pvGq z>$B>lsW?MXb{a=8Rb!%N_*hb!MB4!vtU(Y|EM^`1esp_c7+eyt-AnWU%qXy?b43v^ zr>{s=z=}D8MzJ!(Jr7&VG6tX={JR-V7?tV zXdrdm+SEuEqZYPe3FLGGn9ucDoRwSx6Rd2{N}ft3(`Om56XmS$@MiREguxD`=O9Se zO3EfJort{KFr}UsdhZ5&H|#toors*1?TN^7AOmbirW27^WK0gsO5ju^k?54(jk|g( za%XDxIg_=V^`!5qRb`j&d3JzjP5)zR)@J{|HJjxr-Yj+Y-&C{h*fX*`IXxoO8Uzmm zhS`F9{vWdi|E*VL-s$9d1{>SYjI+2MMKgLB@%6|OI+br-W$Cv zaN+vvwI;Jjh|LA=nf0dWXNdR6YqY+az`BN)+raaLmPw9-_6}_!>(~Xneq6z!Bu!Dj zf#$ScXILq0P5rmMRKs_+y=?vJol8(|vH#IjfENm&3pl&k&^!HVzYKLFwTnu#2$(U$ zSw833^!6P3JSl-AKB+$*OV_(aFMRzf!R@Piguiu&eXF9G+eKV8oLYY zhN@oBrn)8oCv32X;i!S`=d=Wh4G0`~VWKW6_v_ucBxenaf6+2*2=M3Bmn~cNrBE#Y za-OS6P6?g&GonT_ur_QzaNhHAIQ6A6<|~`$ltYg*X&4IhP>18l^U<$i$i?v+%`gtH z#-=+`F<;PlDJAqHI%R6)5QCsHo|y{cP5QMHvxecw^7XEeoVZG<$)wnUaRkemQgS^~ z87&4~!tdGP($PS}IC*lUv~TTBV;)k`-r#<^c4XYF1?!ntI&qP}wFI}HDDE>o62S<%bY+0;EII!MT)AiH(erubjl0u!P65K zn!y-@Lu0%X*lbl>0x*<0*3g3iE}gYY-~RqA`R6aesYtkg03#@^wfo&4Ul-urW1k&= zW_vCa5{3;}K<|0o<2Z2OY-uT$*_01fr1yks*!0v^WlYHB8_5Y4GXIO{?K!Aap8hj*Y*8$8xQm<2--^6FYJlh=V+UH7fJYBoF1``I4)D-oRiU9opivYiHgnog}&+FTnCbxNF0cfl8Ro=S1SRVN0EBYX>b@TFz&ma`Lo z@|MBayNnn|B<)-&;u^J!W!GGJSBtJAv0CF+d~jS0>)lGfAWL|oba$!*701K#o`0;g z{v7$3xQ?7^K!V>yrn;x$OO0;UvU|g74p~}IHJ_*B??3;)D_?LtR1NubZ)i2v72a?M zU3AVeBztc;!>?~sHILa;1D6me*2;!zo;8lV z)s;G;__YWXRV9$bNxE$;tD$-s8k*B_Pqmujg+tiVthLT|pZyaY=~k50(bHfijV=Vt z@V`HgXhL+aOtmafdLDdeImF*OW`624s45RtW@@M6?KXX;9U(=vDS3*18 z1#7y?UGZwmU^#Thr(DMYZ@XMt6@YYU=`WX5;ahO8gE1QAOXE*Nu~&n2dAby=j<)mk zAp_g^Bz#%n+~olA_5Mo)@vk+|*A8eY*4g^4f~@BvyMZXns`}AxI6jN-^dJ;M>)-|* zRXNjtoWQn?(*`IpY&XeFyAHD)#g8l(z-j;@(Uw- zVAStk=B1cKTYPe@*-F}hl`0jOOr`oQtYeU^9RbC0Yh3Z5B?-|8|4_fpw`oArgTU#1 z&{%ixLJL6oaJ2|-s~m|e&Nn{l@2U9|J3N2*KIhRz}$Ok)3Ifnu@J=8al-sh3}v0@fGXD zncCP6?%T(5u0C_>*UH%DdoQS}dRpUaN;8(CoUk#8`ig6GQujNuTa*va<~J;yMpE3u zq!ufFt58Jw;xxh+Ohp$1c1XPebd zXF~wou4IL8w>Nnx(c3^ttc3$Bq>fkQleJK5;uKZi zfIzK4k5qr5%`IwH{dd|@Ol{D8&8l>|8BIx>&LejFB17&1l#U8ocb+g;bRuNJVksy4@o zEZdP?`&_|qcVD=;5B`_qv*4;moEi}$Y%VJkg~ci9msJ!NsFd+itX)hN!@W-NvJ5h_Y|~Wh7-q4I zRygA$wlU7Goi^4f+ErKI(u-P&anrMZQUO1X@qcyMzI7z@<8|&Lfm+yG=QHR@oHD-U zIK4D$UpnqBfH(Pv)O9+mio_!Xz0LDnRNUaQpF2Y2IaU~~G0T9-xSGntHSh`i6SH8T z4?Sa4!t8~rN0tf*jM!Qg>K{S{8{;%!Xm0HBpTb9@!v(YPA1 z2>iq6Koxj7mrV3>=Jm=stVg%*%U~+s2XOp{PaOXj-M813xuf^lBXa95Qpa^{GX;@Wa&>J4v$3NsL0-- z80z<{?E5>IPL7eLS6X@>T+)E&c>JHH$1QyX)1~}RVfq;N{mbq93o_7GFrCV$>u-jA z{|Tm_hw?Q&w}Ke@0JM?mcweufM-I$E2Bhy`dO5@mJ$JNuzYkyNxt~ifwds#wIvEh1 zez~PjVY;Lr3o^XFA_E<>m2*i&IcaXU+wtdNT7Qp*ANUWSYv3R1yeVU}J(9b?=CccT z2I>pC97;{8o{na^RE{wj=(SAGck^(bO@A@df9KM3UnB->r+b-RKX)?K`Y&$f)+S2D`pfArrUzLHr zAp?DT26}I}Tn!&dXG)kZ>D*+dV;z|`aq4Ft)1^MOoatnHL+Wo8)1~}vV7ipQ?Mxr( z=5MZTpWfY&kf!`hm-086>0#-kuB~RB{Xwd(DV@9pwGH_;GvC9|hbgM-Lh;=j0Y22X zb)4R4h@1Mp#Fl3>(@%8iwTh1ERIonawf!OK%)|WT0bc=({rJqp|M~KgdoD`Dyi8&q zy1Q2MOtX1Oq@m}*`!_NDZ%}5F2H5P#>sVUld#qxfu=+tz<}$qlJCjckIFg?!7HuL0`WSAS=bjXA#?q&KKK!E@C%N~mo&;b%H!*#(k3PZB zy&g#Lyb<_6U5`9=;`@0_FLvqmHh&|Tj%{q-k6HS(4D>}z7t7Cdshw9dz1&UT@nJVH zy~0O7)0Q9ZU&be_{_+6NWBL?i()B~OH%6F8R^- zGF@uJwHeZXCqw#O_Qi~Qo|OL4Og{pdbR9VJs4*GnwM>`B>cvc#Y}LJK=vb#cxQ(xY zco#R1*w-9Ce=EzWb9r*rz=ZA8qbCwPPb%jCrjJA>T@Q}WP|oy|T{`5Uf%Tw11N~a2 z!}YyZ{w&*!Jl}VQ&Vdbx6 zI(1b|k6QXVruTQ#cjg?MnJ&#YyVKw&uVwaQx>UZ>4D?wU=vQZ;uVi|$TRz7RTgP;% zziiGx?@kpXkA3|z&`X&v>G&+BhxJ#~kE=7#S2F!XxBNVR#rM}SUGhUWGd*nV!}q%% zf`UCydcPmjrTQ&p`Xwkt>nA(8JRqTbZZIXA>8w%Iv&7u$CN{2`>4jKG zO?Uju9<&$$>#Z)`@s9^EUCLiM)1~^YXL^_pQ~NXhKG&a%+)d6L((h-wZw!KQ3P(bu zzr=K@k8Wl96kqw&h7Z2q<1i%1^PmA4=;cf=Lm|3;b4OAGp!|BKOa15C4D|b%F4g}_ zOea&X(|7h`wlZC+pB|V=crL*I>GyLdn*0x7deo&m^Q&^EOY-Y8$iFs&{QH?s&8YLo zYgR~liRnY#{PF$-lD0Bks-GT*$6Vhb>9gY;M zISXd-CYB?$^TSLpa%DLCL9a1gs)tWAyni5dP|uU>#9*dNc4iXOrTitBF6Hm04DUb8 zbSZzYF+D7Q*bXCD4{n5bYCGgEw)V0QIV-Y0j9~gnZaX;pvs0K(+0kw7teY%gdRRU& ziz}Ee+0!SOF6rY&rdPW8%T)sc%HP3sslWEg!xueoKK@VVKUWRN`2L7IB*^okDS1Q| z-(SG=DadBzTYJ2M>Eya;x^uqZ38qVPf{jdPtP#ndFmW&x2gc^fO)hWwx&FXL^{v=D|BJF}=}! zU-`haGk|&6mLZK^v`87Q)0jp9)6=CfiFqXZmSFmLH=ns`GK+bj;JA7(mL zyQZIG>*+P7hqX7BLF?^l5ckI4D=mLm+VU)+I*AO`$jNb>NiuE9;UC@&I_0>+3yuhm-6>S z2KgH^$lsAcejhSyo;TKQ12$fm{}D`={MjiP=nI%G+1C|JAMU2_?43Qq^dgt;_(2<) zF4g}Iri=EUHl5@%93wK&r!ZX{|Cuh0&nuW-A_(tu|bjb!xVLI9wlfGkP7BGFLOV3rCxF~-G)1~}9!E~|zGyOuJ zeC2?l{2fe}`fndvOqBP4M`WN+$v|I_fxd$2gWdekcI+S1sTjI_oW0kLOqcrCjtulZ z)Uav%AErzG)fA>ne)H5vR#-u-j=`(zE)j{z6DNH}vrGtFEAGm<&QvX`P^h4bD)z&`wWlu0& zvM(E%PQ}#uueSQLgXvQL>N5Znmd`+qU^*=;Y5D5-sd_Qw{|86V=lJOB5E7h89swR? z@QHY*3ixRPK2pF>7jT@B2cL)+4d4jmICyt{tq8(gC;#O4FfHPp8IUuU_$^L-!0?v^ z_~#P8#Wnx;2_KQ8)8i!u?`6h^u}rdY5dV;I8WSMSUBEjWQpEgU%KTa$)&Sz)Wqb^W zZH)gp1;?KeZ)_m{cLI{*+4gc7B3`-slO>{@@d92U;1dJ*PXWmR`D8;uau*{V$croK-6C51HM!a)18~!v;#G@T<^IJzE z-gyE)?Y?U^6GprX1pcW4exZQJ1bmu+R||NJfX4-VrhwB4VN*^WiFmUG{ssYW5^y?2 zsKuHv;Tz;FksPD$=*pa`HIW?DEjBh_@)fkA#T#?Erog$RRs$ zm=fgHNW@zl;8${%2>4Y3o)qwF1pHb7zfQoH3HXfy{v82_KR6xoM7&!C{5Aoz%xGf<%s z?=gY@mjeEzfIlVRPYd`n0=`DTpA+yG1pHS5{%Zk$S-^iI;J+1c_zUeJPsGE~Huyxm zKMMHk0{*6e|4G2#67Wp|{BjE1}_+|lrU%MO~5}AaP(?|PsIC&fPXIF zUkLb@0@RJ36gn*wa;HL@rC;=ZW;86iTQ^3mv{44=47x3`{ULoKU z1zdk&n5qxw2>g=;{CokQBH$Ma_%s2ZF5oo+UMt`;1bmi&UnJnO1-wDPn*_XBz*_{o zRlpMhK1aak3ivz$zeK<<6>$AUYN}pcCh#v3@GAtoUBH(J_>}^Fm4M?{hruV}T_fPv z3ix#bzD&Sx6!7l|_{{=-tAO7o;NKPS+Xeg%0l!ng?-uay3HZGNj$fw+pNO|o!0!|A z9}D7oA z{MQ2hl7PP=;OhnaRRMoZz<)2`e-!Z71^f*G-zeaJ7Vu31{BjE1}_+|lrU%MO~5}E@J|H%9|HcFfNvM@F9iHc0pB6u{}S-81RRz?Lmrq{3V5!7M+Cf!fbSvT z-35Fv0pDA|dkFZx0^Uo&_ZRR31sr$%f=|TjE#QX=_+bKmxPTui;C%)BC;>lOz>gL1 zegfWKz)ukH69xPv0WTErK>}VZ;6nv`n1G)w;3EY5Q~@6;;HL}t7y&;+z)J1TvrU;&POPIxSUcj=DMv;ba2cy$0jm+QK5Q?976qkC@b?M-D1hhehtS6Xynyge0{E4Le;UC5LU>MKUvCT@M$8T1KP0?M06*ja6x}U= zM+v9DPvT4a4#M{f@UJI){{a3L;k^U+KMDU<06*?POzzMCekS3E1@PH~9}&QRNcfQf z{9VHH0(k#}Fx|cZd>Y|L2k;*fJ|KXjl!(_vUIfP72j;)-PZN)TCl&TYKvLKf@twl6 z91*&u*8bM9h)08={%OL9M}w02ts@bS1_Axkgb|O7wE3+g5s!?e{%OL9N5;(j){%%u zMo9lOVZU~>INu3NOFSyCo)C6p#Q1xp#Q1xp#Q1xp#Q1xp#Q1xp#Q1xp#Q1xp#Q1xBQXd( z5ic)*ArSHU3V6PNA0^;H{}ve|>bAeYI}!B-MhJdf07D?6Zm|f^B3}OhrudH+@Dl`l zfPkMU-~$ExBmplF@InDE67WF+K3Kqu1$>Br4;65IO4v`1c>0uZDy~lnr{X930saN& zPZ#jf0zO8-qXK@0fS)Pgr2<|i;AaW=SOG5=@Noh@Uce^^ zc!hwUE#MOcyi&j?33!!&pCjPs3ixCJKTp8V7x3Ub6d5Dl6oG%LfL|!!F#(?@;MD>? zUBGJuyjH;D0zN~)XA1Z%0k0GAiv)bOfY%FngMc>*c$0uP3;4wX-Xh?w0-g}?HUXa_ z;By6ho`BC6@Jj?dxXyr#5ihvTpzwu)oXZ6Kasgi?;8zIvw*|aiz!wX6a6JSWBVKSl zMB!Hna;_Heq<~)|;2i>ft$;5T@aqKpdI4W1;5P{PjRO810l!JWZx--d1pHP3UoPOc z3HWyf{B{9fA>eli`2Ps_odSNBfZr|P-xKhA1pHnB53aKzW5oM`!2d%5Un$@}67c&3 z{Ko=*zkvTlz<(;>4+!{!0{$}re@MW8F5nLf_#*=T3ju#rz*huTfd5j!pAhgT z1^g)ie_Fs-3-~hv{;Ytn5%9GF{+xh6FW@f-_&Ndqm4Lq};J+5|mjwJ}0e?lnec|3bjO6!3ov_znU8mw1<2MmL-r(mk{$s#-OfEQI&X&@g7s5n`1fP};e?-SaA*#CDhQu!a1cvRE#c=GT=QQ}`1t|+ zR>ChZxL%oh9q_)n_WWUYmj4gpxA$Xci?)An%>Q*pj>b z!_Or-!F=9E_=QIPfh_DH!ea*4_48}Orv>nL39mMIKFj%v@EZ+&9OH+53(H$$a6R)G zMEFdD7cu`CfQOZ9I`P*T{v(-xF5$BceiY-66MnP72QvPNz(3?r%x8n)*YeL7@XG+- zAI_=j5oQ6)e-7}l{QsKde8`!daICk1JybqUBYz|FYdg~i@O#&i$fN2Q?CeID?9@0mcl2Z5nR{*fvC0|3wWnhn3U z=d%U;KEhiJKd1*iFADhkfMdEBak{!bzaah>0&@JoywozMQq zqWmCUPq@AJOtBRp>!w=nxp6P@yGWcl5zfbt<2A|FN+k}7H;3qP^2Q(b!e+)Z@ z`pf@Q8Gnftd?xc(F+LdlP<%ivJu?}gPz+N|1RU&68#SFeD5ujALxek{9Ta$xq#;k zQ0eXk<3B!H{xHTb;R?8cQ>p|!th|p3_(v?KM;?jy@b4buGaMQ^OwJ?$zd*phE8zIY z7{cTqFW@H&c%y*d0eHT*$<)7|q5Yolw*&Y-FyYPj-U;AQ!tH%udSRf1iSI{{DdHJNGw&S*GV~!ml#r()RN< zLCy~Z{tpOG8aXiCpl3K3qI@qnZ#|drZH9k;=I04>mXqbM!n)c#*+1 z|BHkVF*vw+^t?y-P=o9G83Brv?*;wv`GgNM{94W$!cR81uKymVqMQ*1zl75rPxz?@ z*Zj8=Ztr>1{C^=l=zn%O4dn;@%%ce(W#k{h>CPj3w85t^{ygBlIPM1Pi9a|F9y~If zf2x5081Q^=tI>xdPUN42+xzIM7>|uY{xeNJb^HF9@Zfl~ig0^RowiFG3I8OA#d=3_ zWUtdveh@Dq{4A3$m=}63BD~z-=QDmg;lXipHR0n8KXh|?J|sLi4)%nJK)zRD_;r7; zCw!v8wcop#@JR;O^|PMv9E0ofekI@o!0_gK=NNun-U|u0_vY#HUQc*%+>m+s@9~3br>9#P@(0JMDB*pL{JzY;m~eaVpSDZS65ee1wcXf4c#FXgVmXCpVY>F- zKaIB%9vnyRBRpZ`fLMAyB|K>VyNyNpa}2-k?}da1?f)3U=LYy23AgtL9?$7MM))O$ zU+4cz!Y?(ruAk}PkmP#{41NL2xtH+E3=Zl_&wGR~GPssgH4f!mVQ?+ycEZ~YuI0Q# z_*Dki`8;+!%DLL$x_##nevQG4IX|}(ex1QbvtC6epqym}*Y$G%;X!*ll<=VatP|wS zC;UbuUylp-6K?Mh1h557hj7fnyK*EJhAYPv}`h_-zK)_Td4-gLdyPga`Zk zM}&XZ$kFK@4Gw?4ce}xLJ|_@v?;F(o-zMDNH>mTuo$$Mj9G%bOCZhbH-Mg0X?-~9Z zxxBw8yvE>K{-Kp9XQjclJt-ypK7;FYn+dn~4QhLJ2jM}x^a|nj9zv~G11E*s!(zb0 z`frTzU_VU~9_**D5pM4}gly1rY8A>4_QyqpFEQmkl=&Yc{7QrCcKj#d4;g$u^WS+6 z%D=+kx}W|p;r3oc&A;?qSAi-g;ICXZ$OOTwQq^1+PIQwj~8?*;A2rG&3B{968x z2!GDtTK?;VzhH1J|G23rKWHDC3ICPhAI153jPQF6uE&c{0q^VC`@vw^OV5A{QGSqr zCgDFYa!z9Y?-6eAz3k2SD})E_&u4_cWaNMv(=#fD>Aqs{6L_3@i176W2eUv=)imS} z+J~nJxA&m#uh_jW2!Ghf(fH76loQmG7~!uPesJ69SqgZV9^OR!uNi(V|F4AGdpH|e z&Zy~_?jH@mmfuA9>ju|yo+A7WgKIf`YEVwF-Q$FBH2m7`-354G&)zSm+x;El|FhvA z&-skhqWoaHBnaPR_>W@#*9d>d;G-BnCysJ98$6%!X9$1a;Mxz_K=?L;Kg9flW}uvp z4X)+fLAbqdQ{x{K{;uIajpbZ06Xm>Za6O-TfN*<1r?#sn&O-jr41W>JxrcCjKd09J z-vAz_C!2}?3nNFjZndw;3+ zyZT>@a_s%3HOzlG;r9MgjsKqTpgxqepqya4-%hx_@APbz-@O(2gY7<#aC;A`ZpXI> zxA&mx@w6g=a)Rx68R5Zt`vc+j-c((0d$yq*dp|0e3wp*9ew5J@ouA7HxA&vwGyfXG zk2U--EYfq(oN&8wINgmf6CRXbM!3CabugFf2Ey$%lW=?QtCs&c;r9O3fy_T>9_A;=KY{RIxh^F< zSgzX%xA(Ye`EL*&Ebmu@+xuNL|ET$xAA7&+p`6coga`BgFyX;;KOsCQf5;`_@*4o} z>z!fRt2d|nIPu$iQ}y_}*QF@O-kW+7^Pf++y*G6<;|~*V?@j#<|y^Ca<)H2ExK{w@nq&gll%`58;Ny?^y&=3ha$y?<5r>%S9j?_WKH`6pb4 z@`LH#0(f7~-cPIZ`3m4+sp(u6E(9!| zCd#38>YC~r>#*9~0^*Cc@fp=^^$DuySaVZ-UCsO`vJHvFnp+d~!{YPeHEoHyrpBT< zh2?`<;}@r)HdnXE)L}r4wKb-dP;BUsfl2Yk)~1#rbK`Y0XCi%c$5tkW%!$_|np(;lDyz#HqS0tE=>cTFu%Nl6t}zj-txi;zjZH-7R7{G6ya3Rk zSXn*1T_24$37&ZwmA2`zvRH!hKCq^_8SApJK$T@oytQU@aWr#5AYX+hUnrq6Q3%S|I;*-R zUK?vpw3N9b%bhj_wsPn(x+ROE(J{r*IV_?;5a9ColNnkG)muF?zEer)mu2zX#ycgQ z-<5?I7L1EGG_}mHXlklYp8~WAX@XGG*@_Ws+6suOqzYr9E#mN5SzrvnXcZO~7EP~i z1!V-i0|Qf511eDnHmoI*Knc&S+_Bh$Uv#wdzD1-onlmT2*_I zDF6<0v6*evEwyEp31#$1vz#)9vDg~wtWsgK4T?dljjgdg1#PYs33_)9S>qV!X))S1?BK38PsgQeazkrrz*bnM+yfu#^=VWvCo2c(Bi0d zIv!mM4i7YHV^d2*b$zTZ5pN+`6=NHMqok`%8D6iqZ*x`nELeuO$BM5TRfg)MF4f7d zuZ_QB05lX*F)0%@QK^!VunhwqYEyG-;gDDv=tL}jaa(nLYiv$+ec9LsWinJ#*cVdO zQP-=k>dc(F1t8PzIA8^pL3Y7ZMb844tug^Ve`T}Me{DJw)icW`#iC=HCromxT~pC+ zYu_UMU;%la0Cr0l*R50587ZA@Ofy2M`W{EDyME_g=ODcm6JK{T z*clO7`@2rcyD|=0{og$rrscu<+b~*+gJroUSQi9U2EFHXunA|JQ3-zQq*14rM`IqW zS=2P(Qdr~Mh6=qJm9lySR2LVKA2b6%$gmWQC(M$ z429Eu9Ou(Y+%UBD4b`*FOds4;XrNf4E;=U)0}=eh5`aFZ<3yIRS_OwR&{CD^a0VUR zT6al21uHD5jnAp8iBEz>&=gXMqLkGt+||#~MV(%C9CPl#miUZ;uxvhWUJQn`)}}^S z0!_@1%_&f&OIb4-U5FZ>{SQslv+PhJFWZhO#0*z8&aH#B_=)k>wuU&S;LCqur}K~1 z>xw9@MP=nG^ATWGG?XwDm)c;VH{_kcjkR?RjN2(Pm&Z0+aV@rB%6hM=J+*2x7?){C z4h(a!E(N<+d^IH?~h8QSoQwwcM=tZ6(YKKE5O!EPEPUIFUy+ImNWJGjsexRCa z*Vdz{wMm3KJ9IN2b@C_kupIZrAvLfCm5A?zu7F`u?cy7wT-njlg}$ZB+WLw)^p4Av zeFi8sTEjH;5$rS0Vi>J(O00GBqLh{3mWZZUCDlt>A=Ot{tm^OFy2PvrEnqCOs7Q8p zf+@plC>x?6t$2ETW?due{b1PCsV#F2ezpr)(FR-^DZ&51`janSnJ9t%FIXLcxynw7 zjH2Bx)(dEQaUCWvg~hdXGiJnF;vhOkrlbK)n2Z|LznqeDwGYM}sQ^(M z4koI7uBPVEq|J7jqNT1G76?IeDiecia45rAeJf0KQnv!Y5ls{hQX2+MV8-Ir4P`Y@ z2xu)HUn`sHTcc=PKo+!Aeb|OVm=qY(Vd4jFE-5eysW@X|WmBb^-MM{XSWCP$p1?gX zD6cLO{7keo%?~M)-&z+I*0qjqYE0D4Y-(czY*m=6OY3$>Ejt6JeT|#IQ9fwvlul>0 zHb8RT1A9y>X&F^4WC0Wu4EFXaU~RM*7SyBDL4!wOonn3Z^$7i{(~D{IM`=uD!mQhu zxSK(`frEMgQ&89vf}uHOlGPZ89E22x)E(^d4?k+jLeel{2gnc3U|^g!{^gZe&J76e7ea=gu|yAkd0kQyV?lyPaB( z%s!0OYb;M3i(Q8bbP}4wZal;JkeYaFYxW!1xW-NE+|96!Po0F5Npa5@bv-kkAJ_9^ zEX0kLa;kmwv@VsH%M_;^64{wasIK;%ecdB^2R&%4}YA z&QDj9F}<>sqH>+c15MjZ9(J}US_cB{n0-#PEdmj4eY*mpFv)dhplSGglTn$^$+noL z&KEcQkm+scBWB|_rzr?nIVfkLv&5hml3`^wr3S#MGuvcH0`uC|&{hw#pp4E*r~?2_ zL16(1NYpjg*Tq}$SQR)Y;G~>Reu|pkYr{}DSWOttfTIq0%qB{2TVoGa-Z|4k#v8uM?-uusAwJJ?}G<$gtxV*eUUZhVq45 z5>PcY&GS(VmLu#9tzAI(_dT6%dJ72Eu=6HFy=+W~+XHm!Sb?U;>CwLgma;1Gpc;~l zd(AF;$Wn-TkIw4BhDMZJCT^IuJ9J)HF5EM;%P<9Izr8ttcE^M&|~bnoHrZ zXMJ2PKtkTa6;UUXwrW!~u`&_ASRZ)-*V2~TPM7oyNKNV{#;aRe@Bk{#k9c0EISegz z)?^6IOE_!bY6#AmcIp@;chi)K5l^;o=GoLn2T93|r)AV)d)P9D3#Jug!%kkmmaBOvtOG!2G7hu<$Q*2>!ki1XNcqf!#3v@KyQ8k~;@bH1 zwwa*ka+;#7s4)X$Xg#-beq#+R{x!mSp$A8Jz#4En2!#;^7snf#6Z5s_UpSZ-0cOaP zM4TU{FGFEWO(d~@!GYPP*4TvRcw<9TEiKf7x(`#S)HT65MBqtTv5z*gZ&%q=Jhuhb zR&bG@`ZVxnlfmMe`lePrd(7??ZYzHen33Hg8=LkM-hqQK+7|fX;UIa0R?DE;cish_$uG)fpPyKed%*eR5q`9HIPY&GaBT zjHONDV^et4PYHBt7$yv4M)VRUv{3^rn8MOdWvn$`4@Wg)aCRL>PCJH6DMGehw9kw0MFXq?lZQGAmt!YFInl(U^^}TZRU)PO znE;E(F0tdk@5IU_{e4F zQkOToh6Yl=QG#~6AP2Bh{cql714}dn=1aBkNrhTKwpGB#Eu2Z09)n1SV0n?g^nis7 z@QqZvjdz$?;R=1s@Hv=ew)zsbFH#X==+yF?T7by#cIX*-tE@DH#h}B0dV-IQ+cqKT zO^RjMV!(M13|%;O1tsp3z`?FS_*9Pe!(!EK^U5mfA&<1LHOTC1WvYabttgxq=;Ioh zUJogO`?~5gc3EoY_aNAhZVACB6GV3Y^CWG;X_YAJ<7JKu*ZCxk`xiVL zXa|T8nt3maodeBlkJ0kn1S+XmpZ4X!S}%}-_XidgzbwF-v`vo3kLiqnb4c3C)m4Em zw}0`orXF5v;4_*?2353DwIcC4#?G~O=ez_MufY(TWw}nzM2cZth3}ogTbQAof1LTj zs!l`GoOl_othkk?6*(H6fUePy7C5B_-*&^h>Kp{t>R>txF##-$a*TI&rIJ6?kh0=# zIPOArTrp{zODztYt&h;A3_LS=dTzJbpJJH5OHOH1SMJ9PNO$)1ww{4#`pEdWeLh4al~=QC9VnW8VQ z@N;h%e&IX?&M`{3**jZRv=m_EsR7fH!BnLU1#UGwP8=4k7~coCtLYa zrk5ESc|-b*q1Z@p$|(V7a@7~>YNkN-8IpYvnr2?CrJ_F8#`Q=Og{1&v;zDHdAOcC2 z1Xnk9&Ofj4Ep0&JzS*u`m`JB2;K+7fNmAQIyU3>Q)3$jyx76mL+-cifzmavD^9$+Q zTuV)D^ZM{MXZBre^ANuk)=1nXlGHZW3mn|$o%9_F{ok);^<^npB#qP1l_3@G?j*R1N+SiE&cLd||)5(9S!6vOwcZ7?@r z-vef9Tpj*R((K3jl%u`wWGc8|!kX;NNdi3Dp9CunD=ajd<1D7Gionhad`d76E`x!S z{@P5#WY?SpS6@Y_v*7CJCQxC}@tYWZiwj&%0`sp1P$68ggxYeKL!BIhfb5n{9B83xK?`B%LLAWl?-o(&HCM` zz}O7$hgceChD=VDI?+mBD)Q-87(Av|;Ab_1opqBG|33RpHV8IA?8MpMRrN$8trr}% z5|p2m{Z!ZbCp$G^qJ&b{S9#KBR#2T5NHGYhoM09-r2VYSFKbRHx;sVFg3yGSxi3t9 zv-zqXCnf%NU};%4gF9FA%gL@ivzS#54L9xC$p>2`e*T}Ggu$(ItDC`~(L7KtZuXsE z=Zl&Tdw7{8gY`e}iJJrPMoZl)-3k6LjlcOXx(s>sPX`cNmR>^{%HUyS3g! zoCTbgbiUr<9b79h_1?+pO3_YDS2~$U$ef6ks?)9l?!vOp>GK_hpr=N9L zR}t(GIN<(HX8eEj{*CjG<*=40t=En#nZoX9a0e9J{F?eQ!BwsRId*TDmr`IG6AiNm zXG7J8;7-0$@073}P2fz6;8&UHmf#w}>SSkKqrh;Nm1l(Opu3z>cAg2?C&#zxtj-(t z?A(jEAF241=a-cB96*4t+jSGZ91`n zVJq2`)ODc#LrdGaHia^(!++>11?PY;bo%Fj?iaSW8jc%7S$;1Iw+gW%9e#!(t+?&=U`hskQCsM7;vtSsB=IX-}nsjum%46LdVHrn7=4@` zgOfEivvyK2IRbEDGap(esV#9*2~Gn(xwa`ds+BUr!Frk<__3!MAJyl)hxna#y*<jN^&32OZmF;(k;Um@{|{@URR#FIn3BdE;{~>rs$;FMbVcj$qL!v7`cXuv8cXyD-8rRQp6hUu|YSrtW7j3DV=jr3QFe+$TzPp z0`)G4?vzlp`?NI$*Iw}TDRK#+cN#O!CG33B!p!q77cEREZr!F74RX6nwz_B+Y&CoK zWUF^26QoZ=X4HYZOUBFDT^tMLWwYK%3s$s?77XlOcGr8t{&M<30=6va%ph*R>T61& z*}!sZ)l*lk$KkT|tW(Y6Qc+S3lZN`U4tPq@94-2WFEw>`)9ikfTEo_-ZFbLvfl$Ax zH;RUpeYaLFe3?9^QEBGxDFmLpgA0K5T=^Sr8?&O7wZdgqb8}W@5KNY(V;SgFK|koM zLs9{+WIoZ0?U5BDxI^RAz3^(CNA*wrZDS{$B}!EuVqg$8OR+q3!T_nJ@6++`8R56M zc1(45uy~~v5`YSW%?)ROr124ML9(OWV7vGMzLyd+w*?o^Pz=;L90eg`3bG6X`M5Z& zSXyX0^>*ly(?Ky#Iym3ztP2n3fvW;^NR0^k#EI(+;J%#tTNrqp2NwZh{Q%cYTJaC; zw9p57P|4~T57<0dB2+i34WTnDW5K1RzPlT>UwxhhU+!2V@8&qg1Iu!3Npr8UURB$T zMbIYzWL@6nZ0jUYHF)&%ojJ`ifh(}8@t=@2iRbf5ossVAg4hlH9L1OxXW+>!0 zo8G^#HhrrXx{PWK1C0q>aSLbu;XXdRD=@3^FmnsSVy9ktbKD17ty1oTtfpp%hq%qR z3}Cr3-=VJi6^>9@u$#4mtV*465alb^O(u4$)ZLs{uy`qM7tREBPmakt?WC3grO4@(%gjPns(fOs)x-Y{IeBwJCmMyz_8=@>0nHwlU`Z> zVJKXhR~|~V0{24Y?V>k>NFM)gnJgWe$l4kQE3NsP)4qnCtr37aS*1=)u z=24Oc*Ikebi={!Sa_`MjhtiCVtmxlh654*C8C~ zmjrD#ed7IXC?(-?Wt~iCdY@bPPB(z5Ge(&+9x7wUFRm5l2|5>_D{)IO7$#+GBZTVfwiEZ#*5);^)iI1*}#H{!=91l1O>Mdm$RUgj?X0I`@n>Gu+ z*sqPvj?d@wjh==YNR0YpM=2mw{A>@N9{4|aj)#9=bn(0$U#n+K4#6Hk-q8_@Pj&DE z9lXK8k8$wD4t}hIU+>`iJNU~E{w)XplY{qj@J}7Qzk_?ZHvh*vc)o+5;NWKqc$0$< zaQLrta97Sx9sDSV|CbJaqJwX8a5tY@9Nev+uN>T!vnR;J2iwc#-`ByrI_aM6;BGxn z6Yw_1G5JH_UkUufXF&=+Ucp|wBgccd)~o9s+)eiZ2S3i?U+v&-yT9b%uAH|W+|`E< z9NewvJ+Q*y!Qdp|(fK^c!3!Mxcn2?Z@Zk>b@|Qb!k;8wXgS+~(#KB!Tw>r2h=SL3i z>gN*t}?6yZqxF+|AE42Y2(+?%;0! zTIS$xzxcj`yY2phLdf!()= z|G_N3x4?glgS+i=664su!=3gT?(m=N;A0#-;^0dh{1gX&*1>x__^S?nu!Fzj;5{9D zn}fS{yIXfvZ&*LBeLld!-F|w!gS&R%JOQ8U;HNqHxy8X;GQ@e~-OvJ-hmx z@8Iq@RpQ|8I5^qC&vf!X+rdj6{6+^abMV_3*X6C^{NL^HyXAe*!OwE!tWA+~F3Wkz z;dkY{>EMSraz1i!S3kS$ZR>5U!+)HEmpk}y2OsC))eb)1!Dlg^S`W<*znlNd9ejc# z=Q}BKYPcS5clcd7KXLGb9666VxNC>kJ9ve|zsbSRcJS^z(L%fD;s-dm+h2M+xZ7XO zaPSe1{Bs=K?WZ#w+}BSXyvmVtql3Hs;$a7O`JZrbSAYKO;4c5W4&K{IxBEU;uU!6K z4t~)8*WKH|Sv{uz<0l(Jo5({HM#Y9?rly)mh8iZNMw29QW}2CrVV=yKGGQ6Cv^LCo z%!(a0wCgFcgqB%rS7^z@5@KU5A%xi0!|wmOKA-D8=Y8&T-F?no?eFz}eLpWc=e*DL zzTVe;-Ou;uoHJlQ2=J+Z9|`y_fKLM)?JEKA3ig))-Usj-0FMKHJK*JjuK^s(^*-Q( z!Tz9bsXR0To)0*#503+UU$CD9IF45{0LSqy0eAxZ%mp0h@tXn1dHgQGaUOXVaGZZP zIq-dUP2~aiH3fj9pHl&EfcWbHp9T1NfX@c}O2E$r{3gJ$-tPeX9I$^9aNMUn4>;<+ z19%Jg`516)hu;8>es<}e${UUYxqxH66akL$PX!$PGyq-=ah?Y_j$;=Aj^o(#fMYv< z6L4(jp8$^S{0G3Xo$tC^Di7Gs_X8Z;`Qd00mt@S1UR?=bAx zrC@(PltSI34*0?K!)Jiw zI^k=}zu8Cs0QTsobB{0p6Hl$ITe=Ev?<;#)_GTa5%VFQgvN!waLml>cmi;n`XSl<@ z(6TrC=wgTc1k2v+qfd3%pJ~~fee?{6eKX)#FFzHW>-R{g-+5q<`{-)`$9?pzfd3Ty z{0?wzw@(6&<9PCKz|nuVJ;HXv?E(Gt1suocfq>)qJQ#2spDzM@3Y6=Yfa5sa z4mgg(_XCdO@E-ujarh;`v0uLd_|6c|=YpI5V%pC)7XGxf+aJIW?#Fl9GnG%AkM;x{ z=c9uF$N6Fe;5c6t1CI537T`F}-{im_0UYOx*8oR9UjdHoaJ#)yiT-+wl}w_|9xV*D>G+;~5J$>Xrgt33-_Az*_*H2ln#? zH*uQv!?l9j>xY}cPXYM38*rSbmRkPL4T~Fl4D8X*A1(UZSd zZ`KbVI_y8Q?B_^4-#F}lu`nXZyszGVb`jj<$F$EL4*PvA zd(%GqIP4Fx>`nX3bJ!PH_NILnI_ysZ9P8y&!R_`r9qh4v)&h?0vjy-gpk2)u+~m!) zt7`?f+ttnB2iwElfMfhiE&rxnJqGsZ=Z}{C)6%Y`l8`?Xds5WpCQmhYtJC zEPKv4jgzMxf1LL0^S+$A%b&xuLAsNu)iAc z(SnehlE%=1qH$2{BxIObs$;FyOa_Di)l^m7K_=;uPf z(a)`b?*QexLvTBv_kum<^C7|Qd@cuj%;$69AM^Pt;F!-f;0N>hDd3ogz4uS$2lG$} zIObs{;PW95O94L~@TUbgi1UK=UdG-#6{XLewnP(qz*gtOB?=SKE$zlJTWpC!$ zmmT(RSoUU~eb-^X&ayZ2>;{MZ4}fF6?08@-xa@g$cfoo7!tts%;5g3?0sNQHu8s%) zI3A7%dmJC8fIW_@6=09?Gy#tNyU4}P$}oB%kshw*}MN4k5{|E8Z#27B~# zhGpN6a>Q4a!@kzCH~lo>us`3jH~sWdhy4|nz3HddIP7n->`gzt&0)XTvN!#7iNpR8 z%ii?UCmr@H0mpiIO>n!Pz6 zKMU}i0G}oJcBISgiq|N7T_CvKZs!WlabkPB32+=Y+66cM@BLBWKDHR_(N729n1>aB zWB+|0@L&BXm4~kYzZvjv1vl~BPw(;7sbBcW<;C%EC&BIV?hiPY_aMQIe^cJW!5+(d zG~f#%&XWP30(iaUf1|{Io&&!c{7eQvw*ijl#t#Bs0rsx|j_cSD0Z0G;1|0qGa!{D( z&h&u(4+H!b&^-ol^nWto=)VGR^nWSfw}Ss00Z0G81swf94mkRM3-Diq|IYzO|D6s_ zULg{}+JY2L68yIQoA8aPjgLc zd@FxXD8<7lf`y!Tt`gUkUhifUgD|%k@6scY^&+{X-p=L;L*!NBctn$MOyY z9Lsx#;9M@&F!R)OuwPGa7`z7TKLxxIaP)Jo;5=@B2KE<$|IY!RKz4kwkMAfbqkBkz zhB$x6%HRUWI?)sUew?}F<9or^%(!gEzK<9L{y{ez`)|R16V(x4?3@1|A?^Nq3vZM0 zB`6Tse2)cS&`1gYMw(uW|pBxL{ zP4IpezNg^*Eqour2Uz%if}8$i@~~3!&^HV>`2J%j=od!`&i1@6GWA<5ILq+f*WeQY z$8lzFZa8$YJ@$*^12hxgMR0D3N9b( zNaALH)KlI!elC&Mro0BfQt%wheu2EkJaGKK5nP6ur0#u!vkhOy|D)l{80L*_o)$b` z9*zBr^4iq9!QT{ogk}Fvd0lAX8|3w93;#}Dn|d(*caZiv&ayZAFU)6GvN7%9FtO)) z?hWk+?fIj=#@^(g?GI4w8RtEskNq@H`zo*>PI`Q??pc7l;_Mq(hxRpKZ_3MlY5_<8 z7XglbE(H8o@P8NJGXcl(1a*H4_NHBNJiHe%`NZ+(KEQDt!*Kw0dCz40G($W~0LO9W z0l=>Z`y&8f1UP<9?Dv2#1^f2^N8L{W$MG$g!0A2C!-L>w7Z{(>zB}NqJY%_rfbQ?X zKh`7aV!b>9_E;~=0LOYM1RU$-QNXcY4h0QrWSTBzQj`f1N+0@H& z@Uu1bg5|<`c>?@ny`XMisK*Ylmt{oIex3px``yzHocA=QznFP}+u;hpF&?bPNf6Jo zU|$0`#-9Kj+cS+XLdb-uoNa%k9dbeZB@b zwl~DFeldPm{i6TZ9sUu=ez6+t5q|@4Y!7b%j`nXla6SVt{mZP!xPPq$+^pvre;4o* zpkCes+??ODJ>s*$9{bn%fMYwsaRa{}It=WKLH8uU{{i?JfS(9B=I4FDv3-63_<`W( zL%^|}tOFeFKLQ->KL(ulEqrm_dI8=mKtp^#z_DI1PO}eSKd$liQ}83VRRdjIx8OK{ z_%RRp!_jR*h#m-2?^^S*gA>u0;4UJmX5bI99Og4_M{2C&C|x=3)lpWX*J_S1*K z5BAe%0LOm%1>o3E_l5q5{j>;hyzlV8$lLJ$qj|&oDVVn}puD&q`x0@NfScYyyAa9p?Gxy?qf$93t~fV!H4^Y2!2U$Q(GTuB{tNcFe&fZ3$v@_&6X4n8=g;6D^T22I#y{p^d%!UdKLR|P zJnRVem*NE)508`lzZoA-b>M>qC)Y8) zH(~tax^6eXabI&Ndx4+6V81uu*dO!maX<&Z_v@1M^!FGbWhkzf{<=+@I{!#ZZ zz)?3R3th}J-y<{rF`tJ6j^i8R*xwPybBTfA2j{gT07rYw55~jy?6f!s1CH&J`>tj` z1aK_ZP{7g8Fu*aM0tb%c58Cr@LYjEc|51RW{RqG@&Z7ZGKO+H0KgT%mV*$r=(NTbp zgZBJi8Wc_Z6@a_yYc%-5{$Aw3aa_W6&Iw>Y19b7+V-DbGe+}She+S@b{{-M@{}SM6 z|1scbzhkHH4GKMmo(X`XpAx`vU5EBt<9}in{&D=o{9u2a1b%QmFd1;P zKNav?$io!CPZ6AR%;!9{^d4VR!-rsfavIpz0e%YL*8)BjaO@YS1HKIGrvd&n;Aa4C z=2e1X{M)u>-1rpim0xj0Y4ZI+Bbte&O5UJ$Nq@t?L)!;Y_P|9@La&LzUBaq z;}YVy4??^dbg>g;U{{ry2fPW9`G+h782m9N>9`9#82>6v?{~X{~0sc1NR|EbL;BA0+gK-$~eE`St zVSm6epV+T|3I1_>y9V$ai04|sv7KB8cp=#1xN6!1nZ~XMdt9I6KImBRgZKo%Zva2# zfZqss4&XSxZB6_Oz(2;1_oFd>#4-Mxzz@d%E5KdjJeCXN!EtyY_$h`wpnvSYw}L&+ z3+;d(2!3t@9M4Z~2fP667XgmCcL0v<>=!+__2pJjlf z?xTQXeJyw3e*hfo@d?0D_esEUobLb}=ifhOfusMYz}__upnn|4&=1bzn1?&TKbH5; zfMfpu0yy@UX8^~3hx1ffEWF4xwgT+2{X7de?(d!h9QTh00FLpmbl`t=;HwjYC?-mmfhW?zK%uQ>dS&jRlfrY-n5<7%=uS9ieQ0NmuE6TSZ?;5lIb7T~6S z*?tY+roA!#HsBb~JAh+6e+L}nSqnJE^Df{R&wGGlJpTZ^hm?!QNnF42YrfdDf&MY; zIuj4$9H+ru`wlGcKba&~jy!UHJ^&o!KMrt=AMM3WP`@9Ny~!i$t^+^jeb&WzvhnlL zHuCea!_OyxW4++~h5h<&vN!q1b7NP1aXUBmuKLj_$h;UF#cM=ab4X4IIf?s1l%?MVtEe&`z|nEnROTUkFOxkW5NF4u)pX7 zIG!8eJdXQx{9G3IcT->3?=}LC^U>FUGy#&?kC{7@#(8Q)Q!^FymVXPo~xSGzo6Y!?I4 zV~0Fv{KxW~AJ)oqu(r9>3KJct2xIkL`e) zKE-c84|q3d4{HH8XLoG>5#VNRz&Q6!zW8lyYW)E>V;$S^Sr%VxZ{|3DT@APyyBMDb zxETW(UktbzlNnzDxas4JuLIo7WsG-`zQgh4F(6le!1=vCQm=V{n>{I0N&sj3TgCn~ zz}epP_ZGm}et(IN&w}`3jP3jJ3%c3?=l8yp&bq?Fr%9(=XWj5`w62`mAM98}Q|L8p!jm7o^oc%uY^?uTy41V7Jp`XJ6 z=W@Lg!m)9HGpB1#_{J%Kv;7w#97_Ps_UHX1eB-BpvwaU4mrVa=Tk{ztW-JH$LI&ho z1NgB9q{r_8KMwHTJBKgW|0uu@0-WPHSmG%Ld^Fgf3^?156#H7hi@<&s;B0@K*e?M5 zc(6BruZrVj`>A5T9PCd3`!#?U1O7eWV*u|hiyroWBH$wc9}D<2z{dez1vtlfro=fH z@RPuP5#Zwie;V)!fWHrT3E-V&eC7Df+KS)q2lzy=9}RdZ;FW-zy#~KM7jSd$nDIq` zn>mN^rvX2e0lD4>e2M|-v9pY49RF#6_XB(?;KhKO&nWWS)qqa}`+0z$0r+CT&jfr0 z;Bmm$0bT}pmu}$;j=vo6{(w&hycqBbz^ehT1biOgRe&!Bd0!UJG~|;B|m60lXgYRe(1Dz5(z?z`M)5$MK&7 zcs}4wfKLLv8Soas6M!!Sd=}sxfX@c}eZbEJyvuIk3y!k|@ZA9C`Z`bQD;MxNU~m3@ z5c@d~@J6sVdn|st4e$%VehJ_|1$-6Y7XtnU;2h^KB+gBMUj+7@b`M{0oNRxi*qgsA z#CR+HZ>|wwf3X4SaT?&40Db}Bmjb>J@VS652mEJ%uL1lr!2b#OAi{5!`0?-TTn_d> z?h(E)<-JSn`vX1??9Jb`Vf&v0UJdqF0B-)i4BP(#@Wo(1AMh1`UkUg+z^?+li>&Y1 z|J8sG0K5(Gae)64@LIsH0en8-*8+YI;MW1Z67cH*Uk~^VfOp+9e8KVG2>1ZN7XUsE z@S6ay1^icl&jbr7v1-usU-vK@!@Vf!O2k?6U zUkUiVfNuc&KES*83}0~k_XC~__!7WN0Dl1RM!=T>-Uj%CfG+|3A;4Dw{(HbT0RAxG z-S-J!aQu$|ZvGAh{p-hJQj1;_a$ z;JJWz0A2$49|3O!{3*cO0RI!hxM z|198UU(5E-0p14o&jY>$@RfkC0{pLlZvgxSz`OSjU$E|rfae0f3h)xZUjn=l@RtE^ z1N;@hmjM1M;Hv=t8{iuNe+}^N`-Lw!{?`G|1$;H&C4j#Hcq8C%0zL0RIr|3jtpTcp2ay0p1Gu$AGs3{t4h6fUgI9E#RL5 z9?J<|aQvSEo&)&jfENP30q`=wzW}@y@Gk*x2mC9*I{@DZ_*%ff20V6P_=4mA2Jjrf z{{?s<;F|z11N`5B&jtKjz!w4j9pFy`{ypI90RI8-E`7om9RGg+?+^HY0WStT)+zPI z$$)b^G=Gn*5%5l6-v;=0fScci-i~~j-@0a%Rbby4>^B0wJ>YxvWmb5$YJNM7U*`jE zz8k{$6u`THpBBJ(0(=4BI|IH9aPwJu_P-kNu3*0j@E-%-sUNda_bXiEt+M@E!1naAe0QGl4#0bXpS6IS?=rA`?BL+v;B0?8^Yp0lXFP0|9Rb+}w*}{~dt$1^cyt_X9lk6J~|^KM3#~z|EO7+Z6)-6R     z+%95@{Coc^pHmTMs3g@7Lccp2cifVTpk z2Y5T+g8=UU+Lg@KIpj4mfKMm;81B zJ`U{H0)7(UvBOv;jB`BTIezX8SqxXCj#CMcq!l=fKLK^E#T(2f`Y&Y z1pfx(&t$OA0sK_J3jyc6B|i@{uz5~hqI`O6U~{4=e5^LlnyPBc2PUdo60w1m<%#mx z!0FA+v4Kf+V0`NMv89ROi8+l`@tXP>4OQn9O&nVqj|CihxBlF^lGzraG?ABAQC;2? zPc)U+B$~%m6b-Gds-KytE-D#YA+{{WG6S0$(%It%<;LS=+f-Fv8Lyi~p~Y((D$Xv+ z8`78tQotDHxIu$#+N^pPk;_Z5mnH_+)zpuvqY_Pumk#3p#A#DQm<&|N%Pntiu4+n5 zYMND5Lj@_#jn|~dke53nZ)iNx5T7x4kQH8D-t_eM0s)F)>Uhf5@OWusysrH0s(3S{ zt*SDfD4#hbUK&r-6wNLg66crcN>QX_^^6}otTfTkR6et6XkArZL(`lhsw^pf>b?Bt zn)9mEck*&8t7g|!R86A#R}Bin_s3N7yO^%=dAXCS>YE#yMprdgj2cvwH<+4FO?}Pi zpeZG&5haGsuBxEm#?+OTQzdY_%8QSwr5CkDMX`Z1D=OkuEfrOb2`c~e+A3~}TkA7f za&$DLl0!->UE14JC!2YD81+mJ!cEwulhWbN#IHm<}Z79Z%BkFX~p<(0QT1|12m1JJ-NkJDaX=te3 zye@So>MVoJXq9RPR_3zzm`z2dqALmm$WsDHuX~{A3h7BVHMIHCmzTOqkK)ExQE1Mw znmj>^w&tlp!Sq&IZTBESD zYArjpK8=kG;*q|@E~BKzkIW3o`6+FJp@d+Ch2HT4;R zW*QEzs?V6iOB4AtWmk^DIe^-9eM3`Sd2PHVQPot=^LNSEx>WE|Wkp4GMYO0Y;QwfO z#RbJkOA`f^HO-0g`iiP8^IpVDpdls%v(lUtFUlKISuy8!&Rp-nquZ^sQ z+6>(yssmcEm5i+|9TG1pDwPgjSyN+Ux4=VXutH3$F=#`QW3?($VnO+}L~h&iD+-?FS3KzDr>zUD9OLB` z6;;j6*)Jf+P0uaj|BD(ETF=wHo0)yHbKmAX^TZL4pIcQS0~c7(?AMCuCZ(;f|ZqZ=n=^g6bp+L%>0tF}C`g<~v@ z>Up{3AW_p;TT|7{do;WYW6Z`YSS|;1ecJHB{xxfL{gCIEfN;^9kpO26jor#B zDos?KWA^bhn5E{5aQrSE3d1+NE5l~yN+vJ2k(SDdIM20X#?C03P3t&xn>sQWoIEn1 z#S-Q>-34vFx7J1N*1I@s>s`>1kbgm_E&J|8cm+L}W?nk!()XIa4rz095^X~2X%jT7 zB0&oQ+VIQ{mWb)~2I~cG^yR^ZNKO)igQB`3S}2spo2zQ+_$b~`Pfi=>gr`R^#Z&JZ z6rUU%FojzsX)-iIg?pk&jiGN|!e&f9fso!2b+1XZrmnHpgGeKl*$80c<-sTEGT7-! z_@QeU&{V6e3R#-qGt*i1bSRgoDX*oYsLCLf@$$&Upt|5MsiP%-Fh3m23+{=O!%2B@ zI?2O=(ZWAYCxVHlhB;%Hz%3vzw54@kx@|;Ds4c49vX&sDPF6|EtO&zSCO+F57h={m zF7^>MZoW<6gC`JDu3e=WuYMzDTV*n3FaANpYOE z6hTT!aHNqw*Jb*qV-@>Eh?;Da^C1rBdGWAA#>;25j47$5L!02dW^h{TAHO9BBl(;P zuhSK!U`Z#qR^US{?vAP3Nn3uvASdA&3q7ZV*&Jqf%HVPzR%>ufu;oY1<|Yy~ZRNn( zyqh_vkob^b`AnmjB+;BB(Ata!$RhPPAvmt(GX?BKoTv1&fQnigv+7d!%%PX-mFBg@ z+)P5fr}7~*9hO&C(sVu}+$*Gx0_lXiu3>i7n2B7xR684EYQLyx0UXT* zsPn9ksMTFb>9BY((iMawU0U4!yW-%ET5=N`4yLJ%EVUHcQOzi?rQ2=XI5OMCQ0b2i z$q&Z_JI$iO!*DxWpD3@Xr%8r4w)Br|nQI#Ad6j_oj9_}SR-<*wnPTQ%#?0Jldzo+s zweQ%t$-}67RnW(Ec#sJ{Yr-o{E7AIHHg9mSQZ<8A5$y=mnUv9j%pK~MEyqgIwx_lZ zZcAm4l5P@=16uXbvc}G2|_&6vMSErFI)tiZbuh77}-m=^0NuD=-_! zbQ7)=zmrkgCYIoM7wiKPdBf`&rqm~zo5s^QTpsOZs>_?>Rn0RJ;o6hdkaT-{2yL%s z(dtVkBU(PE#p^v)@~)RY%*4ktnQIMuIhvP8$8dG!E%D%frVLH#ZCmb>(|gt=N!tus z8>e?|vt_NP6R48fI%*kwHbezY9VMAXMrz=5Zt`f;$s<|D0_Fep-ZE-Km6otnx$0^h z_ewW=qcrpGcd)N49R=+oYLTyWQ+;_JOrPQ1<8jmTs4<4mX*`@>!k?(j&)EL)^a!gg zj}AniVoBR$gi|f}GAp)-jwpW8l1{$7I|{EUgJnH=%KnVVtkuJnd}d>|&ym;1zy)^<=YPFB+U)Gq|8m_m*7+ zWwD+eW(_5o19Ns#R0$I@Kk1c`h3uT=H=kSHNTZEo{hpqR?A!e|Nk{f_S*9d=he}$w zx_qZ>?E2nxUJ6O=@QIGQy=S#sNj7*KLc&IaCqvp^WLurykZ-!L!XLL8l)0(FPge0g zHcoezpOYFK*SZ5u)BM?ApARuQ?(^|zG^3&ET)H=I9V%ij#|Jwip1SdS24N@H?&l3N zclr1}OmM?E99`3WI&P=Y5{AwO_!c8=U$kwDYs1DLj8F0=($Xzs>p~IpR~gcmIJ~19 z6HJ!0^^y02ZQd4dX=M34ia(W+YTQACl>2PKE;Z1k)rBnX=-yj+_Z7yg%oLjq!Iq`i zu(v41w1~(|vAq%5)+tt2KDgr2dh_ z5*l_>NAq%oLidw+M7E^Dc>z|bSuLrycnz0ae3h6*OBU;g%zJOaq&+6(gSr{+Nt>If z^z^|Dx;r=K#NcUtLn2O}cH&KHO>;aktC2qIR$Fs+6)!V+4Vp)2aDJLP%c76NjiFBk z@u%4e%)9jcs+yU>C*^{-GIqBUiBNh%YLrVA7Z)U3Q{*&1|Gz7?HHgr}oVod;VO`yP zr5b&3gSI(dKdZKuzRAPK9!+ET2y;3f9#FLqBjYY*@Tnc1C)5>%epeG_7$V{g%`I7; zP;1KS_9T=y<76~kTiD|X7fT(sD5V8yQ@wQ@hPUK0w9+QAj5Pwa;9y07E|gn_>PMRC zu+)8Y<}3`4!CP1u`m_x$dT4}AAMLU(jANRG&km(K4jLSO^@Tqel6Hu<`At97cQAXH zxdsnFtO~1z87kvm+h&if@?=?8VCJOcCot-R$%tnox!Y>>WAEJ%-o(ofZsHY9q|O_x zH`UEP9W0vrdA4@?#x$$xY4DaFX8IpxrbWBCE}d%AbtPl*_8uQ;)TQp&Z&63%sw=Rj zb{Z^b4Hw*?31+tPK~(>_MO);bJ92Q&NnfAH_z(zs%xtX5uDq4S*1F`jEVkyKkN(}o z=Fu4Yd=_7K%iQSRU@&2vE)g!AOF!M*LY*2Sl0i9xglPJP3Uh8gN@XtMDRDFgnT8R56upW*jGlG*=xiaPe7tI-wVu(a z3_kEf<(!z&O~RquvfQSjWG3K$uW8Oa?FJ-`B)IPo-l$Et;4y>RSJm`}cy(26BQ3nf z)#MgYIt($`rK=^{HZ5~Ve2nYk$=gPNwl4IU5}s~(!bvUF|Mzu5-*^~ZpRsP(Hl^YZ zCeUC2r;zm15wwzZ-O=N*igv8N=bR-(iTWWn=g;VcUlik8zr4++<4N98H&>L` zmNx~Trl;LOc*~cDTiS$&UvM27+$0@KH{4Qn<6bRgInB{$WXJIdNXCyo@bH#eVqpQS z^~qKXsUxAjNy}0zi@7bel;raUYmf84AIB)UtXraftW4dXZo8tV?e9z!l!3I~OTTYJ zAEs;K?|E)BW!}B;Z9TI?b2nZyQ*xTdX58DS>e)6UemGH*NYH=nSzN6K6Is#lq zvpgNno=9$}jOLbWnm;HjpQ_c;5g7lv!z}(1dvaj_Gf*0th6ox`?60A2e)-1V6bYw+ zq7n*FpQtFc^!qO9<03bH8jzCXaxg0Ml7+IqnM0W(T8@Ss{st+ff#x|{XVdE5K6>1G z0Xp8;tNu@UT#GvD(HH!aXA=5?(9;1};F`G4R*nKmO z;G_@HyY}IWyG*9vVYkZ*yB(#+bU9^YarZX9W6U*{5e2X5{QKMHetm`4Ls@@3Jp5L# zvaF=xCFrfDl$%+YWjR8+XaCeK9J`L_*0Vhl$yy<;d2$4fm2+~uQ(6G09`2+JlixMd z71B)8PW2VfT&q6S5PS$Y>SuMlPcLXkZPw`FXVo)W-h zroDPr0EEI{5a+wA^gDv|V`-BnXW6{zM0E4<$;_AOo~kMRh{)$8I&(hV8gKgYc&j|7 zZssZ__dSo(bC~U_F`FtfzoIy)^nY9L;myshs&@=^TeKmoemKG{sGqc`wt#PbJ(Pc( zaa&b^+N`(DvVL@`eW!+=ZR48O1@zJBn&!ra@Y`@1Q=T9_JK&X(8obF2d!n;M9eFwU#0<1?Dd z>*yP2Gn*P_HHN<%y|w?YKP?v9O>t5vSXX&W{3sUVzeC?yUYkc3f1KEL0si+M{7J$? z{U1E|Wy0SB^m|bmOx)(t6@NnE@8`jvCwwgbe?8)F6F!zd=Apk(;dl1nFIM)m z%D(cj5~Z~jIT*5A$^eDgPwkiUxuf2|UKR}cO=;dhru?!VLNcjQgG zH;=CNZ~kTt#=obBzWJLm$luF@Z~it5^7r=Oo4`##dbHn%6n?7*zeD)gfA8|(uTb>=<1v4}s_^RwGkG(QuKcZ4 z;=kJ?{&k8z{|#xQYaU(t8x?+o2S50OA~g16|84Z(cconOh5U0o_&tS>^}mlt{&E!k zCJ+4q3cuNdU!d?49{kY?f0hToMB&f&;7=1i=KnB{{8uXc0UrEXCH`Cwev86C!h=6o z;pchq+Z6sF5B@@hKg5H-SmEb;@E;OB*8edc{N)P2(1ZW968|U<{wjrkoCkl6!avr7 zzfR$g_TX<+_$PSqd9TM8*58R9eBLYB{9+G&cj4psKh}faTj7uM;P)3kmVbf=pT`=z z{3m(v3xtp5KgEO3ZG|u7PxauJDExDM#(#y+YhttRFpsYFN43JAf%`AkjV zjW%HLbzKjop{U-+26%RKaR75?QO z`~oF@UQ=XS|DzTDvmX2sCH{FH@lR3opZCzORQM}B_>D^ZfA!$EDEcpW@aHP}FM9Cj zEBsX+{Oc6{&pqmIp~AnygTF}lxc=fbZMN-qiNb%ygTGwi|ILHHLMi`i9{g1b|8)=k zY9;>F9{l$e{u>_r4GRAk9`*md!k_QK@3KSY{BflRe-DNKu1EQEgpcF@KRo#T75(=; z`1uO|0}p{?#7*dlde69`P?z_&<2?pH}$)_292k_&j-KJAbWF_?l8lU z6VFzEqr%_bLqE1-=J>m#2fwSr-^qjDQ{nII!SAQ=f9a9GT;XH?;WH)rFSyL3YyVZC z@Q?H07b@}Z=D{DQ@UQVG{}hGa!$ZGPDgU(|`n8JwULN`_3V&}8{#+&geLeW|75!cw z`~`}BZx4RE!r#w>zgUU?01y5{ivD#T^}k%rToWu z@LyH*Z}G_ATH)j8r*HM(Z&dV0dBh*vDXeF46SSXSd+2u&K8{~S9{PJI{1ZI*ISRks zBmMzO`R91(=PUXp9{fT@|2z+Vv7$fGgFi{(^F3YqFRT7vro=zRgI}%iCwuT475=Fn z{1%0On@9U;6+X8A86Nuc75+>Q{sM*1_r&SHtm7WaSH!35B(_${|*oRN`-%?2ftC_|J)<~3l#oc z9{TeX{uLhj*C~AdP7%4tYW!_i_`mhw-=pw<=fPj5@bC8EKP`Oh|5tkCZ>6Gtl?Q*d zqW?<|{#r%<8V~+@MgLk4{zip=y$An$CH@;c_+54h+bP!HjUN2&3V(qIzo+nV{<+D6 z-%rv1l?Q)-!e8jY&sX^Oc=VqU3jbaY{%GN2{%-S#zeM5R=b=AM;V<&guU5)`hX=n= z;ot8Oe~S|T5)b}dg@3_MkM^@(;V<>zZ&LVw@QA;&3>=uhCp`FFg^%OUlOFt@3ctgH-%pAE zDG&YtMgKvM{1+(vhdlVB75?u%_$3PeVGsT^h5xKa{;C!Ja~}MJ!hggg{#J#*%!5B) z;XmrZUm$$!KaY9v+ZFwLJ@R*t!vC2Ef0^*{{P}T@_*W|WjUM`|75-`u{`8 zzw+QOQsV!Y2mc;L{}T`XGDUxr2fstn@1}hIF^7g3>eey&xL`Qw(7!8%|LguqD0m`Y z%TI#6=oy6-D@-EclS1As(Y0Y{_?qv(k08jj=bj-PJD48XlHN|c&K6!8NH0RKDRclm zWuG>`Rrtjz)8JW-^_tDt=MGlE48O{y6wG!$?gzx%`C$+2x<)(Eq2z zzn7(Nmw%>%|DEvJKVCL}I=!jIU#!GGN{Rnk;X* zevzVo2k}|ouD>awe}oZ|N4CYQonC0=ZxQ`mG3NR^LD64B?`!&(XQ}_XLw~WNKSt5- zPMD_uaF+Uc!nc}lY@P7=%nUEPJ_itAtH0F}e>De37w2yrJvV&QFGh;KU4NI12zm7L znrWA~OJc(b(CV-7qtLQUELnd%LAHJwVQgvZza{$ng>I61*2~!G1Zev0qHoH{`Xyv* z>+^34uuofm)6t>ixTJFUgl+LMziCuxRp2s3f1;w#p9yB4+gbeLk)ewD{X?T^;y16K z6TTh)6ye*yjmlbh{h0tQf2)-EPa#{QZ}R$;L;oz%-xI~?xmC>At^{yQcKofvx5rhIt;3O}m;_&0?#{mZiC*E#qLh2JMF+RXM>O?)l?#UF>6GVPql zk8*%Y>WKUMUj${#1bR{m9r zKL7rtt9Z|f zM>_b&3%?GH>A9W1A;j0(&uWPu=ii?x@z*=_9}@kj_EY1~@A6q#QXD@oSM>RNYHVrG zAFqo3z7gf`K^3C$KQr-%VI;lS`Rhh}Eq^%@KkqsD;`P%N^xUrhqa6C(P6&C$;Y0FY zTYsoSzwz@>*zj3@zM>y@=uZ~?10;SEfUQ5(p}$P@vHVvl`j&Fh96qtj|AK@6f$*co?`MgxwVx7+AKO2Ff0xXI%i!D7=Zv|e zcKh!>CbSzEJ|zFO|I*ekcj%v)rT#RB z{sKjxzrSef&vEEqDEb#AGao*&^Ph0&FA@C~;dB4JS<%16q5mJz?-!xJ+`;d2Vn~eQ zzw6*1FZ@#@^w$tyYk&Q}q{7mL^ACUj$`W_)+uEEe?KXIk>jZukHLVAikFWo>G47fA=co|Fc8CM1)Hs>YqPnsOc9g`uzD( zyZwLS&~F$0sPX3mhklEq|A3;uM_0Z5|4a0zS>-1eY1gjA*UGG{7x_>uW{@Y{vIziD{# zXpcW95?{;z0x3U^zfUUVzt^GvM3(xC9r{ZY{XZ)DJMOCIZ;j~pjVS+i#MjEdLec+| zqCeiDADa*gN3FldIP}*l`h2g@ZvV3!`Z-zZH#+n;Df-VS`u8~Wj}iSVto)OUwCgU1 ze)q3K%Qms*@%LFpf4xKhS<$aaQw!S)LLWNx+eN<&^q*JsJTIdBr#tl9l>Glq(Z9-}f28R5iO~PK zLw~WN|GJ|8s6&66=${>-|Byq!L&^Ueihifv_4cF%k2MmJNVZKf2tL*^)hxR@wNCD zNc;st;P}@n@qgyfUtbyq9;N@0Lx0t#P#E>!Q}pwC=VKfz8jG`X%zg zw}r`)hfnPJ>uZPpk{?2S>_7aG3R{0HeZYn-?fHv;*OIS6;Y0FYTmN|CYxQ3!H{Nmn z=3C;n{+$l}{ilYKGL0qm!%`(*#%^=yPZ9lGG3EZpw?u6HPaOL7qEFo`?Xt^%C~Z)* z{Iv={YW;f<@wNQTQ}V}0vUdE(IrQ7J)IZvx->&HMk$|m#zC*u5^k<}HFLV5z?a*H) z`enlB`r{?3t$(*e|KQU@qD(`{DD3*X)1lwDw{2@UUsei(#Qm=O=zSe$@^9r`Uw{c)4B^}lxLkCcP6sQSy_Td%(o;YaoVBZ#ln-(n?yov88J@z*)@ z8$~~=|5ZEmJ4C-s>Yw}Hj}-k|9Qv<`epLTk;Ly*Jof?kc+bjBia_ILzBMdyM{yI?y z;Wlmezaxbo)&IVC#9t!uWB==-#6R1i-yr%?{qG#&Ywf36(cf9ozulpKtLQgK{hJ4J zk#^nU&~Fv}VzK7-!+#UmZoiux`rXb9iBb12zjElWP})y7MSs*jdi`G|`gA;!cG=~x zbMSA?lE2WwzgPG_OH&VHMd(K2YxTcL%HJl&od4aG^1tTL9~2LXqfv~W+vR`Bp+AoY zE4r}%?xE=Kysw`B*`j}7;3f5stv|`Zzft((Qq(Eo?% zAB|%4+}6Ltp}$Vi-$&8!+e^=1PFW~S$IEG#t-r*$23p(V<_g==W3f|LxGz%iGx2~`1E$#Ww-yk$v~_B7Nz`$D&_BSfL{LAEcLq)U(;_>^am*V zO%DBAMc-ck*yXQt=r30E2P*mxIrLYGKE0iG+4}c8^g9&&Tt)w7hyMC3_5bS7U#;j5 zQuH@C^mmya7W?qD(uZvYp-&w88x;L~MgOQAz5Nx4et#6B=j0;os>;#%Wx|i@KV`(% z+HV)$?9het*D!+Y_3NLA--+z({lkT#9~J)-#Mkt56#e0fevbon{X0e99=|wNykwl?3zfkncgvs?cO40A$ zSI>VzWk@_FsT@8b7irfn#MkoQF8a8CC{pwGi)t^bbuH zJ$PpK-@S<6nclJUU#93!RP;wW^uHASsQxqDq2HqDPg3-+cj)KH!DUqYZ*%Cki9U`W zlNJ4!9r}|+KdSxz)uG>^lz)n%-|b+%{+Eb;RQunF_*(t1R`jPT`o#|Yk3>JJ{f~0! zZ&387Df&ws`uooe8-7&#zuTeTMHXZ@{+y}kzv9p@cIcCfwCe?jevYDFrs$8P4_I@b zw%3nMqEEMX(=J>8GY9`N;pe5XGv|+yv_RGL9}#}k=f{QqLC5{+3AmtwP}XJ6nnWc87l9Sz$p(OL@(Mt$&L{KbLQo(1qhS z|JI3}za6N7uulm$;P5|lMc>{(gkFG|7V_+ujUcPGA9{uZVD38nlq9s1RxAGLnI(ZO%Y zlE2cypC^1uZ`x(&{~6+I{`sOGmA?-h{Huf? zmA`i#@psuNY-u?DU97}Ej1EY({C%8d{ss|W%U^+_f2pE>nL|IfF08nHBkJ$Z4*rS4 z9~Qy?@ez9YFA;uJ|C#FG-zfY8Eq%NHoI-ppf7Oyd9RKGj`Fq5nzfAOpMCdPd=(j5R zS19`bbm)JYrT$un{sN`_%vbaWQv=u9Z=d?G;G?XBlZ&)#An~>QuMquK;q&}^m7;&A zLw`m?C=)gQ-{#OS-X)9_>%UFWU+&Ov6a7JHrOh0Fzjg4R5`I+w`PvbGnG*lCO8mLB z!O-gO-;VgL>Wm#me69Xk6#eTJ{d$M~VU1zp?ER0eU*phkQ}l0C^gnj!PZNE6|7PnS zKozXzuTl6>^?!zgf4lIb>VFFHwfrwu%73#`{%aljZ#d$&%YT(azeCZ#MbUrBp})&H z+139_hyH3s|JRECNrUzJ8zK5p^?$j8Un2Z#BKrR&#MkO?gT&t;R9SK@!mq5q!f z9~hzknnQmHf7pvI9RC(6`UmFg`Rm*i3LgxS)CSS&S6JIO;3enFMQ!f9VivCQ8 z{$SC!_uqE=o$k=zq?CWLqJOhPf1>F3NEJ1BW|#j32fs@Ahg*C*{%0NWcmHvyg!Ol~ z68{dgKw+Qu{`(ryx7ROr{GG@^tH1t={=JI+IS&05qF-v2-`1~n=ogAUwx9bI{dXPu zy_!Se-7S4P{|D0nx0b)dg+C{Pe<1O-{54AaWy0t7zf_5Tl|%o|EaTsYCKxULWm(3* z2l2J|S1IxHZ_V20KPw&jd5N&#QRg?$IP`P*!=iLy`+r2y|IVR*rReV!hLQZ&uK&Xd z^!)us_}!As@QKa8*uj55`1bzG=6~YguM~a{OFzs?3f!M=P-yl4cj43aF72}Q_aVMk z|HVrF|DfdmREK`IS)nW)&!=5>{1YAeWs3fjivBeY{e00cNmCDVh0v7_{T9(L7Gv(e ze^m58ap=z#{iytZ;Lz_-%Ks-tzu%F1{V&Zj{|69XtN+!C{+|{7xI=$+mieFR(BGis z{~1MpnM1$-?6BgF#mc7VVIP|BA{s|HKk2>^w?iT9f`0=u$|D{9!Hqnn-zl=Lt&) zmt&k{DXi*@oj`moe`ONCxCws0@^4D~zjx^OnG;GzL7zntB}`p5jesl?BJXMpR(UcY}V`n0@ByF#xi=q`u;0@3d&raXVG zQS`rb=pT4qNF1M%3ZB{R=TnFN64A%@|Bj;H`xw3ajiTSz(znaMH}SRnZ&J#?R?)xQ zp?{s|9}uB`i9^48k1$fq|9gu52M+xQMgLe!-_HNL4*mY3kK^C_ihgsUp8t)aPwvyM zuv96izEIEq6s7zhDEc1~Uu*yQ=Z6}7Qc}S)TmQX6UB6M$U#IAwK?&CMCyRd6`STRw zYwd5IqW`g?f00AKRrC*tDF1m5{dPrvy`ul9L%%&s{f8X-%M|_36#adV)AQdU`b#6q zzbEmv{I3%I4&n3ow?WZA-=W`hK}hTqp})bwzb;GuC&bs{Z{siP(S_s3R|MJn@3Euw z{5>rCds*>!vR=RH;I9<^Aj>|CDh0ksd@cUPO8nm_@t2I&B$5G0+}AM$rOYYd>d+{*e*oA4z;ozssIsrcD3k@plJB z{|SfwBGI?UAG`lO>d?QvM?PfSFeQgD(mz-qyGCUr2mSzeLgJzkk|A+LLJ)*E{q}M1TK?_SfdnuU7QC zEBY@x^y@{xG(!Kc4*gcq$MyH_ihkjVdj9_?`W2SGo&TeVujPM;#p6#;MSqq<|Bcp=7uEki zbnw>;KkEGZJ>qNi*MG0f@w1l_f5kXGf4g5C#(!w|ko?!q-mLDew(5{SkZsdp?{X>4~)=%%%Q(n(I2Ad?@*%We?gY|ortgHzeCX)6Ds^^^3FAA3}Vs{#tv6<0tBmRP@_K-yXkPML*9fZ|F4z{=%W(oqw!} zu3RzY@)s)l?>Y3BivAJ8HxIV{+YbFgMgKTOKYya0|8-gB{|Mr1`EL~cGARFOMgKyF ze$LNAqM83p`R(>M$DzMU^l|-myrREa^yTMDj7+Rh^l5vQcIBkKqTk1*4Vu<|+p>)R zaN=wE>$z`O(-{94CH_lA-;TdS^b4(c?f74H#2>pXBu3@`MMwN)O8n!L`1he3?3_P4 z{v6S_=TAHSGlg&W-$}whHY`T+U%UUDMtm**i<1FI-n+%Po8zoPQtW%K_(#NIUD literal 0 HcmV?d00001 diff --git a/infini-qwen3-moe/build/.objs/infinicore_infer/linux/x86_64/debug/src/tensor/transform.cpp.o b/infini-qwen3-moe/build/.objs/infinicore_infer/linux/x86_64/debug/src/tensor/transform.cpp.o new file mode 100755 index 0000000000000000000000000000000000000000..26305dc98df9e88530cce9c149c1ca8b7ae57f9f GIT binary patch literal 264688 zcmeEv3w%|@@&8Ezf{GH~C{!=;1uB^EMo~eynn-|1Kv2L-NG=e~%j5>*`w_8RLzGr+ zX{8lwYtgDzs}(D1P^;8p#cC_o7xht5w58Uz)KdM=%+5Z}WA06&?XSPz-}@oV-tX+0 z-JPACo!ztN+{@ynr{w13pJ)1Opf7OxMWDao^lyQ_#OcjUUk3U+PQL>5RZhPK^mR_Z0dxzee-HFcPQS(U zZKm%q{R7i?nQjI89;e?2x{cEx0R52De+2p`PHzYLXHI_v^e>$L80cR){WqY0=kyMs z|KRi|K>x|Z4o}Au=X)mC?IlU{;-8j8F&_0~r z1L&Td?h7=Z(|a-9o9RA4_vQ3{O!sGc0MLG%J`m_ZobC^F0H?nK^k7aO0yM_yfj|p5 zeJIewIDI(KBRG8|&_SFY4D={Y9}V;vP7eV(l+(uoE#!0&&|#b&4s-;kM=~7+bTp^O zFdYl@I8GnWbR5tVIDI0}lQ=z|=>(=H1D(j}NkHSAJ_YEhoGxZs0`xRaPX=1b=_x>` za(WujGESckbULTYfzIIcOr~crodt9@r_W@17SK7Io(psyr_W}ZU^<^^1<(bYt^``e z=_JsFoLdY}!QJ`ZRkr<<6jm^K4l!s+vwwlKW_=!KlV z2YS@)3=zu z&Ga3hf8g}HK(})GJ*Mw7-3IgnPJamWkDUG!)9py0=zXAO_r*{DT z2d6&)`cF>p1o|&df6DYTpdLgxIKN<|lgsHmpglO<6X-6S?gg|rr*~z#8`Ir^_Tls% zO!oxZm(%$`_u};4K=BE^G0rW^t4+1)v(?(qGSE^^PhmQh z=`^5aoIV}sbWWEuodI+vr_W$I3+QZ4p9%CV_?rWNa~Ye*^lYFBPR|Eg!RZA|D}h#V zItg?krxyWT%;{>N=WzO5rZqroIb8>|p3@CL&*O9>&?Zi&fHre_3DEO7-2(IiPG1Q0 zB2Hfn^b$^A3iL8gUk>yNPG1SMmD5)NUCQaJfqs?K%b2E_UIVm^)61E*1HG2h*8yF@ z>8}C(I;XD(dIP6#1o{n5f0OA-rr!ekZBF0BbQRF=aQeGIZ|3xBpto@Pdq8jH^!J(m z0O)O;{vp#fKyT;t9YBA?={uR;1@vxC{}|{!oW2+6eVqOY(EB<40MMUu`az&;IsFjO zpKxw%Pk{cD(>sCwi_@P1{fyII51vDq<^s**bPu3CIlT+eUYzdDbXTUk0o|R`eSq%4 z={i?$7B1nD%3OAkc$2-Jj_IpkLwi!9Wk;bPVV~P8R?@ zl+%YXJsju}oIVohAWjcvdKA;6fgZ!@AxwudJr-yor;C6NEnTp0??B=JrU?6PRE&^0`yc)7XvNf^l3mRbGnr26rfW% zJ&kD@(9_{>I{cM0HUsEPPM-mE7N=(eJ(JUC0iDC?xj^S}`fR2Npz}Fh0dxVUE16aS zO>%l6(`umSz~9W`^gFoU;T5;_g#W$br5(*Ze}<{PKYe`O6UZ)3U$P_J{BhgNkBi&l zJGS>-pP$|o$;ruIVeqZ#=C>jH);2F4-${x5w)opQk7F_o$(<%A%0|I$fPPFFcBGA< z)(`Tz_^s1F`|Pu&9er=?xT+(y4_7Ar+v4=oTW0}h+ulBC2|z1;p{no5*i)1-1ZC{y zlQCDx_)Fh|KS4E=7q@=U0+m@=oNiuGoSwO?s<^Gk;e$|T>G;;74rl|gp5BVOgEVB} zbo%k)_V}h^RafFZ9bdz_mch06K<;~N?)d7!_m1NBneYd=6>(KZ->WbTSFkQ9z7{Tp zzw!Cdji9N%KkayEPpnS+epFujNF=&mly!FRU|k7@Ha%6Geu*376_0(Q@AaVgzE|Bw z;z*9R7%dOlsY>tdUaZQr&3mF#sM3Q1xD;Ovv1nz!Zngb=5YeuDRpDaPfl^z} zhT^f0_x(CG_1^4F^qv44IpIT%97vP?O&gL20UjB`4k9>Zja-zfGb}%rsa;m_Ocjdak~Vr(5C8$Hqgv9 z#eILgiArr7Ur4pWLhYDN&n_fx`)t&K#alb_i_@=b*(;P{R6|xk#;sMw!b zdXwVx6KO43hGz)f#-F*oa~ zfXZ(!P5+@Z{io9O?@QC$V2}cPb#eOn;xsgi{ZLzc8Ds=EHU+BlD+42#gnLwITPkim=~d`m87XhRO=s5Ex%EI;cN(h#eSyrJ91#@VcZ`b7gTztE&7f zoCvs@%=CTs59nqmt8)s6Kx{ z_4kcIyPz3g8N=~Qk$L{1&{cFzYX{7LJQl_huc;T_U-Gu~Tnqn0WYzHy&Y;4y>bKghUz-PB%JQ*JL4MyKKZ(K14K%k+!15wvq1O27ocE}n zZQr3$9nqEP_~oFE%N<`24FXRWUxqgZ!D5t>sY>5bvUJm|J*3rWYUe?1rVWb0!BuD} z#n2Jt7H~chyMXlQYJ_^s*9&bbk5c~{tt_93#@A59lM0x^nvh_z9x#0e&Bj-W*7%y7 zZLPX+>`7{rwKUVmG2plJ1 z{1NnX#px-VfXbL677`b<31D>>d!c_s7!;FZWa z(+`x<)N^Ao5Vs=AD{4biMude7Umc^E6kv4j8$%R&Aimr%Y|R{${M$ z@!GmI2ind8_ZQmT8oxZ}z4K7%a$`fw6FF;0FPv3dGlY-Xx(i#%>!&U$6@w>Ks%tOM z{$)y`3&En78Ls1`qm}6sq7=!3WJNB^*#=%#TwiuUiy@!XY18bIdV_s~%hRAY{31(M z@&`X&oK|kK>%AOl5ZMe0$G@6Ho6tfcC($o1N?D=I5GKImX)sWfVjUF5zji1?dqPtr z=f#L1oEH>lZMQYP0;3cY83j#$CGQ60!yFIhzcRx7Gglxqn$h)SH!)Zsm!w}UZkuOI z)Lxe3Aw^t2uGVtxM0_!cD$pH|p-G@}Luq;oZTuDg{m;cZOi_!;!1fgecJKZAQ*d`} zs&OW^)3Jv2pJU^zZE)fA;OZ{)U^s7}2Un}$!iWzA7seby!3A;BCQJaqCVh`i2jT>~ zitkFWp{D;Up2TUKf(vJ@Q|PXt{WE&W<|5b0GL%|;_7Qfm*9qL z?xheM;qDD}l(T-r(9h@}yeOwbSK@9FWOc8!xy?{^s)_@5?54ioqzGnL8p;lr0wDLh zznxfFTo{4Rg?OCz>Pb>rQO#>ZP3ODZjhlH!Sc*?#VQ`rb2hGeAWy)p z(H0vlIA()e>$C(=lP!fLz&;YK#a6!eGL2ze!$d^n1jm%3t+T}@=x%X4m#V6|D==JY zw=Lw@T*=}-wc8AqYx;28fNq5IYJ>&$$byl(YvO8$A`rFX{Hh{$z5Csbh#gAJPhYY# z-MpjqlASrt!>+hw$7i3xP4>aX?I+;WrmE_jL<6xX+`P~Z;!(U5&zyQ&S3!(9&oIDi z?9i9Bm*qpncNEjD6Hi}?^`Kno=Iz98qeVWX$M=Sf?$l5Pj_=*~r{1`uQb)j7;87ls z40){Wz+NFEcbsjh=W)oySehs$eftUNZAtZS-o zJT}!>QP;GvzOi;_WkZA4R8w7<99omCi=7xNj8!***qWxK7jJB=Z#*tW+*nhpqA?X4 zEz%GQ;VychMd^6xS}CBXz(0W5s*F) zOHtoc!AJ8qtfZ+RvcpQ;pSC$)-8g zu_3Xdd9i^f#2QkKRn-d@CQ^fR7Odl(>Uo1>M;V#=s(IcCCs6U_4K>xN7%46`DY>9| zQLLm6)m%Zn1)01VjTcncEsCY;V@;ITs?wX*SiPvauA(NUIw?j~J}x%LJ0|AM^ClUc zd$G#q#>Qk_inM=R%$uG(uQ}P2N>&ABHsRIuDQm2+YOYMh>KDcuO>R>*^Cy{GD<(rK*lQ zZeev@)gZ-MkOWl?8hi}rgkHxl)q}C?R1Sh{PcMPSfi)kOn38BrR@BtgS0<|Ka2TnI zR94j0)u&>Wz)K~;*r7CaMXaWxaZ$1ny1b$;R$I}MP$uNMkQlf!8oTM4Thmj}rFqSx zwBu>>+M2;vOxX%^()J1I?RxQ+ewwx_$Rv;`l(jyE6w!-tkRz0t!;~&GFXuJ}$8EXW``)`}dOa+OuIgycP4CvaJqK<}=Wcsy1^Thz=JqdwUJ=eG@?B+C zqjl<_Yq&e!x?`WdUwr`DmhLES-*+jsvYlGj)*~1826^hp*>vJ(TS(9EwL>1-WPYzz zr?y?JcAcB?Fsiq$ZY|vJ_qHusiz?hUU-8aVyy=QJS@9+*-U*5~O7V_Wyg>>asKA31 zxX;>rl-Wn|)Gl{uDu5iH?_+ePXEn4OCq4z^vU+v%aKIw+Kd+3yyq3o44F&?wqa@wN zU{L9(dYTz}=5}Y+MQbAK`Ruct{f-3#IR1dQdc(sY%Xj%zZ*OOB0D!UAw{v>JvlRE{ z^n@on+H)CPozru5E#NfeK=K%NC+}Cn8_VU_$9{fZv?>9YP#4msDr7ru* zf@i@zezM!uIk)!p-tANP?Y`b$_qYhs8~UmY11o@Dm(%mZ9`H;{PuSUbICt#op59G8 z2HoD%ySnETs?ol2$h5F0JhXCuPEUCBR5$qTQC}c;7_v7(vk|D_>O- z>#^J9p73QQtFL?3mCgG8b9^&^Y%m6{TUjeCK!U-TFFTJl$hOitD|H9D^@Bm+Qa8X9 z(c!vPy3;&w1DJ4!cCEOEYLCCi)c4AT3# z1<=hGTV84UJve3H-d4f8hUt@gqnn1imHNi~#BJlSvFgS=R6ty#V2NR_s%YzzgZTDn zMmxdgp)%X(@CZloiF9!y(_>v!bYZpE2)5OgVM$llMz zTMO;+cXjPhfiM-qck9~xZ0F#^;oAM*wF`f*Uz(vDPLr9M!b z>L$OQ3WN1ltGvymXPm)j!XJ2Nr`ojWnS!%+A3D#KQWL-5OUrU`nbbwWx^<}Nei)9x z*x3Tj0;(;&AAH?nztRz`8#_Qf{XjkXdbkw&68qDG=>{*s6(EwxJg#1SE&z8nw->K} zGY^iv1ge47ekPY*7p%qwF6H# z1b6(@VZzN&jZLnZwp+iZ92H!w9Ti5WAozEVE3re4SUM-7lefqj| z*Zo((&c!}n5w!0J%W*qWjOe(oBEOBdOTb75YyW2)b8l6IGt;afTn!`g{rf*vM`<L=s-8IlkN750Zyb)d%X0_SKSF3CptU<)eaua&TgkT zV2LsX{^GAVJ)Z_12~fGW-G!Okq9#YD-TCAfRUd*GnSzvcnE1e%sAI13tI5)YW22D2&By^9CK8OLXe*ME|S$|Y z8PlDpoj(`c=x}Q`Hm+}DJQ69;kq>^lnJ>1dWYaKDxbo8I&7px5jRN%?V^Ey#r~wcC zQ6NGRjoquV(q;MY<*0iwwsWvu-+{!Y1{nXDlZ*>?kH_5 zbaSD&(wyzE8u8bH^0j#BxfVnV63yKM6J0k_n$oYzkTKo0OaihB%9PF3l#tam? z|8L#f|A*rS=1nzAISXvN1~FWk!ypu*TxR0Jul|()&cVLb@qhCw6IvHDGc@esLzIBi z{tkbXFhW6ku7r&m*wd#Ky`ACG5L(cMD%Z9LKr?|g5^3hk7|5#8Z%V!hIl?AgzTSq} zO1lGaydF*iZH3SSC-5y!=TR(<%Rk?~SzGA7wYwq&nYxN_Aty8Ky>OoYg8Dgur{ApI z{S(kd8w9tcd}F_Mzy}tj7%{UBC>{Xq>aUqMUhLG%gN*tr7?sm@1ayau%CTah#dlyV z@!pOtBElYMvjRHMaaG;zcB9r%xgr0NhQ~SDz>j3oaOfi3VahNTxKv?h(yqBfwfhfg zg;#gsuBDY7ImRw58f|+r$^9HVi{$S?-PU?S9rEw!#Vj5T>JXZRomHV>XPSnMow~B% z9vd8!v8Uc!cjVB$+0+!WIf$|GDLp=Sl79NLUlZKg(I@(<1MF4(+^bygXJ%B|J_g~m z_Xqhv=(zYq$!`va0W`>x1`L zeaA`NE(^C2kf0WP-`##=H|x`@=9@hs+?p~D&MjN#L~TktdTW1lhx)+NoiIP*3JWsh z0t@0FGc%~KN;z%`+&WkmJN;@W58sS~n+Atf7<~z!3eH?7rHXJWh`!qhg(|v8WaCQ{iEgVeK77m!QsZG$*gKcTOWV@GQU&7O+ zHM#%cvP~m`l@1tjc#e0+hq^fOG>tk_>JDA#69@q^c^$Ow#bDiQz`9q1b+2@%TXkDx z4WE7SIV7<7wIF>BZC}BL5G85$M6(2GdJw>)bnP?Y4-W4-2BG!};Od!c{V)$9@T0y* zv(brcDsJ!dbKH!9^ycCm90^d4btq5Aw=mZ!9-sS$lJy`dgc7lc4)-zkG?dmF-_lzN z)D;h038iA32cyn;03NTTmd}KyOwTVq8eSWL`;~1|pa5Gll?BEp&RN*#Eyb|C59zgV z)W~TPwi4=ImcIfL&dpqNt7x%%C$c+cWM|4R?dVH8qPMJ2%%Ns}@D*sGJHSvKr)wwQ zuH7=Xwx<0&1cm1I78T_8>{PUG-Zyo_JrpzhK+?n5B7C()!j)DBFyLBsGj1HngJGyg zI}SbI(VqNP|JcP#dUM5r1^49&{nrM<5OhqCFR+0P2$vNIN>DkV{=oeiU)-vD#Z0IX z69Vp7xW1dV$3iR96;QX(GB_L$N5*!6cx;R>CQ5_HL$@5bgPumEx-k(b2fLE)E^RO> z0flWl3Fk)#QL^xutM#OG-$~_rg3`{j#eu*F4PX#d5iW0re%4z!y1yg(4E!@hkoI{K zwKpTr+7RZ@Sc&SDe|zbjT1xs&%Z^7K(_!_$v@IT&ovm zUe)s>cwur(5TC(*vUUD2bHGdqOI*h?pi}-&wo8i6Veagze}Q=reHEy#8MH|PAkDog z?J4^#2fIV+x6mYrnVsDe6UWfljm)hp*)Q&GJU#3(`}zXgu3U*@ zz6?5k+;5}q%t2SfUaami^S$$XT)y^2STZyj3kVC0Yzc3o`I@7;4p%d+NG$WF!cnfuH?R{Y|l{ru;<;Ey&yr*?pYUh9U~YCjTKEYg2At$6Sgxa!8uJ$|Dook3q~vilC< z;R4W+1Qk{9e1!_zzy}5t+(l5kd^Ypg_?E`KeU}5WWKS2_)&su5fQ*cSik&DZ97F2H zg>Q>v`!>^{2y7v^cF`pp^m+c^p?KOsn6e3+2h5JJhNjt3cjhIyi!jdr@P@VniZ=W) z!@tE=OjHw`+sth9r8WJZw5DK0>sn?SqQ1~E6n8E~n^+8{vj#lOYVa^$-tn;B)OWAM zgOLxT{=e#u2Rf|H?v$WU0ctzP4M#XlAG}w!T`ilm_nL{`YhS^8@!AW_L-{=1{XH6- zIYlEFuHn;7!+h=#jSj;s+CJXmtRWN(YJ_cLcedq7$633EohH77(5;Kw!!#4BI;ahj zZTC6jAaLxND(IWq)KO;}P5Fc1P@o#DywAOV!8L8isKHLc5jFMATb*68#6ZPczwFf= zj&A?g_bSH~{?j6UQx+FkFgCZFrK4|;AQY|eI0V%-!|IgoMS&&hf%oN%nUNm5_#Ce#KZoPkX z+W@n&a929G#Kzqns{q%Xb*BDSFFYTUJX*TDH^6af_Dz)7ST&R2oUER2hl$vB#QGh% zv=f7&Lf>og9gi?3Ffef^HhA)CSBwwdI}M-AavlN9l=t8+*+eOjU$Cr+6-L@cmeN@J=oz0l<&lun0 zGfSNF9j}d^JG3}(F5t{=SQfUU?#ReD^+1bt&bzhicT#{cJzI`WP0coYxS>B%e5c3)J-lO7p*IOMkTAMHBgP=q0%i^<2{#@$D{ zTj8RRiH!xNLn=CzeM4mcP!fR@4m4vkkH4{F=XhO-ey>=SPaj~Ih$+UT5g)PFzYJDkld z^~y0PynuJebmm@2DEJ^_0DFM@LpOBTW<_h=wt}1O$J_bXj1ESH#iXl?uW0p_Q zE8X}7acOwRTzed@VwFHak3v?@0B!L#AbNwk{S3r-bEtYQT)N3X?Qy7|Vy}Q=okQ=w zH<+1B&Vf%=v(_i))63nBuE2<5J|j{`j<@(;e-CbfwsHGond)pWo&SKD$9Nbw2DFwP zEC;TD7)0#}0aZ=0re!A7=n)f+sz4rVYixWwy*`~^js^+^`R)TwT5rD6uHROWKU>EDv&b97bd>@C;y#?Cm#0nvO{TI}> z&{(2AfqXRBo9|A-A<*s}_BjxItuHFgn4zBytej0jj}8KSI{NpQ+WB~tervLu!6;g? z*#z^o(2Vfl!u09*cJf5HI~G3D6}>&?o4@{pkDdQ)X~yrV zn9TSTH{LRN9N*|VT^-ACSQogb7q#n-W5wqU^T`{ce7uVXZy!{n&AfvR-quU=mKkkg z{qTC(yHgd>tEU}uWuYx;1DaG+<_kX zHxwh!w)8eaoU69G`<7rS;tD-etkZRA(4(-5G2;WR&A+NqrbPo&+G4U-hwBE-GEpmx z>0D^3GP1y({Lqz=5m(R_L725dJyL@0@sDMw1din$FqFt=QL+yveb@P%W~V$-w|ti~ zMnf7vAIQ3cA4SI(_NNM1eV|opgngXO{LoEC75#@NFK7P>D!$nstOX|l25(Nig8i!( z*S6Mf&w*!%xAxxlCS0)9*SD^7>5iyOR%$gJihk8vLy# zb;`wgc`c_aY`a*^Vk@$d9^;^8!38r;THzLY}1c|z^BK@m=0oE zNYrY=mV+|{)XV8PJ(5&yqi25;hr+e6ZQ_A>@XE5=uRk_UK3t=WB)`D(poO_I|9maU2wtvyNweFQ%oS^<|KIN%l!xQBV ziOTxsI{0y-1pF)${BBP}O+{t0WYi4$F{|>5lJdetymac=L}F20bE2}PrD#~9w7&A( zL_RlBFoe%#=6uMfqSFYEljQ)J~2@^m~sMLwmG? zcKk0!$o(j!g=tOm)ypm6-S9SMci!}hNBZ%#JJz$belQ+HYzVH^>5Z$d2n-(IaF!V` zI$)!5${PT0@$ZOiFjaBiu|!w%Ajd|9Z!x=dicz!lDNoiFj!iYUz?uHkLCUS~hkfnT z3j&(p%Z%~V!j7oNJh;E??C0}R&fJ!97LD@!&b2r%8t0b9Fv0%DqW#rYoSVeoFl&S{ zkg8aex&Xonp4S^L}yIbUZ_*2PMmmLY|zXFF#MWhV~36yS~#SrnZ6ZWI;?PL z;fSHb1}mJdz@JdUBMb0*GTy0EXX00VAteop^By?U88;3Z2lK7wI=gy@J$e)WieD=rxG=GY{{$j-8a|r%7 zf#siN`Ku6z5B#i-<&Io`&t?9xc)f1_LVNwi%s>bSC%jFl_y7j}K;DXN!_@9=Y`;57M4D;c4k}TipyV=Yq-RkSlu<~27T)#5Q z^=p|w(JlW%D}OWdhq(MI%SSlp!m?7b`d&BT*Z$h}4}&lGho7&(V1a+OjiBGW)*d7X z!{Sop!Dle&)0eNbJ#TrB)}DF$=9s)KDyuhJ$~#fz#r|8#19UcCqxF|NjbtHzE%QmH z;ny4fX68%!-;ssij~c^cQ_FYyZ%h{c>@5724E)2OecxyMJOJ`hR$YhFmV^59dSDNr z9>|#iy%heeo|;JyDCYx~^LnpHIlt`1a`30g@FU3?!S?h9P+xt}9-hxNdM{%+OHi1u zb8fBSH!xrF2`iXSK2KkNI$iFey{+SV91^sl61KNJlnuoo?^e!R3~67T*@t1CK5Pe4 z9@Oz+yTE_&ITHWVbvj$s33Xh?GOmHNuN=@#?!_t}%Bf*FLm_SCIR1D!^JlsI+^I%) zcQRiZ6C0U7-n|~`&5%r-HCIh?sE?gl z${4UKWP(qWzG8qEvpr=hV?N8c3vb>`vAnJ?MY>MYlHFrVfjBY&dl^DWF@=qrDo z;d}6V#Q4D<0k`!_8-9%WBVB$Cv8f-%GrzCPKgIIrXW(O-FK7Ni=tFeped@N^cC(Fs-9Wuj58iR$*m*R>t=1V@Tf%%dTTfzKltc2D_ z?tIf9_b^}T&rQsi`g1$;rT)zCg9S(1JnC_rV~;&BqplSDD&TxF;6id6IpXF<-Lp2IfcU z6YIHx`H~*)Vg698gzhgVhSYo+N zKME_Mepy%;mFRjsdoo`E&wB+X?hx-ko+0{c4sn zc4jGKz&=>Gr%t&DWsK)M(s-ZG{P8Hrk)hWlmop!Iu;DvvlGV(YY`KH^(sP!+dG1Y+}AtzwOMAvNLKwZ9ICO#4lvNB)^RLMW{rrf4*md z@*9{hjinXL7yF<2Qe3o&`Ej>=XAg2a^NHyC@%}2czdv5+d86?^&9Ad#sF3+m|CTX- ztb4r^S2Zx7T#}aW+!I^Dd|IYxe%zLS5A%<8E zU(&%`=1V%bnE59_eT@!z4S;P}#r#Y$eh24?+7Ck+AF!=gLAnS0nR0T~1dqDxL5oq( zqflA*f@?^ET{4h)vDnDaq#>2lMxXwCOV^ zrt3!o%k##&d?(f#!~Cc^qWsy+mu$O*`I&5c73YbvZ7gFQ^Q+xDI(x;hGGEg3PUcIo z)PMpM?0J&>am<(G&t<+O|6=CT5Z2?)x!<*l`O;jtj`?@wKqd9{dAHC80P5#e=1Z~2 zPUcItKHyMnz2`~Sk7K@c{aogkqYz!cT;*7>eit*JxQx zb?b}P0z0$t2ON$UdY*LsIOa?BpUZqHR=hZi{8d@xugfC;Rpw8^Mri$VoJ9T4e5rp1 zP{T8v|H{Ij%Y3Q-E@uA8Zv8H>^X@9 zipH{?*)xxz9fGx^CkHgXOX{(`I3ItF+a+FVo>i_nJ?{! zVjgukQr_c37g@-5ax~1dJ@6XsODM+8>yHyI`jb+cNMQfd&67KsIvn{m%$M}BocS~? zbs3JozLWWqzBe-8*Iwsd?gz}5?6uF)SaCQfi~s5J*p@v#^oa#oA{XP^hWPP`4_He>w9>l;-pOPP~3K^QAg=WZ=i( z`Yp`A57MR$xEJz$TgM(0z<8ci$HTJlCox}|N2{1W%&mj7A9NM-X=_&NfbA^?*L-!AHibtP)L;f*Y_@!C+HCgz}nLh$6p>6zJ>!^ZJtuG{ z^P}cttp7&lldIGE;d>+0e&$bf<-u>LlVN$@^)5g68l%5sm@mamrOcOnSPk=~ z@w=S)d%NZHSq-e;oy?cwnT^ap(7nD)*$+ZdpW|{dcKP#XQM^3wCW!dMpFHn70{&eA zzgfU<5%60DTt5|J>x3C|%+n&ChjABvaH!`x^{4m=F!b$+>+gI!&}(9REKvOLZDQ|6 zNMruf2{p?l?+L~)VqDAH#`w1xpTx93R21dB9>DQyo_AZQy>UQtJgZMPL!S3T_bW?8 zIcq}r43hK195#Iwr}7nf9_^f(>vbZ}yE7yQd3oO50)CHx-y6a&0wf3OKZXm_8NS81 z*1yIdWSrt2xJKizFn)YM4t~w^exkE6SCKsL0RhKS!k;|vApw6_z#kFt#{_)6fIlJN zPYU=`0=_}OpAqnl0{$xj-z4DA3;3@E{6zu(t$^#N8~uHj=e;cCM{5ax^1Rms{0#yB zy@0)mF5r6zcwYhEOThOQ@O=b)KLOugzz-1c0|mUl zfPY264-xQz0)D7~A1>fW3ix0FKU%7uAUL@ed1$?A{j}q`P0zOv2j~DQ90)C=^ zj~DO>0)Dc9PZID`1iVdyV@aqNqMgjkZfUgwrZwvS;0spRm zuNLrI1pHP3|ABztCg5uX{0;%XQ^4;Q@OuROJ^{a9z<(;>4+{8N0e?uqe3Gs{!0OWR=}SV@aF~m1p)t!fWIW*FAMnZ1pHM2 ze_g=e5b!Mm{-%JxCE#xh_&Wmru7JNM;O`6gHUa-o!2cxR+Xej30{)SJe=OjC6Yw1Z z{)vF^6!1?4{4)X1$@7nN^)1iK6Y!n_-b=uD74Y2ze0KrgL%{bG@O%N^Tfp}f@cjk+ z00BQx!21jMR|NbJ0gnmzKmk8gzz-MjBLsYqfFC8`#|Zck0Us*hg#tcIz=sR?NC6)$ z;9~^*H~~Liz)ukHlLUN%fKL?gxPYH3;3Wb+S-__V_*4O(Cg7(Fc)5Vj5b!eue3pQp zDd2Mie4c=xE#UJ7e1U*h3V2e$7YTT^fS)ViH3D8I;0*%aDBvjpUn1Zw0)C-@Uo7C4 z3i#y$ex-n4CE!;J_%Z=c3-~nxzFffD1^hYz|C)ebFW@%{_%{W7rGS4+z;6=pRRaE9 z0becP-xKif3-}KN{D%U*M!@e7@H++kE&>0sfZr?NKN0W;1pGk(e@MU|7Vvcf{-}U= z2>9ay{-l6ECEyzb{22k?DB!;m@aF{lc>#Yxz+V*b-wOCk0{*gqzarqT3HTcV{(Av` zOTga|@OK6LJptb);2#S3p9FlnfPW<5e-ZG%3i#gyT)&Ofzk15^J`wWo6!1?4+;{B3 z6q4uV(6L_N{9pl(33!2kA12_33;2-&K3Kqy67Zu1e29P#74TyPyhy-@3;0L@ zA1&Zx1^jpcKS97x67UHEK2gBq0)C2s7Yq1l0$wWMQv`fk2>&OXOu%zrQ|VBi=h4|l z9v*pKStviU^1SH+K10CI5ODn#=HPfdOUOT0z|R)&gn(BFc%^_>3HZVgel+a7p+0Y7 zeKMg2=6Q=k`ISDa1-wSUYX!V6gcs9JVEFjST*7zpD;~b7JBEEixVodUM+jFts2 zwZqmogsYtvIzj5=Eu@)oA0LKa)QW815Pmb|-!FteOE{gd_VHAt!7mtcIBI6prz|bw zP|_s7oWa80IKuI?tN$~TaDB?s@~$PEA84|$_d~+@0UrnNAeGBtW~ z!9n4-69_N$VfbYX;jTAQS>l9G4dt&Sd`1XgM)+AF{4T=hhwu)|f?f!n z8^Wg(UK7HvB)m3+KT3FA2;WY4eF#684iYa3;q`={AHu&!cuNR>itr0V_&*5Or-v>3 zE4yO4)dHNFL-=r4&l!c{{0aC7s7uS!VlODp}RwP3E}sJ@QVn)KZO5`@CQTq z+k~$T;d}0m&_f~oOu`=t;WrchSP0)t`1%muk6i6} z{F4yggIsw|Xkw@%JU4{jPI%7{{uJT6gz#Q_Vv&PF_?d(c3E@`|UKqk}Cwx>0KZvH5 zF(G^^;gdu7-GrBh@Z5Z?=hP5>F5zV%{7S-458?L`K0So@+6(2Ehwuf2pB=(~PIw}O z527ivB802=kXD88rzn3igy+x{yD)^GN%(mod^zEbA^ax7^)JL=8`bAQ!Y>Kse~a)- zLwFud?U#k{BMHAMgjW*2G=zVT@T)`kJA{8Vgb$$5;o1<6t;qAvrCFGpWC*OsP1;Ss zdffeL(r{nZq}@bb^l~3^)8y%9rF5Q0lTQRrlS2ef&N~7p=Nf^N^NYaAX+_}VR3dP4 z1`#+7;s~4uWdu%xDFUa#5rLDjN8n_15jYt|1WrZ~frsNs%#`Pa<4J}03&~KpyBe@r z^3*R#S%6b{UO4_#a>DVa!uf@10ZDo4*8&14);}DtDmk%GUWE@7@B#rpRKO1t@WTcC z2mwD*zy}HVU;#f$z>gO2V+4GNfDaY$V+FiWz>5TYn1Bx#@DTz&Qou(E_-FwiBj95N z{5S#EPfz*TdFGehf;hd+%n!o1JnzI1hCrTol7NpF@CgEbvVcz%@JRw57w}UA{8Rxi z7Vr`QKTW_V3wWu3PZ98`0zOT^%LM#%0iQ15i``yh6Yi2zaG{R|$Ahz!wVmA^~44;MD?tj)0#l;57nXE8ukk zUN7Jc0)C!=Hwt)@fTsk!S-`{l445&`J737(BH$MYczFK-Gv;~W{Rf3#EXcV;z%Lc> z%LM##0lz}PuN3fB0l!MXmkRjR0{&G2Unbyb0l!AT+XQ^MfVT_ywE}*ffUgknuL=0q z1^jvezd^un6!32d_%{W7rGS4+z`rfvHwpME0soGGhxg?$W1e@jkbkv+-y-1O6YyIF z{QCm_0|CEHz<(&|Y@FxZQ=K}tefIltZ8wC6p z0{)DE|5Cs=3iz`E{wo20PQW(_`11n(YXN^jz+V*b-w60`1^gue-z?xS3;6E@{1pLz zRlr{p@Ye81KzE!~A6Y%#1e4BuOAmASg_#XxQ zPXfMO!2c}Z9|`zh1pH$G|EqxiO~C&y;5!8T9|Hb~fd5m#cMABw1pHG0|4hK6H~L}! zKPTz}*#8&sJOS?^;NkmgSRU;E3;BBqcy9sURls)>@ZANxkAUwX;Cl*qUjfe-@Vx|l zZvo#&!1opK{RDh}0Y5;%`w94g0)CKy_ZRR10{#^NKUlyI5%8FR4;1hM0Y6m04-@di z1^fsBKT^QM_np-^8Z6{LO2F}r(&0~@cZ`6C?`x}ahYIe1(92O~Ah{;MWWI4FZ0nfPX{4zbW7=1^inA{_PN6 zvmgH6!OaFAh=vKD%K^_=fN$gsQd>*{sQOyz(p*7lptrakiU}fJ54>~Eax&o&PoBl zN5FqB;LihI;JmN+GA_6O0f;|u>T?j|ivW+(Lkr-yc=wq6I?t^n=U#*F#d4k^{3iz2 zJEX4@{-D7R<@^KsVSQdP_)x}Y5+2TfJ>d_T{D*M<`vm-X!XGyIb-DWcOwUf^vtat>zwyhr%^20xzhe8^DXy=d@a#s>o))$dmd_AmzWqrKhV zQwh*J7gI?7AB}u%{|gEK$l&_U-yMX1Y;awl=Lz3#a4qM3z_HxdcTwejJI(9>q*X&Mh<+X&w9c?3E{sI^6w=4sZjm_F|6kfgCEWE zPa`~>{~W?MnEbjvtpdJ^@Ml8#pCQ~H1lMw267u&Pi1pkUl5;rWpBh}(zg)=QMEI{v zeqHWz0sjHu*k8Ns>d{9(DZQTZZ#Hs{RBZ2c!tHyfFJnBX0PAnxJFWf3L4^0p+)KIK?^FH*jeHnB^!YpC{SB_?k>Q{yl(UNEyxvRY zuK>KjJH+JIl(J zdVT?Ltp9~PfoMJKWE|r@{OC6i4mtwMwePRi_BoPpd|Fn0w4Ia_9`*zC2_IteYdN)q zA8T+er;YINcw9+%k;y-f%YT>f;RYYi_%Ij<*xt?k6tINxGXXE~HW@q7ac@20;dy*H z;bV;ay;)8N;iC**#`tH1k2Uy0#t$2W z_=!f2j`#Wx#&W~sVg%vpDPt6^>$8aP$ssw*2)FOI*Y#OL_!N_0_wU1mpKfqHj_`LO z3%qFt*ZKD$JRGl<5?*fdYdiTG;r2b);O^+Nh43>>e%-EP!66rTvkb1s`;~;BX>gtY zIl|8}_=PNgzhh92dI}pMP&<7Rgoo$%7Qz!oPG8QyhVbw_zMk;;q5OX)JRB#C1jQD3 zl_ozlpFWESUu^JG7{8J5MF!_#@BN%``@ZW(Isc(xNCjTf4b;llNQ45`?0k? zA0oWo$bql)!Qan|icj{1fmz_4ZSre5a|u7s;B#2cU4)0@hgS#>`!W38x&p7s$Qj0R z`WGUuo~B2rFXJZ>Zr{H>mhn2m!{hW?!o%^vI>N*L@E3%iAF9tz!dnck>l1^8bAfk( z!F7G65gs0=^@Lw!@}I!<|1shA{oZAaKSTIsCjZaa6CDAHD)6o_xR$e!@K%FsIad(A z)ZkjqYlMH*;ClS_ACB^u8GJm~CrNml!L{A4CfvTqTaU-b2)FM8hiR8S{~)~G$l06m zqeo!5_Wj-!!PizH~19B zhY)Vx2M*H;eWnur4U_*?#(zZkN`q_t=Z!-7-!`}&cV`g3%HTSGBjMjQIEbas9fa>^ zaNXWN5x##2A2b@vonUaS=edNhHn`SLE#bEqTvm-y__XK&mR%~vdORQ@O{Ed46f@lV?5U9cOm>GayAkEuEFDsA2<=^gyYE)!ozXn1%y`^Ift;E2L$}jgl{wX zb-x$EK!}PnD+&M5|pYIVKjvF@-{wI@vf6o6I;pZ4!+rw07c!6i%_r53Rzk~2U zoBYQzeh>t@1>Q#n*W-N_;U60urUUw{ApCCz*X3>|e22kxzwadcJ%b;_az>tt<-Tul zE&qJN&oj90-wlL^X9kCEqR%wK?R)9BGu}dYULJq-bpGEE-qYY(Ki+95e^-O+a!(|@m%-0ufwhG1 zW^kSVR>F5TxGr}K;d>ZdkJE!EW4U`8T-X0p!tHzMwVdx0Zr@9<>;D|#`x`mBK6}6f zQQ)b!X<^lM{l^g=j(aa8`~V|I&m(IIKhWShe%>1bp#rbJ!56ZAh7xYyS6{{WEW!^l z`S)Rb72x@tcFElWzD~e5lKhyFul3M>D%R7!_r8t`X(jwnlm95jpC$ZogKIhO5q^Zh zwVeH?p?v$^dp&-~6MmG*uk}z(_%R08dbpeLAqLmw{)_OT2G`{#Az&-;?0fHZ{eMij zeeb=N^Gm{qhvfX7@R0`B^)EOb<&QS_eypFxgooq3?+~uu(uP^bbN){V566LHr=$Gi zjC@_sTEdSvxVFO|5pLg;ukEmBIm$W7m)mOw*8fz4>vE4Je6qo{{6&OMG5EnOeU2hT|j<5R6M)~3RY8K(MjC|-m`m87XP1C=+ zKD&T}E%3r|R4L(S8#x!SoR0|)&*MWNpe^vi@lP$`^Nk#BKi3d$e@9>#%l{?el_tN& z`#=Ct;DzN3COm2KYdcI3zR2K1S^l>OuQvEd#(z!txdzw$vd3I3x5nT{aQ+hruM6R| zgxlX2(D|<=-2T3RuFpEcQ%24xmcQFPEH^y=78Ab2PB*-2U!>*7M*5%2#i>#HOCi?Y)F>`?~{L z|F09CHuCr8{GSkhjls42!{?*?|TlGhUdG*gs(RFwLRQQ`1cI1?csI8?e7un$MO$^fT6(qfysX`C*K(dA{7!>wIUf;zm%+81Ly}mZ@Vrw-_>WEgNnFpn3Aevfp!L7! zLX`6plOMtl`phN#fe?NN;SU;IKUvX#5z2YU;4uBs=TXALetXbj%zviIukElB@PVFv zfAScXbCZz&w}j6za*pNvdqIar#R&%jUf|6$`L%t{BK%>4YyGzo9`@%O3Aewqa0HjT z+c_xzQ6oq9?+GAwcZY^!E&E7xXyne;r90zz-;LA0O8^B+p8AkykO+? z##mA8yvpU=T^dB3E?^Q zm_Ix&8VP^R|{Vv3E z!+0IxVf+6D;r4efMsvk_T!izBeQ*C5#!m;lz_Y)9p~u(lgxlZ0(D)w-x4(a({e0|V zlw*JYqMYTo5^jJ0LhJKR!tL*0fP10O9|+f{x;$?(`>S98-&~6-OzsT6~gWBZk+9@&yXuozWsd-EoUa-gG2H! zCEWh*#z>a)AmR3RH?*I8n{fNP8`|$pYsGTI?Y)xlm?`%pmj5-v?eB3M!uTVE+u!5R z_-lmQ-{a7D&Q(}$Sbjd?_IEl)u>9GC+u!LJ#Q66Jx4+X-!uSV-+u!Mcuk@L^6w9^0 zGotHr4&nCqJG6hhP00T^;bA@ek?^n{_PiR)ozR1;Vcy6;h48R`&LiCZK8UW*{e;`! z2RVoZZX-Ng|0BN|UG74{!}4z@JS=|;-~&DTdm`F?hAcz*_V+#}aD6Tz-2UFjv5Y@U zxc$A4iHskaMmhHPKCWl{3c~I0eJo`B1HcD*_V-G3eGa=OdOjU3;PU|==#`p&(e`r} z;r4e=v>m=fc$LYY&-Llo7F}*3;<=p3Atwm=XA1ZNz@zkhAK~_QPY&iX`Yy-(b4@*u zV*E6~2YM9-*Y#gW`R(taX#eva!tL*&9LjQj1Gv7{X_oi4fbZ6h<=Wp#(fxG<;R{WD zw4Iy@c+3^7(}}arDlHn5NGwV%ZcJ8GB`PXY)k~6zh9&V~g<&++=p}^wz?eD~uBdBH zRJOE`%w$XD;)=RO$wWmR2(GG%mz_~v9n3c@k*G-3*H%|bIfji)B!ZF>P1SXaYLfA? zuso=Pl_%#LHabFTRdrKCMXGXfgxq08KDmLM<*8E8VRe0VU3DtZK)kRjxv-+SCY7kE zuRJ%=P+wDBc|jbrjY=dMno>1mlP$^0=2Ufk-LNG^rNf(&=VjnFR5Z%mvA|6<*JYG2 z((p$>)tXWjDbPJuyt<;M`a<|tUq>BQm265iHiJ39_d?Y3sio6So=^&xG%rY$BvOqP zbxjNF8*7JFHZ)*k%2Uve%47pxFmYsjiNcD8EdZ-JKUo0{Pc^zY2wGgxn5;s)WYmmg zT~mEysbfM#!xE*AzJ|r)ldROj@U<=*jLOMSRH&p+N;Xw?s{~hoznC+Um8ts1#Da>Z zWJzs#MM-Tu4)tcVy0)Rlsk6b|+*Tu_mXIth#xkrxy08X{sp)hvoZCv%#fS`U23#?H za(Pigb)&wDnYLA$!Qw2VXoTs5%KGNIR7oc_Ur|z2R1PB~Sys_lQQH(ZH4iduQMwMk z3qeceg^4&`U!E!&Wehyon5e04N)a=nfsrz7QAMudO=hcwGEm%-czjxUqGU#b#wQdS z+2IyN^Ke>Mo7J+MYWPJ*nuH^|Miw3HH#A{$WlC$-+C(krtWX&RnqfGrwJRs9pbNM< zL#9Yc71~wUqAjEqWaptGaPxKGq$+9>)v077%nXesWt07#Y+4D{j#g6yR)f-^rNgPE zj=14<$@3ExHQ>leluwYUQD^h(>V{0kCsrG>-*CEit!j-P%vKUPl*rP2@?B@DjXv{m z4vdIW&?Go;Jtb%d={iE_DaUOlswz?y+PGwWW*9SiOms)A9WPGNQI=F&Zj6VA2pf)3 zwd*fJ?gA_Dar}Wokk>4V*)H3WDSE2IUC!;7C<0!Y-WD zFl`1qNYsbsVvys85!r?hOeM9*l@RtVd%s9s-dQ$GU?U_ipt#YAXkO3pn(cKJA9GyjbWvVV4QGN!wa&`mjmHa zmUh_a(}rpR+Z}-lOD?ESEKb%mBpXXgs|({$1?`%qBJ*e~(rapq&LPUeyCl1AO(>>5)11G{#WwYVc!cYbXdHoiPHytcXy*7egT!_+qn{{dQ5Skcs!Y)s8)Y))2} zO|FDheYMw=s!BFCdJ7vGtLst=Rk48`JgPO-&(Da=tm+8^3_>~BV6HzO;vKsi2$h4q z;tDhK;5M(?Ka3JKNeP<)^J}nVl0x$N!mhdknE$5W{&S)+xsY~gjea3?>{_%27feYk zN~RLcO-U8yp)A~Ig}vX(#pfm#R#ewOHF3kVXe zxXPoZExji+wh1?GF{;AC&DMx7TE@TszGe)p?5Khuy$N?%MH~McaKLdNG__bGu}Y& zheqaw>9|Vew9!~yvUA&w@deW-V}uv6(uP?9Yf?UuqN2%Gc2wq0?QK@EW}Gf!)l|2{ zYf1SdDsdjbZm4NOhg$@9@*44;Pj#xPq!Lz9fDMPWE8IG&hY2THQCm_8JprK`OeJvL z)bje8CM^r}Qd3gdIVq~O5C$3{Wv5IpuP@h|>beK~6`ra)el%338tz<$dW`pvxPk8E zAEEM2X|bV!3t{&(vI@Pf!rak5suI+bO2z}u*&QI<7_!O|t|_)3H$kAkZD!rYPU))7bZG!wQ4bgX>(%)0Zd;WqB{WK(l(5-X>hO=CV@ zTZ@J>6-KvjY{VxTH>Rcf_~;GLx*;1>?{nUz(M1L8peH45BqRM*_Nzr_f+~e?>iFC2 zc44Sh+C!QC+PrE`~RRSCH#+WFrUDObeyadfpUU)zB8 z#HN@aL#kioma=I(-jOn2f=!FQG)qCEA|j>Ph6WVDxh-F}S2IG2&*7?|m=BC9EhF6{ov!EKYW}@aowUe}SBQt#0Vh{^)UO>EW z!pt&PMcm1N$N}Y|s{&WoT-#g&uCpYYyN(bPuSOYgeX6>lraIY#cjjRpg$2-rMETeR z@9>4H@3e4QKsl^GH}0{Qvw~cv+ht=LlTFDKo<9io3SX#Q6g!KEE{>iE+(NoK8)V-< zC>q66kJ(xXPj6V&a!_%KZsi8e0c|8~pEgno^Be=}q7_8l8Hk=vC;F?4*T}TUk1iK zFq~WYG$ZM(bWT8j3D!M$9XN26wLCQfma?d<_1BJr8w*#QGwN$s_u) z6EXP76qd4jvnUL=T`F3F+b&^`2~!LOemqnHGY5A7jDEGNFoW+p;_(AD7;VFFtxmi6 zI#7q`sy1F5hlN6UqA6Je=Zg|>n5(kB;R0JEjE~4E9;)SZsE|pn7VP=JvbDL+p1Y}{ zf)f?p>(R1W=u_*zr$f_V@Mb(wU?YKmtuh1dy(hpWryA=oC_(W{z-EdPBpE_dGWnpj zT9E;z@kq?152vu2n#jVUWtuYW?GPS?&~+pAT^n8`=yfv3)sf3(8WQ#{7i49F) zZ@RecW8#;t@0S&VWjBIxRy1+Q;uv&y*Jhv~7*py7dRCY258X|dgDs#IF^-Rqz~EO6 zw=ZT*#+!sB5@nzJhqj zV0ox%>$sv132tGok+=Fg@at7cJwL@^ zhE-E)pw~IT(F-7QU+y^;J+n#;gfk*i1NEidX<#Qd5`qn^iEbd*!fD_rRtg*(=N(lT zr?5U9TUlQR;Ru|$FR48Z7UwYWESU~N6i%$Tr^*7+jCC5;vFWIBG|jClCai|X0vb~# zT5$F?HO_qI3eYm;o_E5cA=g^9E6Xm_^NRfPg!GFHXG$ph+#3`;}Gt`~Zg-(Bcs zwglKYjDK!(LLt%SL^os1Bt)Cx+Fglj^YRp%U`^Nr`Qq#*2yRa@L6+fc_d52;%z%A{ zGgwC)=0%5?EV;Xutz2|GURGl)+lgPbUhIumZA;lqGwQfFn|;maHT#Iu&R{9B+1GT< z>@jf_r*M&K#OrTgJfNdNB9># z5jw*^YA%o32nU8d~tpeY92>Pb_(%A=d`?$8NT&yAmT1-QvV^|V58>dsKTK%CzxnVoC{pXvuI zz=yg5$NFK&7bYT3{KHNO99tNnCQ`5^++9voFYL@r95tI3Cd6k8ATq-_Hc<`}?8$g~ z1A~w5SiaqMDo>5(#W{pCaIAQG^llJ(9km^#;+^Q29ip`1quu@Bx>QAV9Vil`3;2&0 zA@Ce8?xG3L7w9F1nG$K?Nkcu$6}Tu-XCmNKoH~day>j9r+*<_Tv&MusmAlu6WjfHH zJA#2Y2$<;Ib>J9n%ofescspo1*?2u(-1{ooJYd8urIs_~d^5ydu78O1tr3!8{?=0# zDbSg+y4y?9rMUb4!CulE>16J9BU^6?I0_56Ju}#R}-k{2UvCf%cCROKK=!t+% zJv8PN8rc#4jV|Muz)*3okCqJ3co??4`3+U`ZwGV3GjBi*hXbsI(t5%t z7P}L?dGK`t+FXO1d7Zh|+=I1#tFz_0yCP&#?iO8S* zM!Q`xL5K_Q@<>(Sy_CSQKL{k?2Ah6tTyNEZ)fSDcs$R%9U=nom1;Pp)SzxGcu7Rh- zlZ)yb;r^oknR8k6=+Lz|m^o@Gf@AbwSZ!fC7zMWks^Fm!b*j)-+Z;RB0b!)VrTA2s zGoxi-(5iPw_`Ili<2aziGO!OEB5HR5_JBsH9SCrO!D+%*Hs2edDWj`5I>D;d&8V(k zh(Ql~7rmDjaCV)%HQ+y$*qK`c5!>I~3fW`9{=RK#k{Y`?IUan<#<0I-;j}n z-Ah;u(TzA-6!7US|lrjs2~6M6T+*CAHVkq71@VorDJa=KGn@i1J!Wa-Wbso?W7 zobT^qcj~%yy*ndn7TKM)1h+eDeBGIC*$`nmDmXP$gydR$!>p4x!k~M@w~gr9?DY)X z>k8O3Cv_H-c5w^MNcrb``v(0J%oPlXl5P zjqrjRT6kB$o01mLv3-6xj`l$lFqq)g+Ujw+#w0)LMvuKgiJFUrLg-sdn<5n1WovlU zo*C_2j8WSpcnpX(1?bdC#jp~rS+|7xZ$A{_t_W_A(A_cjAXYZv5hexu<8=E$Y2B{? zT;rSD4w0?lI~1x$_7+9tHbd0i495fNA%h2D{1(Ayngxq+cla|~L^pakV!VI;9*z>K zdRXtYcfW_DWM}GO*o4T^!)i0Wb3GifanE;toDKLcyvfC^q%nMn+-}#$Phk2EZ?~%v zZ65&ANA)r<-_7|El!=Rm=+~%ZnHi&;gL2=zon*EoGtT6l+LDAJ;JEyn({Oh^)7b|- z)S_HIU7O911kMTAFoq>NF5gj8U?0Fa4~@5(RouW^8Nq1+HfZ2U-xhcyB|Mq|{lf8( zeKHziLwE@%IF82pMinB$3Jvyv{M(&+gCE|}h=*3$y@5A%w%^>g2Wkwv{W`AaW7$xZ zK#jF-r5b0hQ%x-oNU=`hzB#^=2R4u)#89pr6dzg*cW6V)Q*@`AscwyK#0?BTP2%pa#V4C8trcu0EB(hpfRdRyQS5%?2No`Ypw`T3c&5HIe6_$37%a2H3fAR`=g#3rb_1IoLa!x98cca1SYkv6c%E`&+NRgvzm(C*U+2(w_-9k~Go zgC%;k4U-Y-&1p>3tP$bHmG!EQcACZ!R41DTUOi*ebUYSnnzI?xg{B4isn!j=Yv}~% z65sw=wu1{1EArZE{S+^(#^LA+94#Pc0qf^rhr!}4i{8O9A(W&Zf>s6uMy*k@OC0G+ zgqX+K;-DRlvdOi$aiCU=pmv&Y-NMgr1TUYZ-cgN>Lu`x7)zSJ?2no$JRJ&7%iV`@6 z4QkA{ArjSgbz(ELS~iNOw$r`}L|fqVw43Wn$OO8+>fpG5gnoC60zIbzqDkP)1_;y#K%thE-3789t`5fkgo-SxExS`UurxextoP0?1 z8%xy?@A7Voz9k0@?A&fkzy$|xV$l9Ymt1f}ud)lydUq${QhnYM8Y51{X3~Z>tKiL< zJOjyK;9o(Aq{sNFp93>?WsRspBXu2exU4m(qY8h;gRyE5)uK#Z&Q zEgXl9_w}MCsBYW|g4e&P_dj*6I+Zmg!ri7asAYH(hTXOW^sdrQpS6yd$>GKadvtrC zu9FWeL|7rNkk$O4rFCb(>-|y;(noap1;E|&2gC8EobYPS#QD{!#dxp>{^2tjftBG@ z2sLmSxVQqIWNBKMQehLQ0e)NJ{}uNw@KF}m+pD6YQbk3J7BvVWDni0FcnwOF2tou@ zRO}`sfkbj^5+EpQs%Y_!iWXb6XsM#LN?TiLi#GL2l`5^)qN0t8iWOV5)Y8`4@0@qe z>^?htChrc$|M&hNyZh`s^PD+z=K5aNS5|sZhDN8o2N{aHw|J+ z-jtkvmMOI&iPj^$UE2BXa50@)G#uLREvCMK?(HU%sR@{9j9bKM_s(Qo7Vrg)qOyh? zrHl<eqVg3ZhM@PUe2-q3nQ3O*Ud<%S1#OpIDUe=y+kt;ab64UgnpK#!t<8N$ zV(AM`D`@ELvfl`Z79VtHOZqng$Xygtc0VyWYYsiqp{g<(@N-k8i(xuHzZ^Y;m8L0H z!~BLDT^T!>u0WH4JL_v6zPd~wPOGOc+&4|nt{l4N$=x(lav$eU{{3k#?O1IJ78PqGnvfEl?i`yb#W)dMJ6e8=%CQ10W1Ba_NamL zkVnswr@wY}7=I4JblK?%RkEOjmiAfe9dbO?6`jDEQPVQ7hAyEKMsyNuDy@2C(GyMd z%vmg&(3pSG1rHQCkzG9eWT?@-NnegyT~!^Qi;Gl8&aA1eO-7@2deeX&uO-_IsjoAx ziIA`P)O?QpVSMv27yM+50rMb25S4Zsmwk|A!y_d>AxBkLrTL}tIb)A!= zKn}xWdV#?wUAgJ-EaS`VcDebJnEm;bFJ15d&YKx05^Wjf6e;a9x6nnlm6h~_Ren-= zPNRN&4N`ASCuyR3gZV|YhKSqlwAWNP3w!Fe;(y`Dwkw`a=aAf9%9UHGW4jUgL8O!s zIr%czT)Ws->hY9$S?HDz)e(>j#= zlCEyR+f$%><%>{i=`lO}D|B=Cezw$H$DJ?Y6{I=sFY*&ns2sEXQv{^Fz#ezkKIW8!V-3Q z0Fl_$rAQCI>21a@d-UZKK&jD}0)fIuMx}ezr!Ucm8G|lQDxz78U(ohK zxC$XR1oy~YJWfa_O3(C6b)vZpt!D5xl=w2krRyXUb-dx7E?>&d5o1>A?y2P5y4{T5 zoQme0)Uo2ZZtOdHKUC0Gps;ehM;G$vFf}Ds4tU zqyr{xBUL_J0%KEjzCo8ZGb8)Vc=JhMs+?s)Lexe&ReyK|q#d_Ij`7h03Z0Qf&b_%* z3IASV*Q`gna&2QLfn8;roV$n;)HGxsA+C&u+M$(Doahy=Dru(y>(6f zL>tfi%afJV<{Uj;x%}$-tRi2Z~f~S1aTs6FeraYz-Nme$~k5tpm8@0RzoR!GLwfR+3(yEc);Ov?TPVcyNO@3x>T*&iI zem<>@#+Y4(QTC=SUS_9u7_5`E-zhbY%4tiBf|GwdRk@hX-Nd$oYuX&Xho4Wgu$+7= z-Ha2ge2&8tsr2-oDOtbB%4MS{rX^iK0MQAzYI=rsWpZ{EJyS3G3BW`m(S!c;aWwt^ zVAn+Av;Uc6@=g()=teJZ?vlp84)~t|Zv?yp@aqA83-D!tzYX|L0e=VZX93>?_-4U( zCOh}1|Aofv#NQnJEWtm8d>)5<_T&uc;PPU*4g>sMNIyz&t6Lqh6N!@qXMes2=}&`v z-Uoa-;F|%j5S(AQT$f7zvmrg^b1~p(|4o3S{dWS6x+?_d7uLO1bRUHDsQVkhF`wrJ zkKzi-^$Mi_3*c`9j`sf(aJ2t$;_|^?vz;sW106>R?)tw#aIRmpe=^``|9MXSk4CvC z66KH{^JxSe?Y|0ewEtGXQTIE7+j3d|-vjAU_ffzxpWg`X`hOjyNB?gC9PR%IaJ0WS zJCY7-r)_tKIruu0EHOZEe!+G(7Vr;A%l1DP0KNtAM!?xNoBn3N{|fj$fMY&C2K;Z3 z{u#ix0{#cUKL-3=z&`=}W5C(B*3SO|j=J5b@8yHPei70S033Bs^5ACz{yOAyKHz@` z{BjSz1n`$3{r3Rh2KWlVImT>xkKQdRR#dKk0zL`w?SRh&d?VmZfd32d%K-m3;P(KI z=^q6g(?09hxfTMB>DK~|>0bvN{qR2E=!ee$$NVqYBVDeZ)R=4>UM{%X zUtR6sJ!F1x1LTAK)zg6Q2Kj#g_}+l`-ZO0v_HRc6zAvO73;2G3p9A>*fPWWoOurg% z)ZG9$rr##G>!1HP_yjwz+ACcy^v?jmdx4#!0Y3oni5`45;F!-sz%idCfFB6?{|NA} z0{$G}=!cDfqn|$q9OJ5=OrR<21p0Xd;8?$t07u<&z)|-Kz56<5vL;4S(|GyB@<26=+1_GP`4TKLES}wGBrKZShIo)-Ye_Vstb z(f^5kOue{%?k+f0VFJtbHNY|d0>Q2QAB+8kp8Sg-AGH50C!cMS&ooayWuAO4g7oPB zIglRxzd&$nzg^#63HhM^uY-Kh{^fw9-|lnr--(6jco@=SJ}&}}_4NkenE(5Nb3I~t zKY{dE-hT;h%e$}Gljs?l742hSdAkeF@ql(71~|q;5#YGqIn$H>*^nN~RS7uef3e`! zej9J!^yGgHcYhrxhj+@1h9+IgzrR<~Sq$3uE7?}dP4 zKC=XOTIL;6M0yxeOKLZ@+hX?cK7ai;e>}O5}9Q(1e07u<=z)|;Rz|sDP0LOXM zvw-)9_<7rdZv!0j>DDVMR#dJdAfF+C^PZBO4_^Q{`k@VQwEq^s(GL#-j(&I?aPCXoo(^mnG^Ye=U$NBS*0mu3Ivw-tn zpY{K1faCo9uYhCz0}o950qbj&2cICg+kU4w_&Vul&WC)k{mumZD6qc?aI|L;;73FH zC4e6T_;SDp0sbK1nEw-i9}DSU0DLguZv&3`B)*#VC$4vTdho*l$96Xy@FAc(9`Jm? z&llXSmvRT+CiPMc`Cz?V2{`swHv^9C_dXB)G~k%eOMqj${=kFx>Yer{w(Da6$96pm zaBSD70*>u;65!ZQTL8znx*Tw9zqbI6ak2_<)LjEO>OKoN=D!YbY`=RQl=cty4~GCg z5aJ{saO@w(0*?7N3C>|d*`)fJd652f+H_KIM2J@fb>fs|2F}b zWk&bGHh`pJOvxgMKd zsw=YR=YSt3=~?$lz-9Q2(mwg9KkzQ3erf9BDD>z{lNE^u@{i+y#R0Bi>zoDbu+K4kkjg`dNT=#Xu& z6kcv%yF5+dwSvnuA%)Kq+}5Mjy+ZIl@@nxLe2e$wFf)8}kKO~=R{o4G0 zAvpSt^La|};qq$JzbKyz9sD))`*9&$b?{?rz-B=$K$$QvTdm zaO*d&&)-9wSU)qq9`Ib!{}IxU7G2KgHNg7;{wm<8%Z<;*C*r$!@SQz)cMrao2j3HL zTaN@Kh68>Y*gpbr?0;|`G7Hj=g!IjTXN+rrkAn2K0iFl=3czvR@-x5-A^qEcqwYt5 z<9HW!*z_G`kr)m6>?Lu;IHvCjIQC0ek65k|pnE*zKL&8r#e7bL^q9{{fMY%<18&

    K+35 zoCf*ZevhC8;^_a=A$?!SXB^;|PYK}Y=Q9Ae;|4(q^aJWnfb@t@1bi&mQwn$~;Aa7z z1bi~!Er3q}{2IWg0=@)rtoN0GqyN$Vb08m#=W_wact&0H|9Oxe{eM2-s5>3-{$S^Y zfMfcv1AY>uPXc}h;5z}1`OE;kKcue!9Me|=iht~lxP8bRD>FQ4&Ye7I4++D>2Z9i1bi;!vk>rkfZO>UK?#iK z&5#~-{|Pwi?!1%vg7ayG{PzYN=Lv@bJ|EJL1{}xzae(8vkJqDK4EfYUdbIO8z;XO- z2ORxi&!=&@(EhLR#FGxjG5Rz6fxP^J@UdI7d4%&T)Q-agKh#I7j<2 z&M|+CbL{^y&aVYKG0w5P80T1CjPvUtAB^+s0mnE;`!UXMfbze?_b>CvZM?-wx40tZ{TTJhp-{QPFAM(d`cMITWf<29ZUkLcEfaAD~_LM<-93OF9 zMtg8vMtdqDAG8O@Wwa-kad`>YkK^)ffMY%04mi#ymjXTp?8J5MRKV|m^fiFv`VQyY zIDbZaz61H7J<9+`d+r1r?YRqZwC7~NefH#m{ojN1xL>y%a9n5I4LI75_)y6I`;gu@ zKO6|@S3r7PXZ-+hTnFt0_yv#;`hN)E_d$AWUn>Ded+rAu?Rfxjvw6JVHC*7!QvDeh}D+>vxR9 zCm=npQ=S4G>-}lKF`p*^KLK>l0em9h&j5ZN;LifS7;wzz0l?QndW^&00*?LpbAXpX zKF3v0skWP_y**I^@#Iitj9MYJ=Wu&0mpj8a$!AQ z1nIFJao&sd*a7LW9^V2S>k;$8dc=IN9x)%R$12F*SC2Taz&OEq1>#tb_dxy=ARgWU zydH2&e<|SAkp33HHvxVZ;O_!{INQ< zF900d*Ea$G0Mgs@kOU?E0{Bu$kMpa(fPV<-^F27O7q&op92Y(U9P|Gx;Fv$+nE&4( zz3ra~N^Auj>+xg2u^vAGJePXJc8&Fj?F;MiN05Im^@#1OtLhQQU+gD+^LDJ4zk~f) zFNkBkY=iVz@BaWC^ZyiZ%pY;g|DTW^>wP=mIIsH`;5e`Q58xON{{_IC#y?H~Ll+mG!O?cW3J!S=Ez;J)^f2maX$(&K(k z55O@Fv3?QX8`5K(UkLue^ft>zSiuL&L8{vqaYvjTTh6$B$NyLE4Jw$NuvWz;WJ-^Fze39*_3;=NQ1JK>nzU^@8h;TsE|&T({!B0j^syA6&O$KDchhd~n^0>m%Q~73ZtC ze#Cix0n`i54;!Felz|_xJ>q--%XJo{$MfCTKMw_69FK+pJ_hnZd=lWpA$>XEBLGL; zksf>$;AjuxXb<9OPa))k_KXJH#y`0@aU9_2Pn_4GKTm>s!EzlB`JnwL0FL$}j`p7j z>Ct|E_KO`)QTJpIUIaMWgE-oQINCD?@23a< z^;d5|y;MN{UA51Kc4B{p>p1MMFn+MV!u}uoE8HK${tENK{tENK{tENK{;CT6lS_Yv z>rL#hPKSEI{%RcHW#E78uW+7F0_icI$974_Io2cQgYka`{c^xDJ@!A7!Ow_K0UZ0EseoggoDKLs;vb&JO#>YF zi_Qc5C`r%$xd3pqXFA}x-b5VF?<3v}`TN?-g^)k$CIQFxQU*BsA(wWF=M&2zAKRa^ zA7%ia2liLwfTN#>Lwmt|a9<1Usf7G-{a6J!=KoDk{H+Y z>)i2>9?vuS`bo6^Zm=J5ocChARD+$kUx)KX92X9N{C)Xgyj=?E(ViO6MI6UF%>N=t zk9OjG4DFl^>3#X2oi~7;qrv~r!8la}INra8^^5a|TJ>5Z|H{)0mpWMcroPP4EA6?fA{2rIOfv=`CvWH1swagd4Nv? zds+c+0DL~+3jxRWi{tghkp3P>e+rcAr+{Aq>7N690pM={ektG|0Dc+Zc>OBk2km4C zWJ<<;Pk3D`jxQKLmqY$I&uIfZ5A4By9^>;0PkO{Ley)V{C6NC@z_DFl1vuJ+?Z}sZ zktcuj1FlbTUV(Tqu@l#QXeZ)mC*o-5x4_Ob!2U%!;Nu`Y_IqW3 zW4z(K=o-j>at`UQ%^^LmFR)&49gX`TV?Y=2Nr3zINpN2Rb1;1#r~86>#*!ZGfXaKY@M{@!KIij#J+Se3;n7 z^Y&$cp9gkge_jpvosb@{55#`E4bo%(jQ#&zkk5UP9^)MS^F2t9{n&EAv0m;59Qy~v zi$V7uNIwy9Y`;D{57J|M!T#j?kUzF}jDO4r;~(=`;mHT%8Ry~mL3+&pe!$V64*1{~)hKLH%`|0&?uKRg0B_78|-|L`+NUoPe4asTIl zqn(cej&X(KFZTa9{$f5j?qj)@K>p~rUjmNx`zyc;!TvRXW4pllSQ(^$4ANt}dmQk= zkp9xb++3zXv=I(*FT)8z-FpCBSXmGX65) zXwM%3$NXOb9P>XIaLgamiWQ z9hBbga~>z@IUddh+|~=@TwgZ5Z@=px$Op$q9IrbdADmab1vvKGnEph_e=elQeKVY& zltFs5zZP&@PqhM$>n6M&(l?G{d5?g6Uf3n=H#^>Oy}S+fkAd{Rg!JhDcOX6XtD6AF z{_S1BCHp8o-y__{AI9hVgu4gFr!W0xPkQcuzH0+PGUR;zLFKYXp1k3-m`xGjyUBZg zDUWX)+xIfQ0PtfOkYgF(tg+i3=8GQy&i*-7((^q> zeAsdw`W5q;ubJn=;>!f*Yq>4X`Ba&Y3BJ~m51ant1I*_^4!&LR5rA_(SDWOCsep4{ z8>A7m0B-AwDN6vib2-LsTyP$CtY@6(kbE%CxX(V zuLr!40Xen;Zu?aJs^{(|0hia#F&N1Q+}=~kcq!l~Fd#=G;C2qcUo8gQ?j13{3UGT( z8sqB$FJeHBt$>fQAiegK_R9930(d^)woUNYqX9pXKL1HL(*PeU7{8PQ&gnPFcMAY7 z7K~q(0e%|bYXCnT@Qr|v1N>dUS$DJjPTD<}x5Q%fng_UD1M=5JfRBgtC4h7OTjaY+ zz$ZZZg@B(4_zJ)$0{#@>cFoIQZvwm&(s$d_eBg4W#&&w&AMnYL-u`Y2r=J3NC8W1= zPX2l!;Acbn2LV3^@O6Nn3-}hmrvct;FY|%zIS=sRfS(WeRKV>$r~Gvb;L{=f62LD6 z{1L#v4*1J}+iN@d>urFSLHb@j%m=o=9Pr_Q&j9>Hz&U=l%6HQNuYmLm0Ivl6D!@6P zgv=9G0A2;@p8|X);F|!S1$Z|Z-?&`WfcFQy2Jo?fUj+D>fU}+UIdrvv&xZ8Z16~XG zgMiloz7Ftuz_$S20C40Ag_yWMM1AH0a*8{!=@EZW%2>6YFe+KwXfcM$Yd|>+* z1AYSFHv>K$@LK?10QjwdF9UoD;A;TC4e$=YZwEZOxr;uto!RCAc_C=fd2r}Zv*^Z!0qn^u>JP|UU-1{!1zkQrvZLH;H`i^0QgeC ze+c+$z#jyB1K>Xb{4Kz_y!Q9w5(k<@Z0AFeJ`eDR0UrzaD!?lN|1se9x&55~PXJ#D z>3<6NTEHIxd^6xb1HAiJ%?Hl^=YS6c{87Ni0lpe=`#gNk=NEuq59xmi_=AA|3h?!S zuK|23;Ew^`v$y%cx{m`s9PnQQJ{9nG!0q$uIiKGEz68=g0r(?;KMDBDfIkKJHo%_- zyw^eI1M5Bm_;A3V1$-*tYXNTo{I`HF0sJ|@9|8P%z+VP@9pKvl{~h4H4mKay{uclr z4)}|JPX&BE;4Ogv9`GfA{{iqv0DlSamjQnn@NIzq5%6ALGauOgR{$Rl_@4lu3it-V zTL6C*@Fjr12KXa@zYh4zfNunRJK%2s-n)0_27DUe9e}q2{uba%0e>6t z)quYP_y)i?0sa}_?*iWE5c7fUe-H2z0Dm9w>40wrd;#De0KN?HzW}}l@DBms2>2Gj z6Nj1)Z0ARS=K=m#z>5I?8{lPtZw0&!@Q(ps4)`a4KLz;T0pA4pHo!juoX4+P8Lzt^ zW)g9^{sHL+0{$uB;{g9B;I)8n2mE@#{{{Gifd3ot^??5e@U4J<2Kc`K=W;blxq9W9 zL|m@_nzzy4;d$l*OB5y zaQlo+{=6}d{~CN3NdH-01V{7IZh-gco5ptq`~<-5Zv}A5>44jN85myxcz4KWIpBK$ z-VXSlfOi1C7vPD*S;dsA2jF>te+BR&!0ok~%r66cA4uN@xZNAz^veOazm33nJK+04 zJ{^GX4|t*1+MS$CDUpajl;D{tvQINhJ@S_3m0NmcQ#`z=$M1Q09Yu7-GmWI+?mqB_z1v@03Qi> z8Q`M;Zv))kOU3yv2i!hGnDKVNkAr+V06!k^#6V^lJ5K;S5AYKKF9Q4|z{>zX8Spm1 zivV8^_!z+30Y3%s4!}Ppydw8VLBAkiGzLuF@9yt`u;4Zyo2{2>2w(XEETVfUg4lEWp zfKLIu=P|5q;%X}3`GB7d_*B5p0lWq9a{*rp_%y&*1AZRh8vs8a@a=$K0C?{~Y5S)G zUI_SwfKLP5KBJp+Xazh8>6Zdt2KXAl%K`5IdrEILFl$ z@?9n1Ga>y#z-Ixz0`O|Up8~uF@J)bU1bFws>2l2md?4U<4$djZ0bU2`s{yYEd=cOc zfUgAH&V@Pub%1{Z(r*F03Gg06m}TO>8Sp`X+qTQ;CICJM($@k$7x3!=xA(?!J`V!k z3hCDYJ|FNcfL{!Fk9^L}*nbJ&g8*Lu_yoW&1-usU%K*O~@XG;z5b!p@*8%=bz_$Qy zW0vjiS-^jd{Z~Tze83k1UJCeCfHwkuHQ2%uLt~Qz_$W^3*bG6rR~2J@O;3R0A33CZGbld zemmfc0sl7Os{mgL_mfd3HiM!+8gd@Y8vWK9a=os`FcEk-!i|kDp^xMv!Uu6#goTRNhYE!iCjOguGA7Zf6jSuQ$y>G?JXhlJOd2@4BQ_Iw*IaM{K<141*Cuv=*`-Oo*Hm8=5+Iz}W1_pfcx6Q6ZaRJ!Ph3LB7!(Th(0g zMRhT+oWw^>X=!LGpHZX)a11YAM6SStTvDW8Ag_zwAC>lr3jTA~Gb2Uv`F=_y{WJm3e*dat! z!`hLNMk0|rhw@gZbFfiL-A;HT9Z8LJR#i)KPIHxM+*!Su<{7GHt~aNet>xlDwL)wf zDV<3JC9vO19>ci>yCzXN3!vAqt?`WFVyAsgO%{)BoHR8*+th;LsrDRoX)%47QWWyG zwY8vN3Uzu_rR7cKboC3j?dIfnL$;; z;sqpLTs&j)_=!`J1*4M5`i7>u^4ergOI1@jwV;xzNpc9e>QwTIb$nVtW8|dR*vMwv z499mtj+Ifxr*cgDTVkW8wYV;=81pswT|&>$GeVeK5E-8Cp_Anm6;;j6Ho`jX{wB=G zFXsQnjV)Of=)`44|4ZHAB&{_ChN$UHqsc->=UZNq5iZ4ApW~)c!aeQYjO=W*Q?69T zku=P<)Rfm!n{7(g)-<<7O`-G7Ld{M!pa4ga@iGKcd01p_F*z|wv0a?iD>;3(1Wy{} zu|s!Cop)?h#uVM;)ClDtFLdqNoU}7~(Tq?|*OwG{bmN@nY8pOi9yWQr+HU;qBAuC} z$;hlW%jLFX6lq^q7}H18rDY|f;>OxyM~zz+HPShCb85?5s!DK1s`P!ON}k0voCon+ zY8q>6s+xIbK|Lf*zQ-h|j80CCALzgvlfbi3JQ{7vYiSOgY-uX5X=yH@W>Z&OS3GnS zO?Dfb8XC)Ik@3lfdK!Wo=O<}iUo$HjI&9|`pB2rvr;x_H^7+lls&C9GuT}aWr&`kU zuiRwF_g!T%XBA?xgk9R=j9B+}+t|;3fxdev4az(%wFASppYZPL`_WM0;fY^thjm|MjQ|Ei{{`iiP(8ZeyK z-|q07wuh%XS|3Rh(kaR2s#;o{*$$FdncOSeZj0N0=0GbIM3XtQx|!N&^5zL}R?GCU zp?0j2FKJFfwAVsQ`Z@LKNm8Zs_d8b8rFB%->49o8#Vxg2$HP-v3X`;D(9+Z}zk~^M zDLM+0QCeCBWcbhy4-%ZIrI#&ywwonoPNnC-$gHH!8v(AAB&Ba2k=43whiNVMN#|tv z7pb<|EYpgMOKWQbtBnY5wPlQdRPpKAaCUg7zs;Vbl4nLlt&Zm(W(>1=QLem@Raezf z*TPF{WpWh02a!*$lnPW;-kh=GgT6tV%H^^QrY1~%1@B1NP>3DFGS?%iE}t3=J0x0M z59i5zmY=<|kX$wkQ;SF0kKx?KWEu^)H!Yb=b%iS+oZQgrNAGxPL1g;fNmE9U7c#?! z8>F4n@OC8T*y(hMx9%pX8+~7Blb9#JITYN+dX8nbH4ou5tiEbqdO%PY$f2z4rj^~{ zWrmjQ6Qw)FTsJY*Y)qtU%`vmnRmsM9TVxL8?3H{5v6J1cSgP%uF&1f*s-*3La-J(V zyG_OH#}Rh-$}XTNt_wy~*36U%MlzaPOc|b}e_2s($F+m^x!NXef-{=tC24*q-ekN#||Z z*+7{{x2R~(NIq_vpE$lC3R=erVMS{&YetwS2{K6*|xT! zELEU2C8{qc70sY|d^MBnntDEGSYqX{lC(j#E*kOJn(TH}pq7iyhd{Y%Nr;eK%O$lJ zQ?7I!YvqdM;>uNv<+9BfMhM*(u4j3Xr|oX0GvZa&jxDmrz@#OepSnG-2!!p%D;Z>M zzB@Yw6}7yL7@eFJ_oe13v=HP@BYPDHlacIwYS*^~L#t?Gw!ES$+OsSF0(+);d&=fOk`NyM+U3tm|oY|hEn5Fe>)=JaKw6LPQvAm+DC7QT6 z4c|#hXWHq6IBx^;WV^iHtp0LY`LdB!bU05XF{IAa7|_TbiMC(E#iY1hILHN6&gBc-;} zNU;mekVv8art?UVOmm48NxGwvVtO@RVrr2BBZDocdXCaoNL*Eq*H*VCqH4P^qBIq= zAdZcw+8vH4X4+-C6HytB+`H(@c`c=_{4yk{gTiH>UHzn{ZPZ1z6xu6%%o(s82Q+Q} z(2Qu@J%52r!bTs>jEvzYHL}!xv5`fh-W^$Z!Z5iub7aBT&ziiNX|7XAPG=u&TUJz8 z(Irv5u1MBa)q6IJ-ClWUK{Vf;Sx#r9U`>MOx?w)x7?<0=^9~qidnUhe8U9FHNP}xb zV>8cNcnxRHO*JRyme-c>gm)sfGukjJalWFNLh_e$nhU@<3bCR|w=jE0d&erL)DGd%c~@5lmc@1# zJ2Of4LN(RiaiMBDxOfLzvhIUiJ4a>kO82l`;YdY*hJV~(&h!$_ikHIlW33gz^YySI4#%xKV{(oywE(N^@ca z3&aC@=B}s|r1MRbVRQ{0WRn`PvU1CLUXaU{?kHz1V447K0p%`7VX1HFW_ly{42L(* zD{rJ;9CzQz9bp&VnV}ZyIgHMYSIkgWoKBaSx!b00&ziRDE={s~9z82}Ba6Cm*viUf z+tocMNB*Q1o7x=Zh2^YaIq$h!$JeD)tMpHLV7e-2xLNmher7hK zfU?b-*>Q7_`DP9U`LI74?I7|VDGo1QHcfuB*}RFFs{hmpe%_(%bf^bTm9?t4&^o_VqFOtnZ#XyaNuC3q;AYAXJ zCR)*ze%8XyVv6d-EfAe#Xg=0{>wJP>$Dc@G`0H+ zb!z+abk(SPn##P!)A>SN*g4x!p10;67m613n4Mfaz2k9VEzB`4(#bI{YIh_qltKot zP+()fF4~X{J6%B)oxNcno%O)ww!ZbwTcG-B2&WBEq2yMhjLV?2ni}7oL|4SqZ4c!w zNxta56g$0wLK^fMWb*0R7N!n)R82jNxJ`6!P)9)HB2t zRoe}TR;7hbr$h2I8H$#N(bZPvLrYkChw{(d2QJ{nFRz0690c}rxOj-uQTid;D2?`k z{m$Vfptkj^mQ*&3jkW5IvAQ;m*=L4c*|xsyP@zi(Q+=r~RB&JTO9~aGO?PRiX!&Pu z@ODY4XfpAkLW|toL&fX_2Zu`b2C^N#ve${ZuZ-+s5t|lI&@xwAIjuW0c7b`O4jaJ} zXqdn=`Eadfr-MFCUe$>lS*T7xDbQhHs6B)CQ*#^`Yzxx{hA5l3UMAx_uH#QTFo;O5 z0|U>&=q9^lbmy76UB>4d&1592f21q=ohv*mtISSreX^V{z?2S-I@t6%s?O~$$i<*& z-&lOiczqd z@mjf`w@aK*N0!}cVd#vHscsqdGDZf}mw*LEXIag560T$+0^ z9c^p!ZdyrZKGZ@-%MCJgUS?Wz6Rm6MiV8YSQ*uW1w!WbyS;EbjwyK(wEpr;_R+`$H z*;Txi%lo_qgho@?^l1*d+opuiArq_6ers-v;u~?Iujm$AI*T-?iXIhUhO5bZt={QQ z56E|OWNxmQo*S3cQB%NM`Z5TqD+PU*lx5Ypc77)#%d9<@f+TMnW$)S8AuNNPGRPem zw2d!y#LmdBxVSL2>ou7!yKt=F8c=#DW!DS0yi$XKnb&mM&p6g_>vaPUhmNdGq>TNm z=<2|X8c7dg?kNqr2MssQ=&Fg#)F(`Y!&Kxpwv zrO4h0qUz8Woz!{zt(PfKxb@VV=4!p2%Vk*)+Z6%U`>V~_HjFk9V1p^6^5ah2Dy!7G z8s#2SnwryDd^#N)t>H_xsmjzOMyaxkL090@EX%E?sPoClT(K!W@8O=jX$}Ef1knxw zS^d&4&1#jvF~<&48D&SJP6>_p|b; zdA=MXbHzH5CPBPBtuCjhXEe`jG1ENqIz3m$S!GgzQWFiFtvJh)%tejq39~LiNe7_2 zG|8lAE>o!gbQdbV7M+<*X6w%Kdal-_7X)>sr}Ww*%62??ilHBu&%%A zV40F_6}g8dgw*!s)(gHAfI=a9h@5F!!=e-I*^%h?No}AakLjdHeQf87WNQJ;*`pad zZ_RbSu~#}tc16GUB=HxslKkLWnb9}LnX2?> zjnTt_w6Z#FlY%8R%qj8P$rgTr*(K#uerDXI-BQ`!j|N;Gaff9td9Wak9TY-eh8;Ar zf*t5pimfg?Wt?6@s+a5Tn2=Jd(I{mW3W^YK?@`O3P)|Rblx}6G>*k|-$8bcVj=%G3 zIO#-2^uzJ$mo+k(C^)5X<$DMrBGSa5ZI3;gV z+BuDXHk#uYLDTCAyqL{;ST~=<&*{+(Usy=(CiU%-9JO*7_|i)jK1ryop1$0{jt3Uf zEmfbvqwW-SUC}evj@AnDfW#CvN?IDV(4~%*l}-Gp+b&b4UW&H=Q$n4rmF+NU#^0<} zGMN~9`rAY^_5sw|yJk=^V=)QPVCnSfY^Zh$(3TG`lcLoikGOQ56rKC^JnsXVJD*Um zrH6v@qxa|VQ*Tn^3Jh{ic{45t%V`INAD1T&ALaAW`XEDV^vtWuIDp^+S7%HxOOZ@Z znUMg_OPNbo9VVM#&k;oTQKVvt_%w{04%1#%h4Dp8Q|E#NUQ^SM(p*tqTTXLc>N04R zH#fR`fF{7?UGo5;k-T=JV#A`@)#C2C%e69U4yxT5opX#~e=ncz2hAo)x)UJ(NDt63 zvuhU}6y1}fj^O{QkQqH5_qj@cpUZ^abVpgOTw0RS#rP}E_5g65B(?jf1AxY_4Dr=2 z@MH_(Rw*j20PofvB|ilsGscY2Wl55W=5mzrc%E~8|T`i9)r6CJrth&$JDEHzjW)dV3gptiY74pJt=f87gku_=)(1?!KY6=`&fE3ZN+T)Oh0%xn18c_9;0b9uc}VYq~Ct3qF*v=YM9e#egpA~ z{GCR5?5ku@Bv=o5ZS0mv{DOY(kY1w0zWVqRBpvd9xl8W+X~OR=udKf&fM2HY9}D1D z3m@%&Jb>S#@P8TLzcz*cbb$W#3jetP{!-zi|9=<2U#{rC5Ws&>_`T$n>;J_7e!HUo zs{sG6Q}}BF_!|`dV*&h43jgr{{#J$mYJmSfQ~0k3@Vk?(d|>C`9S~wHGscf;cp7yZ&3K}2Jkx+{)Ykl%?kgc0RC3tWBq>|!2e9) ze-gm&&Od@p2kO5T!0)B-e@iyme#^f4`ma2N|6BlnkivgHfM2Na*9Gv$D*WFC@Jki` z3jzG;3jf6bezn5qHfih4zWVB~MdANGfZwL@{}8~xUg5tKz+bBH_oYJHa@$v5`Bw-Z z`@j7H_zx=j`v>q>EBsyo{HGND0RjB=3je?W{zirW)d2ox;iLas0{B}M{+s~*cE$eQ z0sL;%*Ybh(Umn2kq43)R_`QUW{r^D${5*w!NC1D3!ap>CU#Rf=2Jpu!{Ja2uslq=z zfInT~_YdG#EBqq@_$>;5Kmfl@;lCWv|6Z@~{}{kus_;_58$sAKK38?1n}1@`Xd7P8x{TG0eqh8yXEIOJ!KIc_SHB3ZdLS01?X>A^a}&{ z-NZnwzvBb=yjJA{>+h-nK94mn|C<4P?(6tK`(F+4{~(3GC_tZm?dpFkfIn9F=)V&K z_!AWU9|iEIDg2WH_>~I(wE+L~eyi)h*8}(q6#ke1`?;=NeXbLMQ?I`EyF?OV{cQ}e zf0@EREkJ*T@G<^R58$s-^v4D8S1bA@0sN;F{uu%M^}VI@m%$LDf(vx@beV@lL7oe3V%uf zzfdXvIRX44MgQCY{se{pc0m12Q|x~yfM2HYHwEyk75kF`uT%K70sIZZ$Nua6fco!H^y>rk zw?8n*jL~9(QP-A4*k~@!0#b^9RHdF_`MbVxdHq$?0EC@l&brFAd-~D*Vd=_zQ%O@q0x8 zf1#p(WdMJ%!e1D`U#8f9bpU^bqJK>Qf0e?&Hh{lI;a?ZPU#sx158%J7@NWpvMRrq%V@Jki@KMb&cnxfBp8uTxx_^VX-cLnep75?`E_zM*N z@&NuK;bZ%`JAl7f(f@t`f0@GH5>S6D75)zb^dC|9_XY6V75>Tq{yK&KQ9$`ODEtQk z^fw6~>;JC-`dbwJhXVAsEBuE8_}#>CwEu4b_V-Zud`%FUkW>5bC4B6EKMvsMDf$P~ znkHBMfeQcU0shNZ_>Ttg3zhPJ5>WoJ3jY@Y`lX8fe-F@~rtqH+&@WT$-xi=>tMJyM!hbP< zzfR%*A%MR@`0|qAy_1&$_#KM=%K`i?3jdD*{Ot<=l>mOXJxrWq|NW-`eouw}Y5>2F z@X`N#PKW;G)c*P_{MQ2bgB1H;58xLn`u_;%|B4j-Hv;r0DEv1A_|p{rp9A=n3jeJD zexvZQ{yPHrt&09T0sMsue^UT|vBG~hfWJ)PzZbw?DSY(b`vLq_3jf0Z{%XbkEdl(e z6#hp6{Pl|ce+}SoQ1t&6z~7|sw+8UHD*TTF_@61||0IClZBNtBV*LC)fZs#-IDYUs zZTgo}|J_^R|096kU$Otw0RA9F|DOT;LWRFQfIn8T|6c+835x!w0prg!MgPA6`elm# z=K=g$h5x?*eyhSy1gswyD*Rmn_=^?(KLh-~OyTbupubY#cMsq{qVTr|*x#=3_XyBm zCw%OG{uQAAvclghK!2lRe~$qEW`(~`0Dqfef6oB^XNo>wlS54?r~bG5UM3y-|9}8~ zPvK+#|L=hM>mz);fBQcH{Cvg!=>hf^D)#pY;Ez@Kd{11i{x4PR|8D?)nxfw~!2U{w z-!FjQsPOv-@E0ijBLet~gpct%Ab`JE(LXYPzf9o|4B)R+_(uisA5r*62k_e!{xJdk zbqfF30R9GrKRAHDN#PF(;BQs<`2qaT6n;Sfzq@pt*nbWS;P+DaLj(AE3V(P2e~`i- z5x_4LKDOVH0sJC`KPrGfPO+b_3C*?to~r1N4&YB$^p6YRS1bGz0{ATo|M&oYo5DXa zfPcNhKPiB}RQMP_CkODCEBvAW{z}FEF#-HX6#Y{I_-hpXQv>*G75?V|^N*Jm{%HaF z9SZ;S0R9%?WBrc{;BQm(O9J?buVl}k{}eU*ZwO{u|Glg z6z9qxsOV1&&>ybwO9S{t3jeGC{se_TIe;QgG;iLU^0sP*IenSAizoOq5z#pXOHwEwu6@GI7 zf2_i93E-D1{9OY2ujvZETL8aW;m-@OzeVBi8lc~%@LL1)uUGiH1?VqT_`3)2S15db zCRDETk5vkPj{yD};bZ%~B*1@9Df$Zn`0Ew@%LDiu6#cdU{w78LiU9r=MgOV*{&t0b zbpXHHKH2^Ew*vS*75+5={5}f*+5rAQ;bZ*m84$nu3V*KvexYLjO#$|gRroyu^h*`~ zEdlz|mGa*jz^_#JUkR|kRH&rvcm5j!0!+~_Fo4D@HZ>^%L4qjP2ux1Npg)}`=b)L|MFr0 zzlZ!N1(u(m36ZOQFX5y99|QP#ivFts{DH#9>t7BIsJ{~w`~MK2KThGl6u_UV@Lvw# zmnr;?0Di5)e=C6BDtwHecLMlrivFem{`Cs~?EwB##s2pL_{$aj*8=ztD*BrP_>U<1 zZ_qOl-S(SD(>yw@mXO=!&7_b1whRA}gHllRhMG{u@iqFyc@|}$iNrL%*_MuhhnO$; z{-5s>WaFW?ffEDhl~c;&8U;-xT8>G*&3I#SO+$m|E$8XuyXoD}BZYy5J72W%W*Hr}oLv6zk&s*dvnh?Hf4h{wD#|7O$B{{t(pQ@PQbqr6MZejj z{~ysmFfARuV_WgKnm#b*`oA|<3>~(-=2Q0PZ_*c${WgKTu2Ssh-&C;qS-;*&=~>^c zzdX?&;^fO|@pza%FqY+4;a}qv%<9`b)(GF#e^rFfvkB$Z)qj+5t^U?Z`OAdI@w1Z1 zR^OKYP5PemboGbk8zRRi9&R~br7t!8fxKBphh!6-KYu{c--$3yze0$`Q7-8}oE8uM zof}Pmn&`JmVlMxKivB>4{z}om9L4CpTmF8;*Ywwm{y5NoNYQTPXy_` ztv?H|^4MP}{C#agd8OnT#}5h6?C;H+m2}uRVgLP@(pr6s^D`MaPd9!_ME|8Im-HV_ zi-&!tY>`v`e9_1F`KhAM=j1t0*MEJ68kJHMqxW1^JnS{Wu6~*5cZk7U{+}uO14Q2~ zf34{ELos^q6f@D60L}kJyxB>I^)u@~s^~X(^cRc%VNovWKd$~o9{p*GKL384TmN@> z^jGCr{v{s$Mn#{0|I5{X-=qI>j^%&bqrXVe=imQw^$(=}o#3qa-zNH}rd=PsBe{%Y zU*c=^zg*Gh-`{feD?R#!!;H+a8EV<_m-OhbR`mJzw_N=jMc*w~wdh~X&Vhr=ipLK; z_WyObNfj5r%RTn5SL}aEv7dj_iA{9d@7^PF)3?9L=Jp?(6@C8wE7yN}kV7>6!8z)8 zBfb_t2|n3D2lgLp75!s9`V({1AL!BVt?2Xb_qgSs>CvyxQNP@ypRefi@AtU+H+%GN z%u(My6V0vvaf&|wevhla%A@}y(T|&duJiCW3jbge#;LzS!CtUp-J^J^G{*h5B(tq6Y^PaqB|L=rVFg{l{}P*ZxmE`qK&xaj@|1i);VK#HXwiuKs$_$Nr0-U+?N4M;i}X`L7oJ zbFHwvy80uCujy|VeVl*t^UGcRZ+P_I6#Z#2`n4YY-h8r+4%Fx8m%IA+dh|z*Ho`^b zCH0qmg~#0<{c(yuKfm17-<_V5#d*5p=Umb6k7D%R^*=x7PUC+^_y@%BAMo&3<*5Ih zhyQYp{0}_*?K$%ICWmU}?{l0f_@Oc7--GyC{I*K{Vf*3dSGx5#%cFms=pP@WKf|NH zScxBgexs{@hev;|=+BPPU*gekSM+&H(ba#;qyM(($IU-Cdi1v{{^#e{x%#`(b0P@N zYCk)R1&`fEks@>!psf9C2x;L$(y1S2fnc*+V_{|6rZo_w;I4(z}A z`46uCTOR#$MBm-UcJ()U^otaIetv+fe=yyruKE88(T|J21BkE1U!&+_{PF$guKrMu z{yn09R?73{ja&X>J^G83^7CGgtAC+Ke_M|J=jU{28e;#~Y{f9jIMWP?)|NA}q3Hbq7jDOyvbM;^K=-(pxm&epUKW9Sof4=CC6F#?p zzJA=*-+3=R{`Wh{5c|jI5Ag6u3O~;NzQot;Z&mE)>$hF|r+V~%F8UK(J3Ei8+3F zOUKn8-a{|{zeQh`(UME{{C5cPwfbKw`q+MWN$KkIeaf2uhZGsfeNxKija&ZvJp7Y# zA@3^x6xl41)zc<}yq?Nz=RHG4B{_{Nic{%d$ z^zg41eq8yN5??F-D!$l=4lF-kzvPzxX^;N$(@go})(?E2o2I|?YepZ(Z@zxX)$d35 z326R*T=b7MW~Ba_TvHzthZ0}YU!~|XG(f_4Kf0^jlIQnk;|CvXBU>{>?nee&&@wqD3|Ho1T)9Qb> zafTSz|MPtoTKz8)ee{1nMZd+PUn}}?{eOc;f3@gi|9^y{|Gr26bJ36M|KIlL=X0>p zf#p9^(eKq$_y6z`DthPYAL-F= z6@B#oF^c}#9{rbdEPts-f2&gdV-@{6kN&nC%g@&-YyR(ksPPn*e~6-gyGOtI4C8oN z#-wV&ZT~lW^xH%~Uzptf3l#mW9{qrm_;so4LtM}M;DyY^ES8AmrtpxIw3{J8%2e;)gL@?b+p zS@bs%UB5G0v44n1|G6CPuk-LX3g7KN-T0{{zUKc%u^;=d6BPU3_UP|B!4!P3l;6I% z`fqsj7b*HDDf)Ztuh(CZ=pSi?<<-(TEJ{kZwt?H>I^o+&BzUt<;hr#$+FXBxLn%kWq{d`6LG)4bCkN#rO zkE_4_bb-3o{_hcfT>CkU_*(syiT&7qN)-F=@aX?Z^yAvk5|4hXqCZ~If7+wJ+eA}w zar2*ckN#ps|4c>y9gqHi9QEJy=&w}tCn@?n9iZ2LvFIP`#IM_bedf_`SM<+P^w09> zH;I1S{BHvBwfNg0`q+M_DEiAh`k#n?U#I+(MaJ>5hu>?GNfy`s#?uAPTKgLy{J8b^ zX~fsc-<=m*bYS~CS1JDrkN#MX{aN+z;h!V?i;Wvne_j9m(_??0*xxElo|pRqW`$0voCJ>-|*-+ivD6Da{bjR`llVN*WZB2 zh8WlX-{RpH3V*OQTwY!O-9UV;{t~>|MThkh+uxwrzs{pSQ}izoUHjtt?^%z2Z_&s2 z`G%t3?`yjMo)G<^R#;wL{plY5HsQzhpVxW#-KH49(J}g06JPWHIK}^Sl=8QG^iLA~ zxbpwY!*3LRT=@s~(f!|+WBL0LUn_r^QvUf$`Db|ae9(Qj4a?-E7- zM;`r8MPItnl)f8(D?R!v75z&U{g*uYr%g4IzjO-h>c8O8-yr(c!sqsXxuSpEA$t8M z&N2FN{vSbnt^SL6vzrcV|KC*fukz@hB>HjwZ}aFkD*9I{`aksO-zfTV{=e6wzg+au z|5qvcU!e&;SCM=Ddn5dHp9F6lq6{qK4BF9<(w{&OTvP&NN;5`NtLe}aeK<2+;YereOAcdq}g z@$e57{{0T$E&oE|YyK}3|1TCkx8EBT|Nq*f|E%cK@*?AK?SI|Fe^dBJWw5jNuU_%k zzgFz;Equ2BX2t$p^7Q)K{d_}|ZJ?ADuKk~T^gERJy;ad4;?bWf`f>ZG7kKy!g#Z1N z<>n2Uka3)or`LbafhJ|Uq-Ou!uGs%1@wNCr>;gk{`%l+@j}u>uzxARory-;L`=yHh zfWEqZgXqW2U;7eY)9=kE3+cf4`HrH$(4&86j^)4Hqn|JOSpH>-{trF+ABle4`sZGc zeydXcyA=HoJo<-EkKC7tYyZ0*uGimj!jEhJ3F2$@w^*_NZpHp%J^B@*AJ_f|dh}O{ zK8|1CSM<;G=&us}xb}aQhrdDiaqYj&WB&%Rzui=0)c)>O?ElE4zu-bcl+z%ROLqUg z*`vSoXcD6X{kKxlKcJr;KWjxluK(#td@X*~iaxf#2NeBQkACm3n$MQd<=>ONFf4k_%^*`G^`i)BYS1I~M{q_1^C;Eq) zYDoPxUQK;W98Y|${ue3wKT-6*@6rE@=*R9Kdi0ko`j063Z+i55B~2Nrd(JrA^7lJJ z_uok2Q}!8$%dhnC&li5&{P|uFzd`tM^WS$o{40bXH~)Q;_}!?4MNYw1i~ko(>3RP1 zOU3^O4bc5JsLT-K_CLmX_{R(XD$%zuJJ~n#dMfcX`^OD31w#8DQ|!OOqu*X;1jagg zj++vfdh|Dle!lS8f4^4rf9}!mR&I!K{(INM-(UD~{_F7A-<>a(q67W+gkt|;G(dBm zxa;Soq95nKgNd)z-%`;>|2?JXPxI(+6n%I6Wn1x>;?ZBL=s%pWi9^yV3-k>%@)UjiN8h$dnb7Mdm?#t^AEj`CnA@$9VKd zRv4MM`NIhw{Y8rY?-l*~J^Hgn-yOf*_+8=AU#{rCr0B2n=-)2-arOVKM}M_a|9@2U z_n-o5^}kN^U6 zbl7&vm{se9eF3#DCcTzpMB!j|Moc{ogD4#VKX;#?}8C@iqNcMgM(8zrmxw zMf4Aj(SOLp-)*KLj*8*`$HPBK_;LO3r=Ie!RLcLMQvTCugFy3ti|CJzv40HlHUDo= z+RsOd{&gPxheZFt82xq+|0&_`?eN|BKYWN@{yzwxzRoyY{+%BFPO}W{stmT#LTD-R zHUDiD|Ka@q6UBe;di1Xr{kZt+@aVVYlNcR1{%lk9kD>;y#m{4+A2C%#txrHcNi ziv9wR{wJazH~!7@=&w@rw=4QT^yv4gHnlv|yrllR@pG?7f32ebZ$7 z_PogG$Mqkl5MT5EIMK)cV^>B0Qjh)^(U;S(sakN$-|EpXQ_8=)qW=?*{&z%wzm&3h zex7rtBnuKrNsYyMxXlz%U!{1U2E zX5-rbxX1o>vETB!{p_vS|4)y8uUez%`p@vpKl{@Jf$PNG|M|D@?lyby4*v;;w%>*P}m8^zr(qgB1PWc=SK1H~KOSH6Q4&+kVz~ z^xH?6lx4!?_VYDGf3rt_T7w~03g5oC`tNx3^Y~^>I7C2h{uMz$It+2eh z?RSQU|Dy2S_;dMdJp8wWFV&E;!sS0od@cS;rTneJWdHY7%0GlU2(A2+8V&L2l(Ko_ z>K{#fO@FQEWB<`l(ZAfIKVS6Y{KwDP)a<`b_;LQb+hc!+;=cik{cm{m|0Vhqo*9Sh zzYQM!X`@X1qW=ae`ui8^+4Y~V=s)1m zpCtNm{`(zUIG`ivLC^`lCJi-I`7P z9U%2@U)=HMS`YtV;TKt=yt@3WJodLM_7^Jlf8^0WU-aYZf6(!||7(RGH~%@3_?rJW zi2c}q9Ix0v$D{v%=*P`}8a?{m`C$ii;QZ%AMgKLAe&rlfVR7@HKYH{Vk2CsZCY$K` zv6B`3olemG|D5Q{GF);oX;UBQ&1W9{^`bux^v5Xr!#(=L=Nh8hf4T8Hg!o$h_dedF z><#*-D*BT=`d=4)+Fs5$-0|}akAA+QU##fg;?ciO^yB=0gGYay=%fEnSM;Cu=x@u> z|Lq?A#Y*{06#Wj5{-N`X+vEKIx<`MdqCZ~IKa@HcZTy@d`f=BcTIq9?ok zyzSw)3ZH-T2oL9D;tk?!@!u}xZxaH~pC&2gzw9Kv{D;psss$)c?_K@*#MktTPRwq9 zXDRys_UKO${rzL~ho7w5KQl-BhY(-0f0|F5mD-`?h^ypWMe%$!G)T7@j z`Z)epDf(}E^nWkd|jk%0Eld?{}(RfBzHxxb@GW#Mk`4LD8>K^k;bV zhc7VYjqCqcdidjnKfv+d&d%rWd+gt;*k7yI|Bgq$K1cmGJ^I~GHkM-ht5@_79jn*h zk{tC9BED9Cd7_WwN28*DhDZOGq8}H(#UA}qrTk5b{*@m6&7wcjsc*OaT;|cQR`gpG z{f!>|LoYSLaq;&jkA9ot|GA2Oce+4`%jBLv7%Td5{@;c8-H7jA|Fudff2*QD#-raP z`f>h0!K1%c(Z5*H|Bgq0rRbA;GY+@?-{#TpQ2f6@(eLo+ZxMZW{BZSO_vrVO3c>bw znW8`7G`;@MzRVQ6Z${}&T7>!%U#tHj(Z~L$P0_#7qdzxC{mVT1(-i$H6#YMY^uI0o zapTu(9{ol|f1#p(!0EdG*NOfq=xloL_CGy|ulawGQvX*g`VAiay)QSyapUJJ9)4fp z$IXBL>){U*ew_ced&<9B@&C0-`3KVkOY8rtMgPDU|DW&SUncyx_3zol*XnP*V*d?_ z{r7wH9~S+%_PfHPzgf}0Nzvb}MEBqN9QAi1zE=LknC$t>&5HgEkN(G^9~VCtdh~lM z`nM|jOFjAn+DzcYwclGk`uU=d{r7E({wE&&YSE8tzvrEy*WV?=k88he9{vr&kMsW} z#MkP-O!5C6O8I~6(SJhp&D+r9{o1a$MNT3MgPAZ{R6KwG3NIFJEuc0dbi!9 zze*|pj}`suiF*Bw7yVL4&nMF;Pn{#4QbjYt2xIqI+R==VI;cnbA@ zrsy9$NiY8v(Lcs1uU_8Es;{@sbMmA`wjNr~~>uIQiT(eHJ&k&T;wP4MWKiaz@P z2}S=QkN#<*A2aG3|9y}C4Wd6Q!~3SL5PI9Aze@DWBr(U|Gm3uRWWD~k zUp z`m04B<8Pg!|5uOxZi|fKepXms-S$^FMX$f!!jH3m81c3GTd&yvqGJCY9{mNPKQqSu zB_93FqMt8(j^E!a`mcHP-xhuAe`|*8zkO+d(#DT}3xAp=%Bx%dJ&3RQulMOj3GIJb zv44g~fATd(FwXuvJp7A{D)s>gim(#UH|V- ze69RzmGZx%=wI*ApCS5($LN2{qu-(EzpLo)ObuM?zZQ!A(J}h}_2_R`^xs$X`+M{s z68*UNJB;|6|9h5X$NvY4{?|SFFXX6yo=1P6qW__y|7(x_Uq$~XG5-IhN54q)+l9~V z|06{|?>xQ!mtSv)aq)MXhyO&5{KdrA>MwCdcKvNt?BD9q-z546I{tI}kNM~8<^Q+v zhdO*?RT^v|zGi=K#s0r5_W#49Kj;Rd(K{^_y>s;oFVO8DD|~nT<(7XK@iqJN#r|=k z&++rAVt<`Se~IYF&A+QX`mKupwk!JYdG!C2qyAeS{l$v@zZL!Dblv|mZZwXg?U{_j z_5b zm4|(_PrCy4%Wj=tM|6U5i_OBMZYihi3% z|7y{9`!Bcrmw5E675&{5{hgC~`B#g+yZ&_bXM6b134i|>|E=)w-xU6+82(O7T==V_c&n(mZU*yqu{XdTQTKQKg`g<$-cY5?M68%Gs z8&ZGW_+RSLU#sZ%RP_0%U|51zxYh_97@k(3|jF9#{*-@QV&|I}NIY+U`#_wdgV{(ubYv&X*{;%oM=7yIov z!|~TgvHu?){RN^Q*Z*ww=xc8aSmkIyy82`QCvAs zM}MiJf3l)KxJLJXQI7gY5udWn^8YGDe~hBv>d~*yQNP)vzgE#dRnh;AM}L{ z>;GLYGXC4e`C_&33mtxU=kwPiz%_o=qsP*6$XcDEcnH=XXsC ze!eIkF298M@F?sMj-=8n`iHNfL=bO;u7h$e!A z+{t{^KMoo5*>UVxh9G1_1!0DtBt+n!wuJw?<$n_=Ir@LhWks-V(0%-~g-qR!{72iX zsl6`i->;)?VZwIyZwz-n-`>%>%uUl4hA81&O1iJnI*`EdW zx;j1TIQ~}&B!E=f-Mu`#s$mkA9AQqsU=SVec@$W173AyJ0gz4+C9|CeSe}$P1*?kn`oX zP-ezpR?Z9YU%3%pN&}HV+D68wF}^%tsKZ1gkr9Lf&7+)#kSKO^kFax*n_F>0oTrHN zh=MkPPsAaF5+*#p1CQrGTvoB~k@N8{T*qw@ue6F?j|96BG>l)%9xbHnOb|o}_!wyf zuSpc-biZ=bz}TYs6v-?rcKabgc6duyq~Fm8aKxYEa3diU4k0q0_+ zC5|BzQW6P?ra09jNGw4JFV0I!Ow%vm5K&{H6te2D`a~M-bV?GGsHnp`1nO`^N5$j$ zyqqWqJEe!z-~^*?fsY8g#E~e7tTM#mqX?q#w${g3FfbD7@tg)i2%?AJ3lKiy5%wWg zO1%hz98b%@6><_|EnPJWcvTA~8Begwu`|9Ci_Yw&b>wW%)+6XVCJK_9*eS>asgYSOUEJA+ z5>J6sBF+1*hzpV?PgX&$lMy^V0_WDq^vDE3ctgl z-NMk6xC%!R8j{03rbCjW%j1!Rp*lPgixB85aJm-J`eNo3A_sxnnAAn!=Hx{rxY?tN z^tIWol^VLD+>i@6q6(ZB5_tBwkVh~NzRMN-l-32PK7NEU*RecB*ogcBh#(MKZ|m&!Ya@T+-`5b1Gb3duBsOx!7i z!@Z^m;SeFAdOT?-*^FTAL1NM7!V_@1^?IbYBt(aqMO=`~&S)EtSjMil#BZ}0;o&9` zHKmC`0`V-51uDsH1?(JzI$R3hs_K1tZPw!yL%$ta79_m!9tuI8p?(#ixx^I73XHR2 zEFpUF5}RD2FBDxBM_x#;f&Fv35pSy+lXAF=cDjf-5#m>|%4G`?tah?Tn-h3Tvp3<^ ztw)gf2E5CV0aJo$B7p^A_CjLId0f-vHREFok<}vf2+RhuULlt(lOP#cFU}N?BdaiV zWicTZy!srZLNb|jd`+~(DLURFmuFG0d3o33?i3W2&FN4Ih@L7B>P*pTXHlt2zMMu7_He6P8Q~9vLLUxp^%^u zD?0&!I|!vPlbfD=PLHtDz#55X6ol&xK3xec{NkDM6vUAD^bB!#JSPhUZeKX0fXIZL zlxL@;*d`YtCLB091R+Jot|B7rBi_nGmJ5f5~o89ige=fgjJ_o4Wlv`IS;WT6!?RqM6rt6im)Tm zxWrAzqnXSJI36Tj=Q#0$qPPw(j4TNukPr@}fP}!a*(IDyTHzJ%(@ZQxf-YPyc!Usb z!Hzr*R}kdFlS@<6FhM%0dSbmnX;jb1+6z zoOCH^ym%%AM?x&3<%cZ@WO;5;lGAPuIA;{31fh`0Qicb*5S$9vHc+Wj9`fO;x4R z6ka0CBRV3?=xjs;C&`5%8+j=u&8~vBc}@lPWlTrJ*Xy$vaMHO(bFZ-|eG-quk22!9 z*~!uDTPOvHZ5p8h5$DC}{_`w=AV>{>#si=hr~}X`)d)NSnt-Rkax`N}D`*>*KL_mq zI}UBmZOwG79<;%f~>+flt6^;0rJbEXP+YnZ@)w=uhAmFo${b zpbNkvR^xD30z!a%6;{K|_5TJuR3J5yqZcrOw4`2uE1snlqzzuK*yZ|)3F-3D<$o+r-;1F;a2n2$F zP#_FI<0xo25CKF2F+ePE99WJx$P@pnlORvQ>QvBlAQQ*}a)DC-8mB??fb&2ha1kg5 zE&*s<1}y=u0M~%)Kp9X0R08*a<+u-d4e;PEuNLx0K+|6y+Ws-rPl0Bj1!x1F10BF} zyny^A&R-1S$?lV6`Nu6qawsR2EbY zPyiGGWnei}AXojXR)<^@tF=J20bM{J^R|N;114B)3TlSsXrC5XZiT50=pHP$2Xz4U z0*-(a;0(9`Xt-jE<{nt?4Z0u84}$sw0a$$qG!O{F>R?PmKtq9}SRD=;f#s2)6fBPc zjRj5s@jxPw45R?dkqUX%Uv)O*IY2IO3djS{IE!gM=y{+JxCj&jmw*!B3UCcTqYU&0 zP!3c8Hvu$mVR{?WO3*t%6>t~x?txYVH9#%qHGno^IoeMXojcZ}_imA0QtAMltW>U)w%m`4`Y>UU|Ll+s z{pg3!GJqaf1rPydfCWH<6;lpSZh!}?*MP1CcmY0OJ+J{-j*XBD1DgPNGW#cFFhz^a zkjnycfC8Wh{2fYAD+Am9o2U3X}om zz-^!sK;sT*6_(!xy$4hSHJJAR^dXivVv4@*5#&u+{RFfHXa(AVXTWpdCD0AL0(yWx z;0=Jr0H%YO4q-YB`VPyfE#bO=di*c%6Xc(0NVjH z3_*4*CQ5 ziPgVA=YV-&0r&$f0th{Po&a!0RD0pVFZ`}766UapzHt#zy+)U)&cy$24Fch zLM{N106{oMCf@%O- zfHt55=mPqHAz%ccVT`FMre>h#fCW~gZ7d<*3G4!F09ybJJ5UF}5pV*W0aw5cK*Jr> z1MmdAF>gPpFK`eD01g3Y1Y#Nl8VrO2M}cDi8j+Zyc@*TaKpYScBmjv(5|9d{0T}=q znV>mXo(q}>T1whpbo3+K^uTZ;1Tc`K;tQJA@d;mrPG{!*3vHTOJs6PSu6fg~Z1-=2_f!_cc^O&Of0_2MT9sD~E0f4_N zMpgj~02)M4W-Mn1tOG0W`co zy@CCJFW?9G1BU=KfSLghSRM@;1H=NyF)smAYHx5oK%D}d1k!*E zAQQ+2a)ISIjpccuXaDl^AwLfk02hHvz-6EWxC&eY(6|m-22=pIfJ$IFs{YFFL0*m3 z4?t^yhgjVJ`Uq$SS}?B_v>kYk)o7m`kiP(4VqPa`7nXN}z5-qYy_oj~(*e*yU>JB0 zd;mTIW56c>jW3u^fX)D4fp5Sp@DunA%maUbMF5Q@Oc4fZjRVC4^uQ{B0U!d*01L1h zU$Y|9gtf9R)96;?*iQo z*Z_8bJ+K#W1kiBC)CE&lP;MPA32*^Cz*+zeKG5}8z5#S2mXk0Q0Tsn^w7nSQ;(!Do z2}l9bfD9lD$N^gbc|Z|R0=5EZD1)j3>VO772GGz3)x~nO)`xsMU<{Z5rhpl+13<$9 zbSGd1>;h~7JHQ^;2RH&QfGgk*cmim6gYE|o0DhQ<`u>my0EdC#zkGCCDAZw?cNEib z&2$~F>1k!+XAPdL^&H(uU8s|X^fFj@`a0ytBQpm4j^|inJ zGRSXWbp_~6EWZtU2e=E|1F8WuYCs=gIn{yB5UA^b2H-K!4731fw1T!_Ia;?v{v3FT zd7YqLSpEvr9!&c{2Y_MV126)71pXUinD-g<3orppV%~DwDX70<-gnR+z)#>8Fb6CF zXy6&C`u9(I@E8ChwvF*WZJ3~D23Y>%vtcy{=o(-hzz=Ky1OPNhmt>JKzC$0$ubRYxB1klI=%?5I?`V{DC;0%xlYx{B^sNM#@3b>1HzXy6Br~w`TwZKE59zdf3v=Mj& zGy#u+W}pRV1KNSV;~CV?fey^?1nmO4fmfKf-2OGxy+9w(4-5crfk6O`A<$tge~0ON z&=D;E2s)1CX#3BQPXLp^SInOU{S7Pte*hf(w>J34?EeXR%!B`KfG`2ff3;!3^3|Yh z06V}1a06=qG}eOh0{p-RU?U&^pg{r^!g68IO@J66fq9akQh+od18fFlf#pzuToF(L zwgSq)Hb50XLk&}PP)$G^&;fJ-G|>ObX8>#m3^C6b)D$qoYI96?fLZ{SSiKWdYfxKY zFW>~Y0RIhF%yYxk3)CC%0S*AZ02&8D1F-xsrqnI)`~QFHAgF^eFAOvshyWsiC?Fb$ z0b+p@KpYScBmjv(Dv$-_0H*;o&Vc4)c>$(HpcjB*tiB9d3S0#~2VG|u?&45$G&Akw zw3F%N*H7KcH`t!}>ixZYmM33ZXQ~Dm4oA(Z-ae3S)^e9c3V*!ixEdyvmPRa58aXK_`RDkj3Kx0kXo8q=# zg|Bk)wfbIVb>9lBbVyX}O+YH_;>3qlE*LBO!Kn*iP!6kb2d}1Zh5f1U%UH* z^aYuw&o0C3Vt)*_<_TXi6#uY?yl*^Qpf4kiSuLw|vn10lHj^LSr(c_9_Eo)icT;51 z+&Ggy)Dz-ocXgNV_dQ3KVtzE(dxlr+IS@qIBAVIGx^&W%jxtiRHy9|LG)=L1mK~`yXq`$}IMa~Ps?@bX>dB71^p$*uZ}^6kr_HsM zi6Iev>y8-fj)~-n{*2%MM*5w~r#0nxjlpmAr*96hgs*2;$i>M@axwV_Wh$6VriH%x zp`F$JEpt-4c=&B8;iuPwkD09&rooN}jNq zi|8<5Gia{2>b@FkSJiL0mOWv%zQUbQ(RbE+qrgIU+bN?oE@x9rK{@-TFV=G!;vsv;uaA9~ zZdLwyfpqQ<-KtHiDCY*xJ+Su*zSbeNVbFK#UErzyU4kM-qp`|YnL-mpRnuqX6@HvI z`EfFN8}eNGk50bWPUc$|$;>r3V=py3c)mes-p7 zuF1t}>#m1&Nx}QKk@2Hv9y41M&GgOJ zHm2u$ma$2;nFSTa;MN=7KYYYJyn1@f?Z{@|%Z)ET%jWL&b&sAA$yYIn?H*lsT*UXn z=`)rA7J(W0$9m__-p(DDmFFuUT9vvVnq8Ow`$*Zo@%1T83~ttQsUPMdJZr8DzE0b3 z=NM2^5vgQ6dgAbgyAxYpzcI2J5ck~kxl_sNm{C}Zd6*`>eIDO%sm{klImPkI@!Q5e z2hKS4<^-i%39jp6EcwwW+q-7V*ln3|r#0JJK95c{Zu6OQt{fSk?YMY^=Sup1!Q-D-U? zuZXVj(~m0~MakEvk{+m9UAt7M{c{Ye-|7l1ocZRb$?=jE^Z8 z`ybbyD~rD0+`GR`PlK1ib!Y3F-^I^6mJpY$BWWRlCoY|^($!FU|E%>(sj}{2%A;3r zH+Hpf49gW>8+Bi}^X7Jf=P1kKi+4ZXw|Xg+K2L67+c#pA?UX&J?m72^oM^A@8f-qS z-YidwuG!zGGMZ-W<{y*(+!smJJ^9v$HPvM5+PSxa?cda+9Bn)`teWbcI=?NPaj~(a zTX(d#{aM}%kt0V%_ZRxD-otoDkV$0h@(fSm$tg{FNh85P`HXct;_1ufEGvCDCa=0W z{9xMrMWOYC+K#(ZlTscg&zbWr8H$DZ!k-+uIB_^LNM**`pZV51l6mp(ad z^lPlYV7ev8`gTIaqK(c;eYpeLz48b_^V*$=ELn!av}Vv$CW=3-w!OwQNcN<3jRFeJ7r`l+?6!YFPvmy4ALkHbc{i*V!O_$quT#}g! zJ{ZPPd&Y2c#mnOn+z(>7xHF!IDkn-B9ZXZse;&!FkiZ!rv7Xt8&z;sNm9z9rPkb!_qq_5RcDSvLP_Lkv&YWb z*oLp;b%^$_#fC>>LfLVf3-!-*J!LMg&r%=Hxwpqt+goyx=g@vRN)U6bKtyoz?IS1p zvPjw6UIh%)UR@uVSGs;io>QA7Sk|~-J4Io`{qZx3sV1Q_t?n)k9W7ZUNle=0>J7(t zW&5mqzUIcbq419xMekq1qkmoxofD{Uaqa(TV>r9>`@8HM?wo_e_ZL+Rzl->vi`bjI zdb6~8n~>cQgKUh_nV$Ko=BHhWhK~l?o6jYW|2%&zBS+TUE{K@s=XOV5jnB<}Yx%KA z24?}6)Q^)FMn=4c%RAzlH6FAUpKHTqZ~f}(HQUTzQ69B;gYHPj*XsGv4W_}&lj|4z zmBP*|$@>#Ze2m-*64bT)uQZN}#)TI)6}?n;^-#5RO?}5qej4k8b#6DFGzxiN)Abzsg1n(^S)0t{ORE1 zDISY~0J^S&X}B!o`rfR3d$osVjyJutnlF66AR_unZR(a@N8i|jT8zn(3X4ZJ# zSLWVAk+YMedwVuTl=A!t63(#gU|fg~t=PC@-eZrOb&i;@#$%N_bGtRC)BHC486nLE zc}|A+Fn)M4`HKU0@^;hjn+Fx;{6HYM~5?;j& zBUc;Ef41tnN4PQSmFC3g< zD}FGdK%QX;PIY#xs%66;AO3ozZr()7=lzy5T6v?w2do6gbbhZIG=Cwq<3p0#CXNB? zv1YYfwPw5B_e}C_mgd%qOZXu7)lgVb+;mh+(mTs%=ls4>{_huju7K)i(*{%F&vuGPv-dL)zd!m zWW-iGshza>+~Czz`{R4NwAi-q=14j6+&?qt+s$- zHZMGDQQPM>j=SIby)(|dfj?|>e}f>t_JNU3Hh*DsoiNd2ug=o;G{)_prhll&y5aeg zwow8f>j)g1r~A3iL(J;RCq4Ch>pwLXjEzrLzarHh;P{n$>vCI=KbOlhy~gjMyNNYw z*V<;Tc62_y?~@a>>a~4O=40vhzOf7K){J?S5viY2{k`9;?H50v{M7hIQ1$1ZEF(wl zfIE{**JTY!Z*Mplb|>$vFJ~SqvmD6#8r*6jPiL>HBU83Bnoec$nexG3-g}!z=YED; zJru>ibso*1zI0$f_o12A-RDf!VW0HAO7)-YzCd?5Fs!GtO@w=_RaeHYr-b?P)vu0a zJTIQ{FK%g**r#PH`#8^FeKlc^@-r7-!vv4v=cXQEms;oF+T^~puFx?|k7r1WVQKOp zrG=c~UcEUd(^f=oTrOC8H(H%cd+mLu&*Oqs%Y@lnUv zEc*RlUBs*gUOMpE5E70bf0L#)oFbpMGi>YF&JDA|%KF{>@?}-6*RwWgP4cltwIpoq zZt9dc)^sZP3;Vg?Z#LUW;yjn4U9>YbT~*Jrg{K(9SAG_4grU>c&Xk-YlKG(+gmjk1}SKE17H&}Ilp z8osUT?yXi?&F9j*u;y!S#}(%ahi#7(4AZ}rFs4-WUhvw!BhK&l;Fq@ajP>Uyc4pV~ zpA5QCxc7=#kok`C`LOyz&a*B;la9laLNECx*GZ?!O?w924VhY#MbI&mlsUYxN}(rx zuE@;sPf@*liV2JQ_BoA8<>Woy@+u9tbXyWWSl-(g>C4neO8&l+BoKzwysh;UZgu+> zHNAE6{-WC{Io8kaBaMpi@0UwGH|dO7&z9^kR14{keYzq6Jo6%P*9`ybC ztT1z;A^R2Mm@5@c29X7f!Plf#G4eb2$iGrt!++~l&gRVDUpxoPU(If{eN=HS_kEz` zyuop?uNVCvnY}dVvq(6;b&OEa`$n2Q+`iPSyNU~cNqlSF*};$-zm91e4LMzP5#sq6 zLe4q~DY5=yqndok8Qp%##uRUgOwYrU>J>+5 zmXb1ct;gm@1l^68=3CYQfs5ayq92vjE_HWu^@OM!9uHIFaU9hX4HEk5VAo&!?a}6l z(3c!0I~gB+OIDBBD3eh(iR6o=ma-ky8@W}+Xvg3tloyzDy1VvqvXdlFw=i_YNkK^F;30 z%YKMb2@I0laZp<7fc;#6mffb`lXVV?4)dmnf-4^q{(WZBIk>}3`O4+}3){TAQz`=g z+`Pwn@Zhjr7n_J&zc*X^aSz4{$)?eU+n0ZC{%ubds3c8<+L90vo!E*-rVgzU{=EA< z1&7B!9pVt?nO1IlQ*AwOdM~#;YGYZI{E}L$mIUQ87mxm~0nHaocbBTq+nva0^AB8; zS1Y!c-a_I{pX08uI%JEif$N+)~UA9u;48fC=j zKe~f*ZFHlp=ht$E5-z>Je%R2zBh=4-ooDjf(u(UtHq!xPdM#T%AU7O4ZkfEbe@pjW zjcIns?$m0xnc#Gv$1I16gWv6=m+<&jchsX;Kryd&I85{&bDzeQlAyg=&xfTd7h40o zOj0AisMg(o{rSOwpxs03&-@qqZ?8R@e=DK(xlHH-ugvSj>-_IH_B}tWvRM6M&Ak`9 zV~4npoN%wo{BkSU&-Cc?_uT_Yv z{$nn++y1SAROX2pjtGtZlNSyO7Dd|`pOD?cs8w@d>YLST<+1k<#x5M~ds8{x$ggrQ zbnEb?E)Jc*h5Sq5`iDe`*PA5?4K)(`ntpap2PcM=4_S9TF{}EC+7%zo%RW7FPTl98b7XTdqjFXzQ=b3q{XWOO$WhDt$A|A6 zI&XJIrv2E43;H~FcslHv#ZFZ1_D$Tw^x<2Raoqez#&wz|=g*{`m8Adu>Xf2^``q|B zL+KQmG$HNWjb~UXRaXXwcHjGj(2s^FAZl7$)A1AI(kZj$Yo~Y_&d}W$6QC2(B5Qqm za^thb(b}o*p@kl2*QdV<4P z7#}Riyjzzs9GrS_^kkTz&)Lq*KOs{r^s;PM@+Ycy7n8(tyet$yEBBlZr`xa7$k}^6 zKf3+lDosNT{=SSvNlB&~rPm9+Z!LzgY-2ZxtdmM{lD#rm*csw?j@SQ%+9Q0{wH?v7 zu4hFRw~7ANc@mkht~J#w=&*dv?8o9eJi6pxg3IZ8z_X9OFjSf(=FCXAaL_@&cz)YTyG4=oh)D) z4$kfCl}WL(-~V;?^1Uq~-JQnMrNQ`hzv9o`koXpxoVJLNohhG&bXMBo-lKkKlCHH#!*ID86U%K`qK8#IGy$ZPhE5({A&^A$vpyswfqz=O);9O>9U? zD9+|5EDZ_e3oP<&zyE?{sbcQ4YeS|@SV*sfuQ4mGQ~sD&-TJ4)0!5=+zkTU*a&5bB zF)go9NZ+90qM~FXy~V1?DVj5AQC?WRD$nGHw?cz%`)aQX_9aX5cI0g7wg^??-tDsq zn+#qtg(mIY@FBX<{z>SyZ71S>wO60NxRkQTgWC56I@%^?*H5i`w8zS6_00jd z(Jg^nnsznGJXkvNXvtXq&E2ctt{ikT)Z?I$1Y7w2aLAL!O^@~dcT397$qAoz!Q zxF*w@G^IV4%Bksu&llihxX#Obxi)V?hOa<=dR@NT8%62%>Kg@59IfXl8|Z#lAM0;f zH0jsgt81cZJ~S7x#oXu8cH%wJ-+<|4i6_r6?$* zcV;Jd7N_r*3u?J#-##;-SI{Ub__#uO)4LsydRN^WS$Gzy9N8!w_MUV$Fu+kZ*D&My zH=|(|M)!A|w*&KVvbO8an=Vf9^Zb@sFZql$e*E4qVc({ZCFXCEt?A}&o@7dN$l3N{ z-U2SKb@}Z&?z@zD=cQHSA2vSj^>);aoYT287vCmp3x6N_)OoFm%E1vIp>Xb|(EC+8 zEAmwYPEN+2*mQvJhJ4=wGLWUzs^)m8BKZ>1`8dMp*V|I=A4ejWyhHB3-#46iVD+hD z!)u=yiVn_NtoNF>o-+(i*}-ssvjlw)q587jPooky)qPu@kk+jl7n}5EuU!l!le{*# zzBSsg$@W$?>ns1~Y1xS5uB`Fy-@_D0{a zQ%CfNZ9G0Wo^+p?@%!oAA=RxTa-u6(sUiI%foma#I`C?Gr?N=eI~EY;MaL ze1&)2+1|5FJGbDv^}{=lJ0I@5o4Iw<&I1NJbT#VJuZOv={r)z8U02I*7pHB;1D#Cg zgK8=cudRK3;IPs6j|RJz#441O?WP>=`l_^@2{~obV>4aI$5Hahol{()Z}8bu`yiWR zMz|%@+h=ci5X`nuU7Hg6d{}NKszPwErl;Z<-LL4Hr_P%^e(SyLIcZ+_UL{`UvWRh~ zjo`XR)+M~&@1zE5YcAo6LmIh`zk4z5FHSGvFH4YQ;r2e$4p$hS90t<$rvw6pd6@SF&ghT|Z`S%*a}j2v^x}Xg(sRfL6$kf6zM*B zS?x}sr?lVKMN0C=+p4aLPo2}eZU&5OxajO&9gv}BpDD5_LAPMODJglp?x0M^y)%Y* zWnAcNcXhA#e6LNJ^HtX;Q{Ua3 zj4JQP@+7S{9=oAQzM?)L`#O)uq0+ZwAiXPnpwj1$DkbMfK2wsI(Y zilk3A8wh@`h)aKdEi&$*QFWn3UX{{?Cxxj_qv1x$-`9Mh9AYW`6Hcd}pR#*=O}m7& zI^)fg<#%qLc~yVPyD&u8WY^|dMLh%YJryK*b7toAO)mp?$rTJ__8k~YT{u-N=0@Ti zLz1p%5eqzWUeJr&)K|Nu;Sf0%)B60V94~XiX|1v%_u+E?U6nrcKPq>rnRp(26ySNN zbFti{jML}w5N^P<)@JE3dB`J3s{6-74b4|ub5F-Tf5pR;gv7F+kv;R+{Y1@$q0st} zJN*Y0E$ad@FHVd)J2(;5$Gtm(b_~9eSfrd>eK+&VP3!bew;SvHhsd=%6Km|f(% zvi8gWuAsF+`P)}^7kxLQv}1YgZbxU1?J{lWFP?5W^^CsGD^8$RQp4w<*yrK{wo7BF zRR!`&3wiV3Dt-=kH$6crk8D2ky_v&VD`;`Q+!5<@0%I1W%Y3_Pxw4KF_{1ugh!qd_ zjIS+|>dxq0{A9SsVz-Lc_`9HXlAves-03gxoF)9H_*DIHPQ|C`*`w}V5G;B$>CmCy zra!gt%R)USwwZJL!S=g!yI+ir$a=Y5jyxoMQIw)dA2RCNXCnTon=$?EKEar;+S`{@ zcwPD$g*wIej_V$(kuvM6e!gpe%)yUx@!L{`6gSto2L`|EIq)hg&F^Qmxn1j@^iPj< z4Ac(0-L%&KF7N^RzP;e=>M|{>?FcJu6}Q8Mi!<68k_8!mi-`asgBvZznM~h08#mwcU)k7-@4a2u8I^+#zLMpw3 zgxC95hUTq3S>)B;GZdlJ7jcIr&96tn+^$%zYyOS7%Gl1=JoZT+E@f4&4oFcL!fO!o z>P}S&OWjOc@Oar#Y{RM56{57Y*SCZ-;^{#B7vy7T)`y>wfma%}>+-{=xDMCaHJDAP z_iVnQ#5%S5_1WmU)nCI8$mbDe#BV#Rd2jCfQC*XpcIsJ4hT4RAXDUlJhrO1cA`(+@ zl2NvyIFWhB88g?jq=nn-Kgk#ev#ritq^HR3NiY=+o7%vB_WQw$d-zNr6snwmnp)N7 z1K*@7^P9=keDizma@$PUP2vdG%&Sw9+xa~|mk)j@7W*l6LQjHG^x8y$G5pWCz1i8@ zu3ZR@Yj0Z9@h=Gb^=2b6^457?a5dT6%C@v?fN=OU|wT;#OWy(fHL?0b>IJKkB0CEqV71@}}E z%{hCt7Ru9Y=7NMDU%PFqEY`H{jF)xJSr$2s$0EDkLwVFir31V|Mt^M2>*q>*?km~h z+^wblI>{|RpjoiCzRyVIZu^^KQoMWDR2|xC%b7lU_mgV`?&B^Mldb1G)TjINy-KI% z-tX@dXq(x|DB_<Cr_q%RVkh?+#Ap=@|wMWmO0{qtXr4hjqHn+cTce7CMdsqC9>na z&*d9LkEM&R-o>`G@f-edzffJQ_FYz8`bzQf2Y7`Z&8nf#TD^ww)s@woIK&d2gP6Y? zZi`g>Ik+wi(fYD=axozHQ%2b*y@zE3oj>mji0_;dv!YK~eE+-5er-ge-^*jX_hm(| zF3=ggvF~wMZ>ZkYBHhl%@Z#gMLjU^l+dDqXj!tjd2ghqc5GCx%cHHXqV@p_N>8$Wh zg-siVH2J+9JAE$-9;neQ%DVd4+bXCx|0u!iXV&+{1E=XT5`(?wvNJYkj<~J1J)@~3 zgL`;{$aHNj;j5G8*?$S`8V;0{3ewZ+qNL0w|u^xD)X+O{j1OEXtn0UKIfifZvNSz zL$dO@`S`uonW=RF&p&MAtqRx_wExWUtYcfpf{nWQ*O(u5Ryjw$>=8o{%@&*$cFld> zG1&S{V1WJ3u(5Qykm2#WXOpZ5E`bjn=1!ZC#9h}u+$5m1vFRR{^N)xEw(tvahDv)g zPuqXX5+#V1*lzRjdT6GwM=Rk0es4jOljn~8H-<_+*Co6t=cw7I|`%2#38?QJ%H%hPQja>9Kk$$)@ zrajxIELClUUyMDUe5a$B?fuj&@>Tau?`Pdp3!Ey=<%;@?UY1Y&1A|1vLvQKi)&=aI z6ZBA=Jjr*1+tJdN{g`Z)#tS;tgY>63V%$fD*2F!Kpg4^T#C%rJ^Y2i~(1`UtH4 zH$YN96Isp?t|I#6Y<+#&%Y{`kChzCz44h{;8hjtjG#fyLqwjsXthFYs?u@~(Sze=lu|C*JjNR~~z06J?VkA-wu-1#{dU zPkzl@%cD$#J7fA8KOXw3y8nWh=$#hcOxd^|>Emr5=K?saZa3XNW}-1D>b)gQh{-SI zSKyB0L_Wof)%zLVYyY@P#l_>Xydd$q;4Ud;1J0jS=UZ+$vMogxr*H*6 zno*uRZC;j`>TMK!F+w8jiWmFP!BJTmheqkTs}loFj?t%lWE5DBNifddew2OeSo^lt zKV2#GwO-d9=K?I>e1=6}z8St5F37I=Ju`XlbwMP=6d`)@r?>j-#VkhXWY_Upj< ztNwZqt}1ZtWFTi5K4HBf(y`}Gm`slFu+WP`10utsx7lKGy))JFOS(L}?{Tr+T@nec zw+Pj2{^q{h@{DY`u8!TD_@HlT{u-5TeUX`KX%)9pB{p9j^~$*{o-+IU+f3ImKWC`^ zmG{?6(%zPSH{UL#`K?Izke4-ppQ4WXV0i>rLM&5jc`7^I2cg^hoTLvzEpzq0H{` zyo6oV`#7c*MPtADUn;!ABUO4TxSGu+0k8h6FzM2T?{==I72o$7H{BgkRonhy+9uY9 z`*B>PvfYCkA@6_w;~$M5<00@p0$P4NZC@Y0w?M1?AS_fu%iotyyFC#WUV$~V{wJJ^ z@HG1oTH$z}9H!lWaWrk;`v`6ST@Y=**Os<#ok-iizmK+`Y)sqdbfGI(OVd$8aVt^3RZET%-Oe@ilLKRJ~4I9y!ediRW>-M#@9!lbp|_bc3IB!+1p&t_Oe ziq>_v-=B8>vnyN|@hh~)rO@uDdxd`9tnj{Wt}vb*D~ubDDee7=tS}F?Rv0&3WjxLE zuGB!g|1Hoz>3{BG@SWHduIu|NTrayf)9zn;g?a87K)e0u5bg86afN=+?~2lzSLpYZ zY5A?pwEfj9T)#K!X=iSD1+U<9& z@P7MNxPE`E@cb{8K)e6>QriB*6`pU^v}w0jSm8XP-_N8q{`+77Lt6ezSd5aEpS!|1 zOssI7>c!FCulfp~Q|Q_QwD!vmpY60h-y2~3X|)eq;r*Uj;W!Vj@SGC{<3sCyqOqU$ z_>8VFuL@Uq9@wzL?<4XnTt_o2+|SV^J86v%d4=xOS z=vo=H#xrz<>!>S%_Hk~4`9rH8+zQV{-&Q!kohy8VHAMcnq-?^G`TakMjn)&;Z1`Wj~GTTVRU=cw+v- z5vt$yjM{$rn!3P1^<&Wm6Ck9p_VRE+qWeXh;QD)I0BH+xdspQ`xKTUm4qPBP^zi9?vZL#eN78%qMVB$7cxp zlD;nHv#+AIm;TTD(kZ1L&shs<|7)=J42jhClGymLVZIB@H*~*?`>5j}i}`ZDsQ&Cp zYCp^4N$*HK9>r4XaW0RKoFldWRBVBWjaWZ~J=FHw&@aV9EYBqR!{#aVTJ=wr?~3_xQYm&v$g3nS`kQ8)Ew{ zh@~FSFYJ0;p3h;h@GRQ?%A0@h_ox8XKZ`Du0C69^aY87;BDHAyF0B3X@r$@q_Zy8Z z1Op$h{{#n-x?h9+)Z*zD_2+L={pItU z*iG$!=M1&~$KdpDE`#pp%Bm=SBo(G+ug zjM~rgeJ1iVwI69%-XD$Sd9Da20_}$rUC<0-x&Pg;SUKvi?xwb1K3{@G)b?|39pK2V7Uj(l$|J*Jz?>)L0RFL&dIGKu|;l zEKwsMMNlaMqGH2dqS$-yy&GF>v1{xxw%9Ay7<+7achB>T^51B#x%WN4@4NTDkHfRG zv$M0abM}8${5ee3%`{I?ap1DPWC{e{fFGq1``_ex)SK&(@YHCH{<{1Q|8cRjQs2?* zj&>jN10G_}fLY-8(;uQykJ!_n@z9s?um*XukA(aSe#}PX`>Bt=YyJ0L1Wy^tr=0`5 z5OLAdQ|&jk{HikOZ!7wVo2tiZ`~l;?#L1WR=M(hj+T>q77yWfBgm&GDr^*AK7Sn*M z@{#q_RUfF<{J#1lAL?tzjLf6+>dbZ(+CaWH^>m`& zzTn4*-mSm2(C*yj=$GUhUg{(CTF(}>z$5XhQ3c>{M?zlKk-&-QS0^KXSUwQ=k>Zed zr#_+TBix#Q_Y3g55O1kIn6L5Um2KOtF&XlkcK{dvT$F`84S|b&?yJO1^L*J6c2;G7 z>$x}^{8{O5GG3>?f!#6(gXa?E5A*`hN&Nva^@yFvrUFk@#`(q8zpC%bZHGRgN#GHG z{=yA-Z)3fPUj+UqjbR`4e_GGdeZY@Z1+JD|>*;kH_!w{C;-AO<0DfvHa9P)W+63I$ z&z5I@Z|HyFC~&dE;M|Z8)gOmcj~YJf>8LJl(Dqz#8S)a}zUMl-fa|Q-r+G&3c=6+t zP07Dkb{Whw_1|fX~t&M^lgZTP>xqZnu#B$enr$GA_+h^IhX>7-yc!g^Jf%z#%%%5QR3c%Fuo7?;WDvXNp*RJ*5CTzmXeimPa6xKo6K*;J}0@pEaCbh z>w9z6kmgy^)mG2;J;C3OekkL(GYInc^vC?vBX)~g0z6k!@QDArWrsekywER4>#x)| z_iBLu#Tdv-{*#aE$SPwWQLHieFX@jCs%Lac5G~6Lo}{thk$9*M|62V@V7rpfSL3=o zn(ML`XzbeM1JAbUw)#(3mz8MyINSj)cGxos`t;^FO1|OwoIH%rW5}P;5%}&o;E{Z( zp}K5C^S2ye>sN;ZfZMGFF5{TBCfaR13~?Zk{H5oCzw#yUmm&U}`X84536PiglQR?K zgK7UT%J&=&`M7G}5x?DW8v1N}1bxH~9tn_-)gN(GkNCrkp^&eB0P=-6AHxr!-E;gn zp750W4tO2z*JS-#(+~V-%0eIE56K4nZ9DW!@W~Uv)8s7VWxfpJK6e)PW#XS_5+Q$t zd6A4`o|nLbYD51vR`1ll->Uy)wwpE%xJnPKXA0M&b;f$MdoS$J_73EmkVk*FQIAV) z{V`zmh@DrR1KzGa^betYgDK!iIt`wx#H(_Af9eAs*>_|g1Nq*JSF-QO%DmB;`zEPB z4B&cw&)6q!9}9i1xIiBlwyQ54(faRXokZ-W_0{+Q{jp>9h#yX_LVJz|Pb=$R)i*Hn z0{7#*lX0Kk7P!L_$XBQQSmsy$%&)#6-s>LPUCRAP1o6x6z*`#ovY(rPKg}%gi~W1F zM!%v~!_MMY_veFW9P3Nc-@nR2zWPYW%e=nHedQ_cD}{eRHOSXA)}6|Oq0b8a;YsyK zzUS=>e9L0s;?J3%0DoBn?Rv1k)s({8{*_~3e?1=h30D6@ic_rHexEbCv@w+>Yx)%Z>Nt*m#&)ulTcKgIYwlJe7A z0k>y6LIT6S&^zmj%$aeNQ;1i#FeURU&c4V7o6g+8^& z6R9@ln#YdoxJu8gr^8>co4qUK3ldM)6nOI=z+a#E%|^hNFn&t>(P^IM&&%~i^22Wz zLw+3hS+ZW#;yRmD4E*loKj#FVaedL>5aMH$18DvqjwZKEK0GIq>bxHK=y8b}1dc^(#>XItWKj$Ij z9m!K(?Z-7R)GI3sZ>&ke}?^aYRTpXw3pFrym!rCvW-Pp#RIw>~UO!)Poo&mw zq?Q%yslFWikBWlFowzgWjni2Nmhrv!k-G7Abzz{^Gutchh@A&q1%AI2`V~h0&FRtK z()EE?C7y@*k0S(7!S9O=_L8>P}gJ>$~anQ}PY$aJ??@2*7q<905;N z{n1?YsITUq=>ra6nvi?4^3zBFX6e@8}iwV z{maUXz%R2-(ue$xUeIU367aYXU$hSRJjNv%_kn7C)b{`K4fIsg+bMC1m(yCZPv z@33VUuV-8@WE~IYdJ)3)BA=y|`erWuydLLo3F2PLVYGiH_(Okp;wi(yFCS*FPkg6Z zu(bSPt}oq*AEKU1Izi7d#9hc!pLG)Dmew<%2KZlcA0c)qrp_B^{t4@0A6Lq+os4$V z)334!tC;`+N>Ol=0Jv_=B6km$9CllX#>`P__NL8S!McD%5DaIrE1qlrLTs z_^R&kAIVpHJ_Fu??}eqmhwh`lZp>H3Kh-tl7XNF;VR!QX$^FPOWB=QICTDns^pfSJ!sV5(}OX;?K@QzGN@RtFU7|Jyl|F&6noD zB|p!^a}F=I0k1`Q#}$zGDg#{XU#cbAJ--s|mZyAyP>h$?PmmWs$vhOeH!ow5de?$6 zz!SM&7XCenXxE;8Rhs-M@}k|E`lJ5p=|jBj1jxUvi*{vRxA6!6b=IR=P(JN4$e%Lm zoF|z#PGjCE^YoeuP}*)AYk^Rwkap9s0*jLA&CI?an}cdOUDhN8TO;o>&vOtnV|kfG0&~;Ns`!A3{F! zP~gsNH|r?i)oTD3ze-(?_BYqq+a{5W{p)&nkn)j=gf+Rm+5|B<*Z|=n3REPeXxt_{8Qbq+X-EP_5;1@qE z+z;}(PC$RLPj}Ye9JrsA{CqR_%c;|$UD1EMN|-f&KEA)mLVc1Ht;UCOpDXh^Bj1zO zVI5xl%7^*kj!@_+CH#+QG(6gXB^btRKaT0hU zu7Exb^kW7KkvP*YJ^b0>EASVie6G@HH;HwAiNkaDLH`o8r?eXv4}9Ks=p%k|uNrvT z8g}@Rbucr}%)Sadd#-{O7 z0nu|xSKv*I_i@wx&~CzG6=2nKh&+vsq1}VG!7u&Vs|Hm2LlpBsnHQ}*AU|xXtv-$R zq1`igV27W{Ke`rp<}rVkah$&&c$aAKh<~o_4*mhgd|bH~cqU{1CN@F4i9E+5(`^cURuSK#;lwAJ&ouYix42wcWxbOz{ihx5Xbc4*3a z!mw@NuSvYD7xesrmp$rx*7eiiD0tj|x7EjSHrnlf0qu&v^-_UdkJmSbp9cqkXAJWe z8ONvhAfJ=#udH7^l){>)cUQy{8J9+^#|~n@RQhK~O75uSDp zfY;NPVXCJn<@a?*96P-L`27_4hwOY`Hg6&9Ao`D?-BR*WOsR7w^8CPPo*$6-e<=a& zRwx0x8kpFBr_f5-DSlD`eCigl;uW~@7+=WMR;JGs7#K0m8OLfh>j^KFTfzx)9^ zC$Ro6^J{(-`kTl)WFM}-PJJQ&dLQ(ZcsQ8*nhVDGUL1k(opc!Xxr1)m9Z(B_*5^kf zUX9BF+=cnoTFSrr4LmD2e`UPxszgHbq-VdxpJ(#?-k>D3TZlYcxF1|V|8yaqGafvd zi=bbPiNARYe3~C{@vBnpfkzqV*-~8t-j4MXSMs!D-ZH~jua|#Ae&+w;=T5A9PiEa) z=KGxTu*1et=$GVI#aQ3)XMIER#&G4R+CFx#ATQ$^$a7!Wc@9(T8Bc%SO@EgDx^2g} z-@gGn2+wtuIBEV-;lQO1@Py}nx0Oe`E$Ki0k?3#E{E#m~`@DJsd}lCl^?zE=#gpJa z!uTfbwi^lgCPw_N%A8aC^po8L#YWq1X0VmKr>LDF3BWK;uD-|6*tVr5LX>JpUkZI|bisFJrw_ z^pB~9evMxael^n8(@g!R=5N^=aZ>WtHV=WX`xf#tUPqgw-RmpRFXi6W6QtH1{XU{Y zA>iAIx6cOojQm|sIpTM(1CM$Rc^R)X%E7cgrF%jjvCqLE$oF0gdEx26Jh03V;1T36 za1ip%xUUhOk*rr2WxZO)t3(LoPmo{aFR*?)+^FBiaz5_m`=DHG_i|S7JmUEi!ChQo z|9dN7Pno~xxPO1m{rh6_tbGJKq~h;@BwqCz2>x^H(O+-MFZ&kwxK+RJ0RdY(@MKG=vezmq?25%6TA{Gj^KKg+8$77|W; zm?QK&#(JsbIbZw^d;;T-*nfmQ&h1>TmEK}-A9_|EQ z&>6VQFSm`rpS1>_gX{GzC7|t7wm5LLd|1z(tgyqE%=g3(-<*Ja-tLf>ddN1~t&3r| z>7yWDnfoB=@6!vwxAy{%=-*%_+TDB}xXin>j2oR8H)Pz`EeB7KH{`|N^0>loYm7W9 z$2Q3K3xK@nlXU^`ldMn5I-YJd@Se0!Q`+Z`{^0j#y-RpDm4p5@7(Xjg{+S1O&hVU= z6LGhb;K|N9wv2D)U67y2bz~sr)AIdk^Cj@ZCd8{&1<%`gz{Nff|06Z?1m8vIwoz^{hGdTQJNzU?t^vHzul;88dAS#l~}w4NT^=O(i5 zByl@}>!!mPTRo4^4!=j(;>VhU{|Mh3irqG(fIi+lHzj$_N|oqpKi|vq{4Uhz1nXKs zHKD((`_cO$Uzh$b<26GiFq-Ej-%m9s&oQ+R(0Clrbx6B=H$b2GU%)SVZcGVzXUfZX zxh?{}qcQqhf&BNHVV!+213Ycj-|Hv(N7$kJb;zsjr}f{*b?_b`n zcigA{z}&2 zPhsFPUJm51L;j-FXTUFzH_x&6UIBbJAMS z`t_^>`YY?&+O)89Z~op>;%B}s;8|53`Zs02y!Hd%Tn)JBQ*0~vC*1|F?Y4 zCF46^<=dL40qZB?pIx}_WZ=3Z@oK^_w7Y}9OUlRiP?`OzVD#&c7U18;_shb+tO$5w zd48Ze^)Gu7dRiCQD@@jfF?_$%mitG^E7o-d&r8m)qU2e74Y+we{HG_t?YZ9+Kj~c< z_-4MZ6o1|`4*Hw%cJh44yRqIb{`ut?;KPl)A{Fz--K8-u;x4Gn>#BucGx+(tuavS(Pjr{FoBdm{KP5_S@4(mx@AA0tjjQ*-+ z(RvaX2Xb=1DgNm`0`m8a-&r?jJj}*;Scp8i)P&OW@mWvsOFlfY0s2*$bp!F+Xx59I z_d{OrmV6K5_&xOBO#O$p0slV6TZtzH?LXYFyz60~JFMp)q@E>NFSV}-eyKm~Q4?J2 zpU@fcY9rgdH68d7H|QzztLAd_YaV~MBJrWeEy(BU2l*Uqw}kq^spe1h6zz)suH~V> z+epaE_;z7_o1XDa*1O^3ARl)I^1_qrGVnnq!SBL$uarc)SBBcQd*}i5EKU1_P`)Vl zYYP$}FZ0xQJmeqMLAybe@2UiJznqu{%KG@&3;Hi1N>W9|7=Y>oiF4QQbFEu00(ay-C5*UwEnxL)fpK2_`-i~IbLMNN0&hislku&s-Y07N zBwdF7V&{sFAYXYH@IGv}KYzy;dI@-4;#K=YK8f}$N<766z!$fG{6OL*8GlL}dHVnr z=(Ijd24TJwp!~5yu-p7z@SjG+YYu__FIb1HPki@c$Tw$QOZuCa@7JpE{hGx8r*qJ5 z#)i;S{A5CV;G>MXYu`@b&$S8qY+$<`+5;c34!Epqr>}wMTmty*DIdZ7^919+_*Ghe z^y|=%ke7Aj@?r3V#Q~RoWe-EUX5PMaKjiE2yqVZ@ekI6Xp94LMQ6I-ukdNT{=tO*I zN8rQ1LA$a(790n>U~S;>l%M=3@Yn5tOT2Z;jCQklpk48|B@=<)I|W?U{m*#L;WuM{ z6wY-j%2=n;F)zwyoO{_*68cPF-Cov_Fxu@1_d6LFpBoH>#}}6%zoY^1 zc9eHe4x@S8y1|~3FLh&{Ie_)aHk7}y2J#J>fk)PnH(b{qb6ty|{M0_sr!C);%J{~p zg;=*+n&Oi$C~O0N!IX-mfG?rZK_rC$!5 zcOjg2GA_l{P-&jGtUt(nX%Gv2${dEC((Z%Qz*~+2F8<%>9C*y%FU(8@ytz@28kh$> z1E+yU^h_HA`NTrNW&KU%3wiVR2oF`lqxHW$9RB7)J6z{_>cu*P3TxK$O(^778TrGv ziNFVZhIVDY+_yIHE#wzFylw!zu^r^QlfP6f#^u^}jEfU-x1Y&loTIzN`f8Iz=r8<7 z)&QW0^h+%_)>Db=`!Lo$#D9L92YnuJ9K}!4GOvp^*6aSP3vJzDt<5h+lQ%zW7lQv@7!><6g?22QG244A)-|WBpA_ zf6Lho`Y1QH9*1<$)2!2G9|`_tN5C)RxF{0(2XWrXcvYPQ+{|wixsQ%D_R(V!!QYSk zqW`ybz+?VCVkpOVR8iPi21R!TjCC+V#*U(wHy5=7-%nhl5Arn|~U} z-|1khPyU*a-)7VY7Pvvrc8uGdsLzgb;2G%v{%?ujRtDGhU&M23V$bTAAiuVRZNCcS zK)aK?fXjUOrZ0HRb>Z6ukT>%{XRa@8jJiz)?q|*4Wvu1Cz9;wf#iUM)oXoGN zvCwBY|x@OblgtcUqKR?)|IEBKe*0FU7Js{=nW2Dtdi zJ8Bbt0vGctNXt$=Z-aQI|diqtO%=fRK17DB`9`T3eKhh5Vz@2E%c87q+bpc+5_&^`%6T1Pp*x?-O$u9YUi$5%S z0(oB}znXI&{8Lz$6FmnrKICM*H#^(i_Y^#@$s_A;8S>N}Xsb^Y-=CWC&7SAJzFdKE z=|g@`Rd~_BBV_Z@%hrHy;<)R@U#rl}^_vl6R%b)cD!J~eHe6^dldZyY0 z{^%Rvcc(r-#ervBVaUrm8}udQBX|x?=Hs1oz#sD5yV!Fi=c6O*b!z)yJ;$>{9}|C3 z0l4}7ZF>*svzg!j5r00&btIPSNEh<|z83N~n?rwD$1|t|QrkJ6`J1eRLuLd2l6DjM z0jHo(an4`yht~;^_u_s*?3|A0WG=G4A$Vb)=SkcRJ5T3)`C<=v+E#|1iq?9HJppe1 zUgLM>vF*aZ(}(h#eu4c{@SKmtne$=L-~7GC(iz~-=#F+}9q}3o`Sh&Y%lwLI0Qt<@ zG2dIT-R!A>Z{~ip5b+&6w`1;iuAhRQyQyac<(u>U=FQURm&Au{tHEzy0(y#H0rHQSM@hf-L_mHMFM6ise$AcrA2n^Q zCrbUH)-y3&4Ul@oo;?l#&%=I+f1Yj#{20eq@WPS6_vAynB`r2PyIaG67p!7yhm~JL zPyeaFC61+Bf^n~BypQPFA3Vd$LtgUaWr4tt@H>{Gf6NN-m%RdcvBP5SkG|vnD2#fB z#DnLmBjhFjzsUT={4P_cj?lC6MvRy6w`M)IW`@4$Ul0m?^Z0q$w6r_Z+oU&`}2>Ntw^wB@%?(d3-Jb$ON ze;e#6^6~XywZ#gLOcJ5<2e_F2k0w&YR%6=65?cOsAZo+u;7 zu>yHsXS&;y`#c=x^qSEqa6h ziSZu9Z3gt8WYnXY_k?^D>o77;bCd_}-4E-9+WuJ2Up%jS^c?yt@vw1W$Y@oX9hi`|j6#|04SP)P#Lb zEk(QKC|}zTJWcpJ$1=pv^Swjd7>uvv2`QO>9y97)_r`-KIuH2!kjFU%+MQ#_=jjM} zZ_10Ft7c(b{Li6XckOlU8 zQO|$A3-VKq{YZcp31WFXNt~HQGJLd{5%1i#>kFk-*<^c#wa83CQoC5B}Q33swT2#n@NoXPh^GKk}VQ zr1iL0G3p`3W&pq72L6HMFQgJ1E#H9euSNe~d!b*e)}mixpZ4mUxt34jezq8SzCR3n z5Z~(wPxf=*59xt+C2rr{1^lx};1T56I2-*fcmcSqgN?^RzO4&ziMRc}LmlJ!2E@IJ za67v^s-U6ubT|t68sz^<6>K%0i|2DB&O3(!KgRc5YFV+K$j{KPxO%XIjC=FFkYB{# z*R-O1nIXXA-#`(OFUPzqk$IP_`&E^pv_6j*S4Gdy`QBTC?wdI-2^^P()MsK*@HAt5ll5pi^VneKv7)EXWAM8V2Y)5<4B&h>?+d%% z8g`hq8hXk)6}cEZ)0oG)ktZ%c^clqS)uPWtt|Lyyx{zT8c>Gzv>O-DiG5{~d@sjoH zuoE9%<_j`ITbY&qwnt zmIY5g-Y-yu_^*ts`Fq0t5>HZ{gg#!6fveXy))RFFJZEV)iT`6417Fhw^3vbWasWTa zyxNuR=GzGTNNdQ8KGTapf0w@KmrAp(XFkuFcjGy8;cxx`_B?0g8+U2XR_VY~i#)NN z(C(1DXjlB)t1ft+O#m+QYr_oSG3RajHU1HB_Zr}@NB*?K!S7tnR=!*uaPwZdMvKwk z{6;;b3hQ|1nGZOUKe8lvLV12q=5M+#&}SphjYvGa#dW``vF@+m2A;M7wtDVTfnEFa z%}r^n7G<6e_6BbLzT+9s4Vu57^>>H7=Vn`;20TCJ%=2TCU&T{?4dp}G->&z;vy1yp ziEpWT!fp+C?nwF@5&}D1+<|@-B+t-=;IVHET)C_Dj2R5vth>f@|I&*47l{L7&Vql< zEAYs;Z2SuPY-|L75W9690{+Qd$Hi}h`Fm|Qo&yp7*H^$}-V@h)7I;#5pk3krtO)c8 znhRX~d}|dD{q@^w7y97bFnd$ITj8RyS)sP`lK z`&1A4eJaVLnyZ1(`Z%zT>O!80bD@tv>oc-`B{H9%!F*o)B$E5J-o}3IyPM!SL%%vn z{ux_<54vUB->Uq5nRyRgmx>s7^PafcJV#OcBJ3b`8y1aznfJuCISYIW_X`E6&&d;5 zU)tRV&t9%C14cnUODf>*l%F{i`k42-Jyw5b#Vguh;3{#m8i`{Bx6dAd*TAK0MF0ww@99o^gH&6vv^NOChFh(XUKc=T%7pxbk>b`8+GGWfw2GO zt>~{CdGhoI&q<^16cGYDUuXPjLV5SfkT-wVaJoBiH|EbpD4&Jjsf*Z!c4Z!oa09=2 zA6&{`>F3<<%l_-j;=mUTg8xhW%$ybN4q{zBJKLQ*2lB)ALB2lm8V?{pd=T`Jym2Yx zk34ZtvaPuCwrI(@4Ud{^#@+9qp{1E0pPQ-geL%uod zTCyK0q|Tq{IGJ`!21`=Lw+zpbu9<^=Ngb~*&q$h_;4fOd~_ z9f_bm&*>-axL=U|POOD-zi6C0YSIrpB~M@+CEv)!{cOLsz=g**8{}_teumDK>8xMZ-zBd2#kPl*>+=})bJ^}4| zegi!vu9i{>zwYme5%9O7l<$-PJl6QV^9I)K&3m$b{SEr`evJOQkY|{3B+dVtb#Jj- ztE12}>b9-jUU1)9lM2$_t(5n{P1^Z*SuG2J@qv2B~Qit@bO^iBXMs@)5@OrALIIzh#Z{@a@!w zMcW}O&r3_5yr~uRFUveZ{2>eDws}AJ-sfnyi3{u?`=-KKz;mbpaPj9iuhFjgJGC44 zfPc;Xm)PM-5ODK;@V(5h%zNMV#-iOQ>M!lasKSEouU9t@V*zmKZxzN9^Ilqqb&xmf-tonNuO0v%sq0o906cmx@EqL7HQoq(^L5}GsDGY1 zu$%dNjhdSwU-l;CYg7K$ztG>x{Joou`=eCA&HIZT0>M9>@ma?40N=+|UxxXm$}HAX zka=dPk!S9XfY~E+L;XB~LxgqaDJCE-S9`ioU{z1SuY=gXv z%L~3A?9T7x7a-4J{@%p=&R(y6XxFPOc!p8FYX{&@sE>MmV?ACYV9%&Su>Y^bH#P!p z-cOgH>Hzwj!>p?5Ey?PXzs>sap@nEy#<6!9wEKv8n&3Zo1OFxC_j;eD0dC&&-X%3~ z^SgYTE`i^(H`=X2edc)sugY^cp2TnZgU7tLJs0&q$@;eV{}$Fyf}CKVR`lnx)xf`* zzn76bwuKMIu_WK`%Ra)>8S=mK{9zdNsi4eg`A-n?3R%a?rGlP+xuIRD`(053r{xPX zpBF!$ya4T*_s3241HSx6*hl#1G9DgeJnYVXHGB^KQ#^;}L_MRGdb(ZnKDhqO+cWe1 zR4vM9ABX+Ec^_-K3D76EQ769>2mV8ap{L}5mpy>*+xuSUPjd*1IBfPF^Kgg!FvMR~r?{61by*4cbnXOn&%j-x)td6*P@zhK@g zxAt@BW8NRPX+CiC9?2u6!S6N#{gv^mULCl3e_Z}v=+_m-F^T7EcSHU)>vFR1tGNR0 zn)kM^Ujun>zAqM@{(T{D-cNUkalYws*i-g{2^GMzXf*V1Pd`b)dav6n$gA*WJx^)3 zqx@Zu#FLgY(5`tea1ICX=jsjqx#V#l1Uv=Lu}D0;9}D{|U_D>(3p0V6_p-GQM!V)c z@~@^t-n>Wn(lX!&FQZ=)2hv{Wd!MhMXEEw&uL>G^zTX>&xFqBIDg<~FBaf}CCZ?7j z*a7mAw=}JW{+jp5r#lGwP}b9A|K;-=a2KA_ko+ny>qX{$e5-lxI)A>BH{p2$vCr*+;D5+EmH5dfp2IQkoliKAac^D>`bga- zPhseD&iI|wL!P%W@2SsF8S>_RqhIy`p3T_*9#?@#_xCQ(iAfwQQU?08W1bMk{%+q& ze&hYrP31_M=NQj9%X)XM8Sq~}$9PG7=GOzTPu=_Iuh^#$&zaMZ{AoP?tvw-Wnd9UT!xsW&S_3O73{aw@) zJYvs`-so@8GPK*3_AKZE`2>D%U&g&uBk((|wdMJvC*;lh0pGCxu*|3@*j<7A66ReJ zCx>MNZr*R}&3*J%=JT=+uG|QD^B(!X)Pk$W-MmNsHvM@${aO5>yQ+Iz^NZ(IWL`V6 z&S2he*!c+TP{HWeR^>>VXP|K&K7JqY#Llpr%*T$sfUh&^;q_V9>cF~|nhw@8H57Wf z{DAT5MEnrXuXZ>MyejeinV{#8v1nJ~aO;l1&3h7CkjIfcb;;9%`-rWq$9fU>;CuGW z2f#0O`(ix!Po4o@i1N3U!L@xHc&x)|$+z@k_wYQG zN_$999ZU#H#+-Qr3ZHBygpL?q+sQ>37 zeaTGloAd z^&8~JIRYO>`}F7re9jT@I1wLX4}4u5aOv0RL(s>(*Rt+K;P-mj%11EHxb8=sDM~|AqW@6NJM+Hvetn^jd9S)X>ywGBPnM+qjhW9(U_K+`^(FI=SYy3dNdNCW z5&BD9o%j-Zo@d@D^R7xu+HH}o-A4C^{muIX(>8~E6yuoWZ$spKBhT^5d`Y;5b{8^k z%eeU706wHI+Re`X#-9Lg-e1|G2HHK$a~-n2_g944{^j~&KPB^bP-DoO_f%dO3*7vj z%gmpFZ!zAx)lhXj&2Qer?!)+*lkrpXmX+K`*Wx}}#;cX}{bYN)B;)+>yvxvk8uP=< zw8Lup=Su#*Tl%{uBlL0R@7X18O#U4_1Nhw!@$(@WF^=Z<^k(vVJFPgs#Qx5lU%ffM z1P@k;v9^O7_d&8Q>{|ldycfPQ>y`(Mx@C)K)P+1-VE#({{FC{Qui?*cR0nmtulYNR z0<=%kkLXwNTi}uYrl3FE;rBen&mVjRyM<4JzsWou+Yme_7K2CnRi+E{Ecy`gjw#Xb z!UB+=!1!E*`m|7W1O2`}i*fH>|2ELSIP1yPC_if?aPvO;;u*ky+8z3c|4d&F{lC0o ztA7=>9_cz;e*@;un^`LK`lch-HhYu;;FQQ1M;ZAv<{Yfn9=)`9*nxgQk2iW~|&OBGxG0gr%h z7zSJ|6V`K;<2$iB@PWh^uZMg@8Q`icU_Dh20ypm+p3nDy{(KL(nDVK5L*Bfi-w`p7BfWqs1RQS#LSddP_0d?de$bH!bUd;ty%ufp1%Yag=@TCDxhGu+E%= zJj=b%-?!Yi3Qr}*nKeB3UX=1xnHPB)d6C0I^taRy*jeW9y2{Y!1mD-lJlf6qHHh;| z^3q%C#gO)A^B%^Ulfj?g75r}0XH5m*t$0s?)?YVK276VE&zJl*PmN8-};GSI(dB6#Xi{uTFAO;`_?x=__=;OXQB9*NI2KS#Uf z_n$-CLq5uQfAm#dv|E#LQpWf8NA7dH-W&YC^Y?)Dsb^dMF2jMp%NR?1r4rC~Shxuj zD?jl~#P<=Gc3TdBd>i9^-|=I>A2QC!{-x4<=zo&)?pN|VFkcE`z9jiuhc=Mkya;xd z^*0CeqH*1TZ)V;&B?kS?!1t$;e@>VMeYSYO{xV(_xG&hvbwTpj8Lr@Q;d^k&D^BNw zKF!&#=rdJ0lD3<9|6^nl@c4dcSK`mU7T~|W3An5yd$U4*2fuSDc5bp8^2={RUgG@e z%)r0lIX#KP`&gf!$NId~6ABfD{;~Xin&csmFCZTtk`ne*uU)KXcNOU4%=bz%E~R;H zJo8}4OZ>Uw3w_MrWu&N^k7H)mqhwusbsXoxF7vxL;^&bk zVIT9JuLAsirFjov-^Soi!FuC%>i-MlPfO01qQrd;Kz?R4^pyUV+Y9@6-a&uGZYf#! z{*ra?+T2f9urDC7+!w@-0FSH}CGUX8iR*>(|$-I7Q3*b1Hn0gxAciYEd;SPi&JJFVx!MdYQ!KZiZlF$U|2 zWZk$!F5>i`hTnsyPF3)9xHj7!{Iv}^uOXJ8-5S1|N<+=2ccVBY0K{+v;$+pJ=} zUG(wshkO*{Lm1^N-2;A#@25l`_YUAWZ}hA3X6Q4`5B+LNp0B2ZXX6>z;RoVQ*U&HX ze#!Ex0HWu+c^~Cj2jIQ={!8?aYzzEgZnRsB{DpYFYg(8JAnFl6>8BLX{Nq|eUh;&E zp1@D;#5$XW<6gW7^uLn;{lyM{Fh3l49z25Y{2u+v&vjPf#=R`ypT=`FRYlKM;NMOE z6uYHQ1Wyb8PDJvG@y~%L@O^>AheJETKPnvhNSu#kUFhZ>;0@UB&}QmhW^D3KiSNDeg*Se8Lu?IgC`%~lS*CU@htRf!3F5) zLjJsiAwQ0OuH3_V8n*#{Fc;+gh^J?r*^%`GvHvSI&{}`$7 z&ZEH1`v*f#K!5XoOb7lBYe^LRN5*Beav04Y#P8C}xc|!e^0)8d-+=s%#^|pqUt7c_lP?0;(sl>1H{c#q!dH<&Cbl_W9$CG)t>K1S()_H{g z7dz;)GZ^~3q1|q_1#W(y{Es5w*~{Opv81zXx{hj^flT&zlK7v@7y^l$dTep7UD9H-P(7 zd**2>oLbMapCRv53p}zu#@0r=Z;gH5UlY+U^WMyj{5{|U{vL2M?S4Okl)V6kBzZm%LCcvAM ze`P7ydFKM);^!xS$GOB=JolxR73=we>#2F4;GtOXjNLt%5#1 zX=jN`DbfR9M86XMxx)40cjFwgLsQt{#8TKn>^4@NKiBX5D@+G2elqF~_|1DQcdI|t z<28f#VpO93E89W;-?qR$5yV%OfqW{~i^QM%_(Gpq++WDLuxu3YApRao?7T2F_|5P3 zul@zN3-dRz+pEvPGo?QIB|NvDLEik{#a@+=>wZNT`N=E3x5{`DJli;r>h*wpHsgFw z8Wngo&uR8o{4MKu&}S8Y*BnOs{4fUcX;FH-YB?^F|NK4{Zhfc^dFM%Zh(EodNtwqk)V4 zPw;n_<~^7XYJ#UE&y&bLXn9)L=MKlKINSAj0{JsMCm?oPqY@--hayIP{)5UNG=6gl z`X%$?#*dIs%YBwIm-Tc$3;BAFATRShZ(rbTxxfCF@<&$#Pr3|wcj9F;0e|8FTkYX@#6a_@vC&q$Lx)J1IwxXl#VA6PockyT37Tl%@cMo@6$ZA z0s5qB4f%r*x67UceIlG;pC-g_a-H4GeX-G$Y|B3sk*8KWT46ejf8W>F?>m#xah$P`<`>==qT69F$vHPrkC~Zxqingc0vD2=eBA znyaIMoA+ie4*_o8vw5o)a0kA>b0>eLrm(~GA7D4>Z(aINhk3ArghEXJSX;nN9MJo zGNAU`0bLM(n$d3Srh zDhPZvzn@&1{jC~=c30+u{`SQ6&m?ub?R{aN#i{TQuX15uGmq!$yHS4ieAxMz@wcXDp?@A6Rht%=~8Gov<-MBo!&HFktwLpJ8`Fl^X zLk;?mS(mQQ`DlJu@ORcN>u`S|cB?ZDdYbp4&)*Gwwny6fZMs^}XPZ%Hc&sMA?w5H# z<`|wAD#g4);zMX9@LcD2`HImFX){3|^Zvo6$G~siM_E>luWr}8FEW1u_#YW{jJ}&7 zKe{dYHIVIo?+Cmfe^(wx+qNeA1K{a-UR3Z(tTTtR&Mfim)*9%O<~sOQcdVyRQSg|*2k};kjkcTlyN`uq zft&YT-dh3ws4L)CKhLzDxg0O^{^dVn!DHTQndAxl`mQwA-`8TEa9DLv>$55b^Rx@- z?Y{5_{s-S{OMgqWf<9w6Lm$ES@^?mKuPMh+kK`>kuYhNKAIM7_UULvUGm?PI{JI?g z{Cn0bB%XYi4)||8*DdW1Ed_rw@3&2N9POI-V|HepVR$X*DeHLuxsYGd0QyMYn>PsU z#%=>H>tM^lh;R8ez@BxjUfbElIzXST{7#7Mj~28C&phTu5^oo50Pb-O{Ng{oBB76Y z4`z!OkT>tabj<g{E&C# z_W;#0Z9UuS05^Y!GN0$%KAQ@D8Q=U%alR;lzf)Fep7s2w7F_LzO-VHGjH=f}W* zS`t6E7W`L?aa=bT_D37%EPdBW{Hx15*nB~>`G9uL+@68)jS%!@Gx!PEXe z^btF6<@=F{v%qED+}98KbY(s-JiTVX4*TnazaHD&ULEq&me`JCT{RK(crD;}4zp9f zH|uFH7NK3)7kA5m`I6;Z_~%u|u`b*X+H*hHmOTA@(e5GsenEvV>zTX}cp<~i_td{? zect-n`u}&=V4qaap^w=A!fWuHsB4S+%?57XGw#6rb1L&s;fYXrjc&KOQEyqM%%ySP zK2K;QnavZ0IlhwXycPS(j~-TCVnc#>{IUi{qqD&(Iq9}v4`<990J=RscP z%ci{GIn)!l_`|tp;2FpLo$%ygJ#9;@t^DVWz+?Vy<(rGJn-|aJ%6i@Z4DczeKa2iF zc^>azDC{rm;NWzSH}3=8H4%Dx)q_4_=b=fEZ~G^B>eJ3G3P8_iGa&Coyl_F_+h+n7 z`#k;{{N}x)qgB|^>we|Nh(BdmC;a6N_E{UbpRG*(iW$MPhWUf!1ILnpoA-D=ECqcs z4}(53F0Ir=*8Ud6c`g1hl>4&GMjhiz+QE}{kZ~Nyde@Yyw*CE%=S$7|F@0H=KE}GV z*eA~-=wsf0?ROCPPQJ&Kco?Y^*80z2J}mmwVjZBUVYhyRAU}uu4v9<8xz3I->Snvt zf~)z>dny+zN7DE?);%Q71Yd@o@05g}WaoJO#P4IA7y>+qc=Nm9+20>_6MTDb;7^SF zs@MSFi<}@|ggk5d0ypoMY{YYSH5unUDBtA{^vuZLO%x;EndhT+zkr@D#2;RQ|9kSg zZDQxA{@_WY7f|)cIHp|)dGo%=O6mZTwwvd8$SXIqo`5mXe>wA8*>8?)4*4w`AusI) z6###J##Pb(8PBCW>IMD9K80BSd_EGmK^f@jhi=(*=el`{^UIxhOE<{3OT>KdM7$I0vF7hhirztgs~dIoM%Tb&-V52` zJaF^A#mBeNueX(;e`dD(wlnyv8u#ti$^xG1Jl`epaKGw^9!EFc3oqlcQaQZFck(-g z5}zlm13sk%>?1rSYXPst_n0o!CtM{Sn#Ym%8%o~0SSf7r^E>Sl2eQ!5bJNdNx@$en z9nr6?{N0M^`TZZ@Su_;-NWGMo)>V{P3t*s zk9H%tU#Ly|NJ`)jd9F$95I7Dz(`ti9g*)q+SQ7FZhC&~S+g(B-e~kH(^mjuc=(%wg zcw}7|tq!K>e%0#)o)Gds>CSc?A+OxhdQz#tsO5usUalnZz2(7g-n+U?6;!pn`MZNn ztRFsUhIU(WzLYo(9+#JBS51HGDbon&K#CgQr>e$tN-hm;$7SqG@N8p!KzPQbhd$|z z^Fk?<0$R_+MQB&zz^$c__bLrs*1=n9fV6yO`oGw-Rb|-Iyce=B&sV>l10Lnx)^m{W z9WL^{gYZX|1HTKuD)v%9`QHNa0_hm2nzN`+# z^QnbKfA>3qzw@d3lSbmi|oCon$Cd|$x**EeC{aV7tXlxWwy?|a@5;4aih#=*+Pkrhi0e{ap*kAVJyBDFqkIutx(yl5GSoSyXD-78UJd*DXCI0xcK4VXQv0L9v zkWXOV>qLESv<2SeG2&hs;uQu$AM<^^KY#b}Istmtqx@{8u=bPcJWnO_D>wNU|o^}NY}aUU`T@+u9so)z3TJ!Ab!=CA#E*mE0y$1UTV zhH=K*5%#nvPwRe=uelp|W8#N?1->K+xafH^KlHE6dV;LiBYc6I_cLz(7Ch^?&Wio* zoMynYf+!A9*|$f`yH|pU*QhCEcbWPuRl4j&HEFh`ar%-aa;Qz=ma|t zI{-Y6{2MnzpUyX72U*v?c?$VN);VQf=d1}litAShdGac^*W+&9!+4PEmp|7p8TTc8 zzhG~?U%2W4{*8R!C;FT|0DklQe2e1HKYB3qmwm+0Q@|(v0bJJQrnkX!is$`ApRxYH zJ)dCy9%8>DnGd8q0=tR*+n$E}mF~a?Qof*C__Q5Tj{%;Scn!X%Fz>OA=RP!!@5{sw zf29As=I=gae_fXA!jAgrm*{^p5&G1tj&=*MT{oVict-yeo)HzHXO_{B7oKy2!4uE# zSBpNyivTz8PrN!0Jj*VDXDr)Y7zo__K1Sq4@R;{4#%~0kxh;5t$TRgb@R;8j{^1+o zAb3k@d*e1M;56y}(aGk>{7(j&-5qEi~-T-!)VXgPx12 zPs%mmDO>`$tf%>T{@j7*&v#S4=2i5|yr;5mZSa`q?N>a8Kg_5BJE-BXo~i1M#jL{oKyauYGuAgkMZ-Ky<91pPiqFZzI2;;OOA?VKK46(Y}r9 zc8ZJ$_6=y)DH#9u;eYW9>=xh`5*88ADJ)(Iwu@URGv za98h!zJ5MV`u{1kVo;2dFIPS+Lj7%6z1Z$i!G5s{_wtD?=j7KsI50Lc+OJ(eOt6Pf ztaF_8f2#jgu53(bKy+}BUsP%S|i|9=1XVc}7o>N&Y; zb-c9Hr!YAw2G@6g^#83)IV$r%^VXGa{m)Fc=BoSur{O!McjM>xPeZ5;uLr4u`tSF> z`!D{M?Ur+D+BG<$e8c8-0|G;Xh0#FY>G<9Xbh?kan%48Z^xs_UgMar=4)wu5(1!Rx zL;t0^()ka}t$)$_Ke)y}vb_JhcZ2!yiL5*QUfoaqgXDJcOBN!^m2DIp9vR)eNn~Uv zPp>K};ze{(0jrxDKhMa(j($;*ox%dUyE@s-_zHeP5g6GeBG%(S{NevY+wUoo$G7Cl zCwHKa%e?1tde?_!k@7!uBzZ&;3w-)tzi&IaeiEaT1=J7zv;OPA=-`0Z;N+85#cj2m z=;+<>ee5n*rCxAcSRe$wgJUAQLO%{AN(%U5))PjD=>Qh(?ZTOlww zI@1G$234ZE#DuEUA+V$RTPxR4Exh~NPhNtO(Wp-UOC7{(-ZxBR;UAhs zbPbCL@>9;+C0yl9It5d~C@{36Ur0b$C!Lse>*f~~932y>)2i6+esN{(g1SVhSoQY_ zl6-3Td~)wB=e3z`)>`7)i(eG01qfE9i~Jw`&9z)v)sN`lPQd{&vMl_6C(NSUzu8~B&u|P=eUb;;Y<4QiW=UtF zetI4Wd9rk;d_-{9_v83!8zG&CzfU{rXv@2BAZY^Ulb_cAIjs3eNL{J2NH zmwB%s!-i^2jnO}6YNqwx|EXHLkZ<%1;URepuF=wlKU+o_UB1Z4iEofRmf10vc7 z`@PR4J(@IBDWbr$K<(`BuTPty`Kbn zvdb{&fBo>^y&Ea@?`BcupBwt`qRIz3-;Mr1==l#3a;HvCkZgZu zqqnV$6SgL5UuEqi^zJD+q0Ev`*h*CVcUy^aPQhxP1q23L+X;^bR!K(f-kkjEMK$&% zS2^c*36gB_EKh&87<^b;Eu$B5SJx_mkr67N>Jk|1q2fgY6|r;`(>2ap?c0OZXXrjba z7gQ1<#?le?f8RSeBQ=$PZ1^X=;XkSeB6yKZ+JJ`|4pX&zC81{_30C* zkDf+W(fX?KWU{BKxdU821SOm zQt~cBzE1{?h^4nGYwLk z=cw*}0oovX_pbKrR$8EXYi;4nVGI9pFVO~hU;B}wPqLc*KP00c3u+(Xl$1#=LJWcL zx454$wLh%qf1b!6{ySAku$i{y{hFvW-~-e4!+(zCi#h+ltA5zT$p11EgSmWSD{&|C z(G7#jYP|Kd`+GrFc?c`A?8Qgo;D`HfljJ!0F_Tg$jt*g;D)UM4S-;GxU{#_~IeTY*HJ9J#av#^)itqn0B$bm7$v@3mi7)>! z)2u~2c`E8`rJ_22bB(hZqht+zFzfr2p6S+9+0aAfhk8VG29Q(mULSLlYiH8{FDp+EltKdo4>DV%17$?`PBb# zj`>MiX!m)SAN_4{@M`$JV;|D%zwfDX`S;bM_zEh@7WLTtqRpn4ZV(OcL8#NiFT{ zuC#M!mV3ACkXBDm&rYxUG3xG_-KB~`t}-B(Q;BeyScIY?h=Uay2N5VohQyH_f`xI6 z>`<{%Ejf z^jfJ+);zy&blOUqw0AGl-MBkF9FK09-=Hw^U5iu`UXUwYKvtn!Kh*_BYt(Qhmb=K- zw!H&`cW-uw;4yE6nhy_V7UcJ%8(&61DcLn?c>{9SN0mX-a~$L z_7bfX)E7tEelzRmt#B}pwt{JGV*yi>g8hysRf!iVm4>=jUB6xYh$oz8%{Fbfks#hZjFt+Y^)_?7K~+S zi|`wlc&e;Syk=${R;#Z=ygh|=;}S4!)Z2^pR$qVqEK00G;8Se2rXUfW50M*geGnd0 z%UXEQW_lJL(tO{zSCvQ%#r2ECgc^gNK{R->wcCudwgMSXO?ehUpvGL{hkr2|vnM`; zZ5U40apG%ESuUy67~r6%9S!E%=su1_k5W=5*6w_72-VRT`fMc#wc!W2r;hWo z)4&ZTP4eohkb)NC9Gmfg7buJ1&e|cen7Ir zVEeM!=OTwhqz6SKjaBxzZ;bjmMWnxW*4x@}@bi_*=LXyeH^mJf5 zR}Cc)#l*Qzv;kTMK7AFZseAhRRmicI_~@L?D@JL!)t& zmD+JSXo}1SoyBfz0@tq2w1qGh(QC_br-z7&?sBNy#1bvp5n_6xU^n6h5qW$7wwl*M z^4VBihh#?FW>WfjMoesbiryi%!D;9qR$x#aHd>2PH<^_PMGjoQuD z?nZ0pnm%mPqg4hucPO(cn%T}cd;%d1V+A?r3-DfhU+)-w-mb|Oh+kh(D@H7!dV3mZ z9ONxKC3bXaKyR8#C^Hw3@M5p)r*V}SeEw9{bFi92y^hg&W5{*~>s3={@{UOQKxUffn#O zAC22Lx`<-X!jkKx62oi|!=$yFkyo)Sp{8d=q?m?g0lbkJW0CAwiOPjBK+tge7W-VQ zUkdhlrG^zR_w5=n8%GF?#Gp8}Zq<2L^O|VYF;;6**3U|6b|vn*olw@l(P&xsqruV_$UXRl~;;3--ZP^L`CTk>ChxIAN2b-)HD>$cfw@1}^Zf~5+QixM`V z6hmVIc+^}`v>n-=y zZIPsb0Y}1aaJR-{t&+hh;YVLZ^P>;BL;`E^O@Ip-aHQ@_A&z*3D{fxC^msZQigyym z9NX`-mUN^w_jWE9T7R2Hnac;n>34%sO3F}MPkA5H_{hbr(_&4$9JI@+pLNtkTn`OL zM@mWRlw5?Y_?ToR)nQt$1>3ONtpt;=xW#hbxN5g$eQ^ifft*;gUBp@QJKyyVRPQq# zmx_J5^RlL?NXcz+A&mE5@Dt4dR@1UC}?dRKl&`KsDy_|Yl2nV+D30QQFo=aE!>;1u!bl@ zxONm(-bQdky4^!^8f2Y0M6mvT>&5OJbBUHC0EghF%3p(k;!*b{x#qW)5YW*X9LZ-; zdnW$>oc3J6O-BX$(-~V>5V5`Oy~)1fMe@^uZXI%hWVC2(PDi-Xk9)%#$Q*Do86M%s zsVYMqKoCQ*+q&Eq#%60bR)AXMBQReg`37(pBXvO&e<9#K(-Y|v`ZH*|wne@?X{-DR z;?mq4>^|B0&g9^3CckK>jP?*1TWxJ&k)9b$5BlvJ6W~*BsUvtZ#lLh$xi7lLj!Lx< zYpnfz;uDbKDFN}-2NHgm8IKZGZn8Y44!d$$V@qAFQs(0vt5X z$;vbu=J!1VP}Bky;t@PwUhLWf6gY+C-(&bD0^!CR7+m6ot(ew&p#yO-034jksq(HXZi5IOJl_Zu!=o_ zYA}8twoQ@EVfXg5eE{s^qyHJz8*A`VRH+KK5Pz_@zkPXA{q|gU+TOlwQX71w+q-c% zod~5EmCU(Y3e?;_DNi0Ey_~bXT$oOW&m)Mes2|}Gjdfe&>hZL!LbiO3iJV6Y4%sQf z(cK;0*vq9^apF`6`BK>+)TmwY%k_#2TE0|>6XkyBha(j(*RAm7x)rWgsNfdrxhvb1 zCe0O|1uLk~heIp+`1Yd;6}XS5N4JI*syGte?G05%Jg&2Zs7G){Py{3`kjY3UL3Q2+5+2HwG?BKA=8f<5*eJtn zfOMp$(UD2I$^&8mcA5tYE*~f!P>Om2&lJ_El}=oyvZocF*}XIDAf#cab9NFQO$R2U z5=0Bh-FblJP*xjjxKT-lGEgCny4+x0T{yB-V8&pkw6fgkw+Ba-w0lDpk)I@PMW&6_ zgCl4wMAGh$+EYm^y3&}mkzvei5WRsQbtl`IZIkqH^+9Jd@KeL=cc(|uY;XGZm2Ou> z5y2P${k!B3z_gS`^z_YLl^dQkiD}KJwOyqom`_`Fo@vA+Z?fttU^Ff5t!ZSlO@Z{O zN);GvwrOOunWA>HDGIO~(R}ks${M1mmE3;^xNWQCe_*PatG^7Vm;~Vtgok6 zkGexS6BGg3&~kRF!0U7LBDX`0p$$yFe5 zK=}n?Qc}Y)@zeHUzGI zyM}~|M%z47556U_b;z7|Wz04iaur%x@AOAL;;s^T<D+tH6sVFC^t48TSlsseV1b~`4@f-@)^BeUMH)nHn_>6LX#4yz>?oB4G+I&8k5 zWmOf8HJ$lUw*$3WuLl+8mP$n4>GYo(;X18LjEO71!3E6lZ8O zNL?3XNV*57ovBg_MY6YS9X|+jw zA5p_^txlax>!kHmbweNEXWjM~Vxe6I6I2;p2lLI8H4A8g>L6Y_YWKz|XOtNv2TF@L zT|nFJ?mPdsp%;+sRXZ2XwJvv5sDhY3jj^B7L&)LHT*(quk+Bp;NN3RGrc_@=6>;jh zidc;W&Kdz1#nZl^O57WR-7Fic6u38rdcJI^QlRQkap#0yb?eGtIm2eLnaA7XfKuJ#2(~pMW zeCQO5dRAmQO%slj2h6mw?8m@5LY_!i#3srTuq1U;Dw~JNtph#~NU3WiBScNt|OpOh>zJ7n673*I+xZePB66RHXMyS*JIcHh2y0O;<8);7Eww9kl6n%-@7 zZ+E&!Qz-Hk6Q7H6PgHVp55SID9piW57K~DzQBKXIbv4G8lWEOp0sx^U?zL!=mt8Jdz<3LJ zYLK(kJgWeV$&O7^enzW~4Qz=H7ZZCVhaK4=%yjxjhO&rm>(N#?jaiwO(`SS?K8}uj#RzFc%gOeaYpWp` zb(T_S(*U>ENsP1WXc-5L>JIts@&u!-Aj?}aE<8!e>e9q1{QvxQ$c z>4mAJVRYAD=}ix>j`6@|1?q(`sj<>vp@v&+TIF~+o~PQ!lXic1bTCEqDUHdN!VV3J zgyd*T>3zC;tM}}cZ;v;kzlF6iZ@<#^GLaBD`a%{u#E55AeSa%r+cLdm7|f2!w!90c zfqs{U9~+;dA3JSfUsN?WJA2n?tZ6p%%y=}Amj&VQ914Xj&MNX2@e~?$A6{=C&9Y6_ z;6iq^hSR#a3-ffp2&?hGSR!2#Z%_vg2*E(1KEKfK?{xQ%JDzRp#nuG5vO2)KWZmHq zt`xV_L$NKS;X}at*5Jxub)|V_aBgJ{|E_MHmwy}hXJ!4|`L&JJbB%TU(h$hn%I4bU zx$`TV0+Dhn=T^^cu56rJ!}DMRRc>v0W##9>#Jj%LiCq>0H|ZsWGz9mdU0V?`lqr;8a*rd4dciexgnJ%cw7 z72TrKV9EQR?w71KE?(6M@4iW;mNcc4y1i&?SSqVy3H6G$HTT`}3i~0a+p`~Bjm-F> zMrccbzHA0P(OPMgMT;jonQ5R&^w#J(iS)uYd@Z{a#l>?AR9rlXjR?C)uND<9SR^V{ zcWok7(@PR591h*%)Wg+gVn$$x=nK0Ah;-mbQ&=j6Zk!=H9Lo0U2bS3xu4Ed~u1KP> z%7gHe^21ShD(S(atgN0dY0;Kz+7BXK#LTF%tiy}o!YGg1)$!p$x5g3NFtkuqGCjO}OaAX`nz9HCYd+Xxnx{)!ofT%G-e#)^IXOYSdhS=sPwW`fLZ3nna{% zeC&~esS!E2Te2L@z$;2A(d>N&uZ|&3MJ=f)HbP&m$P}|hWSXQS!E?2e>mt=&et?F2 zqLOy1FI!k&*Nj_|Oau3;kT_E<(SnH{5L7_VsM+vN_$WOC-%nrrOvJz@-baXG4p_q& zk@*0Ax?4zJxCxID`7c!NJ8%)%>Fr)^;r~0k&)|tggvh5}v&kF~`Nd6TH!a8|LyrIDMJnDd-fD?dD|Hv6a-*@$KshP{}}g_54#A!ga1DN2SL+`3Fg!ydUj zM%@v6xlzI#A`U%blhyGCM8kemj&U=BtYja+WW4+O0-c$;J^I%qo*AX!C2p+Ch|zk^ zxZ{&MWyaBeMP&@=IH%&eAvAF5+wOqyFjaA^pSc|z_25jSPV2A-9i_zqI48pNoKRYK zaG6H+qLXuAugED3ieK*+h&yfF&gs%SL+y=D8}^Id)CJAg?H(@{i|tuS2P(`2^sL`6 z1u52TtV-?;>l9G8?Q?QO;1OgHk4-#Mw+U=;pS7fJWrryDh==6~-ZjT{E7Zk%gox@*bfj}J1M&vp{I`J1&7K1|5;hcZx|4+K6!yuR!5oc9B~gXUXQt7W zSWI+XVnwLcbD}sCU2}WdwtAS_%>>*Q+h>~twP$DzSzTR*H|(RPS`p0=i2O5o@5)1z zlN4e@n6nSj-qcZ7*-e^Mt=bRnh{;f*?I*dr14_JkGeU_pe;~F>5H8lkGo1R>IS-o( z<293`mNE;0P+yAx70g?qf?pZSEp2Qtvn9Xf%~!?gsy{7Ssm&F6Qf9w<&_3=@RYrjx z!qLAol~~v`nH<0WkRb&-gKDI3fB>=_QzA}mTT^4lPDXO4F)T)szH;Vc5Em~T`6q7X zb@#D7iLrsrHabNMMl45QYbiDPGGO2cZ>>@2Su9t?mM>nXlZxClK2R)@u*20*#fnQ?=)_;T;p62&2H^)g!!+5&%Xw`d2r#;SoV(hvg{9*wK6BeDn*HQ8+((QTGVM_BkU@bp6FIh^ zqQ%XI$1=lMbNjzWHK~{g>kL%CPc)VK1vPggNj_xdqvq;Smn-V)o+vejtsb9K# z`px)`v1k;PZA7j|^t=!gBBu*>Nb)eR7dsIi5Tpb^0r+a0~-ACAv*mIHQeY+QDzGi}X=4(HcO|P28 zs)yv&;D#k;^g=h$ND*TP!4)y9sCuXkZH zc`7I_o}!JyT2)1}s553vL=%RP!7w?1TEzMynJvtY^#xQDYjP%Hp@s@*gi6^Nq)8)E zxCTW;YDWkmRw}vvk-}A%ovT(`ZCo6R0u(H_6*StC>2-EREY@e@->wzE!rZ$@ucyj* z@IrFUT@P-YS!p}T0&-RTaO`=XFOpj7T*->-?CV7rnU2PYTU)Q$Yp#fv?r1zZYTpnW zF{1R;)1Jn_s*dFHMjG<=P`iJteP_}_P@R1bLB#TqqgXkh;Itcdan%rsz=<*zo2IhM z{He{`5QUUasu}r~s*A+H=J`+&M=*{|8yU~de1Sk0kf|KM$C%OKl_b1pXEQq?09E00 ztIRw3a!)DQAPV*dINZKkWexy|sGCfIX$k z+KxF?Yw-k5##jY^^VcrgJ?tdoaE%lx%BKK#I+)Y*@9)aH>{7hP<=IVjktxPl6nKAF0@5n3#$|dJ|Pbaxj zK)3e{G;Zcb=C-t&uM_c$agih0AQ?v?X{M2^0`kdyv$8~u=4=&@hHR@sr(|W}@H5B2 zQ*}}VYfUAYGdR_ZAzE9@B0p8Uv4M4y*A2)HA8^fz-B!nqFIlnaiZWDghEYojQF^X| zS-eoMqt;V2JDOEeeJce8Y9M5R57a=B2t5kPLLe2=bT?Wj+o~Iclo`ECq*{10lX}Zi zPxgFqy&z43?lE7}hUBe$JVNs9`>i^H4M?R>q@C=-7lkHq%5rX_FrEpN=i-^TQa0kQ zsAG6a#@dkMHXrnvi3}}jH!-0^yL7-PGTRaL0)lqM3M6kmcgNjfr>nv(Vxx_x@hBQz zV%j^|$k3uT*6N^+cjF9=-jx8>qH$iN$%`|WRt7+H(+|z|h9{#J5hi?B#yy9?TeSzvRV(4PtyE}ZISmcPHHkr1OUsscp)U45_h(y5Li47MEt6CZ@8kMwQ z6*FmxjqMLs@G+;f5sPfrqs-To>}88*B$Y~S7F`MWRuuDZi(JYnQVGb%eS08Ik?M7G z_)X%9t@4#BWeWDX(*7tW-Co%e7bg^%p;H7KH{@LcA+m%r1Bi9*8MHD9J(WWA1fVGh*8!Em*xEnfwde)5q2498rsm8}G8I7FKERf!D7H5)M`B_1=1_MYjAtdsRo`M1gfgiB;620upFByCOa z4I5&e7roe#wKD24o>}_<8n3!+K{L0qCYXcAm|l)F!TV=diptQbsw(7Js+E)Q`f>hL zEj=RogMN%Zuq=_;@7P0ge?8JA%1WbRA>BIc=C|woCn=jn!uI_BEEH%KG2z4zQb*+3 z4MB6TyAi&st{``zHi|E~JJCPbe=tDgVrnMfqL1p9|dWy=tOO=@%L=K8DFrSLp-^qzMKTLp+(@w8_9mPojyi3m1hyJ& zXl%?hpUT8oI2=QwGmRjtE5TyLwIzikFced|KcbBcgg|h^T@|}eilatrDwbGwI}Bn> zE9{A|yxR*|4ipp)d;jo-4tO~s2gt2%m$0ruU{iCbSEQTIE>sP%Az$cEeB(4 zro`N^R1Nub&1B(A%cm8ajtiD1A+J2`PJ(5{CiBv?8f_hkFzCr5JY{dTOuPz}C^xn0 zGFUJUaT}8@qJ<~8{jVjrdQeYoabspn>%OW6ai2!_RkaW7r2^eo)mB2?SMG?js;a!o z5#8H($fz|;UKP(-!-^xrHFwa~ocMM%!Mv84H@p zZ$LGwDjP7fhwK{FySelw<+0u+*!dJs9oF z(|Yd7ymRst9{w)F3V;LI@`gwi{SaTIZ1Gcg9OI0;gN@gA9c;8N?6>9FS17~pjTiH@ z0oEo@^w_%>N7{j1-jGT4WoYcpQpqK!&d|JZg~4f)3%WLZiSDkMYa)WKOL!NoQ+eK{ z0;>Z1A_YF- zW$P6bt~xausXzP+ttiH4(8X)W1g(po6ZmsSs4H?Re7IIh7u=VuhzVbT7uMdk?M}-_ zwDc6UAk3n*KQcz#M2sVF5>J7P13<`#_G(Pe6!U@8vO?l2YR4Qo5wT0zzi^hW)tRNz z(qmpRkHnc84TB!FZ{ob79u<8nJGY`vM!>f6uT&>+Vm2z!(Rxezz{sdCnklOut+7QJ z8yQX~j^r)IcSP8{NsuUAiq(Vfv ztVdteWvLQ{_31oYfS=X~Igs(fg z`oT^uFe_B5ZkbM2VVoR&(7T>_LVij%QJGhyT0pmQX4UPthV>whJ3|E6PY~EbTdoQb zR?fHCL5sm4uo+xN%!o1^AeK+ zzFqIeywY6Ux>K5N^3_3_1un$CQPdA$fqHu|ai+sifyuZ{S|m zR2_W-7e=;=-S)l?N)$@ID}*fJj-q|S3Ht2EXpc$8huQX6mcy)o!(48xsJCms((TAMPahSG$T(eXUMA|QOG5Sx1k07 z9%Pc`-jiDpVIZ??!OUs0inWej;p4aq*LG#*wS}ytmi8Rh5_!aW;2pv~u+HV{x`JTk zkUH%u3dL}?lgwQ23A&bTT9>Xf^H@ur2;u4ov!9W->j8?Of~A4}^hvc$zIqT=P=(vtt?Ji6lS7{}o%viIy8okEcvbE>z@fE#l0Y=nv-b+1+R z31^{@zgcmV34EwQrbZa{gaKy9=N7g2;7LqPC^9^-Ds~uFB+ZWKcXfuGSc;;%avO@v zc~z4KVW*_jXkW5vmh@q7Qw2d8D7l5SHsj-yMF}d|tGnA)u$e7^kwkW{w!n%d+n*7U zW^0MG;;`94Q~Q!UYHe|DX%L5FGR>3Fz2uHTHRqu?+gKV63GvqHqjXz{c+sx}ZTtOT z$5O%PQ$G$Q-gVjJRFUT3Wo5CMwss&dl(hq>M=F$M+^m2#XLxUjFfzhvQk`KScv!+g zOG9342bHY@rh+jBsBx59x_SfSsgN;M6K-O@BRy}IGQAoFilS4j1pGnpG^@)u6)AtM z#V!nyPQY?7DCNqMZ!k|)9lQ<)b0z;2D(N2OVQBI9n#of7Vq z=cQ${DWk5skQ?eINWwg4%0(0#rp#V>j8G~rV7YVqcB9c)0ZsNg$h9=?4R369&~Kx2 z*mmiiI!zeYkQFMHEy^XJ(uGxbq$TOPqNh$cOi3~Y1G6pM^ ztz0SF5gcK@RZQX@Nhl<#pbY+>lIyY|6u0{oPp~e=4yt&98mA$x_Lh3e+Xvi%exL4Q z*+zF%?k+W6AY)&{wvgK?w2*;~9$^;8!)e#5S$W&7w62UgFSd?G{a)uzb4469u^gr~ z&F2Z#88h5cZ@FG=N8QW#PwO%D#9zp9u! zI31D|>TMPy2hE{AYXG)2B0i=jHH?}$d-vEXIfpqxQfk(D7g?1937lM?-0sEeeXNsC z#W<7--|9EJP)X zFs-`?=*cfanwB{`jFU_%7I%IcVS1XdNmy@m2we3HLVBNvF|j@=QE^PHSc_)p&q`j{ zJe&;4X4;V^MU$l|({UF^Tx#Dhux`q+Nkz5Gvu|QxQnD%rk#X;>nnVaS<1W9j5xadq z>!R4eCy3*cx@T)UAQCM2W_Q@$6T|bt*7eC4A|TQr;iZ!;r0KKvjlOciQ*R*cSA0qG;Q&U-c6-(vSMoY_6ZQ`^Q~qq;yVRWn*-9 zx+oXLAs9J3(i_H}u|{|;WVHRr9%r7vkm!9-c0OT_rgI}&pu{55hUh-j8!>vrpV?Fk zIec{07lMH7N$?~hI&6NLMd4iCHZ9IQ(_3I7Ch;V69+KcvPF_7jL~ppz)E)G%V`iea zp~Ym#k!3XkyUGpXxM?Uw!BYTy#EILBp*fKWRDPSJoCpy*!r^cNB3VZj!ORv*2xd0h zhC4e;rDIzoCuNn0OgD4vG)i*u`~e{j%+)TL^@6Sg>7yXtQK3i?ts=sRsLXH|B5!)# zj=h8WZ%Q;NZl{nRx=t!Ilv|-BkwpiTlZ{uv5p~vylBJF|rl`$p*H>R_cSIk9uueF{_2A+#P zO6jCao+@TY*KI!(}bgMAgd>6}@^Lia{B!3{`g}Jw`O{FwnKwvqbOnu+xm}74=eMR*{}( z;07~u|6w2!WH!yA(%m{^Zv`ceFFWh_d<$xEa)XPQT=3`zYMtP*4dieL4LeR~iSic^ zK$O2^%s{c`%k4nx&JGl#ijWkUxMLHH%n#o6LcHp^5^Uwnkb^+pn4L+#Z>sG~HaM4? zQFbO34~%mfC$K$q>sTPRdrGsVPtmn9)p=rL1Q!M`>_qpZ7F4y+lcwi*M}p z*=tr(VcOu|gbFh{DQq#xf{ysaas_6T!2}#Ty!UxhW`%1LyF{KQ z%juOma@A{-DK%0`omAF!Scxbj6^kiq9u?+7c{Gs91ef2}?%b?}wtTJVZL5S4xmLpE z#jc6&3RgTleq9ns%V=CVHzgb*l3yedK0NC#&AQuVv{Ytm#Uh3)DrYj@hMF!Drk?Cz zWbU<9GGT|ViNYIx$T*V2$nAcSZlcSaZ6c>22y`&wl|@Ek9f-TNooyt{+bJ=)`kn`- z7K+`WfR$@UOHbCW)qqK8e}CMaOhyM=K|mVrAMUyolw!^*Z{doMrM^{##X`4hUr^8% z`eEMHI5MS=arUWjVAUg#l;lMdipi|i(-rG5>TqCFeEfo!; zmTP@UTPa6$K}G0@9?r(Oc%xw8;vE)qPSwT5oL6(EnDKHo-i+W^l7HgbmqiHFe!tew zQE$Ew=$sJUqowzfM=(gm1TGgvIYZ@k#d4O61NnU!xJ>i8e?yUJ5{BaLb@y|;LaoC- z5hbrf>E*wO`h=lOLh7Fd-x5{*)!nOx+CEkj6KC?w#1Jg^@pPP*MckG}WH$M6ovroS zna!16huwz(0;llr$9ZFX<9DZ$3Cbjw$RAdaxU#PHN^)qb*pF51`-7)p2i@Ux>pJ{@ z{IlCnA3=+Sy=t9#>?^_YrrNA=3C3wv-lO)x2~kN}^px!A;M{E0OVCizDdv&Zc-dkk zEVD|;s7zye0o7BKAdD+jxjI?13!0fZYj#MXaN!#s_xo@s@4?@Eyd@7y?}3J}`rsHS3z2tptJNRfKwLRs@sYi zhR$|ayFS7dcJ-c>5ROT6y92C3SekK8k`!!$04=1hyD>eq2Hd)n(k!_*IT{&$1e3dp zuqfTviC~IcEDK^&c58(;m>9TgUluh>&AJ+wwX2cyb=@AHUe;`!8Ez*2;(*RFMLV<` zb49hENaEbGfeI0|mW_i;sHP-LBj;GaYn&TMXwJ0nI~P6R+`!odC6xF~Yt>dJ@IfLU zoC{F_zNXdjrJi|K_Byljp-#`n6x7F;L{!wl*ED01w~ryuj4#9}1o#qB&OHFE0)@6p zoc9~Si^if7U27qB9Wvbxcs<06#Ph=t`y`7QkznY?Q#E>ERk*J4L^lr?AB^$V-t-XX z&#m6{P#*J9sviih?9eluPR7rJQbg1|Y)@L<$-z{oo&s}$5S?)}*9Vtq7hw{n7n0d4 zz7$nP;a)ecf`v*9pfKD)DIASa&c3lCuF26_(C*2M&W-VY^hp=`fAjXEIO%JV{3?V! z5Iy*{iOJ>pR;Znq6EiHJUws?emwqkH%uMz0+7GBc%_$e+(W~tg`k_5VPvQD5(6c+u zq0SfuA=@_|-ij0rI~=*_Q`s|%2xcI<18dhl?q`ez&-;-T*vNayTM_)7vZ1Jp@_mOQ z&3ODFpPDjWfJ_7P%R*KJELNQT6}$0CUl-EViDBmj^1)3akP>-Hy@|>4K~&oyUi$gN z9w#KE>Ci~$A~{=y!VbLU12$GLnYg30SALXCFkW2?kx6^&2C}rn^f)?#0kJdg$xA>* z7Sj^ZGCqU=qbvHg!}KPOw~^sRrj=8yOSBukR?2xzgvj(DLtkfoR;0X8*frhA7R`w7 zgtZ)1?KFcC=Rgp*M~4HZUE_8{)qefTFShBfq^AFKZ5m zpddRFx$oIC(HNNmw*@}P?00+mr1wRS7@MUgX74yMhNJPI-FJ_PBUeVmo0T-cTs=d( zaWqsu9pcsX>EJk5^hk`?q`8athtcPi9g`v-MRfwvPBc4KIc1`E`e-G~Vo7@f=|o?Q z@GV-W$`z~hk`y?_R9_b++Dgm873RGP)OIQIscad}sc3an%bhoV(Xd-Ic8rya1W6in z4j!4Qr#XPDLF-^Rfx0s~XT262=v-lJ7mF}*58p9rmI4-#^#`N*m&%5rSDbl~Cn*+& zrDe3r4KIB)O)XY#$b>2Jds2KOjo@7MjMU@hKq!1d`P?|sX+5ON!_rFo*1=yvBy}bI z?T~Zb{qF)J<_dX$6wy!0vB&enT>jp9rpaz(Gt~jvc#GhIZ8Jc6x0FccNt{A)Cl^V}iAb-tmO2j96-gh^ zlGpOE+dqQoTRmn0^`?<7aJf^NHIXSN(dZ6N?;U&%7D`-aF5wV?(~KOMN>64SbHI%1 zjCoeaoB{y?Hx}ngQz9{_e40)kh#!*ovJKrATmz=v8>8`^t?h!=;w~Jvx!~dRf>!b#9BNEJb%C=_MlFRaX4Zql zJ8*!OYx@A?923M}*hUIK#v1qxp{yQ@UUg$`0YVC=dNDTEC=xefvw<>@5Jm@u^4K!> zXqe!$Iw0(MD;6FM4rGX`H75k7;xRpPiBU4w0wbS=rp zaGKbT_9ery3}Hu;mYAq=pXClS-$4jhH_A47?$q1D&E?8CD%R+<6zRaZHcK~fzIaLD z3?|#&Z9YBJLem_HD4U)73?}0Id+O0c>C)KO$cdNq2gu1$otG@chNd-2+KgQw7%Zgk zP}abvTy8xz8X^^f-yEoY~}7SO_%%?V^hx$Vo*>@)X<0Rqzb zb1kHLX^oNhQnPDT)KZrH4J5hT?@fE7Ve6zl?nydJ%PfK##s*rjB(W?5F-izFL{u?- zsX8(97|S~&jcl%u0>W{(qs&Djx4F$FD%&>7Ws-IH7Qrg52n_+rhB>#|u&-Cuou=0$ zr${W?`PstpJ*nn=ue^mwN|=lHkuHohlg5R`T52-aY9wZw=Zl8t7N&_zVSlc}pZP&X z=wqI~&&b257E=n-ds4$IdM`KNdq~HKD_fvrbm6dhwHLTsLX`$ZWk!`I7p_1Fre8R& z)u?q_IAue`2T5xYX!%q`e|B#9u{Lpb6=?X@0u7(j9|VuH7$cIP(uQb>2XA{tMRu`{ zBC0=0X~05N2U^b@sj;Ut>ZI(-w`g=C`ic+zme!VtQB(9KJP|k`W-^XrK}AP3L4Lh9 zZt1w)W3neIKRB-JNIE#ZbJT6Z@dv^saqiKl@QM52; zF+>>w4N>xb*3meW-!LU~6s1_>!t=`Lx$5+d#1B{Sju!|i^R|e@XDdbV+$UoN@3hD^ zp9$E0KvD^Bg6jH=3YsY|R!Y));}vrj+9&Bt zVmh*kY-(-1@>PX8!3O7LHGL`xQkkA1R1T^t`ihQmuZc#syEk%HsNS-?UEsIZgnCrCgU@t%=sa?Zu#)P`&IIX?ao`;K2nCkXT8dc}D0h}xV6Gd(rPm^0lo>Gh& zHbN<)qCE>rkr5E3`xxPY~uoJK)gP zn^6Yy%wkuYOmUr0USC_h@>s@2=45D*rCCw|xu&MH`6rV&vrW8*2YI;i1#G0Iclgazml7#^f% zryw2e-9+|hT*jO#70g@Yb-62uDP&u2R@%EXM!BJ{CogSq(~Bxa&xT%eBg9^`q$)>S zs9~v`v5#~Gmc?eZ5BGxBj&!6-cL(NdyxDVFmsNS*J7#?VlFqxp;g;|L@pxkDYXYN- z^3HS*<7AXXryB0cY{SK-x(XwqmbmWKmzA>6(ebI_8OHzyO$J)df-H+F#+1NTD{|t0mI3TQQNsI0v zBcC6+rXI9hx1{3`b?T=@qcqT;iKeh5`XW` zV=tHoD>)iQ$uXmXLzt+LhB_2Lof(rv=w(B9n@qVdAesg6D7OomO<%3Fu0>z@f-2}h zEu_0@WrxG}Fehr7rI=l9KH z>iW)Ony4fB(0S;_pE+IchO{cfGLc(KW-XBY*ARA4LCRlT43^84lfhfXfk}AP48dVa zo>v~_audvSeYcj($nU9Z#@!f}lA7Iss@(=k(zsVOq$~3U*2zlSnMJB0wOSbmD;>@t zEjA~s)EYgp2dye7^pUH6IHnD+p;M;NcJ)J1oz(hJiK;^SRTYAe_2hRr+S2Vh5nO zy6ZaRiM1G9l1)~BL4-C!N`#v0x)a@pS%dO&z`5q7c}1GrB)!?KRkaI6t!IWGL!7DI zG1v7_wCH*YX|c}hF%{^jDzb@06v_+4=kv&i+UsSF95_|Vvrf4=XSiIdNksywv!c$_ z%<7t{xg$4A>P2ee;N;C6e5r?ej3b(lK+DX$>Le7&xXlofOrme14lUv2cY6x&L=DEu zm^>&qgc{~b42c#CX`yH$X6+prg~)R+C9TMt0LgVqb)eQI(R5;6Y&fXzD?TDxKN}#K zWu*yJ)mXkwEZq1QC~eT1Z|3T5XkR{y52yYM#xnskqZA9zM{I z96{$k-IQ2|R&RJP!n$7D!i0<_0!QG-@?|jdR@TRSmVQrt+$U;YIQiJqt++0ak6kX^ zkNdJZ-MpEo$$e#f{N+fi?e?$)=Zo9z*3t2}+dAour^ju1Q4H?`g74zX?e<>pq#=3q zC%smu-=0iV9GK#b!Zfb!&Xw7#79R|D%g2h%VP+6I!y@98QWxy!oHK06TQlftzUTA; z15Gxt%avu>p8s(}Guu#_X~^v$Z}JueO@3-dO1BRw>XT z@^DPyT%#2WX){lN@05w=hCE-DvGK_K$ z+LU1=f6%5ZK4ACINqiO@TY^U(uqjK*=Lc-c50dnSY_~>kzbmVpaQGE#M5*{CcQ-^+ z=;0%nbaLFEB8ToJ(k=o{Wb2KGfEpcFM|~>pr^)&pcGrlx?#lh+G&f8HJgdaFW)`FAVRtNVGRVsel*Bd2%?oeAyFnzUnPx$b;449nUt4vR>n%YG(e=F#g*a~;q)DgS(v>+xI;%=>t}B-;{IHmA)P1uJe*%+9sr>yL9>JG z570iUK=D)!dby-Zc~$~R^N>M{n&;yFfv+lKDK8yvMOQxK6W5R=p>c;er2UI}7m2eR zJcz6NXdO)t$9S=|rBfUp5BsCei%TI;f#Az^eizRowr_M1LZR^Hem7~2?C<9~XmQ$Y z+LYJJlf(8HiDBhwHe1HL8F!*GD=ondx;KPp;B(3a%l-Onclvbqq}S<++mx-@Xz1oI zOzsRjc&q8e71y;DT>DuGQ=-1vnW*?#IXrO&A8i+nR_~EMUD=0BB}UNH<#gnlP!+tO ze9>0B)9FqoK^EP@F-s_g3a2DHcFmMBHq%Ej|vm3OPmQwoMc4#uNFYkG*s6mIpVhk_nqmQ7f4 zHtEfU;dC;79%L-mox}E|)twwnRd&-Y$lk!``iXjb;ly0Uku%{Gqm?+N={~SXU z(^SnhS*g7yH9N>Gx!D0_osRMYi)Z*_E|11r z1LOq68`l6Kw^kp&Aaf)}29(+o28a!-3dzUDGBw|rrTNR;#QMH^MMuvl_<*y+O_e{K z$2*)oq)jte;7@CK265m{M?U1aR4B}0N_mv@ZYyyRSr;Jb?Qe00zUYHTvA3P^w6k&$ z-ayhi9ALm5zxh2=lX7A<*4>le({Z~uoyhyajr^0`khpd7ewv+G&b9sav~AGm*V<5n z{ivyCbo7*xj{TaCdq<$1EG2Bh4@1P?NeAuPs`zy}9^Hx64QQaz=uMs)4X3>uqhq_Q zeq46Yv?+2Q*Y~$HTthd1zVw>{YEV8I(5Tu54Gu{-j{ z3SLUbmQ;Jj2O|u-7(t8>_jWGJ6G;)my(WWfHsQCfM(W#&6^=H(ZjNl4E1xHU8He4yWLTL(v&E$(p?*AgyuQhQm=v*G^%!|v$bHw4CaD1Iw^n|qPl!e z@7WBiHBjQRwqWFRGKDjTlY!dKB~&hPRoF5`o3^a2ni(Uy1~`2XXp{drr&B%Pb5RO5HggWVnysi<9m3SvHcVa<{3poYM@RQu*<(O z=ni&8Ya(KT6CTWgk`X`@r*XuArIRgbY4;^7FoEUxS!y1NU?84Y{zPjmD@m{=#JY7E zl#1&sH{)e7yk=ndpwdWDR|af^lg7Glu_dzP_T^IPPPb@~nm>#zX4dz+d&f7hm>@LA zaaQy0L{u9`!u1R1U08bcszLhcrV3DJB|8O>om z$dDD6C#=_>vEI!Y>-8$vdp@(YZigU<&EMgH^KMC$EbcH$2h}>BbfGjlb_?OQy`fTE zJv^UK{OHT-)}b{zrVuu$ZtdEUj%RH{w$wD>kA6|9C{3%Cfz(Z<--M8|+nOLr1>TUs z)91n>T$ealzyObj(9=$PcaOU0^jql9Xd-_&% zKG%Z~hLS z1!vOhRwj%_=$J3^o`B-gezLmBxg-(>KZ}RkPnHMZ23#^oT$t1q1RF{34Q_F!2G_Vj zd4NwYb*=XC?XB&;ywoL^vsFWe<*#}~hTvcIP+Py%y47vJs20bZtHt}+inDsUs;*+A zC4848JJ4;-F5O*oBU{l)*U(Mmzqxu9q;y-hZ3S+7Ma4wdO%{BcTfi0p7X{~#_o`8d znJBoEQkpoA8+nu1ptosJfxVKEw-MdH1JT?Ns0$zunzoW~mGEqp5?WSyMq&D@H_AoD zbvYx;Bzj}Ll*wYp`h&)r-0Z<#NOew{WeZYejYsshgBWSGm(-hm>Kv3y@=bFhM&CVv zdU5qUAJ{{@5HQ4%rA}t|rb2Ht6gF*CW{r*fDPDd~MD{Jta6ZD#Kx-O8U`^p8dnPj) zH&D20&DFbXN~=E7tU1jeE=jon!igM5cDJI&8Ok8#C$XQwpqEgr6>8P;Dp@6Z?CQrR zeCyG}H$uKFPnXn=oy5AjHuLin0FWAoqwxSoq?MI$PUa8t;Lt-IttwUTBN(+QHuv84 zWhI*!@kH~@>L;Axbf{MQQnC&iNp*y`CA;m3QA*O8f$3+|rHxE(eS5mtl2q9q*@X5) zWm^(+;N#4-4WqL=@^I!Mixh5O`fg)5JLEu7=#yf~A7iteLKTvbxWwSIw2v)Jr`#+p ztW9pUkH81C|B9S*VJnpmxTdqgqkbs$;q<~yO~k$xEA}bOM{b9BYpv0QNz2|G)xk9L zVPj64r{vl*+i3%{JG0=5r_Je)G#RZY_CgLW5{%|5)hcvgV2K+ew$YAfJm&U;o1-k< z5)_P`!BIG_XOxBSDD%y+t8kn)|$FzHFrdNkUV#A7o!_jmJwH?Z( zh|+1!gIY^`vZ~Yltb6cu6c4H>NYqSiW1|%BTU>3FYC|rvP!%gpv2Z}6$%ejDTlc#A ziWaN&3NfanSK+eY%~su_otV-44l>ob_04>iy)d`_ofyUn8?AM;mdkXi>N*GpexWuJ zKSJSHXEQmNN1q;QG|p!dqK$!@xxSjjmTsrts7dN-FuEq>_zX$^ff$EK)Yj(F5gq zT35uK;AqtEb?!7*ux;p)Td}PQq)>XruGm$2*mS*)Di|wvP=Sy7H=IQlwV@$7);iby zskC9|TJpT{neghf!)EY&vwfplkP4mv=?G4-=AG*KokSw%8pVdU9DwYtV!2>0+4hW@dqS`4dZjC0@&fn5Kh~QH=)A&Km z=ae$7&Uvihc$ouS9Qo%nV!eIl)0Un+_9jFPvn`{%$I?OeacFSc{L9d!nve^2AGcvR zDC`{{9DuNFOIhilw%t#hEA&~{l-IW7M=%H+CxXn31~vEdrc@JeyE}!k3>1a@nQ)TS=UUzYgb`4e-hACiyCh2_E{*OK#KyO*mXYii z!7ZrJv-gj1vD?Ml!Vj}x4O3n4N>}ct&aUIh*IodRxS0@>s$OH#K@*i}-7Ja3WzL7s z&r=+9N|ZT`t;z#*hIhziv-AppgQ3RTbM;#e13TxlK1wqq35E(@oKun(ZdA%w&SYl@ zjHAW-&q(g4F~#_mkA$;U-xPkqPt{Cpo#)VCxV)!paK_l(8mQEKx%kDcIwTG%NdyB| zAJkl#jXQYev(fZjw&LSa25|-NMpYGt^?Mr+YMo zhoC(qpfm8mTp>Du_K-EkO3YNbfv^_;K{pV>%o&Ws2r@U)&x~$Cq3x?!;Us9s@Cub$ zAoyT9n90G(h-)ng{4y5-gRjjAG`n3DNF|qi8@=6V>YyDkF1#5=AyFpjt^blS2_lin)!mjc3Hq=c-9n|+V_vj@fcNLR;~Pj{Wcs!XYaD1ckB%pY zAXtzLT%$Gl*&-$gCn%z#?#j>|i2*S&d~y}45l0zmsM(V?VquiZO05$xg7er(`H$#xMTiu6Xj+-8)rLS^WW`D4$9 z|5!`u1qkC-&t&i3L5hZzYY&KdjJ~mam zY;1jX_*q{8=g%{!;2&$1j+WpJi8fXcvD$BUx{AAQT@nQE3~UcP9(e3%e3~2%IR6IC zL32go7mvoHqxKE)2QeXxR%5T^3Zm3RBO;=s^;_*bNSGy|PKx2|PKRC4yYz#3M|aqI z2cFVnQyQ(9A(~N$Y8ewc*-=r-x{3VO-RQKB+MS*a+szRZBqsAtsZ3D(2&?(%X+lD% z+h71Qo{871=-(GRQgnKbfw&>UmCDED12KquYMDGGrM`o)rV%QvER8B66!MU2QV(@| zAlG%>9b8)#5+Ox`0pufETJ0@W97$u;@Kkw@GA*VA1kj(JoTQNwa_Z;F$?VN?a>6!E zk&|g7NlxbbtmLHkKFFzGpPVAG5O9-~Y%l%Ol2fc{Y~Lg~_3M$76t@viIdT#ii*%*- z-sQr0_fFq2%^U~8P0h;zxvwiKHnw}vM$!%B9Jis~ zCU=udW)3C~HYT^#bK|aZn}pW8JMK#0ui(_B_GhLYJK(BZGcaN&Y0P@u+uVd&a`4AJ zvML;m#<$wz{Vh4r3uXm7X>?W@yUc)3Vkad85<;HqLxg#*)j90KrKsN>BB$x)j+`)_ zLrK6?j8`Tc|8)k9XuK5O*R`<|+0z%-NXRL;3*%_PK_S5dD==W1s6oOLs5QO%74g(( zJUNLag(pYr74Q^e0&~HW9Ylzyet;*nC5*G^$fcOlM?Z;Z10~xO320U%iV0{+@FePp z%J4bSF|*u|x2@76#|3a^A>rt_jf52KDVz}ZK@RE|7I)iy_!jj+C04x33O3lHMM+Tc zc_x4BIQh%5VWkl5d9AKga}Yz!v7x6834xC5A z`hH0))#aB&?u5W*Vo6vySkRKhcN(AMvoA@k?V=^Yx$FeO_>$170eH*Ha@GRJ@%^n}K7(wbm8qUPNbCAx93*W3(5&-hy0+KlLw(7*WKZELq0* zld4>bc>bM9`R;nA5$|AgW&{G-3FeuP7rxb~l zEzeOBq!e)t-Wnrzu8eLX{WY;AWc#1@(Wn@uxr3@iCCYaNleR{XSvtXfKp!^M4I)}8 zW=Mq$MVUSbTdW)*^{P92G%^-#U@qp$3hGe$HAZ?R9L1RD+;J4^q#Q>vE?gH!vPrTy z5>5-b!rZ{BcJyb z_q#@;wE~S+NXCu{Lk9C4*hx&5iCRsJmV&i5W>wuS`0XbyEkuISn0mhSEORib{26FO zT|EnFr08sFnDBRqXc}{1K>BJuR45asWiuyf`t?|7HMW3M3^^M=yaF;u*zaC%daDNk z-PBN_zt>tPHmASDewz#HF8GN0f1pv5-4gVMd~A#B?c71H-*2f|drCnNv?d&( zP+6T-53v%)GLBCo)pwE1r;k=HwqbA_j`l?_KI!@n3d)%AK_u~o(ESMRiS6MHj1^5H zCow>)>yqDw_Hh|_@u}B zwJ&Gjs;5j*gZxZM8$^-?CrbNA2v%C0=Mgj1UTaOrIOc=A3TT_h*X3SyZ7j&#f*XApeuQ#UDkQrx+F_GWzu=+G83i8QiygOL^7-`R+gnA# z>gb}xR@_gwUNwjHTGia{1V5e!y?WQldc$y|_rt)2YVD}q8zZ+V?kavCs1wrXu!`+2 zon^4Kt1`wmCR3A^L~kz`O0j9gKuG&WZ>|d!HrHuJ72kw+aiVtx?ViH#g~j&U^ejh1 zy($N&0W=Kt$=vb}P11h5;F(^E{GM(stP4*+i|QiApccPjJR4gW?x+hZQp3n8BK1`6 zOp$3`V`ZXX ztIf+qoijmP~&yYt+`y3 z&3P$Awib#=jl0qNU7fq<1P?d~`orjM5P|KTm41+ul=c%!JyHc~Eob9gZwQ5Q4_e{z zjpG5-N4QEDzbt5ia$imRDn2RqF{_SjoL46=RG-gxZ|&frJ>+N4gjjerSZw5;sPd{Y z4mI*_B)+{=c#$vy+nT7&u#rVGRwbmU+G}^I4vd9~HU2T9bgH@y9pW{N;>>7NxtxAR z!Qy^lcaTP-Ke~mi{CjxaXCLWDY<}p4b?DYu9@$cc$n#e1Ip1tq0XCt69nX7Q%c5k@VoSS zIj?HwA?IO5RT=X2BYAfRUFua`!r0wdGRJiIf`B*v9G?*mo~GBhf?4ZNOjPS;>v^$) z+IJO{ITUf9*J0(Ab$eQo3mYC`9t?b6Q$uuy?gokMTZ{ZTGu~wE3!ps% z{B1bZk-uG}$Fd6NjeQ-|s=`IUOw0`?S;q?H{+i+0P?au|FI$49S~lWv8*PHGiL0gr zuORg=elt>pUC+3U_>`Ukb}59hR`e8bRm4TqipMW@@4&NkaD=yi<+M_s3h|@SR!=W} z)qyzdS0B4yOmEJM4O4ECRU9zf5U`f>Fp!J0Q1FEP{F+t5J@S9%R@qD3zO%*K&{EcQ zOCC{zbTZj2cQ4ZHmk=&o&+RbkQ0eJL=SV3=&8KYIG0(GHVK*V#8;hr_f*o`=ou~I1 zoJxeHr%8%$Aal3nT8hTTT5Lat*zVFa9n~#(n*Jnt&=uq^H`xP@2+DjQ*PsX6a%Up) zt-$~(%{wo4rxWcnO{}4g${eN)T=sPqma^1HqO`GU0}ND@IbQdthM zMi2?blmylpOLS0u2r2&&PRI4 z8@%SzuG_U(jQ0LXTc!7^yCac+ZAT@fl56 zFgC%ew-8C6+V38mNnapmSP+q`RZXSHoKj~|oag{PP+M{+CcINNU6nw^WKz0&^E>Hu zql;1P;yGz;+FtyS>h6HWkEp^vxy$B|QfghOmBf;+Xv!Y!D`&)DXLE?0UC>uUG?Csh zL^1aXxI0`q2i*a@xb%7TfG<@_slUCg;-s=EU$lcdorAFpR7z0~dD1ES&-QvGI!#F) z>e9QMosbU4&*$*0i+8$@xod!L3OFKEiKJ@?HKug6I?h>y3>@x>PBX$YyXgT+5pkh?esG#6-xjKCm`#yx1^~BDswa(26>7K{*NhCYAQMmuJQ0GdjSP zFN35alw5@>D~29OO&QfA$q{W74^;(zC0x+h3S?$9sgQ?rw&tK=z(RKG(QqL*IYNO+~}Dp@j!6|RazBxHL|UgHYwH^`ecnSWnMe0D)4UlvoN zkh+c~VrALo*j$P-cEhPSu0V@5R23;W;GIEBo3ftAP=n#8EeTl91drT(1{SE+jx$18 zgIA0l#fA@KnPlW&ZN)5Fl^ZGM;g(C)-Y6K`N`vFG$4C9Hyz(SER4|+qZ9hi|d+^x(afUZDkpP_= zk#bcL1J87&5y^_#n$j#J?4-GSv9)yxEE|u$;4){eG@tlqsjFPzxnbyGpzm^8C!t`H zJwwN{nNW!ZTS7%HIkDoF7Z}e#Ez1_?;IQYrI_K@76Q<90Zr^2VQItMWlA#7n!s-oOW-F+kJfnY!9bfTK1On z63fCerW}_A9#M$S_7@de&5W55jnGpT*fVm>K|)1x`t57#r@(njNFsy0@%%nyfFsg9TIq1VnNKk^SM12&we3`4E?R#(=J-J zcS6#jsU#wK5=_q^W}r9~2ZB*8VQ9%59C;)0srr>Y*52ROkEI7MI^{cx@%(1c0NAn< zH=w7+r9<*H7EG*2iTSgo-uk&R zSZ-Q7=!!T$X-Q_mmM?5W=X7Xyv!2L&Z%yVxF5x$e0iK#oqg58RndeF~myYiMHVKYTZ@MX`XjE z+8DHN!vrC0kSIWQ2VQlR3ovKJJG+C1EGXqqRakW3mG7ioYj@)=9G8^swfxP7TC+$~dzEGo zYuVE*BP;YotqDaLIAay6n&gxPS)|wJYv{;8^BF-cHU!Y_dEwtdHnZ4k5E37y$OWy~#$`da>qfErKBFr-8sqGUWaCu$FbpIpT?Ug*0U}lj+Klpn-y#olw@Py%4cR1wpeOT z9OKI7k8s$?;UDE?a_nc%W745lZl_3}4WYvYAY|Ki5|X;#tBYPyQ>+>XUHRJk-75Yo z<@x}LBqof^-!CB_<7J{OA}c0~N!ko&D>Ftp^*nW}yEkebcKeX~pj&LXB)_;EsCAXp zOdTAALc(kJr-D(L(~W5oa=Jr<(K2f^?6oJ|1w={H)7)4LY(8bnmxU2NgX`f4rh)eG zMpwPP4YxXh!lDAV+_(DnKiL#6F-6kwn@;E@pK8jdp4p?^6zF5m*wnd`A_2ilQYZg-w($i=SkPBr(wnNOStp8!YS z*@dL)ZF!jZ{vRr(owJ0FSCy5|SoX$^051D9iFXk=E z%u*Kl{=oBEb+t!W!05!rikQbCi!9Y@8;e+)jLR0CpQz@p9p##$u!qH5_OCXyLjIwvC2O4u0W-EnJTz>>CWRo+9s5L?1pLAdCohjyM~_9Afv~ zLd!*wvp5vmGx2Su%ZEpKRF_Xo3geAg5EjLz_}ok+tuP58ql%D#`=p5rlggZd(bh;q z43&k-pddtb`gSu?;hH`*>$H11a`oCrJUBGdyiV7DX8uNMNY8(vcMXq>3xo`iNKW6m z&|zehrN|j9p(dNg@pg1jsx^h}LR1Zk?UidHk!^1gx7k3)nqJUS4Y2iLFQ15gs&n?v zPTxX?o+GB77=MwYtIG4xP+=7LaGk2@sFB`fjLGxK`pV3eq>e48+0^Qoh;mT5Bx*&E zT-5Y=D5I=@VaeVQ(b_4LflA$kyL^!gs+V`Szqy6pL{ zR;bDys0BbCAq)oHs-&||wjf?2La&nA)sOss2uX`*NiXNPxJ~ zZ*(^TJxub*75Gz|6Rkjz%cU|6G^hc?HD=NDz#yr7R-2Lp5hS9>*=&58Q$emxh3lS=6Xy0Jd@6>xg*k$IVoW3 zmqn;6yW8qB-LLYu*IM_6)Ah}2u?IX^lk z;M4*|+fUL(1c{{N1+6`VPNWfvJLhsY9> zSe#$snCO^IFF`kuh2DSE^+g5GU_~2}FHA)o*1P!7Y_oI|3_kCxVgo^EZ^8R^R8_o3 z(I0vP8`S_%65T&}0X@+s6U&M+sEN;;b9y6|0MooqYc?HWujni+I-R0E!<@vF_FsUl zG;Aa%>1!Vb@&YJoEGx3-TFc)XOmo^DMQvQo8ys7mMG9zAPT}B^LU9{9hlIVM%+yPz zC^MHgGZkoft(7u=1wS?GG(%&O&<2`%wUa@i%kr>=M(^p`#2IK?^`7a#LvWr1C&9d3DEu*2 zOWupK?VJT<4sG&Du&Pc1(QW6`184dCoOx-+Wt=PykOD1b&BQ+ojG$2P&d(~3n>Wk| zmCf>(XgoHr;*M=L}(h85PZ z08z21OubB(Go>cNx}d2r(?wYu*y!Ljr)FB>s%lx*yci+cj55N&?w^s5g_d2Ukc4YS z-`0X~8eG1D7U_O5mng=-n1}H#w=xO>Ve9&BEzudws3*7#MS3-*1zZJhceojgc+)p<;oz_HtYY4of;m&m&{*G7YC@)!R3@s*YE z*T$0i^C|rM>7`FCef833@IU+KGk`4N_fHv^{PbD;m6D$Z?5pvweE;my*QhF=QNMq+ zDlcWF%%{~iX+i2fq)JQS&nbx1{~Z2GKc6%IO6$_L{P{e7l0SMJpIQ0>ewY90-v$1u zrGG({kbm^pb+7WpLYC0;XV8}ZBDFq`Z=W_}lTtE{&#Le8n|#xMz7}Oav-Cy${S@XQ zGy1G*OV8(P)Xz)!{&lLvXHZ7wW&cR2&*B#ui?kyD+PQqy(kK7>1OuO7;1djdf`Lyk z@CgP!!N4aN_yhx=VBixBe1d^bFz^WmKEc2z82AJOpJ3qr#eir5W%SoR{M6FJ`1w)% z-Tsz4Us!^+un(C0`6}oHFPZPl`2G97|387k@N1W*26p@}^`IX8+|qYb`8_KC4ZYsi zE970IpF+o9T>1e6yYn}H0^|9@(w{cJAFA)4U;0V&`-%Gf zp{2LY?``$_*DU>#`Tc+Xi5H>L{Or=Nn%}4Y_Mdo2G^oF6ejlsyU%m9(=J%J>@6a>< z-+KP<{zY_;Kd7YsJbYF3rAtfiW#E7C4{P}Q8TiRtTJ!ut2LAv2Q{Cc+8TiTX7H#v= z(nlHiPabOc($9qB`N1F6@P{(+cc0hedHvH``)tQ=e}CAN|0?u#MfIrvc_aaU^k;Ry z+48p&@JIjKc=^ZvdkvS-alemIzyFR`P%kYL{B=DgHP8)${|7QC^XCbI|D;ut}AmwzPzf1CRKLp1*P2>xZFhxZ9inA&-j;NMU4eU0F4n%C9e>08eVWN{S5y{6oU?2En&z zoKFz^uM<8m5`34}o#4MkczA)}-$UhJOTb?x_**ouHwgaygok(kl0RR5r(In9_dSAt z8_nx|g8wcm{{g`tCHRK~|8u;r3H})2d5O-u$Ecl$2!4m~`7ptMAC-TE;ExmhQG#D0 z_%gx2LhucO_Xz$3!G{FDNbnKCw+a3sjsH5qf0^j&1%m$(qK89*-=g-92>$a_{x-qC zk=lQm;6Fs`@(RJfpW1(w;J-xeyhiXJr1Gy5{HJKXZxH--D*q3!Eg8w*`f0y8I68t@a|EIKH-Y57^QTY!D{?i2ikl=rT;2#nEX9)g=kuzm0zDDpz z-q!n3w%o7NcpfGA|4w}1F@pcA#1EGV{toT;4TAG|^aR2GAkoi7fF@pa#jc1wQf04$sLGW*(cAg;kX9<3h;9sF}ZWH|f zqJFOv{2$XkdV%1t5j`9d{Clb2BZB`8YUeh=eQ5log5PUCgJCZ z;IC5qw+a5uRQ_dxKT7y{h2X!R#{Vk8f128Pjo^Qe+IgMe*J)mF5c~~l=S_msvMjwt z@IOf9-zGRscj+C1uM&N}OYqNAh4%>ldx*~8C-^R{`v(O7J2c-93I5&G??(jx4W8J~ zdHP?W@jOKEe}%^LFu}h>{XRnQKTYjCO7L$W`gx4t*QuRlg8$33UK<2|l;-sW!T$=u zFB1G)X};S8|6ZEcb%N7ymR=zEi0JT;;H>94BKQ|+y>1iySsKsF1V5nmUm^GpQTbO1 z{yiM?5rG4=Z!Jndae3#(gPVKx$ z@J*tV_X+-IXuUol_!p@BhXnt(h#o#7I2B)7`X4?0{~a3tLj?cp)bGOtf05dMgy3yz z=TU;+rt*&wyiNF7Cb%A%X6hRR|1HAX69oTpYX2g^{}7FTo8Z4e@aqKsIcnzxf!Cxl$?i128|C_Wg-Xi$pG~c%g{(n$A?-2a+w7%~W{BKkF_Xz%rG~f3L-lg^WfZ*H2XFeqO zLp08h2>y9$XX)oX{eM9HK1A@B2@ekwe2?0Hgy4@7ojgkLzf0{rMsOI8-JfNG|6LmA z2EqRqrhkGzN_eOWn@N-D;HR|_>;6Fy|dz;`d5q@4K z_}3Dhyh89Fp>e)S@PF>G23;J%zew=c3H~G0{u>OZ`Myc;U#I!LMex5*?YvF!AEol| z5c~z2@4E#5A!_G6g8wMZ_kDtY9nI?lf{&^F4+;Kz2>%}u{0g+Uc%2d!GE0E zxlZt(p?SSP@ZY9!9uoW;nf?j>S(@){g8yx5=VgMo2>-7T{68Z4e3jt;h}wUR;J-=Z zf1Tj}C$;ki!T%xA)tdzW0V@9%!M~o$zfJJpBzk*?;0H9GcM1Lvsr~l|{xp?;pWxSN z9X}xWhgAMUfrI0HuSB135&VGq zeVgF_F~Q#<_#dVA-zE47(erx*|0SZ6_X++_X#5`#{QGF#KP32HrFHp;;J-=ZS^5P} z|36FpK1A@}PxSmS!M~2?`v}4R49)9Nfy?0 zJg*Y`Kcw=n5&YA%F0T{(4-kIdAowFR&Nm7EEVc6%!M}snM#9gd1iwb}dW_)TNc}Dod`$asgW$hK^L>Kg-%a?xNbnD+{cVCjOyj>!@D{c6 z0>S??f*%t6Cum+r1aA@EZWH|52tO|qe23=y3c=rJ`X~4g6CPe8_&&A&I>CQ8wf_de zzl+BICc(dn@b(tLAES2OCis6z<9UbRze4T5OYlz1^Pwo5&XZQb$poMed0fl5d6>5d>r)P z5xp7j+fC>T;reFum2lq{^rLY;ThTW|Z$p0q`_qm-2YLs3Bj(eIemnN93w>VnZuAGS z|2^pM;5_uA|AF(_hn~cI`q3}K{tTc$gnbx9e*k?5{ZH(}F#07}*9iJ^xPBDtFKaJ?GU|mh^veNC*l2mML(Ui2gIy!z0e!G89mci~(Opf7_yh`uEDVF-OX=0A-79G=Sv`un(k z6#X$=KZbq`)-{fPJg%QWUlZ${L~q0Ms(Nke{6CLzs?nFjKG&dcjeV#^pTu*oLq8JN z*Q5W4bv2-`j`=jA?}F=_(3i%2o6(=b^)2YvV1HWC_rY^-Lth`ys~!D3>}Lo1g_wUQ z`d8S8F7)3p|8DgEV*h*4(-^-O{dvr%4}ERiw;w%%`wpOw;l6|DSL5|Ng#IqZA4XpW zr?0d2dt|aeOt`42K^MQw-$W^ zj8ljHI_6)G{v)1O1NsJ-emyI^1^v{R!;DAo_Rs-^U^J z8F)Poqc4Z^HiEt`<}-?3i~Sr!{~7ZfN8c3ZZ36u}+;V{cJ{m3gfh(_u#&*=+zj%4gFq> z-;O>%_Mrp)D_q}+ei+8@LcbZk8+`!Jw+HJ-l)hPNjoZB(<#W2n|`mtEo1p5E`Pq;POB>Ib3S5^f)}asK`g-)EG5-eicHFlS{Y1>O34IehmuB>C>~jlx2KQ}6 zzZv7Vp*LXP+R+!m_#Noiqj#b&kLTNk{u6pP`s$cZ5Bk0szZd;D?0+A64_+_*=ucq$ z0rbPMpM&T(;G7Skcc2fW&x`R#&@;Gx6ul4oJcfP<`Z)UOc#ae3Z=+A5r?C%JZ%m#4 z<Ob=9HogmLQ8m&JV>&|`Rxjp$A2P3Y4wpJwzQvHvaTy?8yg zqTi2k+R!h-ezv2pfOFe{{szuxC;FK)kE2KNd?(P?!}FR% ze--0by*YLM|AX~bqc4bg)}Zf+eXd3S5aZXOKa2gZM_&u)p#gnGT;GV^k9jtsH()-^ z=wrCP1^o}Ks}=oT%%=_gLiBd@_i^72^e1uOPV~9)T)NP|!@hN+AB_9YFEeg)<~f_^sUIg0)W&iNSnOssbteOBCe z0{v(7N%TpKU-j11`QH`$Q;j}|dDfse;B{S#ekFPx`Zt(=J^ISn&j$3DaIPBBYq6hA z=(}UR&FHIO{1)`tv94D1CGp(b&@acj+R^JUpAPhSaD6BGUszWc`sSEVH~LL@z4V}O zkKT*^F2?Ue-xk;RqkoS151_Z;oD8DBh5Z>qKOW-@qi=!vjG(WKaYoVS#&aA)KLPtW zj{Z9CJAuABp8F*Ft2pOXZ%>_ne*Q@{`ra772K@qzQ;WVc_N@;6HS~J)y|6zG=-Xg_ z8qv?kb8JGt6!UCGUlG^0pdW(yx1yhb`?jI~iTSsqFOBg#&>OJ-o#?M){4Vr4FrRMp zCooPA`rR0(7yW9?rw{!Xy#Ms0Psi(S0R4T8KZyPk<~fAkh5HVpUxRT*&=14&8byB; zvSVsiEDQL_Y-Ut$Jtb{BMlsTaCUadJXyuIJdRvD`7r$=+EIf z)}wdg_1=KKGRAL2{|)=oguWm4zZv}o>{|=^8@O*P`Zbt;8~R(gz8!rHtg8dPi1~D) z?~nO(q4(jObfd3{@q5t!z;o9Y)_4 zue%ZSb#eVD`d)axW9V;UoN@Hcu&xR8JFu=v^j~nEtKOYD|BaYWHF^l+*P!2qdDf!; zfPJe&Um5489{qo~Zv*-pn13Vs*BHME{YcES8GR+pvjzPXjNgj>J)UnH`ffPS?da=a z{vGIRkjh@23^`Nhh`}U%5fZm6GAI?ud`T+U>`T^L7LG(MZ4@2mS zV*bPE7vp)2pvSQPqv+4!zGLWXV*GLRqp+V7=qF;Hlj!SUK2@Vr=YK=&TQ&Lyc)ir1 zFNOKkqQ^0xI`s2#-+J^LFirz{6XxHDejR!f`aamtX7r7)t`_w5F`riSB{2Us^lH4m z+Rs!#ru%E5yP3Udt zzv4Nzqo0O-=s-Ua^Y27o3hV7c4`cjp^bTC#gC4^?d(k&R??Z3IIQ{5LVV(o%Ll|ce z{Y;$OA@uDr|6%l}G0q5jC(g+z`Yf2w82W1H{~1P zp15xt`a!t99lZ|o=|Jzm{&%7e;rcH0<#ByC`nR~g2mMjZrx*Qq>_Z>=OsuOPeRte< z0KE>cmqGN|(TC8l!T7`IgLv*E=q-3Iqv&y5KZbr8t{+GL3HO~qKL_JXqThvH^}*En zKOV1(YV^Oc&o$^L^ZcWqf$Qtg-^ca!=#OGu4e0aWxi_M3jrBI6Z-(dIjJ_h~(}F&U z=iZ8bALiMH{xx{~Pn^Mqe27=|TSi&#M=GZ(QGpUc^54 zqtA&xfc`o5Z4mttj6a0FJ;oVE-yF|n1bqwiQS@1`KV#@Mc-@VoFNb}YK;HuMnM8jU z&#~&msq;S_>#9Z%;T+bWZ;AQWqW_8e)}b$q`_`kcihXWCzYz0jL|+E`(}eyfu5U(v z9{bjUp2NCY(Z9s_ZRp3~`gZg>?0*M(Gsfve{}cPxg?=*jryKn!+_wjPSFEcSeOAn; z54{(!uYUC9ao+*-IdT0UdLH{cgg%bvIE=n7&cg`$?0A0|Mc)PUA46XN>m5hm1>;Yk zFM{hQ(NDpBt3H}K|7T+U)#y2_w+8)T^jh>;F-{%&sp$3S2V$HC^yRS+jp!$1T}|j` z;hZ$1FNpiLpns0%+lqb%&O;me#yHRI==Wg!4)m=s|4#HxFiscx@|b@&dNW=xJ?PJ2 z{9g3=G5m5S>3D0pD{Wr{K1pP={KZ@Ru`;MU>fcuW4 zuZr;}&_Bn1PNGM#-l~tM&i~>VzZ!jBTwjB}8qQ%Y`gF{v4*frP-PNNXf_XNe-;Qw_ z(ND*@Z9-oI*EgfzfPHR3UmxpgMc)SJqz(N^yw9|wUx;;epcgQoPW1h8eHZ#^7{42R z61@j~ZS-FBm#{y5=$E7SqyHE496(DnvuSdTf^JzeT z2)z+~BfRdK(6`4to6*n1b8JDs5c|`Lz9jD3hCYh>wxfTGb#hSyyu`qdb}3w;#l zryG3-tg8q83cQYb(HFpa`_OAKpMLabaNhy+4RPKE(bvKCL+A~-ei;3FoX-*TV=&KA z^z$*#G4urXc^v%_%x41qWn4drzBu-+YHaHK_u)KLqwj!mYS0_dYthfcdh5`?!}ay( zM`GU^(09Z2jp%n^f11$u!+tiSkKp1G5-$qFVH*DW4Lb@ z`jWVBH+mDs=|OM7>!KHZdAv^h(BrtiAALvk0rZ{F2hr!k`58ig5Az&GZ^OA7K|c@k z97XR&A49(m`#+BUJmxupzBBqHdIIyQDomaK1F`?r=$oV0pznh5Ytf^4zIEs`u&#Ra z&oNE|`ePWU5q&f4a})aMSXVQ82=i}2zX{{FqPOGvHuR4%emnY6c-?iN599hy^jG=* zkG?3b??!(Ny$5|k>_abl6z8N5eHzx)kG?DRe*pah?9U+j6`0Qu`bFr&=xgJ-kD$+v z`HZ4Jg>lBvPsDu2(XYlnOrZ07vL?~@#QSsAr&H&D0qk2f`U>bZ=nG<9wdgBh{&nb& zV4n5pFQ7M|&xLc+h@QpuP3R$OQ3h6 zUx?m?J`MBfMz2NhL7$G^i+(%KNgsME&Uruj?&t&Pe_|g7(f7dhL+E>=52H6@-$u}n zz;hWz-wyX3Lthzv9Q}KYKY{)d`Xu_MO==pAWCQX7vA}x1j%k`?jLLg7eviem>^Yj{Yio2m0nXhn?u_ z;J#hxN8!HR=vj>4gT5BdLofO@SZ^QtUO1oq=zF6Nps#}Q2ho?nIT=FV6YCmAKOcPr zeIBfL6g`ab$Iy4h&t2o_^J9M|(6__&ljtkq{8W8Db^iCkK2)PGhjD7qPs9GyqA!i- zUWfh`#;-@;5c|-8{sPw3i2eumvkCngtg9J)e~jOPelqsC6@5wcHuO`_+tKgEd^*s3 z(L2!}!}~)Q`o7qoZuH%8-yZZfT;GeHK<`681mpChr|=vH(4WTs527E0K7`(beI7>t z68kWM{x$A9ihegE57|Ee#h&VLm1tVVwp`&NTKj9!a= z4ECW8eLp;xdh{PLp9b_tu&ze*YcYNk`pG>1=quv-7W6ytx^6|k1@mb`zY^fE+(ci)NBk0Fsy`$(i;QBH2D>42!`ZQcWfqoIr&m{UQ*oUgGrq2Jt=+)?d zV7)cyEtpR&`uBM5b?8@OKK1B#;ra&jAMtz}(J#ewX+qxy<2R#+G5;3y8Q8a0^fAn* z4gDMJPdoZi*oO}EHSzlDL~p`;y3psuy1LQt$2dLc|HC|c(N{q4L*Enc^Zn@USl0mh zL)gzj^j3^BguWln^Dz2OSnmk>iP*PM^i9#n(Er4I#?ez4e*%36^hxy3G5@OZsq=px z#;-;{0sC2lz8s!QE&66Sw{_@Mm`^==65}+W590bp^gQ;b3H>MBw;BCW%)bTwWnABi zekjgE8~Rsxy|kmxkMTRuH^6!BME?TgccGtz`E;WnhUMdb2Wng0mdIiufu%C&_Bj~$I%}|pFn>A<4>aRginO8{j5em0_&BMW2rG>(GD4JnPZ-$GK`i{}AIeqEEwpo6xVqIL+vTxNi&k zFF3cY=-c3ZtPTBI%%>fF70jmt{dYX~PV`@K&b!c;zJgiU>};$8!=8Z z`o@@l3;Gc_53T5bzCS^`l>n`46Bkir48N`fAwcA@mK=htZ$Fd`8gU!T(N;q7UF4j-k(m`HZ6V!o@>yL#r3u5*I+;E(3i(J_2|`DR|EQ`*oQ{+t#RKb z^pCL5&FJT1y)Eb|^j7p^a1Pth@5K1+=+~ilpzntDcA{r+p1aU<7^fTkBs||9^c}Eo zz38)GfBMkx$NB6>KMM04K))9IKZyPw?mL8j0@gK*{x`-ML9fDmM$wPO{*R$whH=Kx zFUEPBK>rK-Fp2&o=3n*g)cOAz_pL@>2G6AiePhh07X3V&t2*=*vCsAB$Kk#W=+9vO zjp!N7zX|dD(W@{{ANrZ-{pgQjoB{MNvHyeUTVtFd^gZx=htapezKx)_;QCSYZtU9_ z`UK`Pj{XqlGl4!k`XqWao>$d(Q|EtA%%>XtV9dV;{UW?BYSAym^>yg8VO{m;M`At= z=>On+Hllxt`!=Efh3D9eUXOLPpdX9zThaHx{M*n^#r5sz58}CWpf}?6*onSA#_2*o z6!+~$@8tPM{{y`jJ&yJEp`U@?kKTbkfPN9?If(uw);olL0iMe+dJ%mDJ%Z;xik`-E z97F#CeH{HY?ArwT*Vvy)^yBclsQP~D{GW*HtI-d|KGdMUj``Q3zm4_Qq0feK>e1^k zp9b{X@p@@QKN$PaguXYPOEdb_7{3MmV)Rz@eQ_Sz&|_FvJNj~%PY3!9*q=`Hld#WS z==Wj#ZuFCJeGmF}czyMvZ-o2yq2G`3`_ccw^#kbb*v~=q&(Md^Pr*Eg(NDv396?_L z^BF}y72}MdzlrO|(NDwm6X@IEyiKA%j$ZY{)cL;z<5#19j(x5{|1ZX=MIXZTb?B#K zp7rPtV}BaZ_s98cL~q1Ao6t|feVftG!u2iaMLfq=^e)V^4Sfro^LF$bFwYM3PjTN) z^zHC|)rEc+_Msd7OsuO1eLaldi@qezb02yN>*_~81mg^#r|~{8h+f1#459CZ`wpXj ziu;bBUxs}fMZW~+Y7Big?Athc0r#CiKLh(YiGC(})x^~Ke;@0rM&AnatU-SV=cg8Z zeT-9w{xHU=M-O4W4d~BdU5)5xU_YDCZ^HgGqtB1tg1!;v*@}J^&RZM$!dP!R`Ue=N z1ATj3--$jqdKdcJSXVdt7|v%8`Vp9CFM2iB+lT%Ko_jy~lNe_Jy#wP6qTh|@GKAiN z=R1u4Ebcpkz6{10MSl?cIfni&#u-Qd4SfQANvvxU{c7~8;?(*70OM4npNZ#EgMJoz zE&5iNe;xXqxNklB>9}tL`hD1+M)V0>--LcP=F^Pci}$e>^cC=YThZ^q`D{a<6R*>D z^yP6rJJ8R;d^*uT!T4S1=i>Tq^qnyO9`t80pI-Exa1Q&>SH=GKqc4Vi7(l-a>l#Eq z3;REW{yd)7F#495=Lq_X7=IM~I9xx5en0Lzj=l=UnLu9v*H5CKi2baZId%T$N3TYI z5$mc!e-Pu>q91{A>ew;Qdh}|%UK-Gs!Z?lSjd;FI=&R!TX7nR*-dfOa#y+>AAB6GS z(AUNI?dU&ZT^;CWwD1m!2b85ABXGv&>zJ<^rJ71c@CgI zh4BZ`Z^XV0p|6MWhtXHS{*0jaV4P9(cJwjyeXy=^^c1e2K%X7!nneEv<5c}Pb^dR_ z^Q}f-0`spyKLg{`qR+s5>d?Q&_4VkDxV`~>3G9C(`aC$FP3Sk{{4}HQhU;6x-C(tj&d7ec71LwKw zr>XPbisxI6z6j=1gT6B6Q;U8P_PGxIcZ^?;emKTyKwli|YDB*Xy$StF>}NCjDj2^7 z{eE2EiheQX--doA_N^WLTdb=C{Su7RiQa?Wg?=*T(~bTN*4u;Lg8B5KH=*~T4`Td& z^vlo(&?7jvgXo>u&mr_ayk3UUlbFv4`p-C5qv$VTf5yW z!#+2ozl3pG(2vDAY(@VK^K3)k4ZR)xF7yucIC>}gmAG#g`Xji$8+~@Hs|S5L_N^Cv z7xX^#^)Y@w`YU*j1L%uloI&)Ras3ebhM4Ct`dPT|2>KN`pQGrr;r(+A{R2G5ar7%O z&II}Z=0AzP7RImob?W@TfpMzQC(vuqN6~B17sfp6(67Sv_2|c7KO4}$!#o?&zra2> zp^sv{&FJ@H|69=CMsG#$!0W3G{Z5?QcJxiqJJ9dLK6Ik*gmrbHFN67Xqc4Z`_Mkt6 z>wD2}#r1vY@8Q1v=n?b*^q;VwgXpuM520^|eHcdH8v8JU{xZfHMPCN9}tf`U^M@-RN`R zzCGyk;GFcL4`aQ3=-gMF_0ZR-55gXdU{ej4__27MRoLoIqE_Mr}aIn1XX{W_et z2J}AMw-J3mjMIev7uMB`K8pFYpznqGx1wKy@!QaQaD6-ayf{A{=nvz$bfW)-`E;S5 zgy-9h-hlh|pnr{X(u;mQ_OlOtW2~zm{Wa{{0QwCWXApf$tZNAUHjFcjp2xWwLEjzo zA4T66~D>d+6z zy6VxJG5-eiHE?|+dIs}pLSGO2(~Q0b*42W(JMP%%=nWa-6qL z^vlq@(BH>?cB9XM`S+l|joyoXCg#(JK8)UvejVmBfW8T?A4ID6!u{Z{YI>J9Q^?F3G@#!&LsM)SXb2_Q|JFs^lJ2Gj9-I37kVxF#@L5C z^aHUE_2_5f+%}-^jQKaBZ;Sgjq0fTxo6%3kd|J?V#Q3e~cVQpe&~L))x*h#o+_wY$ z0IaJM{Wi>}3;hqQs~i0SoVOnIS8(56^hI!eANpx{z4xPEhU*8=Z^n8D(ci*6htQA4 z>u4B#HQaXueOBzlDEh(ZW9WyWkE0)h=Qx4BBi1#E{tupS)t^)6|0Il4jeZ5zRfGO3 z_NNy87W6vwam=S4{SrLK2K0L{&qnm8FrOy$!!b@X`Yjlz1^pv?CWf-R& z{Z@?Af&L`M=|tZM^Xx)j1LJq2e}U)HgMK((U%lv`;eE9aeJ5PskA5KTJAgifc@Cm4 zf^#^8{y3idF#2uiBj`1l&nS8f=VT21c8oKQ{t8}?6X;K4U6bgWV_j8$O`ZR{vCq}$ zpJ1Lf=rb|TTJ&G=-0RT$G5>n>jWB)#`V$zx5&Z>>(}ey2_MsVl4UFG{ejR!%`mQ)X zZRqo0{C4!Oux}mccVM2K=$GUAF7#FLyt>gh#Xj_)e}&gwFZwv<--o^!?%R)kE6&3J zdOPMbh<*jeA41;%eHeWj#u-6xLmx%|6whT0{Rzx}9K8n5Wdi*Uj6aF~G_J4ud+Pij zkA0{{p9}M;L4O?M*P^%a{G(rw@$1nOSXTr3YdAlR=!avUo6t|fJe$#<#(i7RPsco4 z(bvHEZRi){9JZtX57&2~---Emq920wcA?*d{p?2H7W3>u--72K{WFZyhyFk8LqGbT zn9l(Edsy!v`q#Mc5c-$c|6%kb`Uv{m_}`IH^l7+$41GVWYaIOu^a=Fqu--}ZU$LK6 z|4g0#8!?}1^jR^V8uVJcj%v}L#s1WxuZ;2Q(SN}F8_@4ZZ$$qe?%RZZ1I}SH`tEp+ zE$GkVxwoQ^VVpMf&#<5E=-t?#4)n#a-cI!YV*D=j_b{Js^gZyrdeGa^d(qPvrw@I8 z>}NmvbX-4xeivT9gXk|}oFVjftZNv3H}nzo*D?Mm`T)isL;ninjH54t{hvVJ68k@i zejCnDRn^>@MR^NC=+)@2Vt;DT56Ax3qCbURhkgs@UyptQ&Q$~Y^O#Q~`ZDNE=ucul zo6+aSeOu6rSXV3h9Jsy>eHim>M_&%FuMYG{>_aE|vKXfe{Zfq6jXo=S5Bfb=Z!h}N zn13Jo_Bbc~=-=S|Y5=_&eGvUPj6Z~aFXl6hK8*X0ps$1RN6~lW`A2^a*N>y`j^{Xm zz6JJq61^S!SrwW(|AXk&=&xa%8uTpITZ^8<^>yeIIJfoaPhp$}^ik|vBYF+i+l0Oa z=HHBd3GUm1zCPX`TG406IBn={=z6tiTAH50tFo3=T`XKt9n9mUUFrMQu`nfo_Bk0SbkD{N6*Vh>Oe3;KT`hr;R z1p0TF=Op_5m}k{2Q|JFfTwjg87kUl)%{Yg(==)%w>(FOGuSdTM?~@Ja8=^O&?~nay zLVpPBYDVwD{hXlkxu9hu(#K>qqazI0NY2xPB1*zZicAeNS9JjQ$ArVFY~;eH8sV?Ee`08`z(5 z^sBHx6X<*6+)koziT91FS*Onb85qAB{Z_2E2E7LRR*QZi#;HSJ0`sXye-!)BfW8Ix zp%FcS_l+j>|6pCs=vU(U7W5r3pH}o`v931sr|^8+(bE{G1N{T6s}p?;?*m=veb~2d z^e(Kc2mM&grx$$_tQS-G_y70b9r$+#{@sCpci`V0_;&~X-GP61;NKnicL)C6fq!@4 z-yQgO2mXKBfgk5v>i6a|zpZY*VD?8(|4&u(*+a+9^3O=~na@@~qZ$3P_NS#h^Pg#- zmHt2TEnQWS)$-fF8*CvZ;f2A!*br9U(VSJ$XR@MMm1Fi{a@0Ot zj@U=YVf#oqWFI9LCoAjkkPCKF&f7=JIr|tnYac6T?BnFLeY~8qPmq&#r<|})l;ieE za?Czij@qZl5&Kj*Y@a5F?9=7qua)(8$pt$l=j}7(oPDO8wa=0>_StgUK1WX3=gLXD zTTa;L$#MI9Ic8rVN9_ych<%Y9wl9`L_9b%hm&*Ek35XdHW_gXWuMm z?OWuGeXE?dZ}TYx{j8j^pOe$}^K#05K~CDka>9O5j@vKEG5cjXYQG{!>{sQm z{hA!IUzdwBE9)PT3wB=4+i%D@`%O7(za?kvx8=0`j-0aJm6P_UoUq@MF=* z+8@dh`y)AQe=LXWPvl~;vi>o-U>D@P{i&R@Ka;ce=W@pWLQdOX$|?IRIcbl}344Yd zx4)KS_BV3W{#K6I-^pS7dpTtPAQvYp>z|Mdc2UmTGv%E9qnx#Wk~8+ta@zhyPT9Z8 zNqbUG*uTkf`*%5J{~<^1Kjn!1mmIeLmP7VGa`A`C`paLOQ~Lei4#|1D`~e-M-~a9M zKcFoA{%@B*aH#bAzg_+ZZ>8V=?f=Lrdk#5im%l)y^z)xRryRHEl4JIN<)}Tk9I@w- z!}h##$evFweqULCja;z9a^9X_&e;pdS-boX#!BnA7n0NV!g9(k{|A+&zkk@Za>6cu zP*CabANFE$%r1YBU+M23_7ZZ$E`MPA%o$;ODLG^>Ef>G5tiMh!*yS(MDEG|8s%2~Vo4|YrIx2MTzdwDr!m%mW9w0^r@PS`8TaeE~>X0I$q?N#K6y{a6xSCd2b z>T>bh%KFP+OjtVqc2v&WYsfi!O*w0?C1>ol<+NS?fQ-`LzwCA8q}?bd?ElGedp$X3 zuP;aK4djTup&Yi$UtlqFM#wJz!_b*Cir-Y$-y|38@_#T`dj9q%a?URQ2V|w!zrC59 zu{W2~cKILDmey}?DJSh_Ibm-l$L+1C z$IB`E1UYGU$_e{KIc}dM$L#V4T$bMd?Nj84eX1O`Pm@FT>2mR_%KE$Hf}N7{_8D@{ zK2y%xXUQ4+Y&mV8Bd6?h<)qy$C+zd&xP87HvoDaN_Jwl9zDN$+7t10061n(gW&J&J z!A{G0`%*b)UnXbm%jJxHg`BpplvDOqa?@#?7QTseYYI3 z?~%jyy>iIDPcD95S^t1su(NXBzF*GS56D^jK{;bTB&Y3%<&^!1oU{k!g#D--w;z*Z z_TzHYenO7e)8(-Jq#UxJl8c{J);}Z{?3|pppO$m>Gji5`R?gVZ$!Ys}Ic2{fC+%T5 zVZSKH?U&@3{jwajUy&pBt8&mQK|c3#fgZ^$|OO*w17C1>on<+S~d zoU-4QllG{bu-}v8_WN?o{y>h}AIcH?BROn;EQjn*o8bPTOC~Df=rqX^+bZdxjjhzm{Y6H*(beR*u-;$zl6@Ib{DJ7so2=pO6c7QO?^l z<(&PaoV949Jc?KL-s#%@srB> zt7a+xJW;UAA2?b1`QM&J&e^ldS$j4)W6v(9?f=LryZix+rQiST@)wqre*UxPl;d{! zgQH77|JmjLV5ao*k3F{>vFDM)_Plb)o=+}*Tv>nlivUXNx5IMYF8_zlrQd(;1>~%~ zpq#N6lGFCWa>`yrPTI9{!d_I4+l$FDdvQ5xFCj|Nxvy{nwEcaxKLtDLZRm*e&xa?IXSj@o<45qobr zY?r_2r}X^oedXeNmGzfDIH&afXD8&mUH+hu()#WF<*a>xoUsp-)Am7f%05_5+T{-h zDxH7(5IJs_KS;3j`nM00qxRu)#4dk;N$K@(A1R0IqvYafW&Is;!A{C~`)E05A0ub& zW95up{=%rz`=5QhoU%`llXj<^uuqiZ_DOQgF8>FqrT0Jk6ggs_Du?az2kw@hzkRw~ ze7CaxF1cW*C`cgqR;JUMQkFURZ)$5`PcGOQId9)2=j@y1tbL1|v2T^r_HA;?zFkh*{c^&- zLyp^b$}#&cIcnc6N9=p#uzjx_vhS0NZ&lVmAQ$YcoVV|nbM^yr)_zdV*bm8R`(Zg{ zKO!gXK{;VRD#z`|1A2a>AY=$L+7>nEj0$wZD}k_IGmF{$38* zKgh+A%K9hdf?bsJ_Dnfv|0rkepX7}Fvz)enkyG}sa?+lZ6ZUU%-2PpT*?-7U`%gJy z|0Rd*zvYnqk6e7cvi_=B%Re6!?D7X!m45%XXOVOEta8?#P0rZmFN!Sv{A2$|PTA!T z;4c0CZ&%9+yZptzrQd(-@&`kge*UxnD@X0Q<%nJWV9nCcfA+j`$evFwzE)X(`2!zI z>$k&l-kx91*$c>7dqFv4FC?e!@)y{Ye*du-k&|}$gYQZ||JsYnaeFa2W|zNUwDj|z zUH%}9((k`^`HPxM>$jJZL-x{g@zu)u>*Ru6{z8z_^S8?%5LWv6-(FVE+RMoqdzzfK zmzPs^`3uEL&)=?>6ZVR7++InJ*(=LYdlfliuPTS_)#Q-9x?Fsvvi=6SU`OS=UH-uP z()#T+<*Z%)La)-_KkT*Tv|aupjMDSB*OilYqnxnIA4pg_|Mq%v%wAuP+8f9bdqX*F zZzPB8jpgFYmGw8t1v@6^?M>vIy{VkFHJ*e=%R__aD3bK`o{K`}a$g^|#0cJ1*z#9p#+8lbp48mNWJ)a@yWi zPTA!T;3}PeyH!ruyUTHV4>@Ma>_nfPTK8q!ahWf+lR_A`!G3bA1+7iBjm7sq#Ux3 zl8eKY^>@exJ1OVwqvf1^jGVQPl{5Bna@sy#PT42ONxM@{*eA+y`y@GLpDai1Q{;$! zsvNdYlSB6Da`A=A`n%+Uos#qR8FJ1(Q_k9F$r<}>Ic=XKr|fg(q}?qi?DOQfeZCyC zFOZ}5g>uBcNDkW<%OU#`x%hl#{XKHQPRn`wQaNW|CTH!-<&1rWoVKr&Q}$JI((aWL z_SJISzDADO*UC}*IyqurFNf_L2XXLznlbo|}mb3OPa>l+@PTRN1 zDf@OgY4^(s`wls7-zmrJyX2^Sw;ZwWk;C@Aa>%|LaDvIcvWqXY9A-wEd2pvfq`H_Nbh&-;?9^`*O_wK#tlU$`Sh`Ic$F{ zhwM+}Vy?3OF}YwDv$#MI4 zIcEPMN9{l5i2auww*Qtx_CIp*sml7RW-I^vR~%~pq#N6lGFCWa>`yrPTI9{!d_I4+l$FDdvQ5xFCj#vgwc0|tG%g8yq{DIr0fB$VSCui(wa@t;APT4ESNxNQ7*el9$dnGw$uPjIHRpf}h zsvNdglSB6Ea`B1E`Wxhe9hLL;8gkBFQ_k9J$r*cXIc={ar|fm*q}?bd?D7{&lz#uW z*OO!R`f}9XK#tfO%3-_w1xcmzZ*MFYAFr&xNiNtiId5+w=j=`8ti73>vCCfwRQmg$ zUH;(k((B(Ye_(9s_g}kNPS{(?aeHeyW^W@$?QP|VUH;u-|_c0$hE`^hrO(AZP3Y<+Oc}oU#v=lXknDun&>r_MvjjK1`0< zMwwNH^F_Nj8%K1~kUr_04hD(ml(3wBD*+h@o*`%F1& zpCxDPv*omXj-0a3m6LY2oUqT6{3XQcl@d$w|9cPS{t=ar+uMW?w5u?d#--eZ3sEZ;(UwjdJm! z%KH1{f}N4`_DyomzFE%Nx5yd$Ryl3oCa3J%<)qy&C+s`qxP7M_v+t6l_T6&CzDEw* z_sSvrKDqc{W&Hzk!OqHg`+hlRKOkrA2jz_Ykes$3mQ(g4a?&1@6ZWHW+PnW~?lXA#@N-jQ7S^tn+uyb{;J zPTHe#!hTPV+waRU`vW;@e<(-nkL0lZu^h5Lk&D^N`p4vgU6Awkr*h8zOwQV$%NhF% zIc=|<0{#uUN-^fw>TRCEXCx`9t<&gb@TpXyZe?l(UMLBQJlymlv za@PJy&e%W8Y5NyBW&bKC?MXRd|0c)n-{qM7ha9#4lq2?Ea@hV`4%z?6#rrDjubRF5 z^GLxC$$5JgIcLu*XYJYKj6J)Yw*Moi?D7{vm45!QtL20}ryRHEl4ExH3(rdb_iwxW zg;b@V|Ll3>usyFFvgebF_g2^D*gO_Pi6gea>0(sd3zZ-XD=&f?d9Z*JxxyA z%gZTy1vzQg%L#i$Ic~2c$Ly8msJ)6Du~(JD_G)s-UR^HUU0HvFT(F~Z-Y$QEW$Eu< z_L_3mUQ5o{Ys+bS9XVyMD<|znIboN-0JHS_zrCIuv)7lScKM4^OXuI-P!8K0$sv1V zxp-G){Y`Sgj>&m@6FFyZDrfD@Ca>m|8PTS=#051LhXP3Xgu=M_Cx5^27 zcR6nFA;;|U7Xy{v|Lncwh`qNQw)c@k_P%oQj>`JmU^!{G%L)4sIc^^+$Lz!8sC~E`v5%0$_K|YPK1we3SJvMl7wn{*w~v-{ z_AzqSE`M=O>GyxT{Dpp{zkk`s%PIQ=Icayw3HwAjZl5H_?33lFeTp2hPnE;=X>!Ou zT`t~US$~&Yuv2p0K10sgXUbXoEIDJJEvM~s~9+PTIY4!oFIL+t*cV0gB-GNl#91k*54-=?2MeZZ<2HN&2rYhMb6l_%4z#HIc482C+&VY zVc#Lg?K|a|eU}`y@0KI>J#yH-R}R_t$;DeL>mQH{c2>^Y_sco^0Xb_wC}-@4

    ; zoU$L0llGvTupgD<_G5C)eq4^)PskB_x*WEjltcDYa`EQM`iJC#os;wS({j#!M$X#L z${G7PIc+~Lr|cKxq&+Mr>=)&@{gNEBUzVfxD{{nsRSw&)$szl7xp-4${UdV0&dYiG z4LN7ODQE4suIa@PJ_&e&hbY5Pk#Wq&0n?QuC_&yeHx*K*AMMvmIw$`SiJIc$F~ zhwLBZVqazb6LP^W%6WUHoU?zFv-VGN#{OAO+rP*u`&T(>Ps$1VH#u(qF30RY^&n~C!|HvtO z4moL8%L#ip|buOxnPIoygk31vlo!F_JVT8 zUPw;c3(G0H`~~Bspa1PzIbkm<$L+=Bn7z0hwU>}1_L6egUP=zxOUuRUE9!_LUQXF7$Vt0iPS`8TaeE~>X0I$q?N#K6y{a6xSCd2b>T>bA z%K97Rf*qCf_8M}|UQ^E6YsndVZ8>ePBd6?j<)qyxC+z>paeF;EX0IQ#osICTHx;<+QzpoU*r+lXkP5u(y)q_SSOD z-bRkv+sYApJ2`BZzo@Mn zRh9L3$pt$l=j}7(oPDO8wa=0>_StgUK1WX3=gLXDTTa;L$#MI9Ic8rVN9_ych<%Y9 zwl9`L_9b%h%F6nCbca>35XdHW_gXWuMm?OWuGeXE?dZ}TYx{j8j^ zpOe$}^K#05K~CDka>9O5j@vKEG5cjXYQG{!>{sQm{hA!IUzdxQR@Ofv7wo*8x8IO+ z_M39neoM~SZ_8=>9XVycD<|zyIbpvi$L;s!nEinqwLg?2_D6Eq{#Xv#pUA~@W&LAv z!7j*o`%^h*e~G|#{jD6azmvoE_j1Vo zK`!=G);}Q^?4q2vXUaMIM>%W%BxmfO<+S~aoU(tFllG*Xuz!=|_V04c{zH!1f65X2 zFF9=gEr;xXkxrPTT*HQ}!Hk(yo>h_MCFu zo=c9||COWm+;YU8M-JQb${~9`xp;A9{WWsI4$FCaemQ3^AZP6b<&3?MoVFL1Q}!Zq z(yo;g_M&p!UQCYJi_1}a2{~dfDTnQ)kw} zr|sqCl)ZwSwCm-By`mhqSCV7)%5v0RMUL33%3*soIb^Rc7cZ==zd;0ZryRBSk|Xxs za@gKS4%z$4#q%obZ<7mlLeAU!$vJy}IcpyvXY2#zw0)4AvJaM%cDtOg50T^cp>oVV zOpe-z%MtqsIcy&(hwP)|Vs~Zz9df}=%6a=}IcFauXYFIC`cgqR;JUMQkFURZ)N(Wk6f_Ra^AjF&e@mAS^IK1 zV_zYs?JMP!eU+TFd*y_EwH&vvkz@9?a@4*~j@Z}BVfzL-WZx(k&#tV$PcGOQId9)2 z=j@y1tbL1|v2T^r_HA;?zFkh*{c^&-Lyp^b$}#&cIcnc6N9=p#uzjx_vhS0NXI0ie zAQ$YcoVV|nbM^yr)_zdV*bm8R`(Zg{KO!gXK{;VRD#z`|1A2 za>AY=$L+7>nEj0$wZD}k_IGmF{$38*Kgh+d%K9hdf?bsJ_Dnfv|0rkepX7}Fvz)en zkyG}sa?+lZ6ZUU%-2PpT*?-7U`%gJy|0Rd*zvYnqk6b*xvi_=S{r+!<<@ zv-WIq#-3eH+y9YM_8fB3u9g$_oO0ZrOODz9m816Da>SlT4%_p}A$vZ#cv@xsHFCiY z%Xxc#IcF~*XYB>$jJ=SYwilLD_9Al9u9Xw^qH^3`Ope)$%Tap?Ibts*a*Kq8ztZl4JJDa@1Z$j@YZp zVS6<>WUnq4PpPcGK`z))Id88a=j=7*ti6_;vDcQ<_BwLPURO@qjdH^NpB%T>lVkS! za@5{Hj@TQ@VS6JvWN$1NPp+)LNiNtiId5+w=j=`8ti73>u{W2~_7-x=-cnB5&2qxt zN{-uG%Q1T!Icje!N9^t7u)Vz;vUiY+Cso$pA{XqqoVRzBbM{Vh*4|mp*t^JSdsjJS z?;vVreUO~850;a5yPU8Ok>mEEa?CzVj@pOI5&H-^Y#%9y?4#sjXJ!2za-rgx|4h50 z^iTBWTY9sqsvqZD>i3eTm;Q+!{ZPR71bl12*9Uw>z!wF4cEG0wd|bds1bkq?dk4Hr zz}p49dB7V6ymr8=20Sg`B?De4;CTX`J>b6r&*SHSzYF+_fIkZO?SNkm__=_m2mDaL z_XK=vz}E*HJYN?D{@>3I_|$-p3;2kD4-9zkfOiRayMQ+jc*B6#4tUjorv7bHLvP{6)YY1^jlvuLk^Fz|#YMDBybnzBSCB;0*&_JK$9Vo)++u0WTErJOR%h@L#2WQvH9Pf56`b{6)YY z1^jlvuLk^Fz|#YMDBybnzBSCB;0*&_ zJK$9Vo)++u0WTErJOR%h@L#2WV*h`hf56`b{6)YY1^jlvuLk^Fz|#YMDBybnzBSCB;0*&_JK$9Vo)++u0WTErJOR%h@L#3> z_=|u)3i$1SUk&)VfTsuiP{8*Dd~3kh2Yf}q7X^HFz^4X$T);;Jd|<$P z2fRzb+XcLNz#9g1R@VKg~|39G&6lh4H3|1wyAc1lT z6mft8k#HGJp@DK4xdy5bZW^f2fr12D!vN!Wu=p}uUWimMTE%El0u`8$Uec*xpalX& z2m~-XC5==-DVMbW?^@P>=7r3 zn~3AYjl>Pa^~5paI^rmCHF1Qvinx-vf;dcEMjRp@!dLVH{ZBkd+)vy~+(X<=+(n!r zZYORdZY6FZP7`~?3F0Q=IB_F!193fZjJS?CN?c7GA+92>B(5M16PFQ(h==f%%|QPX z4-)qi_Y(IIcN2FJXNcR0+lX6#MQ(R;ws`w;tJw0aT#%lcnDwV4)j0qAaOr&FL4iXH*ptnhPa)$jkuM#g*Z*@ z5hsY7h~vbK#0|vt#4+MJ;wW)7afG;vxRSVnI80nd93mdVSNsG0PdrH6Puxq~L)=Z= zMVujSCvGEdC2k>36MMu7;wIuaaU*d9aXoR2xQ;kVTumGyt|G1^t{@H*ml21EhwzgR zf&M2RBS*94BrhZXm8FjuF=pM~SP6Bg9q2mBbaqVd6645b+RxA}G-R#Dm2B#J$8l z#NEVQ#2Mmt;x^(|;uhjGu}7RBZX%8oHxf4x*AvHx>xiSo)x;6vD&k7w3gR$v8F7es z2tOGW=zrot;(p>@;vV8|;x6J0aXWDvaVv2Pahlj8P7pT{$B7$>8;I+PW5jjDQQ~Ui z2yqp0C2<9Dn7E8ML_CC_5DWA_@gQ+OaW8QXaW`=nafY~^xQ)1#xP>@P>=7r3n~3AY zjl>Pa^~5paI^rmCHF1Qvinx-vf;dcEMjRp@!cWQt`k#1^xSzO}xQDo#xQjSL+)ms^ z+)CU+oF?{&6U0r#apFee2I6|+7;znOl(?EWLR>{$NnAl3CN3im5f9-f1_S+1JV@M6 z+)La;+)dm?oFQ%}ZX<3bZXr$+d&CLiCgM18BXI+9J#mbF}@RO5){wE$J?kDag?ji0b?jp_*w-dJ!w-UDyr-?n{1aT8_oVby=fw-PHMqEc6 zC9Wop5LXdb5?2t1iOYyX#6$QA(m?+c4-)qi_Y(IIcN2FJXNcR0+lX6IvD{%{Pn%E;w5H}IWi5rOM%+r=LYyY{h!ez3#Bt(A;s)Y+;uvuqag?~4I6_=S zTuEF(940O!4iOLGPhJH2pLmeCpSYK}hq#-#i#S8vPTWS^O58%6CiaLE#7)F;;zr^I z;(Fp3aUF4#xSBXZTt!?-TtOTrE+Y;R58+Rs1p1$Nkhq_?m$-+xo4AWOL)=c>M%+r= zLYyY{h!ez3#Bt(A;s)Y+;uvuqag?~4I6_=STuEF(940O!4iOLGPvQjnpLmeCpSYK} zhq#-#i#S8vPTWS^O58%6CiaLE#7)F;;zr^I;(Fp3aUF4#xSBXZTt!?-TtOTrE+Y;R z53QyCCmtm3C+;QgA?_ybBF+%E6Son!61Nbii9O;3aT9TzxRJPlxSlvhTt^%wt|pEU zR}ohdR}hDZ%ZNk7LoZYR6Au#i6ZaDL5O))I5od_oiQ9-NP!IJ#@gQ+OaW8QXaW`=n zafY~^xQ)1#xP>@P>=7r3n~3AYjl>Pa^~5paI^rmCHF1Qvinx-vf;dcEMjRp@dXf5{ zc#ycCxRJV@M6+)La;+)dm?oFQ%}ZX<3bZXr$+d&CLiCgM18BXI+9J#mb< zjyOtOO&lSvBCaH^APy6k5r>F}o~QmN9whE3?j`Oa?k4Uc&Jec~w-L7zw-Bd^J>mp$ z6LFllk+^}lo;XHaM;s-tCXNtS5mypd5QmA&h(p9fYpDN;2Z{TMdx?99yNSDqGsNx0 zZN#m_EyQVJk2pcxL>wn>ByJ$CCyo)<5l4xui6g{S#FfMq#9`tx;t=uBYU+RDLE?Vm zUg93&ZsIQD3~@Vg8*wXf3vrs*BTf)E5yy!ei5rOPiDSfd#8KjE;s|jSaV2pDahSM_ zI7B@39Q8l(AaOr&FL4iXH*ptnhPa)$jkuM#g*Z*@5hsY7h~vbK#0|vt#4+MJ;wW)7 zafG;vxRSVnI80nd93mciminJ~khq_?m$-+xo4AWOL)=c>M%+r=LYyY{h!ez3#Bt(A z;s)Y+;uvuqag?~4I6_=STuEF(940O!4iOJ^Q~whW6897L688{y6L%43h}((Vh+BzU zh||O#ae}ysI8NM1+(2AU93!qHjuKZBM~JJ4D~T(J!^CC8A>yIGQ2!GT6897L688{y z6L%43h}((Vh+BzUh||O#ae}ysI8NM1+(2AU93!qHjuKZBM~JJ4D~T(J!^CC8A>yHD zsQ-xviTjCriF=5S*94BrhZXm8FjuF=pM~SP6Bg9q2 zmBbaqVd6645b@B{)c?eT#Qns*#685_#9hQ0;&$RT;#T4o;xw^GoFHx@juST$HxSno z$B65Qqr}z35#lQ1O5zIQFmV}ih z#MQ(R;ws`w;tJw0aT#%lc&Ll|pLmeCpSYK}hq#-#i#S8vPTWS^O58%6CiaLE#7)F; z;zr^I;(Fp3aUF4#xSBXZTt!?-TtOTrE+Y;R53Qj7Cmtm3C+;QgA?_ybBF+%E6Son! z61Nbii9O;3aT9TzxRJPlxSlvhTt^%wt|pEUR}ohdR}hDZ%ZNk7L(8fEi3f@MiF=8A zh`Wiqh%?0P#BIc_#4W^WVvjgM+(aBFZX|9Xt|yKW*AYjFtBE7TRm7FV6~tlUGU5>N z&@$?O;z8nm;$Gq&;%?$D;tX*+aT{?faSL&p*dtC5Hxb8)8;Ki;>xpB;b;MEPYT^iS z6>%kT1#y_Tj5tI*w3Pavc#ycCxROsl2Z{TMdx?99yNSDq zGsNx0ZN#m_EyQVJk2pcxL>wn>ByJ$CCyo)<5l4xui6g{S#FfMq#9`tx;t=sr2lYSk zAaOr&FL4iXH*ptnhPa)$jkuM#g*Z*@5hsY7h~vbK#0|vt#4+MJ;wW)7afG;vxRSVn zI80nd93mddQ2!GT6897L688{y6L%43h}((Vh+BzUh||O#ae}ysI8NM1+(2AU93!qH zjuKZBM~JJ4D~T(J!^CC8A>yGYssD)wiTjCriF=5S* z94BrhZXm8FjuF=pM~SP6Bg9q2mBbaqVd6645b@9x)c?eT#Qns*#685_#9hQ0;&$RT z;#T4o;xw^GoFHx@juST$HxSno$B65Qqr}z35#lQ1O5zIQFmV}ihVM)v;(p>@ z;vV8|;x6J0aXWDvaVv2Pahlj8P7pT{$B7$>8;I+PW5jjDQQ~Ui2yqp0C2<9Dn7E8M zL_G8->VM)v;(p>@;vV8|;x6J0aXWDvaVv2Pahlj8P7pT{$B7$>8;I+PW5jjDQQ~Ui z2yqp0C2<9Dn7E8ML_D;F`k#1^xSzO}xQDo#xQjSL+)ms^+)CU+oF?{&6U0r#apFee z2I6|+7;znOl(?EWLR>{$NnAl3CN3im5f42^{ZBkd+)vy~+(X<=+(n!rZYORdZY6FZ zP7`~?3F0Q=IB_F!193fZjJS?CN?c7GA+92>B(5M16PFQ(h=(4f{wE$J?kDag?ji0b z?jp_*w-dJ!w-UDyr-?n{1aT8_oVby=fw-PHMqEc6C9Wop5LXdb5?2t1iOYyX#6#`W z|HOmD{lvY*J;dF_UBnsUcH%bTR^k@oG_gmVAZ{X#6E_k!5Z4pOi0g=>#MODO^{P+7 z=UOjxg1k-JI44}23LW@;d`@)E*=}=R`J{Oz__y4>Lw*mHk9=fZ0by6b*v)+x>F+${^S@ud`^WRD`yUCFl*IRs)izgud<8z%Ci~}% zt4&q^4%c(j=WgDdyKwX7%_~D0aA`gA`t?ixb=M#zMNa|+~&=x@h&x0Gb{NLO1>AJl}zk{(wZ~!T2g*f zw<;Lf_xEF@^FF>Dod?04mxw~kQYvc&p8Sqz6&WXjkl65;;&u+-G(6bh^M-plTyD6V z!$%A!J8U)F)#06nQyks`>^N4mM6FM+Q@x<+nT?{Un%eXD=FLxu_vQ9qiR5f_9zMfu zqPM#D1^Vtq=SdOSS5;=AOIzgL7dx4SBJ&_*vWfBeR!_^Pd_$*d#@0?-R^Ge`8PIh# zNb1nM$oI)dzFcZ55@$+2&l8ziFKxY?&HejbK}$rYJC_2J%W74>m!$gBQvrdVUnpuR zNd5coeDo{+;peLO`e4Nse#J0aD}MDyfr`%!R@{h+o3&L+*X3krr4kd>vKSgx^FQ%} zqN_6}!X%Weee`olaX#Jf8~uz-+1y$9+^k+*x^DL&$!&u-KPIxd!~Th@$+fkqy0sV9 z9x_gYMfpunqsiBG;zeSM#0ppT^!X-@q7Q__`l)_v-}>O@%>&yOuIF2IK^v|gR=BQR zhf>ZQMShmZ7s?Ssv$@o-3zpr8&qV@mjAwHf5BeJZ7~Y)w#}nZBDD$yen79S)@y~(7GDiPv>pUgHHcr) zKKYy~K`B+*R?%EG`Uom7Nkt#2ofchPe)Ghw;n|WwVhS!=4<7xD1r%g<(WLGwM3d-~ z%}q}tj!1_fJiPfEjU#`!%;Lz?e-x#brJ_$-A>02Q>_xI71q~jOWG}M9;eCed9eRdu zIK07d=Z7rs6^5_6)LDjW9G+tMw!>qA9k+{$s5Q~AN+F0$xYQ3K3m(k}k-uYDWTQ*8 zNn%S}%h{O!Hv(=ddS7k%@6n`q>64e7xL+!|Fcm{n=`i@k)z0cAuI(O_kcOJSe0Nc> z_~%QTH)D=K)>Vb9$sWpjw_eER`tB;q`bT8_lTIzkzhX9#qR0P7I-|3>sk6`-r~g_z zqf)KR=BoeT%I>tzX!;x2iyR?RC}DR=_9D+YEH^ygVTs|X4*&JA3MU-CWq6Xq7YvVg zxWe#T4%>kpzi^$g@oiNKopJNUerJqn&v(XIo$$KRS{x2KvEZNw#>dl609bRL&gA=~Q@I{w;mf_P5PcVGV;gP_O z-?^5*;Vo4PEkEuezvYt{C^nbA(SgVuT~fO4!%yYIBc@7MT z{abU>-wI88<1e*oXR6NG+$FzPD^k(FS<^0h6YNDAL<&tiU6Q@XcN`vJ_#=n=8#X!I z%`oR$tIY6Pm%2&k-N+RV-!+`?un*X=)-~-{M2X}wH0{F|`b~QvMvyh_O&N(WecB{3 ziN^@!H4|y$A_O&=wWp5_+-Q(5+6SelpqDCaPTej> zT8t<(B03LNHgB9zeoH(0ISJ3%B_R%o&E{FHV@gt6Z&iCYIi!;*o^mteV2GH~${*`y83vb3`-?Nm!U(bA5$v}#K`)Y1;J zv?+FR54-p!yEtKqZ1@JoSz1UnDwrfbDZL?v%1}S4R~vbO-$ImrAl4q}gi+J~iSfCm z^k=WpF}Ja{xpSh_*8$#={N<~+oc_C4ww!+3@-3t{ca9Yue)RGdWRQN@U$>C%%J{}- zl<{{gT1K_gwZ*bqUE83}NAVf*em4{bHM&Z|JI43uj>^5d2%9?%rRd~T?O5$>WKTXT z=D*rnsP@M11*_4JsM2UuNT(Jd6fjs_pnIl ztWCzSfT&wtd?b`RxD!^btxd(SjnOIb?u%iLC(_y6lk4$W_kN0=%-6}Knj!gD=OVQO zh@`vc_jOVhGM;vsNb1M2ui+zR0qia9AZku1ce+gNHUSo*MZk?QyTx$S_SRO2m7}-I z_pZ~jH%gT=G3#x~i#o9x8#R3dFWP!Ur@3xnp~g)YZlOeIbf5>>r1L!Xn%i|JZ-KSp zdM(Hw&@@&K{oyF}*i5e0KaVX}u>(8Ds~;g6yZK+*>{HpGI1R1$-x)0yhf~;9#Wn_{^FFlo7JcY&SUg%V{eM*7RLr6{TPYmGZp5hbu(?GSY=`|0 ztXgZVBk)dLVO#QLKSO&+`xFnK@I#_}spQk<&-;^f7rC2!zK42xa-cI-X~4 z-@S@_cCxM~m3T%lXlqzMQD#qL?GF=k0Pi#S6z_u;>>j^0>Weuu33>a==HEzO;X z@fPNj+0D=E<>trV=cFefyCESls5p_GoU}P%_8qCe2fqA1O;#j9hn7+0E;NW%)x!+x8ah;EUUq zHDS)Ty+5L5NNGt`*u7Fot}%-&leuLgx*9gf$xJqPjhxBp360_unLl)7u0@ESNP8UL zM?4FC*InYgOb51V@z~OUcNfiNa@2+?EG7SO3&4~<^F^s1s-)^MjoI|`;=^)HHy=)F z&llthPDUj#(~@k;6K9IMs7HWV{c}jgHgB5#DV|ISLmwG@jQ~=tWzZ!ohTwxCtjlmQ`mL8RSU9>@qcZZ(}QAD;HpX<+7>Em4beL{yuyj=B_mCtgOfM<{3N%Ls{nu?`&VP29}y96$^2wT#H7zKA17i^0-94CkG zzjI7UhwLWBcRQXHqj6d%N6qkZy_qH-(P;UzZHWCtnjsloj5}xeA4Tap499B=Vwh=e zoYUIn=Tbe4QQDW?A1^Lb7ziU z0Qth|xBRA`p?lWdffrjdV}Bz}+T7Gv5FCmO9W_w*S(s8)}=E!B5rT<6%r7e&%D4&S4N$%~bR( ztO^XXxo(;nc^(~gPCgzM76-A>t9u817tqXRTYp)=Rmw~g3>=qWFZB89Ite3^>rzuw zHDe3$Ra!Sl{wdb}Xk0wGP)L=iG6fmX{qf=3LX2cD`VZFyKbHwqF`J9Z9kprEwdHro zP=y4}OB*(CzUfsw*5Fki_R}#Xxvj3%Zos&$6W{lv+S#&yYF;q~?p2tPiknya1W#o^ z)xwfs1=?u9j;8{SH`#Y3S#l<201Rv2yRQg%>;JreF8aLv^K~p-M(iIO6|`Cu6AL|o zjtCm#Ca$xuMtkYtc22H?`)fAI7Pp|}{v5N-em~cKL1Dq!W*G)wp=m$oY;*8}qSc|*|EMmqmC$j}6HuUi(lo7po`5E(YX?ESQ=s1wY1D-h>d!JIp5&Jy?=e9oU> z*8kb%w7`pYa1+eZkGDpPcRI}?3#3|UCRv4A6;hDSCul*n$fnJF+~;Wo3xx6(lrk(* zPnF?Zyf~1(rdt;GUasn41rp8qtAx4p_2LFx#F8YNyL*l(7n8NfF8VA0s`>?1M+NWD zK^O2|d+w+xme@KSw_Q-+EsN!bU>DZsSQ)(j)I`jDx%V@AlW<1zpEHvG$c39RyelMT zid^z_x%P8}{256M>TGVMw5r4myL(^odRfWZE(ZQA%?OlR3BH8*aV^w^9=LCoj4df{ z7+W=|iZ?+4+rlV@ZwBtWTvm#*?7+h)Kb8F`p4md@K%q))BC~E?B9Qg3if~vD%L-3s zW1W>=cnDG$6s5SJUaa-BiSV>&-<3O~WWDkUj(X(E)jQ^|tizQ7IFIBChSKtTA=|v7 zHaJsH9x)CeDi&6pzr?hN0?}r{7|`+G#l0+TH&F^h+o5h-`wk!&)}|(^I%tU36!Bmj z+G_2{C$F}q_Im_{7UusEV}0%EBgcAiY||wvqb7KasZL(ES_f3Bpzb(5WNQ*%J}U3a zYwszx5`|)-p-5cNFt|dgN9-z|s$E`mY*8Q|sX*-%X~$y!vw;bdvbnvUmu4-2Z=h4q z)3~&gUJ|E<#Q6N+`~<5hb9D&m%2I6e&s9I=v#fmQp#!CKF%|0%DOK_t#WvEY3G`90 z1)i|O3V(T%Z-Jt9-)EeY44dbMP1x7i5n!MWjPl}f4Gkx2&&lYQsb9PqjkX>qu0?UG z*w-XD92wm(=gW|V!P#8N!)Qvp-ztN!Il6J;z^lc5+q_;JZY_5h7KZw-9IYwU!V|N( zSiKYl6+A3O5nJ7$Kkx?WtP&n(XTZMEl(a3SJv3>v*spo9Eoxb}VjQfDZZ0*a*Tp}s z>(qA#0$8Am5nHR{$0Rl0Z_{C|^}J5R+1zhVw$>USXs!G@($_2BGWRx0%g8ny{}(;bpDeSZlN}?n zs<&qS06QaeFh}*FR>{FpJuAY2s2&zY0#SVou8bVjr==nLzlrKI;a1jW!Kj|Qqqvtx zjOtODbqY~k?l*{8`KZ1%^cnRWZ^RhB;)4HoF?^LQb5g+=zDvL+^;)<|Q4kJ{{J{Qj zU(A{#jB^C%zy>J4YLelLebCakwyyo~uX4{JiQ(-9rqh?D$3~b=ulO%%_`=MWkCif8 z_*aWpPH?q&XK}z;D@(WfRVchu9C(x43wY6&U4;2m=U+>m`7G4HI#^S2#}k3zjokid zD7%u)-T#2LxS^`E5V-A5Y=f;n<&oM_S|Md(uVicYY-K za9-C`?A3V>sP*qo!Fxx&8$6ny3=h6XVU@pyII0`Y{H7B#k<;-}G*ci~aa`RX+ONRO zFl?vzIp?^sw**QbX}MwbIc?6gy8Q4;4f^Zbr8wUoN0?c4w)`Q)5W+T;-z6*IP;S{d zU~zgm>4!x>O!Bar>`x{-K@XeeXJaNQYNgM#W9l`*Hls*ZFSBuLiGKYE(MF*aP)HZgo4BR%ux!y zo7b2@M<4kR>Z%LL4@OjGNzDF*)K0g!tkt8AVT;Q(8jFW5E4x> zF$>|vmm%})OQ%u$eU<>^xmiUYB!T=n}E48YJmj$YNK+#T#l9(B={CqCw zJ)%{~VPuj2jrRjrU-?PwZ4d?D`V`7EteII6w}CmTrG;BFlkH5J-vWe^w)Srw%jIHSLy+9LZ0e zLBIW7z(*@?4A{3L(1MN_Kl^k^ly8ZlEp$fv8K$q%-PLu?M`)yC|J1(8kLWF_{XguR zQxN=$`sP?g>YH8!f9ad?(l_Hr?wh=44Ie-EpC1@A!}ZJiji)z;@~w0@rZy}Ir&+^g zbBANsP^9QRh28#nGmxUqi}kzsmTI|pBoE?+k_1%r(Trb6*>bDXPO|K@LornS{$5@dS4l4lE?v%Hk3R>*2~bW!4HL zBdK(C(dZq?|AVUpr|tDJSq!u1mono!iL)PNdsg9W2LUKJGom+(ZZApKP4t6~uE+5G zCUkGnCV!109ur=UN9^I{vcV6QDJEj7@!SKpdR$n%$97girZ{K7zlvj`Hreo7yW5@3 z-BT2)MFT`%8;%DBH!^Ib&ic1pcYFz@rM6RaaxVsX2D0f#Zgc}JIQxi%mv zEj-0{BU#zGu=;O(;=n)E(>pqQ=+M+ty^Gwgi`G?RQ55u&A%-XZ@(-y@$FW6`b|dy` zgX~XanD~SGL0!5O_I$^wgzot!A{u;-J{Dbu&D*E}ec`5lyBjC$I!E8Ju(OB%oUEk5 zLp4u6FF5~=mi(9YPne1??2P=S{p7}5qo0;P8vXRv&~N92iXSSpFY_;!5R}ZExmg;a zeJ0-CFW)JFSSs&cE#8aIjP$Nn!i2NA6sC#h4W)?B+dmI-4O)*$r&V6!S{ws5*U|Pxj_AW8ECsgHA6}>+tV}Uw8Pj;Wr(w0(L~4kvsTC z-gAI&uiXW zKdN1bxQ3bDjo7U)=ioD<7rsx7EeZ~p0=LV*F>ti($YgEcLc>Btr-2EDiwGat+!GQ9 zT``knr(*|I$HT31_shxI{1p<8>C3tE*NdSpMvD)fuGa13Nw6M0ILr6pG&q+hp4~=# z$B#{?pCwgibKkp9-epgc56k}zUxu^YI#O8P{ZNrc#B^bLQ@b*vXdseaGs6$d^Q4lO zZF8xP(BqM zo)_8s6sZsaIvA9f%Fp+Y1r%m-h%QtqxtB`AqUG-q_=`H}ah-_xM<%Y7{nk2t)-@I@y)%kYmb^%TQjJ3Q9#4-O9j zc3i4Xf#*B(yOz+_Q~aPbKc0`gFJq?7ef6heq=b{^jJy^iZ`*g($fcS!n|no0fW%0V zG9!02BNvDujC@QayvV*zl|_cHxzu|MpL6&l!*`wV4-B7msoyhP;jrGY+u=#Tj_*1n zZl<8&CKE9FnK}q`_@_RQ*kvAACZ!K>=2HuQ?esQea3ae_wiLnwZPDd_; zCcE`;B(9|83dUI)o(xY4Wc)K05h5Z<5E*Zkmacn2^bCs1dIO>*2zD}1t^C7rr9}{# zTSh4VQp==C5Si%8AGn$dKL&EC5)JWxze5IGd{2^?RFJso+d3q-weRoRgN_KSoMR9X z>7w#?J5g_mtJeba!A1}BK*R62)JcXhhvN<7PPo+YbeHqJ2vZX&}}vjNwT>xX3+)kGkNmY ze>Q0rVgaU4nJI(k?NgH*FG@yVE1oBD_r3hF=xa;+##cgoUNZWD;>PWp`z9r$>jf$| zV~>CX^5&-XCGl^{QwZ^~_(1`jYYtqNjjsPb@?m3ES%Uozwq4t;Y1q)*xW05?V1B~} z$(f3-U)uMD9isS|gAGgHD%)NoJ;eDm~dg8E?t-ZVKLbycOGzO&ogjXZQsUT_41EtxY8? zUpBV{n+e;}QMf2i^g!WSFHtfiHDq(oVzt|xkt#~cZ#_q1EPnI}X(_GNR*a9!HqT2; zmMp2X=z#&8zSvxS@Ips7{(Uun+R}^Lpij4z*7&gU6le-1wE*C; zh6|k?hZxRtxR+qT<+@G+B2(|}@{Z+bVxjbvdC=I|_?S(ci7U1!b%}`uB9pQ=WS+Ob zWJz=eub%yoUhR{sUfSMF#!7$s7@ZkwP^acZLHpnV2wkrx?zXa|YOGofld4IPcMEDD=O(tfbsOF5k7FK!LZ|gMl>h3lTy^l?nPD3JvIU8H`sn{rDZ2m?BVZ|De z@FH7XXt>PqKQ8r;hU*;u!tm=(IBl47sW%(G?(iDJe>l7Z*b$ZJfm$EBNAQgQa3_b)8gA|IDZ{fau)K>6$GFs=8-C}WQb8MA-2CMKSZp~?Fr2{{AtHs3eR;Eb$9(zAz zW$waiGI7)h={FXo*GM`JKJXUQTCS8g|3NKD3-}{P^*PcnW4tFiykCCiQmIe%!8U(b zhV7BTLX4_pL7);I70`5~oU8Lkm*+t!K^*~DDD1=~ihTy+NeNj?PI zR)cLZUbMr+i{Pq`RAZV9Plu0+a{3)DRl;pmQ?vj!FM()j2LEekJbG{A2 z{Vxz<>Ltkh1$_4%V6AhaW$K90|($@vzA>tkjkfdkz-tG z=SXS3u2t@8RIaL(I!l<93#BM8vaOTZBam|&cXbELnNlZ^&7ws5Z=9?-2hOuL{D-Zg zbmsOVZ@a5&BfA{ahCI8C({n z(zb%}A_usV|2#|E0W;EFU+{zVoP*`!@~7m7vvU9NeYEVmc6q8?9*=d4x$d0^FM883 zE&8E*JLSzB;+-Zx^s1Zh_Y;wbJX~*mbE04S(p~-9|1*`!>noJEA66rFEFgpCT&y5a z+C{Fkq4R;wYIEtjwb;|=T@7LX7$^eN(F*(n&Oz;*bt*F3IqTOa2Xg%0miAUT?*H0i z+HA|__Bc_67piLb{-(@%h2p>R4XkT#!F=fN{QQ;USR@r~{`>{zhG_75el59{uGdJs zfd!p(eX_g0h_26a*B9XWx>wM1!&Qn?mW3`$JzbyYu1BS)+k!>8b~pjPFmPU>idpXZ zl}F)v(XZF`xuQupxt+DyjQ+gD=i`cn9qisSYqZoZ=F`@cf>T<5Q!Yn#(Ws%E=+ zMfz?~(z^q%$0v0ie|*yWu}Gw%c)UQPB<)&L36W04$D@{??1tQl#o*?QycX7hfBhCl zPb&1!2{L4|xozsjScEc>vfz1~4YQ00LgH`dLIS@zdA4DX;dW;jt~C7A2|s3dwi8}t z_=!us$M9bce`NTv!yf=U?sMzFj@wi(XgXtOe{la9W7N)ECP_5$5=%Z6scfz$t=Q@C zn8OwG{_qDIjDG(bPIldWhhdGwgyAfQR~ycAc(LI^hi4jgI-G9U=kN%_5(y?ya(~0g z4tF!GaabmpO{{_m7#y-TRYyqS`PgZCwY>Qw@e{;q$&vCz5|gX=No~q`O*Z$fy@2>N zZjr%UsSToGz4su!q6tvuJR1vJEb9pDC<lI8uRvrtESRiLd7MIBpeeR1rqepsw; zU!kl^gJr#g+3SDXx92S=uA@W;jrw9JSjWx(hdLggTwKTGLLEQX<)pbF?A`PKP{*Oz zK^FS9rcg(9u#OH)lUR<(ATM3_?k95RTgECP!(?>FY8ltgLu{?|EaYthNUC5eIzE}W zP>N5vD4DoK$V<%Di!N?>8f5Im?-`BziFu=?fA=)azhJcVABf6W=3BZItX=c~=3^u* z9+q&=d77Y6%30IAY?j-pTrbnGTNU|79l2&Btlw3j^tE3+ZoIdS@VgLm~YsOYaM$kJY(N50vh=^b%4Sx7%x(o=3RE#EWK1w*YqClv6W`OD@Oa$P&WpiU~WNSw^PYq6Ze zxfqNAu;zLkHGMiqmKcvfs5v9Ag(b!{C+QL+wSz4&9z0#j!4gBHY>Cl#vy>M(TLe+w ze39@X?>oHG@RtrRFucd%>4qPkVtJ#6ce~U>4R3R}kKtVoCjvWml6!Qh^&C+mDR)iR zZRamBE;=K>#CXy+jXKmIT{lB?OlXS(Z@?S^)D8k(VOS#l0i0#1-48s)P`w3stl>g& z4&Wh%^BnGFILl$!u*TswhLatBXyLiUVb0K|ecZ1G2A8kpH*E#$*KI_f7p=3M)+Zc; zrTOk7CLJV6V@6EcMUrxG{ow_ERD&dWQfqk9>yji#SHqK*OVV8UMRERgFL z55GQFuAe;o`YgF#w4<2iTA)_0Pb$dgJAdX0xL){s7`_+G#Y|A-1(zD;z2I!aKG!)X z8g@E7%5b5>D#LjWcQ>5nu-vf5VTs{nhySw3TjKC7!@L)~AZYuUjR_3g)6N9XUm?4W z;+ui<*RMvK<0csv&w(3axFMJ;mubuW1t9rFNVx5f!^M{-%cW}rm!`_4ivpKw7Nyh%*Ny`7bCvV8W^YxqM^jkMY zua2tUwBh>QDN8)u_N;YJj_)1R@34S=PMK_Oq25-pKhPAiY0e%^OSvc1@O(f(!%Lio z-~0>>??Uy|f*;6&%YI0uPg>`u3E)Bt%JU(G7Iezs`_pn)FyDrq>S+3?5U$O3y4)*H z?YCk6YU@6Fb<{TeH>xKMPc;qwN2cG}IiTVDSjQA;Xv$=B=N4;dQ@uV$3JvEv4L{P6 zl=s5?)iIx;;c3G)Tr7({nT++@ckut9hNlEIG-a~6$5Er-hPK$ib6@f_FS&*-`sDEnS!(6G z)t(GqQ@Lfu_86|%4f{)$XeSiw6mxeBDmE#onCXDu7m;UM^Im7mD?B4A&xxXl{R4{W zCB2`7`)AVMBL$ROQ7g4r|9-025|p|`N{x?~ynaBr`IJD5nGV_9yRx;;D`xAu!Xw^4 zmBpLiVs}fa_wS{bMpf)x)JlrYFvS)_vBgrvYvlpO22obgC@>vxeo(BKEhwW4QN%(i z;-h1W71K-Ge1>A*8m`zMHEQc~=y(J@n>#M3*y%yVOowbP6Gz>y#pF-~86Q8iaLNDW z8w-M|d6($2Hk%9WS|p@b5q#u$<#6fb)V$eFE^~CToLrTE;w*uOOLj$n7Cl^Y#xdW@ z|J}cm@hG-=++HR48-SyIK2w*TnI)@HSF(LE-jeCx%B!~suQQT~g?Ph3-+g#P&Ew$( z3+&@$e7w~@PQk}{_Himc&b5z+ud)HQjD!TyvGy zXjkyOnjP3^PB`pzL#rX1JHl|L)Qi;p4aYm&%}}jIYMG(75O9+ndrWo0?;0*}*e93@ zoi(&IRFxf0`6l_KcTw(`D`~F(WgsPHzEFH+a!Q@i{qe|uA^pri`ov-Bvn*ZST&{M_ z0xt~Y6vv$%xFQaI$@6~g_yabH73#Qo^`;K1Y;K?carJq2MMe|GvyJ=2twk5zg7pVF zh2Za1VI`DI$j=I;w6-Jl^PHpc)YqaKeTHvJb4EsGNqkZ&U64cbdDX8Um%rmPV;deBo@I*tc2zZp?Of3+9ug_4cLF(>;xUcwezJ=Og*S}k!nOufWd2UI47w_tq$>n~oe~#B&YxNha^FJ|MIxM^T z`3Tv6|M{}3GZ)o5*RmDsin~21Uwynt@MK@mRDzWL?7I>gLI*?0j&|^Sc4Kfzqu)ex z9yxD!Sf?KO@G<6*H|R7O+UKKud-v)mZSU{QfwU{rTcCxaR|_%x*emiLi@&^-G>6Fd zMWR-Nr(VP(TE}EC1IkT|*_s{C18#+~y>>zSo3S&JeKV5p;1?nwv{;1@g7CM0v$-Q@ zh-hZ=i}L1a{a#u`T&JzfT`L>4e70=vud*vDRJDWovt~qb*PHuFasFdDxri*I;*)B< zvbLHrC6D*YeC+V)>FlAp{-ZkQo|g?D({zHx$x zW~PqbwiVWx)+IRqlHF3FZGS&bG2gb6rzuKte%t=|Yku3Fax~iZZ`QVdJ3`y`W;A#1 zTI@!pZB-4l?erU|ZCA-^LEH8=$V(($IRRHvp)3BAZ`(?3W%b+Nd|%Q=Z`&VzZ=|*j zMucMjmAR%HMdDcsyisE}ZQhh`Y}Bc8skF$+?}8yDjnL?q>3y?I&FGE3XfE;>9C|y$ z*>3C{v4f`O8~vfffLk#4-tUfNZZLw8xeLtPyN8>5+ooc3RSx$ew`lGUYDL}5%+byL z(=`Eezhsy<_r&c3=00hwe~ok_jhx@gl&JG)aa7u(xeLwQMZ?X#f2i18mBadXi{}1V z4qUvQfZMJ z-I-hTNHQ~fbaP*oX}jReM;YeLZ5$gg_w}!B+1zI@8Ohu&_HL({yK1<(OFt<#SLIS^ zksICI-^p={mzg)Zx$EbsL5tmBtuxG5xz4>V?1o0)5e zoB7v|i_KKIR9fUlH!~||Bwl90=w^n3W?m%oSz%td3yWl6-psL|Vdln*KDU{DW@c`< zneTs8Y^KVk(jqsynSFXYEwgZRGj}(;^1XTEl>svshXQ8q@m1g!dUK}{&D`Qx-C*W! z9B%H`0drL@l@__t&Hb00<9M0Hqnmq>*_StWfy|_ZUR}DiV!l_udN6Pc=I+^mUe%|T zbgzKf@daeS0asyZ2r79N7qf}6c%!Lx&yB?$s&c8c$c=7cxmnmbx`lOSQr^Pf%|>wr z3*W$sxyWy4P6ckk!kP;M7UDLcn3hb8H=8EnO%9ZbDmdxGVw+Shl@__tZ8{P++R$sO z)uyyIem zx!l*KkMq@KbCbR$=+&q$W~~>VbeOEkbnke6qo&U+N^kDOeVA}*LFP!4nH7+k1(~-G z^~?E!$;>X2v6~)VczHqU+lY?PW?r#Wr?z)^Q9_XwSy*5~ug+qrd2j^ScyK}LG^=rSvD9Ld+NB`XgqVR2 zaE>+E9BT^Rwl~MhcVd^eSi2&g{1&-XTI7lyYo>N=eq;5+U~IcUouox;n!mhHHL-f9 zEuXVvZSw$59pA`#n6C%^4vnm=#N9HcQ}%|X`|gKGjTZO|S6A(+SIxb2?}h>v=p3j| zicG|yys5MNrfsZJOUdSz)@Zh|(n6_(zM+aEWs`w17QEVyRW7bhw5ys3+0BS&_I3Z8 z?nHapxy*ItA}Z&Zdg_4jr(F>F_1DDz@b_h_knWn?^HjzQw`!s~`F^@6bX}r5j=AR- za-e{-zNDBx(>Ose65j$iH9)(<5_8z`R@m~4(vgc=wDS5B_Jp4>?_0h<`1dUvh zOT52nbH|{zclGH~>Ib^3sE*0!v8VwW+>ReDk!~s9p`^rVQJ&ZefRgG(uI~)dzIwit zib6|r2cc9v%(&W4XtEkm%u(~bu4h$tO7#HF08%A*$`RuBt%5^S3jUvQ#>0?E!w-Go zgiuc5ruh}Ed~efcJSK(I(t*35k}7g9)TuIG6ve(NirHbO%R1T5nnKq6(E-Y=@Dr{? zLdQ51g{uYebLA?&@ zR_`d&&!QnTA$u8oeItn<7!dDR0YMv+_r8miX05DLe%WZVnfHqmP=pOI3=ug+wc9xG zNc{1Y=5gD?rq`3stweuMwsv*KzpvCFmo=z;1d8x2) zxf=S-@b(guipT8v6CkC=t2|l`-v!$5RJo}0gHwDWrbhjFl)_I}#xZ4K@wx`)X5_W- z7~JfAq>Dl~;b*_3MedW?L>_|^DXZWWoh~A$i6A7d)A=TH;=zXBHN4-Yo@@A1hjoU( zcfwycyxpZ9WO$3i$%YFY?gZ?>LBG~|x+o#0hu(V2O-C>1&PVszllJm^W4ZBlV&v9p zq`Vf4Ozo{k-v724`Iz*K7%5U_Bk{7vL1Yu-a`!2GrQ{`sE$6e|*hHVZnF?@Qe zY-;8!*T_xPq{ zul}kpv|&K2&ke8s(VeteO9Ebf|_&1 zmdN8G2wUFVSMm7+4WBpMW-r6#h95cMM+|p$!mWlIUFw~N|8#hZ;fD_A0z1xiKKsNu zsuu?4^f!E;T?wx?pWW?n49sh$i;8B&36U&<)0%jAd0}1{r3QK<^=&E>{%3L+mvKC3)>LkXeRKnYB zt5jNw!+?^C<%hK~?W&F%w|fP`OlEx^*L(gPpidcS0u=MfmP*ZOATDg{XNqx}VJ-yPet3N+V?v3?> zpBLYU*2&G+auHU;>yDNiTW)zV(v67;r((N9HzxcRsIeBPEwq5Dgz>pszqWv4Dxq?z zw8)L#0^4DNgT1rW-VsiW)biuigtQ4IiYy+9G6;UtF(fgQ~*TXakF|bKc&x z+0IAKmV12COOd^`h2F@W`5#XGZvgVz>=M*wb!x1oB$weF(W~x6W1Cj5za_y{N*L90UG&?-ANk?ucI5N76;D|D zdGhO~0jvM#at9!w`Man_Uy+z7l2%3X&p7DwCUE}A8a1#Van%K+3I*Hq_-=bP)i3zd zH^oAE{0W|U#XqI#r9agwm)WcSvM=6UEMCLntEVSlq(-=YoBu^4>?60ibfD&qi@0J& zG44OdONQd13zsc%L-#D~>5_?YI&|f=Fm#Xq5=K=jbmq%8bZ29X>(CV`8@g?Xn()MZ zL=X*hl1O-wNp2f)wBc1Qb*kY-4)-u@b_=$h4I5nQ7{hZMexU2v$b}C3fgQSkMXggr z3F(&5qdor6Jp_Z>mSo8TvDj26EnRoqrx3LOruyQn5NDNYy1W)lpZ-NPz5XRLJucNt z)DkH({ST+8tp|u8lsL^4%RiSi@NkKiXX@y42x2 zz>a%e)T$CC)buA_^i4k;=ecJ3ZP-#`^k0nMDv{B@0lECq|F#t2ReuE=v)rS{NfrL% zKW9t@;!!TlP{ES`x#DKv3&!E2ZaQp^Dn1GL6+Eh_xUssCa?cHAbIGmUX1_2tuO3r0 z&rCm7DkuZE@ON&>PJec`fwZPrjf%_n@{1e!muTp7xwbSvk>>eguZ@Ac{o@Dq-hesR!-r-bW#{;hCt`a4* z=SpAjd+r7}z4hFSSZwDGsum+Bs*&Me{pz=;kwC|cdp?xF7;Bwr4G+FeA?lOz>Z%!BX2t?Z{%*z`$paa z_ckNnl|7zUJqr=Q=9AlALx9aDO0W@?X%;@TSUoPAJAEsNPScdaK0@St8C@vX;qF7x z2LCcxU?qP|N^aKt{yw7SFs;7SufAM_aJq~Oxd&@B@fbhxy;oJYvBPy+QVLObSH;fH zLaKazZKngo4%Mr$pt7HFk#fwEqj6Lr6ttun8_6O|7Ea1ra+&DSoRA!)>#p96w$$_W zs}3zH0quneVcm8*#QkBt5VtNqYgqqytSBf4e8Y$J!g8&`%*p1yuQ!1G9(n)^sG>z> zerCECPhj~=$jiTu5&}c}HAE;H+JCASuis6*US11c|HuUO`cLAlEd2#-*?N<9~$PPXd~eyTC{crG!5K|A=4jf%AdA1s(K51x-7Es=Tf4EvK0i*?;P!lSd>V!4#xCt+ag zJy(A(I8!!v_;p9TN%c^2ri)SIUxp%HyCQzVPvlZ*ksEz4IHaqb&)5qdJ~Lpg(?RoVXtY#^D(lD&2dug*lx;`hCy-a>BG{HE6MLz9%4@+r zr;b5pYkDGWWbkH|>!LA_(_9LL|IMGXj|4{)Xqf)ZGkEby#M2gE%ja@;nYj5)yWOI;K29!u91(kS zt^r5);@)eeCVbyjA9eG~zxYdOgYa~uHfLl@GOj$aG&oJxnP$Tb`L@$+ zNDdqe&a(R`hSWImXH7h4h(&g1$bAO}3>ngg*?dEOrcu(}i_hlXu++RE^N_mk$)Ev0 zo^IuFPb}1c^<3_#paJzJKCCBhbmC>7X~3(l>s?PA6*Qn%BcN}EI zE;Ve}=x`gus~vv06GnEg8`(#Q5@Kj*Rn{NbQH*^X+4o9B#jhJaFS==8$Q@VQH{1%| zmP{O=hRbWg@cJFp@b9lO!;`2@4HqfPyHfW@ks1+%r5B5Y7wP=6;hBa%cB#`1Z*+Ks z;eBBf-rw*Bm%5wb9EW9wO%6AWhv9EI!;cXq)bQt4`i9TgC2#lxSXSe#R9wzCT-U^x zPQewq_iU}<97MkZ?t5UX%$Kli@ZaWvtuL_$zTi?)NHGtralZK{m-2m0ArHJ2>J*%B z#}es-scR&F$!nnxUcp)wT{@_i;rpR5T>C(ztPhTrZJ`%=QUsyt+qxKyT<>Oz7Y)a5 zZ>e2|pE%)14JYnmslPJ(*rncWxWQr4@FR!vJW|J2vSL83nPXHhXnJCo-v_JUt>%2Y z%T^sP~kA=UZQie8Ziu6H(uHz9|Pw+4y^t*yhXolo#3gzU5MOM2b1zOy_(@xs(rHkQ8#h z1p()~T@QN>N4qv>f9Vm-&5gB1Vci z-_No5hVzNh2mnRSw^0p0N)4CSg5eizqlRCzT#Cc*6Uv#6IG;$F;eU}Gn-}?p2tuM! zB)rJio#E#h{?w(OY?yFZWB7nupB-#?lS|#x@H&SR44WNp4eZ##8U9V(Hj3drOMSyn z#Yi#37wK|o1n0ZrE6{eh^Q|p*zK0GJo3pw1u4J&uFypI@e>rz_{hg`{b8gAoKZvl4ff*Yw$anawp}zF>=}80v)HbjK6Z9gGl>1b)IcH&mFGw=4`Rf zDup`7=B9l^KH`9qZ;B)Fc&vd?ZxXvdAbXGY`=v&nJaMOAG z$U6T>PA&`W{Ho3*emjr*{5mf%o$ns5^Ia>8+gar>6t}4JO|rnkUoaV2=a*zBQ_y*z zpw4GLkLpI+(aiOy^w_>A0AAZ7Pd(v1_LqEARd2+V3nrMvm#53Soc3#f1t! zC+E+<4FBj-XB+;~;kON!yXDfUhOI93IKz7#e$DU~4l98j`??5oi71hj>Fqy{`w`|U z9K>3L`Lndg@aEQUT=)(CH#u*E1qZxi4gR;M^9`;^7T0&qLvFcAydQH#_iJXz&S}w88gz!W#VOt)&7qxJX%pUoTr-FY=!uYw!w@ z@FF+6wb71-$4W6!Y_kscNW|d=!&BXI^-aUCxzy(k_j9=1@DPWO06VH&gMS!Nz0lx^ zKlu&5)wcNtKU3l+zDNqo;g&@Usi!bS;fFcYaCt2l{w1*|6*^G+58t1c%EfSzGQ*pt zPA@X_shPHysTXmEhYhE@X4%H@D2E?v2#K8K7KJ&(BVFn%h6g)*+VBX6PXIf<>I~oX zHPs86{&Qr0#Xi&|dfO%X)5k2MR!{7vT*X{o<#qIy;NQy2`;Tw;0xKedkRbIjnCE9wg43rh4R881oUM zdHVcp^%tA^ONOgo8c<*5@O9-as{cU{o4-@sK4z0Bkxb(tQ_C?aeRh` zP-ald5@U(7j^$8Un$&Q#FqX&?*%HE_lBG;ehmPYgWXKX(QnECtL>kk2l5J#Z>|^bW zv5X2)>i2$M*Y#{?&P@4y{r|7u@_N-d^M0=DUas|Cp8I*uTQVxi?pLI$n2xkN4`<E#~J-%b>4Q!)g5_xA>Ii*K;njHp9Lt#!&fPvP%IeMw=;;FAh-YzBOv!WBl1D*VFW2!%fwycDpq zhsDjb0;+;>(`AMmH!~4g8aLZ;LF+1B%rXvVnb_aLg_+E9xl#8Ns^2WdMrkP&nnfH1 zHSH*iCj7pe>{8Gy4rQIJDscbX*}w2_k5*+iOU+)6SZkH6%aWw$KHEW@ZVgkz&TUN7slK_WOmJ;9IoLj>>rvijqN zY@C9VO9g-IAQ5~gGSa@+#VP(On|C_Ka9 za=^-yP4J!B2@!nii>}}uaTi+s@g_P%rZMg~nP#EmwoO_t-Vqr8-^2Dl-rp^{;ispY zoqreI93-m7_|N`PH{?>?L=F_){Ignhv$skRwT<4k3QuG;(A*XZk2d(H+;#|^Ws}F0 z!lR7(iNZq+E>+mg;2glpQ%pDO4{&sIXN9Yqtq{#N{!9AJFvjh(J<9pUe+{A>KBw!> zKriGLAo@P-E6ScsU5NCrH!&(Kt(F-j%z;5?phg*Cl!JtlmOLlHN4=c+79NPe>&e6h z{Pb*a|0=P;G?6Tsn*W#CfLv;W4f~4?ZXhFiIEftiB_rE2zDMCAqu#1;w!!NZeq__C zixs|N)cy*eH+a0lSq6^;tQ>7NxRITRF<-Wv?%E&DR+e6eB@HWw@|@L$q2!xkRuUVVk?&k3g0$rOyL^_Z&diX z&6%!HSZUOA70xi&PvKmH#{yPHOz<&Sr>Wq%FSvr=2~StSzu^(aOo90fx5Pc?^OmKP zF!BVRj4%6GTfpu{8_B`$?gpQyg*q$jPTk=2k`-RbPbG=DtbKwd4@fUr?UkGqEV*5< z+S6Hl6fMAJiTy0u;j#G$KGj)p&g-zNrDr4`+O)xg;rfVx?0!>!!15zOMqm{mlaDQ|qdCaS?*x%Z_R|bKU;? zvVP~m1>wWHWys#-Zrb+Svq+*?=BB2$`F&$Qa4&7d>VUz{#7r8!cN%NYDG=^OnlQ0X z|0iLpRhXK_!i?L;FO2Zv8PsNlDcVki8I(zwXM%ov2j|tEFnu`B#hqmn=JikhCt+$; zm{en7=5S8zT1NQro|tBZk@xfAU5S~5`8*g9bFnSr6J|7LK{&v}lE{m{)~GO;)2MIB z63tz=1>cO)pdzfnSK3x>%a_@R<6OdrcfT|%!tXNOnOo`?Vcu(WT6kFM_BW*06_&jw zNH%xf>&;4`iY+0;e*Z>@toI-}UhjghG$DGvZxO-4w83Q?9>i22%0M~DwN z8hCEiLaAay2yqFHQgQNt)y~E(k*7kmZ!E;<7yUvAAKoX?tPoeo)NAgf%tHJ#H(-fv zgF%V{P1`aYV}`-^vDzJxo;zSKMYv$TGXNM+kA8#}VvF&vcFE^+ znGyNm5c%Y#w5dB~I_@@lB?{2)sxlE@w;;qjld^Rpjkw403RGKN@MYU4B^D-Qri<|f z;^8#%;o-DqwRM%u+Y(iov^AgG1)jEc;He7l$b7FkLZyHdy!;F9rP!Fbapi4+vBg#1 zP8?`Z#|3h;dvUy621pr3@z_```gUTh8d0bU>?3pbsQ4=i9Uz79`*1TN$`P5Y?n~dP-M2dC z>K6(_2H#b9tid+`D?hdA=Vt6gjPvfYC*A31>))K|=S7mdqLRGO$M2S?>Te|9l*hw|1qW6!PqGNHoyTJB_5gfANB)zn$7{EX>CM>0)^3x}-|Y(Vb*I~h zNw*W#nL_7wJ)$Z!=yoI~K)y(Nb8pe5GzRM5uLLCQqs*!!ystr-9sRyT=JY|rY}9yS z=#J#n1$tn57^Hd_;_1O@?bv787boEgA8xuftB3u08VWaEMGt`&;c}ulN*$UcW4Tmf znV()PpZcA|^5DlcmPgCoBqVxrX)MpdId%kZh>YmLx#UQMI@s#4pTaRlJyzi`g9j_T z=NB!TukbpfZl~}vgPSS5*5E3v`zuox%SW&iF~;)3$J|)%foRvIvn$r-lmwT&~2yPT!(++68JT_)UUtgZJb1CAQ{nM717iFsry<=~=>##+bCmcCtn0sdr7oz3Jd#3mVU96H&NUV^Gy6IiUc7)y&(Mk3kkxl9@Ze-L+-L6 z2+5^Ec;1dwZ>WZh=*$M*@JWQqY@5G9;oe^=^=pOg3@%f6xXt|*DBRJgGZeNmI9Xv^ zgZBegZjPt+*eZXYOoQ;!2i+k24g#Iu zsVQ=^|~j@CJEI%*}aAW#Dl$;#!${ySbz9du-bsN&79P z9&kP$F{mN!1nv%8Hi7aMrj)_S;$K`8f$}0Jo4%0F6IDr8-tj52G#S!oVj_y?Tg8j{ z>3Q*A$?!D>&4q4<1Q!HLGlkRA*bmOPG>VaN^1Au_{v=eo~uVFLSHfJts}@Vwh7C&*TVx5d)n z4W7?UVqt%zc_U}-lMJ&7;#vIkJZsL!;#u#FSI=7UIV*u@kxM;m?_BY$N6Co3Eh0xE zRN0_#mck#6Iz{0Z1}7-|yH=TF3O_gMjS80;yh7n;2G0eo+`~Leo+pWMS$5#Pu4nOn zuzFS&lA;@CKAlW>44q96v%_$LBo=){l)+C=883V!%9t-r;P+&3Z%C9uE>%XmZABUP zkP*t5O^!rpCsW2V3coPwLkg=6#ufhjm6p9(;d@5CQsE+l=P7*G;K_iMdzmus<<6@p zBmW*(850qRstm4YGmf0EaVh48nN9z)^=_9p5FGv{i@QGARuX44PaF;B2NY%*{8KWC zJNrv1bT$U-s<6c1-3kX8ELB)+@ESn8 z_!PU4*>%r3c~B=*Y*2w4t;u{(C(~WXq25n z(P5^{C}E>yNg1SD|Cu%h+ljFn8C4diCStyb->*bwjLhYymqmwEOBTI3p+m}jAG3W7 zDdf^D`aW&|A#L^{BYJu*ITE40Y?I?6g%23@bcJ^qEK)edCSyk^j2U%bg|`@NuW+2f ztpF=eu`Jq`osca0Y23}CXJS~_Ec!gxknkFDQk}Eti&k1-{vV4zC4%$Q6a4D;MDWO+ zD)_@6nc(D7!E3e@!A~V41iyhCiO@?wD!g3bqeeYP;XMZXDty5vUdJdbGwMMKZ!@^N z!n+OT0al)Ff}h4th~Qt};R=2p?(M4Jx}ME2OS|bOY6Kk*KJDAZ?|iTPNtv;}epSC; zd5GU~n$=wFd!tMQMV)VwQQk62M9QQ&9}~aleEAD$gQ;QzetI^je@ASvK_pA2+F_L} z&POh_0pDK(rN2u?^z`=+z>x?Iu2cAf!kyn$xLjdtgG&_dXS3{A6}B?!G=-ZPd|2T& z2JZx{Jl|~ao~%AOpBWf)ZSX1FQ*BVd42iVV$oXy;!JiYs`RNJ1@@)}(3ln@dHqNLc zmkNFk+r;}{$q2#MamP9l`q_3hepa}LspU(BZ4ADr@Q~$Nwo2g+MtxD?RtBF?xV^#g zfRz`T;2$#?i}PJP&K3M?xTgw!r0(v?J_iOnZ!2m&CpC?okH5j;X1;E}Q|4sqR^Gb7 zcQ+1A;(Shi8)B4vP}KSI&G}9;$`(>4&G|S9@tp6;d33%Cu>n6l8(j64*x*)?4DZwD zpiCQ(OKq@xbFslGWJC`~k|Pm%+%{7#Q#jG6XDci-*hk^>HYGe-;oU|(P+`>IZVK-* zm)Wa3tX|SWha)Ubp zR+gCHXR#9^_=>12_{E4675p_>+GfanW4ILboG%e|Qdaw!6cp#<%lXYT=Xl!s8A_Q2 zN^p$j{d1OiM9QRjT3Nu;ZfKywXN$u5=_$O$o1*acV!vdnfU^NAoLs8#E3wLfhF2_6 zrEM#_J)u9pSGeI_6i9>)V>ReYgTj3beyy@>*wSg_X@lgEpVvijetLp$LwVvkyNmsjsUz7q1t*sZK8$T9 zLcb)n^KG?zSDD~jDC}lx`STso!+{1<3VScpvY#kCz^F?V?rCt2!u<_C2UvN73I6K_ z=@)uB3Wrp=EXR97{)R8A;Kf+mW?Z7R<53XL)0}7X2Kip@l23y@_7dw)rp_Es_u>Y8 zfhDr2P@ET1_mCG^d8cAn<~rHfa2FP*TUD)4RdA^8=2lgPg-)ibI%!oTv*xX=s(JUa zsx;T01S|9T%7xg|w>QSbexrX^!F0Zv?Tpt@*qa7V28_%lkLLtiEEXr&SyY=$9s0I3 z7nhn;)mN`)C84@SYREitAkW$kSGK|st3kI{z9rO72ESEUWRu1Z74B=)MG89@oTYFd zgHr%2%ghOW{73o)Cm1=#b%NjFfa(Nt{s7bBM&AF@GOA@EM8?aQ0)n@uJn@`8Ze?|N z{)3Iv`^lw(k6@dLP{TqMt(|snfC=7G;nAj+zjzxb5jxo5j|%(Rr13L_2O0Hkg?k&+ z^Nyhd4ZZ+s<$WgjZ|fYvZz**JUx)Zm!B52`JEQmSB-0*<7MNl3hr}T4bJJOAGIep8 zlTjU)zgH-W`IB?_kvsTe@E}-(GkMRUF1C7hMLp{8RTetEjk1|g(%iQKX7ZjDcnovZ zSab>0?yH&SOCoc5(R0;v7Kp{J7Cj_Ww{vra79*EhY~fmVFLW#!A>Jjj{Sx}YZsMMy z@D8J%pm3DIE(#yAnO0|oHyd?lh1VL~M&V5c*I|>S@+q@ecXlEMk#pX`(*%~UdLvBK zVw0AGjCo8WSN&7;@HT{r%;l#i_=WRD@DaDDtKP*O8VXJ>75uHgMergrLhzxohZ7oQ zf)7-9k5NxjIM!fyg^$}@wm@OTsCy{9!Ql1^Z#B3%VC7U3{CIXk1pjc9EBGl06BT@l zEFd#X39gb!lIyCzQ&R5XpDgS@@dPZlq4W?`HT4Hk_Yjkp3ydd(FFuic@@slu;pf&| zuX;bsA8<~leyJA=lhBSLS4K2bT zh1TT`Rq(uqZB$@>9rIYwwNvCohwD~V{THx`?GNhPA(Pv-an-Ct;h($57X`)V1nLGF`>9b$SnUF=WvhWhqhvv4(;!8~51V?19@a9I--lBN? z8n)v6sGr>6#?dBA!+&9zh)-le^sCH#y^WR-h~COcv(Q{SAduw1k|@`D?0uH@Uj6s^ zi)J|QEK_CoeU^Vr6)By{AArKqmfmQ`c@(9p60yHVs_kJWiU*n0>O`%-2MzW2p8}&w zh63pHgEO~q_63gcleYpnI7M#fC;uo_&cl&7u^E=o+`ov=r0)5Q_J*0`@K?Xs*Vs#7 z)1PwrBM)iUNF4Cfi-SGpNF0PlY8>>#u7kt@xm1KZR^d`2)PjttZhvwlLN|Y=a94%J zMs2OIr@{4A!aT^PZ@(%$&ZxBtyBK_5VGo1z0W1HqILKiqVoU(zN4Rm2hbYrH82sLP zz8|+c-rKP;*Q)=WT}pGU&1@ad^Zt!oOWqIY-h?&R+Wa$>e<(^caxI-Qnrn3v&vaeu z;ip-^_1mK*W4f+2N~w-(?MABOT4htcm{QYBXM%@K{(Q>kT}!`pyvv!n*7(WKf~?+E zXcN*`{p34^?DMX1nY;_X4)!MAWt{WnPn%_Vo>qp0te*C{cp5)FPkUvyc-mXT)ziMM zWIvE`$))1tW1$AudWejWY92Wfp=q`i{gT2TjQXU)PYm9t@DH1PMHSW{UMV)Jrs&tEV_?o3LzkbqZkuETk?#?(C zIo?l}^@wci6q`l%^^^aWDt#6yYtABl>-`p4zZl(-*sm3f@YAzM$5+H6T}4j3|8%ZM z9*>esEpitJheW6)86nldjF3w$ zvhHVGN`yv`5mHT*;W+g4a)oy(oNv@y6wWkQqHvkbur5$I!>Fe!oMy16!j}ym4p`aI zEOI?NAr{HG#$S0#5m9>8*84kZ2a6xa=6*JHSFrk#xJ}f&>84|oute*$0$>(lbj!y z;7A)uX4|^=bo8WA->=pZ!M=~PlPG&;>bqtH>wAObN4M{%Xz!#pxWYd4-m5EhYB6pW8%NqCm!{1{DO|$)Z^?-229pD?&9hq% zXDW;vHLP&7!LAA)eof2nr*M=}cTsqq!Bz@yHu%p>7~}*?n}zH|j04%bSGZ}j5B9b+ zZ9cvb2O&089m_2JB@_EC{+z$15 z!$r3Oqo|Y-Z>Z7IrB?b0f9c_Ar8_n%UCYvrHQ9bD z=*#yN9edP`G`RL15toe+nZdztGH z-Cl7VqHI3M5%GcSy1(HN7*x;2t4BDfK7+$y?x6Yz0POB$l9qLBGA&!hO|&_jq&sV2 zd8Y%j)gIUnWrFM=`=YWju5)_Y$wuyDw7z$@zJG=(f1|$tA>QtsBeo^%?RE#|8I)+$cg+IU_v^)>+=)lc1A)HZ zASusvl-HE%boLxllhFdC2TNr9(;e2n*F+}QyhN7LROAdy?e6y_vL8Q0g}y}g&?sOO zl`@jZ8hok9-lX?Yxvz$)T7#3VB2|U7$V+~fC|GoXP;2RMyq`Q+YQQOI8_DL**2JYP zC!7wjAMTFKX+2IqF^1GOg6ZvsY9DI-3RMYGv(WIpgdGJv}W_mtM)v4{`6;Aa^FmY@&Z?I&L- zHTZm>N_{|EPBdj%)4&A`+t(8P$da_k~sJ0#&CI*aPHR%Uwk6@nwI2P z8M<{dIbf5?Qv;IRRw~eG>4bR#MFB`2v+wdH#`rrTmK$b^EMu3lBiYq*Qwf>K_rT*; z+!AGWl3OyGp4!Satp8B8&=|VR&;uBLZjzS+OFN-OCiC)IHh~!Mo;EsmC`dS{gvb9S zY~RNuE^$-Q;-; zvVyJ#$Lp-halSSQpN&>v%M&A0b2dtUD)nOCI8HNCiClyI?=)I7MjFOl*33qF@N$qc z8|l~*8mWaCiJzX4o|z;@nk%9wQy)ETT_BelX?rdX6QMiF2qV2rjznmN9f5jEVU1DC z6~1NgHifBaTK0N{ON}~2VU@wN6uxP&H(=##Gg6{P`h{7}F6X*Nnt(K+M*3w2JLTTH zzdG(S(s9tPP9GZVoJ6soyq757^ZE+U>%*6!*31h1`6?>3r6`o2oxNX6N`~cnu`yHf-i>^aY#rD zk5l9t=GQTwRe64huj2OJMR22^+)v2z(q}XJIezj%LdI+1Ci32`oAIhW+xNYc3P8q7 z+xD2O$>KF~+{gn%tLx<&5*$q-HT~;#c|)xBJMi3b~Qe8$6{izXX+g0jIh@G4_5y z%cs7;(907I^V8Gt7Y~buf0KUWd@7}+FOf?%d_32S_`y9gLc=Q_7g%EV3ThO#!Zr@6 zZz|ki%&#fj^=YL(t8l$hA5pm4U_#+P21f%{zG)hs_kr{ax$I>8fUKRqdmF*28r~E0 z_U0lceMH%w+m&e{vzuJQ{_1T${dZmDgiNi)#0h@#&O)a3oX0&o`N?ZhCDUkH$mGdA z8QC@3>0)%G>5$3k`Yvz41w-chXHZ7A58H@A`RN(7)k9*?d{I*})rD&{N=hy@=x{Fn z5}{uoRHN)DH(NsO<}2J%;Zdvx23;ff7D5LZtW((MDW!g@@Ia%!rEo8UuPZ#j;PZf$ z@0vjy-j{yCpriY{23-s9Q-huYhxKW7*;t=VxW3TANsi#5lP~+pw}I>*?F&3@KYJla zUaZ7?Q?lYAjF9%CP=0y}oihQl#&h12Zo>b0=Yt9*mnw8`%902@N=Ec%k=z>zU1c|q zXDR&As8bYvVQ_-NzipNkQ~0@2Z&bL<;1vo#Gk7jwwH_%O!t#-6|(Ox_aqpr(YS;jC)hXc;ZmaxQCMa0 zEQN0x>Jh(81KD<&;-w|r%Zyg#jy;F_@p`#$d@3*7 z?~-7oOZhmY|YZcA9xqSGwX)=y5o!A>Rh#-6vqh@a-@F`KJ3zToy# zzIS&z*AQZYd@%t(JrnFxE+*(Ef+bU@J}jc*wxer;ySXw;gtj6hOi)0MMCcPcsJMs1 z{zlziVc6j23NNuq;A-6WLD>_G`n|$q41TQec!Np6%JpV~t=WkfHyS4tyC!IhuvHUW z!Au^H>iQ7#%aP1MdyBaI^u#^#J`uMMeoKJDT*M}sgUF>GH|cHZ-zA95!nbdtdp3ePlH16Y~8rB0A{VJAe~XHIs--DjdB z?#)x>Y(;;(`D9}!-LfTs(ry#9e^NU*1A{}O+XU~?+tW|FoikTU+$zIKbB8F4C5OAW z30}N~wYewVPL;&#-X<8Y)C$Xxhm-2wK5Qu=2MArdp5og-Pw}>0$EAC?9zq z$eP~@EQ;Rqlb;Z>FTa(-5gO$;FQ47ZwVni1C-E+RdfxTPJ zy^HHXCNMHWl5gcUNvO^13O`iX@?M3D6s|SqSqj_R3~`FWzl}OUVS~Y#!oLjO2w1tB zdDq-GrC;c2pMI`)ErGLYFztq%jhl>r;)Nd@^R6c)Skk=fHA}c>JeR?{uF(=@w77c% z?V&j;@ZKoV=q8lf&2B=i6z6ok>onNX=UuBM-MHQrQmW%!ACu~O7hhx$7(Fk2m0C%r zrr+dqpIcvZyvuQ)GC%nukkz}&%)74flY0r-=Usz1_ibWTmvcW>0-mRZF}|p$9Uz{@ zPtVh?z8e~g=S2Fdr#*~Cuy`7|)YEFXa7~0xB_pJ|fgFj@ajz-7T;ZceJxAd^2Ky>} z!6vrHC@eGTK?-j(xVyr;4dww>b~R5sO;*oL+h6x~J?%V%g?icyE)-ph@ag`5MLv3k zQYBLzZuD8CVRpJj*5BYGzYnrnWRh9rD?j-qA^R*+(ws$FU@GrfWY$DfEwNu97U8F7 zkuMTrk>5m4tp9MQN-RPywaD>{aS7``GD4~qxQ_%FepO+O!d7=Fd{f~DV}4EHt~OzP zR^fW1KB927!Gywp42}k@>~0pBhlQ4oEGM7nTI6lqT2+he`Gmy&TD(w$u^)v=+ye$9 z0T{EEtfycb1HmfLP6*&#O_rR_#t7ADWErj{g0b(MzwLw)jbdM;%3}Wy$(?TOp9#DA zVt$<{%3vf7VP`O zBLjUuI@tG7N_9HBn^K(&lY70--M^XXM1$k*4I_Ny1t16W)BAq%Q$qGdLs_P1@FL(t zI<-W=;SvG-^djK1JD`nt&aWqEPTKMwwvEw`Tp9trxT1uUkP#8^J#GU5{xM78#|pQO zD@-c<$Czg;+}UP_&nR4H)Q1$VG#FR7*5J*6m3=G%-k2}_LQnf0??%8<#E3>fKJqR` zzrXNui;azblNn3t5%9Vt%>k1$c+quQBIrezoBX<;nl`rzM8IlsLN@~X!xFqkiMHy&N+pn5*11-e*9v-|da3jf@wq zpzz7mJJR zWU#BkXAJHKSb3Ql@5)N)7mT+Szb9#Xsv~h{SdI7YLm=BM=l^Vfpv^Tt!>yR^7|sc+ z-%EVtB_Km38{Z$e+ufom6cg~%Gr<;7F~QEqY5W|DyH{cYa;XWfo{LKen6awHJaQyL zui8<^Y=t3K0|TwRL#Uk$eygy^=8hjK+}Eg!6m~E;OW{5SrvO%7ZzlLr)^UuVk%g`a zen*6<3Hm&UX1)D^U59XgpqGftPfy%-lp>z9zt}OE>d7YQapY2QOW9H)^mjxB+Cw{a zqph>HS9rWBV{?UF4X(y*Ae24B=6K&L>|)f96&_?TsqjdHvjHn_G;!C?k$$14vF@(8 zTe#w$ia)n+X+g zN6}V(aEN=Z7AUm=YJPQRan~u;8N--gBXYtZyY8t)?v`n8^zoLMn&LNh+;c_WaVu5! zNBNhWD|%PPGD6?EqB27L1-CFvj9*p9XEW5GT1KMOS9aZA zJWXgx)!y3i#}z&Uy*bS>kfZJ8e)2Ss)fiP~jM;v2T*yA{S7g$@^8n-*7;ZdE+!4jR0axF1=oMVZ#m@fN*cqlTD z55995((8P8Dk|Q887B(<2u;C^8ln74BJ@1$nB#d3hGe2mltftA7on#^M{EWQP)r9~ z4}NL0ESR8*S~~k8_Q2-OwnUY7w>y9RhQYa`tMKd8DBY;`k!&ox8r9`v0wuRw8u64U zPduv!Purw*Z-`q#Csgpf9N5B-7noCB1WibwNhzFoDwPvVa^h)UE_mh6OSN%_0^Ayx z*}~;}I-&}m7*9}`MjJ_xLMDVov9CE@W?u&9UNui10hb48#=5hub5TArm!F<%?lMYT z^WbipBuKSTMCad_&Q)^t-0og8PYGfW}=Jhnp-1{Xp$Io50YwBTt(A(L`;@lH=YmE zcyoaEgE$8mBzoYdr-vOVUOZ>tuBwOb>X zb0S}H0E_40IU*soauU@b!t5onyogY;>tFGu#NLcNdrJef@a(Yy$3aU}%F#%yOmzki zR5-!lZVD$E%vCtsU=|FBvQ-9udrII`gU4w@6$TFjOhy~fQx%kdN$UL)I9^VOPM3MY z&OeUTX*^>M>atXLERT6C-#zSOsrFcI_E=8A#&rA4{I>HcBc#!eiOS zUrT|(YJ3 z%N44Qlc{qQR){|Fe4WBFgU2W=w1!@|RcbCU_?W_cgJlZy431HlW$;>s4bJ$cu-4$k zm_tF{5^HzS6oJJCf0(508r+x5VlOLo(z4FTxs$>Qt9fgMlMJrCT*{Ui{7GS{!7mh+ z7<^aZAcJoxEH+r7u+ZS+3JVP0t1#bS1h6a|FIsYO_^=ioJ0WOt$4=QPE30lIK33)K z_P5@!cb+GnapC$6^;>$^<(VtE$s=jkOYC{Dv1!+f?fTHP>v?wl^t9^@wl#lr+VyI? zzIWR70;{K0+VwKKz8b#o^*{Di>aI43=XjEUfR>`TGS=X0))swD4`+(01w;!RjYAQT?K#K=ntU`nvtB209ihUUmtE z`nGQEj+rX%)kMWZgB5>)ZKS$aGuE(K6E*A?tl?qSaDK)bsxZ;dq`OqFfN*EAhHW#} zz~4L0RKwH38rrdjxotD);pir6xHwqDN7yf`8=A3(zv11P`mkrPhViUnmy9*cZlZ?T zo&i0a!Wx#g$t1;yCTe&zSVJyrxFKT=hc{8fxxpIVxf(Te%2>ln#78DQ>>R8i&Kj2Q zkV%SS%zWye&q(SIr#dnAc{njq66Tl%wR(q9P}c>+whc$>@jlM4QaATTFMgok_qU3I zdA9FBrh+9APCo`!90^z|FC}S9foK(f9wH$ZYq z?fE1~Ju;H`yrN(3Vyo@3!hjHY6k7L+1o*D6<;g(R;(rD7|eOo71(6nQ3Nk*}W3Tx3fYIWJ?8=b9+;A(kU`yJjqMLlZ?FXOU0w zh^S9{Cd=_n6uF8;CS)wKLlZ@gVv&9si)38|(q0)!-_sQvT4(CpQj!*DBuyvjk&LAK zNV+y7=?0R*8A*?ZGaKj(+(D?@B4d#&nkcdti?qe>x;EB!mnModV0l{i0bcObxX9#Z zGI#79tf}hm$ynstCW<`7A}41o(!Gfym$FFfj75f?khx=fQx=@W&LGPtQ-ejESj*=6 z&`YtWo`)K+L6M<`2hwUNHaC9bD5=4zFb_3kicmhU=)dk9D=$Uyx*80mjrHF#t%wus z{a9pjpol|Si6xxB6{nLsYbPtL&u^u%mDu%F-B9x`ftsBn6Vi&5S&=Fh`6QTijdV#9 zMG`FXbjBiwrWJ9tbT*3k%=GNJ5HrKbzv}6r_)hU%P~5K;N8XEAk1Y6f+$^@9A9Y>u zdY-L;_ri5_Bg#Nh{%V~qgVpB*uUFYB{qwZz6?T0guA{k024^ZPGx)T^ zQiBgDEHOAv;UI%IC@eO3Ibhv1+=XlGb<-yYJL%|Zd6$3*!k3_iAsK79vWXhXgEefy z8ro&7VgDv-I4xMi!i!PEV(b}uQaElIJ3A;vg<19*T(27&DB_ULAcM z-=M`-TLd#l7^uWx3#4B^o8x{zVao5nUS|Bg(k7mcb%q`()jAz}eq6A4wHa~p9q2G0 zdi6T&6q$ArNE3r3n+v^nO0YJkGnb;1kj0s0O-*BSYA)pRW&eyd4^J!N)Z98)^GRuJ zPR&1XQSyFHKpgAyE@?%anqNi{U!PYO2SsygzLPaypRwk7X+@lx&kEN3P#T+4^P#M{ zL&lmfO)KKmya5vuU!M!p*qoYcIB!`H7|I>fv~HrvWELsQSme`_Gi&Br7Qy4(J{NLY znbJg&?kuuj#v;R-DAI~W{tOP~X4xKTMI0@Cg87n9OFO5rIa+#&)2fLXYhH0uP-Lg( zTY@!LrLj3R4`9tbgEd>&{*2jXrf{Bo3yS1rEb>|tMXqI$`!WuNw>MFwJBrlzZyfSx zrtv#jBA5I-di;(aF6}QGE-|47C@eO3qCzKwbpxym1?NR(!CU%e)+ht=FMt%%d}xhUfEjrDzl9dK&CpEZxkSo2$HMVy*13fBBo8kY}!t63l14`H?bJr(QiRyPH^;f$oebqZd0rc$4rDpEAq(6&@zt-(183(Prp zRp-nz*jizO&C1qen+FQ1Hu$T;3WK!@%M8BHJt5x=u+UoUdYM#GU~oS`KU=6#ZjXUI&g*ghzy)Rf!?Ex2=wx0vHfl{cUWEU+@sDR5$|fLU z^a6gpgSS8B#70}>kK9?M{E4ebL-v~s==`jrNB0>AVy(NJ zt)hxN?8u5+=?_#k>w(=)a*X-UD^O>wTQZ<0R*+o|A zYyt1r?#k;$71?!tRr3?TC4KJ5KF5k?ha;nA<3`_d?K5C18<(dB`!v=1-_VeNk7LD&q6xBjT{KnRVo_9uw;c72)uGnZ?$~W|K}0kZ~rVRB}VGc;5pKS95vpt0eT!m<8{mwsh^ z5h%T4{l=oQPVjvCso(RL98%Q_r&;1;=szhOA1>P8Fq3@oQTb;U#pZ#I_`>-WZ2I&3 zokU5Z2!17}2U^nn@uITDrR)~{1r$8r+d(hG=4dyW%Erpxp-twpiLvI@GM^ShE#X9X zfJZ`hcrP`#m3Vxar-!ZQPyuz%;sa&I^^Fg~Z|m`UP$9p=@c~^F4=B*@;!gS<&gXZp zcwu{fhvOq54Dra8N-v1_Yaj2Ai^C`2!c6q(p>+aO%PtU!=x(Ic1I&4{};a z6c53l$}jr-2XnBn6CW5madL*0U9=>e=wAUsnY@b*@8`fRAGeaSQnoKdsbtMjQVDo* z9_4Va!h-X>>)*lW{FPgzy~{G=IM*>9nB}EQQu!R&eNtDx`wXwbZx8hsqDHKG$5p-= z4Tl@GCso2hEi#v%-f_0}r^|7+1Gdwnf4#7SBgfgurM@=iPPP_WdzzkK+mjrL&?+{9~tux3cDJ0xx#}DE>YOU;H!X@4K3Bg>*CTc9A~>D*FDa* zCGwsgXB#NfL?jM$;To*gFdmmd6^<@0Y6vI#je{*s-O1Fh5q3-q@cAcIT4YUj?p=Gx z2yd>a!{Tn(3&Pv4;uFOy%uhaHJ5Z5YEM7RMZ|t|&r|3XK5tThul+901+0ULT%6>yc zNT!w#5TWA2tIBReITN9~$cUWZ&moT!cUF*fWmhSj#D_qm~T+{woxxvIN#to z3g0r=7qIegQ}*3=NWY-$4%@oQeh3bz%5K0&nT%dc$*AIT{-yfsIzKmY4b+Fzkhm7U zs2VD+u+eBKB7$!aa~!;Qh;ccN`pF?|b%8PTO^kHIHD?Lhp*sH9mn?`zrxKaArqy%7 zab!aN7%$A@Jmyk%SG(pk+TsdfDs`Ap?i$WFbUf}$4@rzF4JQ)Jo#BMeCb7@}>T$+t z_?bI;#2NUez-RP0Gn^P+c1FJ#P9P&?ht{{!sJaUNgJB5?Qr#+1rhzJ2a=LPZhF(}} zw?scy{8XpV=v#kuS0#v4W#fD*8m+@JaYM3b9AwMF+}AUNq)fL#B95$1Yez9>#yDw1 zzLM#}C)8V${M(Q}F|t$N_+%L)pc$RDVIJem+J45CmR6Wdeb$EjtUQ@I`C#S#ydb6;@S|pGq)Qgq>tb73+u9R8gsaW|+d7C{a|+`pQuitttbR zMVpKz=j!+)Ol1XZPldd1fA(d*dP2fE$y-z!&c>r3_|hN0dqKQTMn2Ra-myP2Y~REs z)o-*%e5VGU&T1CK``sy1LM*%F*IO}CeU*B@J81)_I(_^)A2=?!f_;9U7}NCl@D%3X zhs&A5Vb|1O?%7Lv{VANrW>j~NbB~%;vU#?$4R$PWYQr;_Y2-T8?!PEao@Y&>^`59J z^(aFvvakdbeQ&Z(%@5!tJzdwBZ(;!i-6T^v5;a9jFuri^l?9#64j)ytBsXHqJ4o#; zrDt0Rzkdt`VWc^_yqX$ft&MmmgLknkDmYyiJ>0upQ_Le-u`Tb8rrTMGTG}{LghkRo zhFY?y%DP&`u3p#KnikKFb8#dSjK8@B1QGA0?IT51+4b*;)uNvb3sK?w4Z;ZFXg_fp zw7grI#fFG))v%H;ZqvA?g>@~D6;>yX?~#QA&hd24t%%sDb&*l)vU1C-a%WYc>@}S+ z3&#Z8fV*6?~Ooft$_0JR6uGTP}!ThI*a_2FFP5r;SE?;)nZwdf)?N8ke3_i-p}#?_dtVvAcsHBb+>eY(ycc97@^Aa_>Q(fbYCM`^-?0S!Ew4SvQr z6s;N_SoV4g5i~-TrcfBFELtO>0z+)87m}&FkUr}75|o^W4l+J8o?^>az7UQr3IAEm zJqORs(&^N&kJv0c#VUhElc__nB(RmY>lMmgl!<-&V}0z7T2|(6SF{H2R4rPwDtC12 z_uyuOb5Eb=`6(WBDq4ppjO3&B8=wcsj=-_u-}#fr8ri%`#ruEPBZlPM`ugx@S+C_u z+6dLR5X*_wsclPPTUf&%EfSAVU}Lpzq?Ic?XpN_6_v$ujSHbvawa4f9acuZ%iWJAx zXLamjc*8|ZU-ABHVw>@LKzZU6jS9>^zsUmP#LzY~3v`6diCEFPcu_5y8H({WhNo-b z;E}5K8WbaU#$epPHeU3d^yD&BoX+wQ(ljJ#+%0zYwcb!>CFbB+bnB(ptF@ZoZGmyH zQZ^&O`)DTK0y^#IX(o&peI6_NDPHu`yj0%cj#;IHV?{p-YKP<-N!=1FY5?rGVZ$w1 zv7*(HQ4PC|-eb(DhEjBDcy46W>a1Z~<-St%W9i$Y>(9_XZm#bKwv{(`Ko*Hq<<6_! zIyRq_t*SZASUoS*daLUC{lJpCIe$#Q)md4kS+iFnszoK z<^_xcn-}DyR^rA*KnG3Bg|0V1F1#-}2mPl2Nc^ZjyPDF*ihe=;zl7ucSbr;>cjTn* z4b~sec|W@aw(z`+>FogeuLe>aCgy9gTrQG07|A^-XEgu1zY2mP3T1%vI`5UqJqW#eh_j?L7P?l` z0=$R;&+CG-9X%5MHV#TF4aa0r>o*B+x+IDoqe)h8DJCh!*k6dDyb9{&->^3(!Pu0? z27ZsZ^5|ru|6_P|vTy9yzOk?S#_BO;1btFC=KLbaQm&P()^~T%TH|@)xNLlf<0b7; z^*C(Ppa|xVxOK(78Ucf#!!l5P+x3ABSn23ZtWr1s2i@%7fR!pbaDY|naiFFqo+KR@ zhw11@+wFJa&{5D3C};u~VVGD#QT+l-Fug5>PKhSC@TwL~ASrVNJe@MUx&_)u^q<~X zD;1trX26l10<*W*PQ_)qg?)>{0{DGtc;45W<0;E<90)toiGC=aOwGZ1j2aSfEJobs zG$j2wi`AQa?^6KposbVd*360A6ZR#fH7F1;6kWJB0WB-vxXXJ=XO=)T)+PNQR zRvhg-r0bw@u69Z~`StPar9OQe1YtI+j}uv;>f<2H#?$msKgy-%fttF)xuumsAHIEe z&gZc2F1tmsnT0a7E$2qFUq86D6V~l#_zz#4JBW)~+bp@Y1vM6xq8fA+9?V1B%?2k- z7R_hn_^CvESpY)hgj~gM+DOaT3iwI_(Ke8rSUi^QZ|xnI=e8%|D7)hElKgN&_rMbg zZg$b-!n{LfR{?LpI*s@5aO+dt-dPf(ci89a@rHIPdcKvOwmD0e6ZVb$(Km+A-}=VZ zYMv0>Y1E!5<=gLYwQ>bZdB%FZAjsyZrv-{z<(>VNySm!TYrUnP zLHoQ2p<%^;#Iii>-^s%SbrCWhJ6c#a*@vbSt*MK-ESv^(lh1?uOt%^6@&+yV_0$5i@~|(f?`yd6v%Y4It08 z1pL>UcBIwOWx{zHk1~qv!tQ4nU3#)1O;Vk%@rR(4`W|uTA6*<(wPq>Lg6HiQWOH=( zG>Tj0Z||KV-_;(GmS3)eP+lV7S|{`T#+jk`-(`mH&ta9CFOUFt z-2Go=h7U{+|US0*}^vv-3n*S^_ z+(S3_GQ+*PK4_GKo7c1>t!9SNXZtckHoKo8GjwJ{8k^f*=?h+sO~!Mk%&=`!O9T^;%HFX_=uCTON*q|38))BGlWaWQLvp z`2SsI__8}XVWISmOxGJ{h5`RBGrV;st8_ENx}E=3W;pQoP0I`&*>NXU+{|!`%mmc) zK0qj-A5LayLFv41_1{aN2;C_1s*&LloJZt3-_$zla!#%7=8AkXS)INtX z9Vm4GsNuBC@KiTm(@EhbWrmclCOXOFb7sDLmCKYPmdKMSs3;t(2~L0;uO_5~t|nxf zh~F7v9kO5tBGPEN|js{2*mFi>1KtqzQ`gIs9}D`-ol?XCX=okZlQ((Ie;Cu zy{4FN-LT_I@kMRL-2}!{xf-7+%Niac`6|=C)8eC`eaj0=IgM9K z#APiZO2sic0 z5|_$6K>cnGcC674r{bTuGVr=}3??#wCD!a+V6IC9*AjO$h(72fO&EZ)H#610Utjd0 z<9j{mt1B1ri%r+QqXEqoU2x~j-M!*|bf#SK?OKSX@wf!LU1_-@icd9YdX0)IA*1<-0ucBROMZLcelc{}?Z*OKu(|a>_7OSl)xb>8KDZxf|ul8g0s@vR) z{DzY%-sWDgQh&p?PpWT7^tZ9Sng1x879dkHP+%Qj?v!+xAya0~s&4+|iBrvw+;izk zv75@B{ympZf1)9sfMjfHI{q*6B=1j#El)n7D~U#tTJrDmWc(yn>E_ATxBFLl^5=R8 zy9s&ndvSyGJlP#Xy(7+e2AY>A7yikL-EOtTv<}dT@4;YBpFds5uDHsk-v8&EcqDu4 zwbq4w$uJ@{ujyb;+`FGICr&}!WXg%}aq6wXGvj=JXq7QKuR*CyIq^Z3@>KO>`yiX6 zvqMqbDnHX-xvSJuyw;EKGiXj+g=n_o?LiHv<;05*;kAwBMBkpv|NSFg^sy-spGt{; zn7jSA5&ztw(5pp!g{}}9Mf}Kr7x53uKE4}TOI!b|h+n&c-nQ|X(JG%qcr&AZyLsX~ zK8Sz2Goxjz0&ct)CTzYK90um}h%aGR{+oy|VsE{OKc0Qb5b@1xIvDYX;vSbf3z~_L z$rSNlaH6ddKOb9MX|tePP%2ZzPhlx9;@7kZvN;i7$T-!?FY#CIM*R6+>%II88u8g) z@qIuIr$zj=2lCqg<04)LV|N7=9OFeFO-1}j-}>l}@0#;GOhWz-)pACx%m87%b}or$N@a@s=`7{NeHP|wel{oW!;Ds~{7Qf2Zroq$wcgjypeEYN zD}E3+*V5wt#?HL9vA7Ss2S6Sz#64GfpMNJrPD|-P#C?zVQ@CwE4t0|+o`osTyV6l~ zORr&V)CF^J@Zdw{L>;*9iQ)89QjUuuED#q?ja|ZBMzyQkMPyediY^Y)?q&nfbPi8j7Wc=&#d|o9$dJ6Cp^j+&)N0k&2dgDnfh>d8WwWo zyz~*T$fGC+@8vvd7-f=-6qRPzA%^jn*g3ln`;QEe>!==7a`IOa(USD&I(klV9KO@f zD3-HGhSS=3{4`Gu$q{s)+(e&YbAc&gB-y;vujDXtkdj|+9 z__P0n$no{v-2~wFeZh}zhM0=2ESI53YPp1gsUdy`sHyKevN`QV;9UhQ=N+bTUd-`7 z2jnpnTOYJ^wHgS>I}c~aAou^<0`ej_dwM`_DEN;8a>UX9NkH=bbBOEzCLsT4D*+i_ zrUBV=R~nW9xrbblfP55Ds*e+`#-09+2V~`E|6M@dOPBWoa=b1EnheM%`B0n%}d^ZdU4FIGma~om(530@6L#-{fP3iJ~`{C{9Hq*i(KZ z%vYQwLZxkB4qV;_K8MRykY%bpbk)8OFM$i>zH;A5YiNLVm}-0v+jXly%ix@s69PmtKg3j?zxL z!s9+TNbenXcW~p@^_hkS{0t2wT@#M!SELFjjqt=lWPVFiQUtpjz6RX(oo6D*t&i)x z3Rlj=$?9`t6#%t7g}IkBn8V)|rJ|O#cCK`>D^>25e7o|3dqrLjgi?>VR~k4s!IeAQ zD>ZgyqNYzNzd#Hy85%1~YV&a#!u>MD?3W=ZaZ~^hX z9?mDe)x&wj*8*p>zgb8PB=%{D7q_7DV)6ys*>i+3d*(1v>{t;s%TvaSGj{0lM;9v#UH;Ohuk+~Q7eZGLk@BZ`bnzsi%U@dl zFpn<2Cv^2fDZhh97jG51`mNAcee8%{CjKpS^>m?s;L#-_gs!n6^p_oayljZfyBVy` zpKyV)67Q=02{o5Vyh&a1m&&rdg}UmSzNe)8*(OpY9$6^&ve*vhlbZ*-U~*(Rc{A21Vtp8f zEC>Ie$x`6Sl9NG}?R>I)&zMe^r3%vk$1V%8YravtJO(C5mIw!0m8B(RxrjFV-|ccS zrk7@yt{N!;yA+`b$1W#g71T(U)q2ZW+-FU@O_1dSFgdckf_0P0a`Xw1HGbUFI-gsXgy42Ko+?e~>G;Zf9&E}t;w5CSm z;7rCb%%uJjTSOV@n>q{m=jTnGg+zKbEj5%f?7-%GIt&}y8GCAyv{}}KBu4-4$mYwR z4wq@Ne$!?Fg04v($WVpE*&Mz0!{+-~fI}M1mCwSxlji98TQyIggjHE{<+F}&o<3AmBVuSrN2aSoHUxaij`Oz1fz zH-xkK&XiwmPv|Dxm)F;MY>D+}Lw5a#!jtwe_afXP@m3dnL#a-E=C+irqM2FpRlbJ% zNBaB+o(ADs66d$LwS&j1_sZ=o8>qRR`<3Q)?q8PMdElhn&X-i=cD}MIxAXAo+|Hx? z??sTuI^=misWegCa&{gBkvC>$SHH@??KsbG;VU>_nI9hWO;!Wj|2EwD1N`D>xa)_* zwntC1z#ER-x%V3UI!snQUV}M*%r_hGcewM1-rvz>!}2=H)kvykIJ-6+X_*zq&(q<_ z`|!L}i}1YK7Pu|J8wp7{8j+8>lF_|^3-A~8i5}SPd3)qUj z6q$6d92W1fp*j|AgSx_r=nnjX6O8YHHyY?gyza|b6Gv}`+X%uEpUKXEpcZhRK)jPX z1YRu%$}&rI>AeG^Q~e(9gvo7hb8_lcWxB23mnaB$D@fkj zv(fYl4u&aXGFAEq`pQRhbnbTzVK!zrc=v;GY`YYfx)!BI?Bp=5dsDAmA9zK|xS|!M zqgG@M+Z{d*Wfd){gi&BWJl}X2#Y(2`L$0dd1>_}Kx6G+e?x)1-L6qpwZPNB~BLmaN zMnibM8;dbKNzWxOW}knj4t61Nf6SclX(Q>Qr<65atsCRTGmD5a6ohfrUcbnu2o|k5~Ovhhfmc`eL5ysm) z4eX0V$h%qAUMK7f$@^lRAML%K^G>QRc3jIca5eS7I68pedlxOpB=ulS2OLkc4xIZo zbtKg7KurzJL)DRmo$OKQ#`_cIWvjntIn`-A+!CEwBX3idmvXhkWJ<1}`Ezv8OLs?i zjf|Ilc7~R!!}Hc`?)fODXS8h!mhdN`BBoH7m~wqX&qHav$pe>adFeg5L{GqFL?$S> z`l?>7)vMGO-;3eCt?5QAaXdd@aCCB7#yScadLZi4CeP<{7ci|XWgOGq!H3r(qiVDG z{w3{4JKQpZeQI%Epq{Eh+QNH<iP+WL!Q)_=HW0p!T_4YSj@vY<3--f^<*Vx-B}*`X+-r@6yG5SELob*+^22T=fI2ODu8Jbj6>{ z8Nv6)1bnp@^GZ{m^CyQD^PHQ~);oSZID4YftId#eeU^7$jiQN&W5K*{HbVraY1&pc zxb-3Cz;{;e!ShCJf%Y5fTVybe&Gf85L$d`M4Xu7hlwBAdY5%~8j>Lt0pqnBecL27G z{I`rvOdgCgWM+}}KJK|NrNsMH>t)Hb9g@_X45AL_)|kcnt%d7&pPzU4Z9g%vf5Rpx zH9}{=n=rzr&^WTH^r;X_I##|!whsK`>tLHf`|J~ZgB_}2ns%$=?}f6eQH`qF#jVPp zF#^6d04#prN6-f7Lvn(#vnjFhA?I!Bss2|vt3e}hJP;jxj`7p4Z5f(g`I3c2@vk2N z!8Q^b?fKYEQ?XI+EKZ=E^HDDZb;nY_oPpSwg>2yC%Mcsc%Oo^Boo3*f_YL01;)|cy znsInB&n*9a4E=S4TKtd=UIx->9jD4O^YIlc-&H^&4_|3PkUJ}9#M;gII}3Ql{ElH9 zy%5%9s;s-b0|{%Bz+@6XDiFN@zatvPuc`f`lkumKks@>1L{{#&bGXrpQzW3prCGJi zz1;3gMDtlBUe12tuG5w4ve8q#pH!Zpj24oJwSauJ9Uw%Bt0?CTiD;=1$_FW%aO)0NspIPBWOmrwR{n8r{U+?w7jq-+J3W^os;%+qFVE z*SGZwLXHEmP5yLdx$7iypl-N%2(Sj|8{kdE#=FV?$^O= zjL~RV2#~U~c3dg|+*#oylVcl&r3TmONjJykCNiQ?Xy!hsGKO_{v`U;&bY07B`DAnu zuE53UV@ixxkSDHRC_Df?!D_Q4c=b@p@zGjdO_b}G@~D44fo)N^dLgQnlBj3{_m4Su zjn3A(*fh=tDZ~O^>T=ohv6iZT(pUcc*U^!9LfOitR`>=BsaM^i=PaG(ikahZQb%gt z)lRdx(ReFn2kug;KWxPniBnxNwEzS9tMXh_%vqL9g2b`t>8LI;m!IDIz#sktR*mOO ze_8LnCwU_SZ_N{4UE{Slm^KZKCL=U4nH-7G5&Lc?@P36$je5JnxdulnywsR4Q~0`3 z&sO-7!9EJ-7(5!Va*{1hZ{aB%ctJd;^-JujoC&xC51HwifLf&D)KOT;;WZT^BtJbN zXZ9QQ`MSeL`W4!SM=j-A6StR^f7^ z4p;b&!NCeYHh3mra% z9P5&Sz82F9hki8Z$!ITpqHS!?gx$aT+Fxu5?<9?m1V*Tp+GWQpyULZ^$T4y+t(;`n z1tmlizSij0aS0AN6t*X3_=vAo5otEN?k9A?M_z!FKT2l*vhA0&@~r%?Z{{8}@5V#W zJ^nGgQzXhBXG1Vyz;}Ym!FjiNR0RL7i=%<&7?1@lBt>tU1=bJ8 zdrZ9SwQ_bV%&s*m+BGJGrts4pGcL zC8Oi;MT>f7;+5P;m0T+)2%6~!%!1;0)hu2!7{^ps_css4$k>>5j>nphNvmyz&SnCx z*Poq%)%s;DT8g5o%X7sda<&nDp}NkM@3Y|6c?nA4SWYC0h(y_S2j~voAohP#G^G`Z zrczHruEDvZtMCvOizgCUGN8%Otnb3r9Q+v{ zal3*!RGW$d$2j{iI(2^RYUwO8GZO+eGLBd5>Ma+9}B-;gp=F zB@Z#4K7o=~=c~R#Qc}+CqNq6k%EkwQd1*d!3{tj}Z_dNZLp#mmhSt42DVYep`I~Cx zTW%G(-q*p}c}F*p@Q@3dQG0H^uas+J9M3u&Ov}oV;^Zt5`g?cnYs}$bN>+}=%F)l- zz7!mbqep_A6`9MHvToTbGFQf-(H*=_&9(}=xa`@Up;Pm*s?0q&=PUR}?gM-gcTcb6 z3s!PHpR17x!-rrS95x5JotxZ=P!H=)4Y;&B{}+2-9$!U~wVeRLxF#+rjtX%>5x2NR z!8NGY2}%@4QJfLi(Qy|Q5FHu8fO5TF#RbJ3cg7VL6cHgTa^oJC(I6-p$E7i-al`1K ze9v=Eb$3;F-<#mf`+naa-~0w|s=B)B)Y;FerP04n8t)YC|8l%|ad+I32rx-9?@+D( z3+hKk;-KLc@P9HO{594F1s<3vb*limO^KIF2^G+|@rD3Gi{T`RlQcKFicDW~~w6{X7JUWmg`q0h?)fsA>2V8|KDOCb=lMIkB|6 ziH@2iD%p3GMWfW=b^6Nv>9RI#2}3FG7IBcK@RSFAs9D|NPNv#*s~ zt2lU_bXD9kE+q8)9?EmwtXb4~yZPRJZu}WOxPftJjT!kmEa_+QK^WUJ2lN&y~+Of3NB%f&4Z09jMaB$UNp5zl;DBg3M`}KxdCKm0 z!xo*2W)MQJ=1EQ^2wI=#N|%p-ocBc!Db|}l4&!o4ILN;5BP=}{4CJq4g#B!ZN7l~I zAU<3(Qa{GP`^|f9`)X6aC{kZx>gPu4Pcrp?M}6cQ_08tdwUJk*nEH{C`Z1>d1lFGr zPHmV$!h^4)e&?{>@BeQI!t5zG`|X)e^sREvTLT+|>}Ew~xUc^~ibR1%Msz+!nFphN612i*Mh}>Zdv>=*>eK=B+{G~9-MLT_zwukoZjv1CZ4~IY>fTfXF?zW3v0njowLI6Ya>w_ zS+HAE2=8QZM`Ce3WHupb&b#M1zAVzQf=;;dSde@*leZA}AJ&9aip<;t1KBf=V(tvw zFk?R_X8HpA09{k{-(szyo|{5(muiB%VvQ54^lB_$!^RTGEL!KhL4rMsuoO?`E;=8X zz62kS#E^|SLD#X51elwp>ui=(mw2UOY-P)5-Z|g&JiE{}h1mM?cJJV!SubjNIGPi$ zpy1nDDZZMxaEWx#8Wq*^VDqfJxzPl%z94v9%?np2N9O6Thtpr8_-^(rK?zn9*O8(t zzEPH=OOYK^sTdF;jx!bRT?xjraYxzoO>e-h?rL5{8-+enKk|Y4A#aHA zK;6ZCB!`1PlT{VnL$Vh3!`O6)^6M9zw*mWFWcc#7P1wQGIY36Kb|GFQ_lNsEUIy8< zd;CvqVYTk37qX8NS2?hS4{-g?S2aEtv=$}U!aqGhYzjk6LmC%wI<{~nFwh#`RZf=Kc&V@SfJKRg*^i zD2DclR|RGGDlGGORG9b^sw(`O#{gI_HnZG??#K|j)9;jyq}6SRAQ7K7`j8#UxB8_T zCFx!^?nSG^IgH`+9v1i1MqaO zx3qdqR{{N}uY)n=W8ymF*>KHtySO0WIw;zK>2C+7zv2MvzsAzXh5=4RP_b;RPZ`}? zJ}^_^cf1XuncLiHKz5xjsH+#0_>N3&9XvGXyB$i{>`)(=IjL0J4w2u1@}({T`yUMR z(cQq}R_s&{p%^;T$8Q3D*hvGNp9*u`Xeb%B{p8eEtl{dpj>I^fBtaL+{cIs>#))yx zdf4^dJq}nEp1$~*JFcS9V8?QVbC8}@wSi8>Lwiwmg%?dc{D6#Zf!x@GkO^Bqb!7e@ zbC;CwqQbgKH2fG83K+E+nD;@N317mFU1a{BG7Y3&>(A`xx6S{}rA9!5!M31qbL>ct z^ar#|n!g{zfF$3tvwk>hC5v&}pcior7PFfE-}Y*R6wl+QV^V$nHqZI5-aygYQmXB! z$bNG%vaps0k8ip*!BfB3sVXyI6YG!Cu5ooW9$JltJFd!&ZkyS!4@-`wT5D% zhKI~ZpxEwVv{^&(jZ@5t*q$jAm2X1P@$p;0TXDSokAayy-jUPpP|Vb7=Xml|65O}# zMyjAe1yet%Mr1@79+K2OGYBT^6qm-B$bTQ{?^r{ZL|_;`wk_1D|3jPd&}}(IvgFWk)Ld} zY><(&b$#{E?pO+~)IUKZJ#m*w^~RqST3d-f^<-SCFa9h?txZS$L?DkOQtcjoFA-XD z@ztFHkojPyew7LjkA(k&S~*`A9M$vwjoJa}T#dOQ8SR!y4ZfhFk8kTWiROz? zs9m~3YCW?7%e4JFV*%t>vZ*nvrgJs{Ee7h>rEo}5yO$4lVF6bKMtw}5OvK3tPHW7E49#Ud>bdoL1JA>&0aqX?F%g{G5U%z)hY*JuP~QlX7vI4I zdxw9(Zcb~m?g!7S3AcjXEUXFTWR!6B#}NE+9lK?3HDc41s{{#89h`nqjyO$oDW>{H@)ZKnHk{xsLGE6AUExv zKAcH(qvF!&`_MbNcY?$7F&H$Cv{+93xt>jt2LI|D3E? z-742_pq_>ab$nXkKlG7!q>)~kIxmxD+lYHg@mk@EL%65v2U#P3pmGF5LxPVTNpKRwcJd8~m1j#yv`DN@rP7vFJczf@D%Ex`tstq~@UTa1)%cE6&Zii^B=)b|6=*D|5MPOskZqVOr$9jJ67 zRF16bOwPJ^^MUlJ8gwTEBcwoXhuls9a(He0^$EN-44GcTYbh*66oMilBZ`cV#zC1< z#k|BCcc|V0H{iT+zD%+er5(@qen)UUp(B-9PJLY_fsSA`Ea{9>G%BwQ2|@{24C@qq z-{kfGC%EDoah!CA(ipnvE?2w`cfoMQ7dXb@z4&*|lA6%vif)X%3vTptm>u@fHJu?BF+11&9QM+u&PY2Z1lYWCl+& zk6Wg3rYMGQG)WacvxO-wd5lw{LyXaL-s@774E486#mP%K@AMadYcg^ekCb@-1UExw zYOp}=t83wQ&RU>+AJQnbc9@`2B1-u`0IVPsZYLR2kr^hUbfA3B-1Ij6C@CidJFciTA z=9wh%Q9JLHBeFXZY?LUwJ`A!TEbMzWh-PEd?kT44gmfKQA^Sn~Jm!(}`b9=C77Y`( zyei6Icd#$0to*Q{D8(RB9Cw9c%v#0n!pnEu4WQ_xTx5Xr%xVJkXJKX{8g^1e_MxX8 z&zfWR#cNFz64?f*HwNHuLD3mvv&n#RMM_)a0R{#ZfY-XQcisw&61?~03};n zl22za?r7?AUw@P42SVjo+Eg(fTI1hLfxT}9w1xt|a(U0Cz-CZ{M$`5E5)@b)TQ`y@ zaB2?qE(HSI7`#k**40AoQD{ttLpc=M37$)gLKA3&)ngmY9Ap6~5;R!yYWMIrIkswy z-sr!e_5^kA&!@o;`?pLYk>=>%p?5`~t2Ad~`|(GCU2VObO#hG{e9dR(a1#Wn;2c72 zn2zWJyQh5kYk!Z#sWhSnn7@pR%|#v+wgBJEMi1f|L8Nh?_46!Z=_k~pSO)2c1sDZS zRmTe~!1VgnvH;HtD6R!K1(Q;#_bov4Sk}k^g{rSOBqP*(7NB#I1z1n|)O@IG0lGjN z{MRhN#rG%D!{N4$8y`*8z@w;wR!H@L7g2dpozegIv179Knz6ryKz~ zSrJlueoD3EH(V|~=SG60ZnfIMweDt(aA{8(E}Ki?io#Ek^+mn{_6@z0t)!Y;$m^4>hZ}OPzQdC{`YeYQ~OkJs7zwv#d+%*cEnN6c8RCj^nB2zhVz4Fzi&VZTzn6c-}-%)*F8vH~3vN8XUY;CXI$jH_n}Y4}#ZR z-vX{Ltxu~3J!Z_QSv58-5w?u4A7Qp^Zme!vo7wld)|{*I++n=6?wL z0|>5cyxW`z@mTRxa{vDuJayL9$zwZmQ;06lD^mtqBhHhofqa_XlF2?F-32H()$zw% z8$}lZ5Ji-vge2saAn8uN$d2qp(obv-uv_FO zFpp&gjaF?+>^E^8YzBkHgpQ|3$3EWU)-_#QkJkV#)&lK$#LMACC7qKWW;yY{guUv*yzMaZ(nBIO>Wnnub;G2`xz62$9j`+>KdH;cf4XGFm5%;bGo2?qpd)Xw@8VXpxh zW&nZhWYbqMq;s?B=Rns0r^vVGyrQQ$q?1h_O`*s*+4QrdO0(&|K8MG`I2b+WeJw@F zrjNT#`h_6pyct}QAx3Q_cO#M&z)v)rZn<+lQJUK4jE!4mG~tZ3-pAiAXY6%Vz!~o@ z24_43tqXXdXZN<#~iXG0xaS@|Tzj^0SQO+UJaU51?5( zfirHsREBK(Z*s;z#-LA^Gd_NoRmvGF#&bxAGY(>c*yW5$sZ!49BG3oxERw(1lK{V) zD*Xy)?9Zf2lrt`Y94BxFn$ZT9>aTycS8nMJ^ed^=7qS4W0P9dUx*oUT;YJmP45-R^ z`?G6*Z@=2p27L6iG`#W~Pr(-2x(MMQIH~bsTnef+U)_hA9S{|OzoM7npiL9Lds=@X zKG3PHTcIM9f;vbB_Dz>q_IwOanDmBlFO%zM_>8Y*yPdj-H>!xaI=_h>U^3aUmMYi-OADZ#) zFqroT_yWEeKb%J>YxG)QLkijjvvupzV$shIUGi!#*kuMS@qGoH4#ZwTr;L;(QlwLti{YR4ut( zo!<$!5DqdrREIKnhAnd5CzS2byoQ_-j|QS%j_YGR^00v7yEB156p!uW5x*Z*nJ0v6 zpy*B92Sxj-=RuR=Fzbzn=Wvero*;Eohu&y|1I#yCbBS-nG4TIbO|U;1WIrR&_GxS! zNY_pqm`+U{nC6Fq3*MQqqk zb!YLCtl5=1y4>4hxW5;53Ljm&P&%4`rc+rykIbG>w2*68|S>-K}D%m%mNeQ6_gKn5aUu`9{dYjx|x4d-Kbb{&Q3|rpW zJGQ)QkLHNMBzaS_xW_Sz%U$nuh_XYf@hikW8EgYI){kboGCIv98n-t~Vc`$qx%$((9y7M5 zHC8ZylQ+u27fO-SkXb#A7*O#evpV`b0uCC=ylT^J1JmCOOn+fuK4i2rEwZps@2>h_ zT2G%GpH?FE`WB`h=u;|_BH&ZV$InZVfD1D|EJYGl=5f1^0j188D}Py4oBpXb9h`F& zq(1Q3)BD<{X@cnycy7w1|l-HK(q}S8Hm<|YZfBU z+Eb;_Q~+z91RC5E6P^o^Znh9ehz~r5p2g?J&w>2aoMn`Xn#pvX&n*>f-3M?hgZE}}GbvJ7D_LBJvkDxgzO3p8 z4{y+_>GW~@slltDgQG5;7)by-VzS%^b^|(4ZvxWl9Ayi%q;Pq8NC#z&3ouRS!a0Mu z_IEIiZF>H|^k?k!S!#HR(Igpn3)0?Fgr*ke@q6}prwLD+2>WAjzkoXO5B|x0Fx&gj z=3XL0$RjjWekwA|xeYxtzN9O2F-`*TT!)Jf+-}zQ1FX&Z(eSVILfmM02gIKy8g9!W z8h%x>W0*}Ox3~4Zc-4)D{n>kgw92f5tnb8$oRgE=TNl1}EVtKp9NKBLJCulbM}Y{D zbZr=<6&S*A&U^1EzLAjI8%=Hq!%`nz*8AwAT)W{(6sG|?4{MT?ks)#$@0X)HHEWw z4wFTL?&PI~U(al3H3UrT?1SGAqg$#6w1E_d>N*BGc)>mEzB!C2WY=A=%iSFT!y@=# zLv)$^T@)|8?5yDAYy2dPv$uCT;D#m!x8UGAl zs6+7fzg;Iso=PD@=i)ivEZ5HbXR~yq%R%-(A-+j4NUkrq<%-`@_A#4V^|t#041l$~ zC994OJ6%ho*&*cIe^Rt1885b|Is;aM|LbX#FoZ7Wb=PpD6$yI$3o>u(Q|&OT<*pF! z4b*oCq_ObbI|lq!r-zi?3@qHSx1L0Du_OG8%K}srM))_X0`l%N>m9>wly0^R{*0Vp zhw!s25^L^E3zF;rn`9;AYJLgSCD{pAB$DhZNRGWinxCgX9212ppIqH22}CtFSKZao zw_fa!&f&+}Y?1{YIrOQE`n)vz1Zf)=x$cW@l6BkDI!Lpjuq1Z3>UD(hMal^tK?$Z%#0$-s-7#efkEp$<7T-}oKkz29GiY=x*w-!146gKxo7o6PQzL6RP zbbbV}&g=NwjUf6B4I+qVUqA$rW@?D8awLNI=Ook@@a+W~{N}PPxR3r30fT-g$}u`F zv68*p3K$^gB}*zCPiL%T{NCm;0vG*<<^ zi0`eQ+R?WzdXfMJ7p4X}BI+G?)VpvEPA}{PN#tMq7IUM7JER4MGm4FCfm(@HV`v_E zV+Sd}1sn=jG3CSZMVAp@;7wq8je#9~Yl~|;hGp!zw;cjYle>(bS7Y3NP81~fMVCBX zJ1QC3+fnV(Fy^UVO z->%+!@}5BN&3+bo?|KB&NqTP?(;%y>_lE!Bx9hzL;Drw9y(dPwboz~+L=$C`q>p&}x$tLaJH@>JdqXT>5$m$l+C=I@UA@9IIhH)toAZ zM&3S`_{UNj&eQEdo~TUuxis;RokW=Ld$ViUFKD}dcg5ewQ=I0g?qcQ9>5Y&JQ161; z9)Zq`ku;xUT85sZb}pd?sDn9F`?nh4f*8uIbF0Bg2{d?iXzCFTWFlSd_<8{WiTS=WhM_v{`~EZS>Jgj% zu6yz$$Yvulf6K=t-=bdG$uD&Dtsk!=fcbr2#>wF|q8)O~+zDI>dgap$=ow0N8(0YC zHi(JP@WAT{RQ}vH^>L8T_+e^vt|T?s!Hwm3{?+6&b{C>J`HboV1r3c#P61*kR=N3% z-=S(E0ke0XcZ2S8CP*dUfqcfD%$Y_zYrxh5r!$#LEG#g$+&dij;{HTqj05+S5HuDR zoG1f@Ix0V3O-D@IF%CTJa3vdwx~ju!x*IH6Ar356T8Br+yPD(I;1+k4B5pidiC@E1 zwa9VXw%?hta^lQ~i9*BV!UVD&+)++^70Ce?Q|)o$_BY7f?abOCwNhns;_j?+Ik5s& zG3Lp*?viEhq73A&MD7l8i*>etw6k*Gt08&EASCDSj3*-e8uf5&3Oq8;p#ak^HR!Yf zqFT>IdK*f)ze#s`ERYR%-2ac3IqxN|Rct?>Us~kOkbd*@1F78RFoVy#fnWW$x3aBR zIeGPJDMDdq2U8h9M+lMUBNIU-jT;I49rEaBC>!@nHD9=a-`hR}1T`bq^J}(8O-mP) znxh$>RZJ8`&ZO8JmuTg0dnU=sN7mM&AG(u+{k7m1F1L*UAY9eK{>C#nsnJwb`}!C7 zoduvepj~v`XqA(CJfAUbT)Q+$pcm6F#WGMww9Aj!Y~ad_@>$$-9BPaLC--D*nI_*` zpM#AHX%~PFOOk)(GpbYa#_M18k50rDh2*?kZ^)5Dk$m%b;vBd8m_rprVxsYIK4rAi zbNlV#?plzv+x{;2S&{Y~_i^b{fQ+|srwN@8MpvD{26;F}TQ8|$>vAwSR63WZ3%1_* z$<=9He(UN$a>^mA(|X*Azwn#K0K0vtOCRib+a(WlyzMl%Z4e&^WDapDRVAx->Rg1| z8Uk$F@rcUqQ#eJr`BE$Ht1O}+|yVNasD>u4BP>K0>Y!&}rZhVO= z?exF4&@+Fm;p^hxmS?~7?hX>fTrQVo1e__dXc#fHd4j}L#u>*!r()z^$U%uvY%fq2>^qxXAw{Jy^R3ZA`~?`WZ20#sg_^Acl%o2 zi23Z31ru$vdC4~CQF{>wv3V>4mRjx45Pv4Z64)SLg3LgJtKEL{LJwi<_}wl3Ak<}v zzW3wUm=w>jwU6rur$p5s(Evj=qS+`AavsXUViTaER|^~4?e>2K8x@cLHf&6|XLYb~ z78YF+u@P{UzcerSWadtA8Iav~1Vlbxdy2Bd6Plg~q@Eesm|`wf>NzMK2Axk4EwBbJNGCYZXTSJdjOJ>nnk0s-&??NF}QNTHXJiR_LJ_ew0UJ6*tR-1aW8!&|R z2`F1}rU{q2N83u3nR_{+-c~j&4{Dp~OQW%)>HCjx2;&Qnk>I5#;*TTndpd_oEw)4(1q&nTg8R|rS9ImdCWn_ z9kCYCm==!81JDdewtdQCvyUYl5fh%rIT?jTYUd?u2ypa1sn+3}k=W#c9U-gyeWnt> zb*Z+>mc=d}^j*{XBbY6#=OnG{vq@eZ`wBeKd^$Ft#(eJ-XIy^_1J&efoTo`pv5u_+ zrO|;~2}7(U+KLNr&OeVrtoFr@QC2G?CD&Zu?%aHA#69#fWUp*${uV5>CprH{zSkZ@ zSu`HP4A%dZ*pl=0tz^EKartwKIdP%xK48lBdazT-ke zn@7Ria;WnXJst6GQf;zhv1Uu!O$SL<6YejsNjKoe$1XK9w_Nw_3jV+5$B%h2>{}i_ zR>!LwYqp=uHs*H75ZjRt&eAg&0)VIDrpbox@JSc%w9lmCqQ#?|QPEgICpDL9;uVyq zmwjE&3m9i$Ilkocn~IBAn1!eHazQ1pb{T#ufSyWaxYwV^m-D?jRF2FtH{q_tQEPlx zQ`V=xr5oe8x4x)rF0NT~9;@bkS2hHFAD5cN3mq}QkhwC_qSUO#J3We_K6-i4&{;Kg z=hhdu+>9oK;RH5bKlKcY{Li{aqg~l9bIWe5FMGD0pUv(Lh+oE&lD+Z7E7|3jRi~Hz ztv+pDf!8vm(9KyUw|3sC5kiS#a$t>`(b9=^O~4Cv*XO7IJI1@?8E8HTZpw3c5VMpV{Ps}?il$zhoqIL zZg8R!Jcq-N5DBA+LCz}pT+GGD>DbyRiGOS-EQ-kA@8%HDL0eo2+87_8s$js`$>ErHx-YhcY8;AKGeE%GJ8FvR5xy*UDL&4UhyZbK(Q}db15d0HYb|X(0-sxZ}@^o^d zDFClngW2rPQ<}Ekn1Qe%LWii{sI$&&xuFk7}w+W7c2pVI~u*&ymp9;Y%IOch^JBw43jKH4S%gTS*`LDQHP5{%bJ# zihB_7FK@nE&V#4zeD|y3lsNhB2J*Q{Zfp+UNpHw)Kou}8`R`+BNKQz}d)Ye}ZKMvLoH>wyQWs6Q$+ciFnq)F(dhIN;TQ4?528)k5TKEnBoQy za^NR?l9WIc1+ZQPM-DFhU)7UECJzA&mTlzT-Je;hNy0$VC1vBE&gvNs1P4piwtx{g z4Lw{xCXequ1KKm>9^`}}`)`j)Xq7mNTxtb0M0W6s$P`+UVTn%RQLG7txC!r;ISC}04C+sWMVf%! zF=K30HDTLd2G!Zr<0b_^32S30Vgt}%ziI<8uvCE!AG{5Jr*Zp-zV(E8YZVe}1GDEo z0R|aEZaonO3s-pz%$$a6;y#05hyMnqTl|qL&B%Ox!Hux&YZwVc1L+4p@sn#HhzFdAa@~THf~#@@ z147OLkJ~ndRbnQRrKHT^FKXQ+UO+(JDtKWo)n(}2;4VhF;jR;GZV;(ds6rU>$k@?G zJ8$}OOPd7jY%A_oFxz+^(y9opuobMe?3Q|{?GmqAwPLINi5na_%y~aT6fAcrW)Fb3 z2C-pOJo}^?hxb7Z(3A7Bp-6bsQ);Vq~ap(^s(1CU*bp)$9z{Hd(|8=x9w}m6}1$-!V;VHosLDUV+_hh|NTMr4lSgY0K>$4!pQNCXI zC#T;~+S=vo%V5qOh#b=070v2U4((Y9Uynj7o3GD6Rs&4s@HLK<<-DE{BAXLpauQ1~ z+$rhCLbDJEy28TP5Pa5)5CG@=8V?43@Tk=t_+Z!)2QR9DC}tV?U`)+^{6fu0?jRM( ztJ^?^3?64|=qBUQ)%?7S%2*fp2v8W5i@Dy~>H6RcHglWT=r!h;9q+U%%6b(QE&MvH z@HX)Ow!msv_RfWC?_xZpZ^;CR39-}Wd)IaKLs<60a&MNKP|nv09^c9hRI3Bj1ca`5 zpV-1PUyN7 z%9eaDL>|Xt!K4#SauJ>`z}7%tS*M!%QGJC@eIF_Eg0)8Y_Ihf_dx_3O?9wUWE%)v9K6?hij6g%S8XbLj75iHD?_M;Atp^r|{-6 zPEOmW@LKm4*dzB#8c)xb53%{D(6P0w!?#93PB0`- z5*8^5^a&hZ5V!ap1UNQN#C#XF#C@5ZvIW)gK2*i^e7r@+fV%RWB%sRKw_wZ%YX8d} zk@vNA$++hBIy6cC{&OYPYG^Qs?AER0hQ1!lt% zi@R(g*L|l=gm1$6#6rm>t2+6+-%(u-xNv4@2!4@wl&>iWAGHibV=0P;TnCn$p&WBkP0-<>9@e9QFe&5jwb$mG z(n+=EIIWCZiFXYwr-|rXedqH~uDVpBDc~@#;gCU3foXU6AW_^d>uM(Ah})jTLfRI$ z;gwErqJjO*R3RJ@TSVs^Nk$dew?S{U^MVh$+o$nyc1#`rB5Dk^b)4Kh_^Wo539?V45-bZH6y- zIer^N;>M0{_!xhbU3CTi3kuHwb2rFGP!SrUy(FF+aNed^A9iW%CGW2Z1wu)HBFRq7 zLf86_#m2q$Ma{0v4|&{>0_D8MjHS~`Z#9c{HM>H+5NsVQy-p*tMh(TwQ$3SLGxq`0@3(J)(pA?p>SsoOY2eDiomQP~2 z_4J^>Q7kw6<&ue{D{9^dYV^Au_#Q@3Q*YWl9Mq6)`PSs1hPhbQTp84qn&}P?YAQ_4 zDM3wNQ!^;28Dwhq4{Ao4nn#f?&xf}3!a2mQH{>j;U)&bIP_}=~+*P3Zu_pc2cQua; z1b9C}Kf9e&B)NvwyrQl^IBaw-QPn;+Wq--G`mNA6Pl|o2PD{HR?V$v{H@@x z%TFN}pyS7(GSIQDeVyk!9~e*+Y7XCjK{C^D%n+7O@XJB2XP~KBI7V>8#tX+hPcZaP z*%E2b#rEfb;Ci(5euUF#Bbd{vNm~f++9?9J=;s2FvhmPPSl2~lWN>;S-lZ7a-?M-z z-oCy2iov~W1NwrzFwc5=2&+Byn5?fQHG{q{YS04y5;HiVWmq|iMXJPl+C^RB`~3h^ zBBx%;oO)@?Ogx2~;ikp&Xogu4p4yj|!pdZQBi{G8)R~+LT!KnFj0)}4>SC6@k?;hC2Wdw(Ja^BO(N;vMRBS_N`Hakxfv3@?) zQOsvr#AH<3c3wAe(uW1&oxfR7QUfxX{c{^U7Mtu}ZO8N6PyA2+Z zRGEEZF;4Q1ZMQiRtZCth-IyrR_FAR)k}tmrVj;9D2c6rX1)-cp&E%@r;h(nRiv#6) zZB5wPDtmCTrbC{gmE_G!o065dEB1Trk1axKv8|Xw*`=kK4=g5jzzxK!-C@g~D>^-$ z!b-C}C@=5!%AOm%;&e>SAF@-+$AyLrbvVYv%%kN1B%gkQ+xV7%7DN-4waAb?f??0p%h zLZDn=Za7n34)NxXV|2mCFa!>Jsxnfizlj@w)C(l@w1RnBi-rpw6-%UJtFTUKA{{YE zgqfl>&TyHcbSlEByqVR@4F29?nJz7#6k?fxC6e&&4ELYTE|BB=Ll)3tLk@Mc*l{~J zTI^61LoK!v`!Q5mHkn}j1NtyAEf$+{f$@8RY_|ze85x5Hawr~9XjmvzfH{qD<8vn#O_y8G1^`^4->R*q<01*qlOs^R~!UInT}HjbB;VFh%svAF-U4U+=7@A_(V~L=KQ3lFHZt z*uy7?Lwnd0h~zHHmtbFEYn@$`gD88>W4OH0U6d8V8pNucd_-r!>_V<8Sn=$YUD=Zx z3E=_;?5}K=4J>;?Oza>cm`pgoJOZ}d&HMSd7AiZoVX_@S9q9kzGfD|7cr^2yAU}>Q*|F)7^Lb}G0DZn6;0LUSC$z@oXZ)`xU(tE|GPJ}I`ov01Nn^HBOeOvcItnGH`PIyNew3Oz{s*la*G7* zyaOfL%m{FR#~AjLGM4Bz=iOIDL@?HX-d#9HLhGK#u}(hOHSRk_4nM~h^zKv1(z&w^ z{_m3?Dk$e2)Vk94bMw;wu)c&j=gnUuAZ%L-e=7aE>-=)wyMwjal9hq^`R6ww{evmS z5SSoI@isnUlzqDk{8%uv{Peu=Kf(|e@L%**0GP*gx5B@X5C8YuYO@`MY{KTnR)qXt zg?}T$Q};(({6Z`Qa*%gM$3Y{ohxNJAr4redQl~Sbc@|55;SVa2hh7jg>j_7$LzaG9 z*cUbQV(0&>vo9y?pRVc{_hn^{xX*}7*E-kV>Eau`n&2ynE5zwRK2DrNq27085S)&9 zHGS~T|I}R9;atxIa2?^)T_Ngh|NLpeT-Q4ce=1!|eY$?Z`J>r|;_%5_ZvoSGXs-L? zo&Tx1mJ+DZZuFd&hf@?mw7>J;fmGgrRGL4DfZ=?>w>09V=XG=r{>#)F80&&9+mX1y zA-6_b`7;1il&Q}ffEH+}Kw^XUGbJxIck3i14K2p?~v> zkTzgQzonu4DNrY&f9FF{z}r9uhXfU7Zw0BHtML3Uosu8=l9-rp1NmVylYu@6=_CEf zv{ZL7)w5ux_3Nrq{iS#jSRX!$)%=XGRKI7cdqCPWUn|vfO!aeI?h~Gq>Zzu>QGT&- zV0xKo>f3$-U!=JlslAUf^{wK<@)U{IpJeK%ej)9r`t>!YzCvQ<)D*wI($qJ=Isy%) z0Z#Dg?0S6r6$A3y$%E0+h$+O)F2!uPpmJiK55E>5iXr0bWNY-)3!$Y&%m}HUW1h)0OOoAa)MbCn;iWnUBhId811 zpCfa82kA$)dg0hYef?6QttJ(0^&^~b_H};2R$GFj+3FGpu7K#^fHOo862E4wUk{%B zJ^7h?U3;^qIe+nm0lFvWO_O5qN%woCxB&M9mwOKcIW}RWE8$9mqbH-IseE&V2t@cr zM>k?EH?33{1I5zd_4L-jM{3oMhTYRhdV>uSxo zDA+Jlo*sM?Dnux5N5#RJz9@F*E7wS|QiNjSc4%9guAPX7A{OUMOqcU=%nw^11b-j8 zi!Yyumr+kKm?q;7AvhH%ij-8KSd;0Fq6o=7B2+o=r%zU{S~X&FX7Eja;A9WZ)J}wQ zp8(}PCdhO5U-b_NL`7c39zp6t2#|%;$kQ5^eIidEE~untxvUy__wdgg#*aL$fL6#4 z)yD;)+UWB_J)aIgHh<5(v1FS6X`e2Ip6=+1a%4v9eOzeV)ti>-fN0GY4H(L?1AF!t^^UIo=hOp$kmm6{WifA$j*U* zpTos%ZyrMi=9|sNl8=d#v|K?AyyU#MSQt9^}uv zOl0J2QC=oH7=R@AScC8B<3-K(Cvxzkn-NE=&%whtzcAGdCe=|U^@{|mh^IqW@0*i~+B0$-y_t^?)c!ZC$RXZBk@6h6x z=4n=o&8GM_Eea;hOru!PV2>`n z3%Ej}yIQxF$Kn}n)3lOzi&W|3oOd*J3q$XcDU%LCUAIMimrF|_~ zH1L+W2;_ZlU9eMBO8)tK;fH(QP^~r-cHNeep9i3vt;(?#;J%%YvsDXB@ft0DXo{n? z*l3F9X>nMMzI&n;>&??aT71?N`)hHIDei{i?B5wF89^Y@l#1Oq*yPvrm3@DYPO?>NvK8p` zAs=U}zB0u(wAj-W=V);QQ=F#7FHP|tEp{@+8?@MVu=ZT1#bWbxIEu5kHgtMU!fSGG z>Xm>_^T6Lqr|u&6L#O!@?2I=k-Q8XZ9|e(4?DGssEQvH3oN+Y;mlQ&YbrR zQA-Vj1_pu5FKZF|HSsv2C41S2a%AXV-VgZY3NV}ZAuSnUh?lc9-_?i5(IbZTYY z=#?P;G&apvZF`Uq|MQnsb*-ZDqbXad#6N%wk=d$s*a{Fe^KrJSx1rcuTHMM!ou|ca zrueien#T8QvBW%`q{V+9s6Ah)#V+RQ#VF3MG{m2IumA<(*E9sge+lBE#NQ0MM^L1* z+!yPkNHNQCQQsM_<|vtxVl&H4IT&Zzrn5Yb0{|BWmY>D2RGnsEVLlp9_cc$KXmJ-) zeBBgH00aaMz*8NWWCu@P4{0mkXb5Wck_>8Bd<;w`?Gw%4wa{ zrACn(8{k=_yrempj>c^tiMj(wY;)k>YA>e%rAvdcW}|@iqfXb0$UK9RwV=wVhn&@z z5ioNFpy}E-0!=xz4`wl|=eWkK?ruU&s29j?9d^vT-BFoN8dvE2X2rRfyU~h)by?r)PlOv3yMaakij4I#WknmKxEMTY&NHcoF6`wRTb&%b^1eiTJ{70WDED@ zo%wRQd)>JUid*i$Xt+d9a5^=^dFQ+yuY;L&u$*^2f>A)%DUH&}Ywj^S6UpH-Kxok`cUcQfz$ce|P-jIX*f1{^Dhwp-=$( zWW&@GGGwNlH{cbEsb>KdHWLexpWBw5h=SM`^bT1Zc+Pog!Ja(rq1<3eM2UBnU;v}m zEf{6HHPD@)u%Qo(V?LVNj}2gHBFejDZu1R-oy%2t08 zSQko%IBb_ndjr$|lsg}C-rWeT)2StOqnCv7=La}F3lse37wPrpJiT85&iNAMsafz; zAS(52g~vmgiO*KeHpRoWxN|>!dVm(UG{s6SZfc5KYO%}|H_+mCrmN0cTw#jeLr2cu zoji(B_aYE!I_EFF5IE-tuIW4HlNq}D2n@GZ2iha>Jm^;`l80}O7qo!>e;gnk@Ea|C zF=UbDap|#brP{RvH?WsLmZyfYOgnIY%nqE&=Ee?$PkoR+rJz`L;C|N*Y|EEbYG4OM zpCBHWiBbB%%gT};u`1B614yoHx50aY8?sfq?IqfEv8ue_l^x(yl8a5J1DPHI2bhc0m#`JwGrw6$~|vCLIS!gG6GVb9nLkAuR`lqAA$^6{-K04Z}*2*F}uwOgH;c zU{$hD0h4(zA{2w5@&?i{?mDVw?D8hcWctP4zjDXG^a}h$*&#drBzcJY?zn7gicHdF zs}3hhK>Z$_IdoNxDVDLA8PDGMblOXJFFn4R#f<)-_EK3x8(-1mpd`=gwP`uov>XQeUAX0Wrse5-w%77t({f)# zgn9ccCww1pha`S4e769c(sG(|V%L?|juuFh1d2=CfAo#9xpYI$x@am~|rL}6f_rigjEqi*yaMWas6r0$5Brv-v>C?fwi#}f06it z`u2xQ(UUAJV*iKPMVe-hOs{a`Z)(i*l8Hz!^oqgUX$UW5JlJ=0&5kg$nUam4)UJ2t z=>n$u@6c%Cyd4Qi)rD+@NtN+&wra>uS}f9Hvw8Xrw>f94-Z90GwRo{<{Eimqnx_l2 zxM&CMYPuGunx_w~G;#40K{W9_)-5#Wekd)k!PXlT7lYV&q()R` zvv6PAmDy3<6J+)xbRCMPaewzmcVVR&xF5En86S1_PLSD6;bI6-i88x<7-JR4>_jF? z(zW9vHl=R#xS|p3;e9;MdHVt7EuCpdQjca*kD}(Ua6&c8oWKUlvGYbmRQcmeY}W|o z^N0@!8Ce>=8Lrvq8`kB#cbMermzq{~vt(ZGCGxY~`t<>Q5I=%UXfp@O4DU+VG@pK& znei>q0_YdnCsle^kzy+MtZOvdypkTEX|fStVbUhZqoJjfNBhICXdTGC=0c{{jh-rb zw5c(<-gxFVLrc9&M^C^q+AVnAT~k=Fye^pYibDA7x_)$ncoVOw(`3R6I(Y-u*mB)uy|%X?Bn;=goT&O^UMbxp1Jk!hN)$T7*lTSgD?! zb3a)uR(g!vbIgcOV&ACY!#zO|?|NP9W@2Ml+;&K^-vb{u<9)81m2;3S%wG5^LQl=l zG<$WWG8|fCzYY%C;^+H$oGNZzEPGERFk#Yzbo~C7+VS0}4WjFL+Q*OEN+0vIkI~wPymgOxYtc*}x6<`1 zQbblzD1rGbB9PB44 zaklDD0EESZw74~=h~gev+}IR*Y4NY7@y1&0VxF#{#m!C6Z8SXDsuojRhT`l6MtScj zzAcsa%)bZ9doS>eD(|dVYbDS1i1UMu1qbgO&?u;D6g9(C8F&;q+cfqg;cnEN_uMQZ zr>SmqQ;3|WdI@s&eOQrmv=~QNwkAc+P0Yk)t0uD*dZ^MV+-xSetrokQ8EvA)pG@OY zEuLT+|0HUy>T~n-3oWkNMtfeY#W&2;*HE0j&>*Kw(t$+Iol^tk{0>Z{$Qedggp&n; ze@13FnhJcALxBE*SlEA#MB{BzflqFu^f!R~K>_*B%;_L4Zf)S&LyH@mVlOTJ)imB% zi(SmqHMF?7>A8(21oSt>Whl;W@aaz=(n9}-f+^DfHT0M8+gXz+@jM1Hy5S?PJFKX= zo+=->sd>+i7yWgoCTIj$9&Fmw``_?)&ij%()asDT@0Y5_8pF8yvT=74QryLfA7GPF zRgjK9Vm1bVO~KK-vKzRLHpDbQpJGy}OyRsY`kILM<0k z`&zdI6vX#vux-cjy%cg>5Z}Mz{D6Tr5#N7%i_c3QdVr|J`3|wURo8E&qBw}gEWmdk zX#{OjGRicc0=9PWJ%>vHKDQLacdr2QcJRG#IDMMw}v*ZXg?2pN43ZwUu+=8i*jM}MW9BkY- zxc&Z16loEIRNV$(LF(G2UT-LNG=0Q0eL2>26DT@u`U~Axf8?BZ_`B%U1@BdOz@ktz z`KWSuB@7C@ryx6b<>FWtuezmQ5QPW1+OOVmociGXL-4+Fo&F6g?Hxmdm(dZ&E+F>p z>>gw9vCFHu%9ws}#xIP!r%%IAgeG1Qa%SoA4R})bBL53x^*Tj2TUF8%6v>oacc1d@ zLpT^hO)hQKUVXwGQnt$Lp`;tn7TKyl8%mDU;xX(7B>amO2bkgzEgoeW576RX=IP#A zJi_$6gBE+3r<kdyxn$+_niLe6@b6V?|bi34)7m6Ed?ea>vvrfdixpJj_|)pG{v4{7l#^K`Nn zFEz!q7RQ;!mum40^Ym;jj@ev$K3RO9wWY-o#W#x&Ic4imBq@?+B5-%#_ zl(b1b{8dC0Fu4Z+FKckBnUY}wCGOwn9RX>up&3b1ckd9CY}GG@jsJ``%f_-Rm#hN~ z#Gk;+l0LGWSio)qP8$QIkO42zmCgs)bEf35yOni2CjoJ|0x`+}u@vqB)sjdA5b`FL zK$mTZvsFG&s3X1Vypp=M^~$cA0Pez9qWo{7W=lq*qfD1hSccDbJBytbH}1cx*?PP7 z1wTjU%STH&`90ufjBp%NQGza+YH_1n{~KlnZiS1sY&5l3C}lrD$}~ys_j_Zk-{N-q zJsH+_SueNWJ7fLs6YKZy^tuYb%l+%iUbiG4b4m zkSLEgm`AbF2DhL=O(Y2T7wyp41+=Pe7J>fe5Jdd3Ogat+(o;}sts|eNwJ2TgO>$KLobC(ylw2VEil>lT?%j@bdpWD0%2F1+he!T9QjdmYB zrNOYg4Zx+&zTy2q@Q@rpPK6+6c^o;*fgDU9zyUjQfpN62Z(vdE99Xz+S!%J5sl}NE zYrh@D0o25v(CUq>T1rZ9Jo31=)a<~dm}LfkqWiCrk1Rz6gj7sut38Lsp_ zU}7fz%x;^Ot$a~nrP;=GKc!AN|7GZBFuik=oDZLUf1W=-^bnsP`mL~P{mh}YM!pps zI|udUf`yLmqd$ngEV{EX?vJ8<0HerJs;VeqToP$*!H)oXatN)vQw(`Yn;V+=t>K}J zt?Q|lt}!lkHub(4YHunk^K|F?sZY4Y4o|VM5bnuRZ@^Qil$xCP`YLIRd<4!Xi|*Gu zY{FXvUd0wrP1nkJ$3wXfOfQ2AE8hUd{xgPq4EECpmQv|8g_jgLyF&p0_G@!ErSiY9 zi7Vs-0btl8!PA%U)T>~KSZIG&9M4(gj%;m190#qSwm^`%A9scu6k?k~7f>F$4s-Ab z>w&(9=Uj2U4v>_GYs>}eA8=;_2O_U2_5T8IJ&8})B2#WPycW7Uq4!w3D>rxA`J9w7 zr#-+<+k@J;UbU^WCC_H2*g=)iFFM!Y> zt^U4{oJzdE1Ec*_qCap?1U6_FOef85L7NcF?+&R$)!}y-NLqQ4C$&jeikN)Gy8XmYC|XJ8>XAoA=I{mI-L<-0Gk=@F!!dx-F z{B+<1Y(+ujx`$TfyiweAl14@i?v@NBQRZ$43zX2*GdXV_GJ`GCF&d6V`Sgg;b10RY z@MSl7S+Z+u33q27yWe2PV zG20iAHbM9sqpJitS%7d>3C1(d%B7m|Lo);$9;^~NZ@6x3X0~ATz?^GfgN;Zu%ZHy7pQ&Vy0|~{FU>%Fqpnx2G_NKt!)@3 zH$5`XT0r^%A3l?tt_6IMn%G*vlsB~@Rx`k;8*$2*D+MaH9x$HKT^-j0Zdk4LfOC>~QB*{a*uQq2DGeE}7#EX3>{ z(umEkFwGzSAcEPK={Pu9|9qI;C4jFT%pUTFPQ+j~1V4`1v$+rBH(+){8|rp3JK{bC zsAHHt_8P_P`OGC`tNO30m@RtG#_R_fpJ#iU=7SbRFnc~^xnr39$Gtv$?O=8}84(=N z0nEf3L??DI_i`~0Aot)pZlYlF=K$jzD|zckhCh!eGeaO|q96|bMSV1#VdvNaU1 zQ@QFPy5K(YsWc+THk#&b;Alq|Y+^%LP#27rSk77(%%=-(Vm>6qYav~*h->)20k0=> z8NI`Joj*kZ3OKd^ULQ&;UZ0>WVdEUuFQ{V&G2av5wU0C+UPqbcQ^3y-Ugse95kOcF zue}23+TquIU(soFK*wG{yONAoKN6tl8uz&2C14|thA!6-1@Os2ucySD0J%h+UNbi$)hNg?AVQ z{dJhh(BbgW5F5&gG?r0oCAiLGllp=yHlmd&Gv9f8m2pLU;(LI3EkNLLIJi?UiCtxJ z+C7YV_QRxXe1)D$tM$HB8^o79;v3>QUyHy6@*Oo3)J}Sov#`Gk7HRMa*McJrDt%Hi z?8@`GisP{W;o^ZLv*zNDW8x%pDFEVT5y>qccjJA#H6-?uN}y|ckgrKNea(-%a!FAArft*UyB zPS67W-Aw{tQ#8I?@(%~UQT#ydgfl7-Ecd$3_pxQPjZsvWTwLxoPz-(SOK6P}K9Cgh zw3{8d$dwX%tO;iDmf^-R1pS(8C>4DwD#LHv%WjK>yP)k7 z;X|kGPYB=;(`dina$M8zw0PUs?Y6Rw%(NB$=BMeF?Zb`#nrrdrmxGI2Q`gpccz6D{ z?7B1uscS19{zv?`)9@Wy{u-I;={ihTPPOF}AD)UhJiPA$uxrD9e#0N>a#Lr9$lJ(n zSnfA8=R2B1E{H_Dis1L3WPjZOjc{*+6pE<*FqRMzC`PLz+uL=gPZDcr#i@)q)GX5UHv*(lhNhxv+);@EZfZF({o@< zUTKbbP}9KD=#}Px*~Ye9p)q(e5INgO==Hawd!Dw)4)&{NQ)8R6D&Nw>klP^StvTkcwGq7y z%$_y}GpdI(KBor4?REW{ZRg7p(Nb@!cneswGOOH^i39W+Y|Oq&J%FhYv<4E_Y?5zg z0FxOE&^!4+3}wbdK+bt{u~0#Z=gY`q5;^YD#xzvKqT;`q)ehqq@UJG-IsR2wIqewaFn@%q3;nC(=!=7Q zLgRSxW2hWER+O{Jc8_|FIg3*4eaN?B|eX1;GvG>SZsm<0JbW zu>TAeCg8sJ6qIFrZBnE+Pb&()y1D>-d>><9it}2IOow`oQXn%CMf4=xV_Ol~u%MrN zkxDFtivz*K_H$ptA-5wMW0FL1#`o8upCO@}ToOnqs>#oA+j$@eH=#-bU*LP|S2UKQ z*o1Gnpk>w(M$&~A6@l~5Otq!4TP;W!k3w1@!uTcLcQ2z^WSPPExq_2ibt&yG|8`Wd zVvkre8y61+6)pUReY867RY0+Tr+f=Z`qEM&x>r(eh}}>&_R-q$MYNn~t_)Yn4dkY` z!BD)eor;c+ae@C^i8tO4*GVQ#kXHLE#{F)IK_0=HnYl9`pS$8d%m5(N#zg*>CqkDd zwG^JleGwBJ%0w5DOsfxoDqbtYOyh8DxD?6&J7hz6r;;`SRwQjQ8!Tw^c--^BE5Pf` zAE8g{d!En2T?C%dD?j1M^>q0`%X8}8$r4o<*Umk};TGC$g%wJR@A6}5dpJG?uyh;! zB2Lxt^B2IHDT8Nn-osO4b1n~fnQ=UfSYH}3xj(}=K)&?k%(bSu>h2;@^oaTk>@jDEegHwA=u6PT<5gD zP>{3Fc`pM+NnfS{{*>83`E>$Qb>@?DGhHxe8#c?kH37_rGX<0mwmt$$JOJLo)>}qe z`_L*BP$uzyO76TUup}UbiZD=c`sD>MhUd@HK7%^Jy@w?*F0TGG6fQaY!uSYAawa}y zBnLfy|F^;;?tuYhMEZZE(v2d{frk}FNaDa_fBl_!!T7~yyb9tW1M02OaE6YQ0=Am+Pk zCp8cVpwSK~GVm%~OjcMD3Q5KN8WcLXi{2`#Nh~-7-(5TJ0wCMO!*Io*pq-UylL?7` zcg@9~eNW}Si>(;3Cq2*vtSPPCdnzv)D+CRAj@$^_CEGV$ZSClSf#8V*j*nZt@38I1 zmJP%030yAkYB+b-mGY8vOJ=8Gz}f%5B@?$Ni`STNOXesr;A9YQB4WG=ws7b*iFkzj z_>tiNE29bY`9H~@6p}{6jC)dSa;d3e!{pSv9P*#W)(=A^f}5w{Z7F~7iInRRCgLCJ z6|Q$?ke|!OE#`)OsmA@7Q0#bh242O#YTT~M`$Ty_TD~Tc?e@p7*|_Q){9vRhOVC`J z%3b0D>+#fNfK{p?JL0zdIF1cv3(}2p2J=BGoov-U>@HI>;XFzi`g_iLm%)R){4*FI zZMR`tH2SMFVyGqKEvH_Hj(3$Aj~y)ap4Se3`bs<4+aAS-a1z_Fo087w5*QhI3^8fO z8J9FpPZ@_=9_W$)Nzqo^z)TGn5X;6rsZ+g%qvPDK^q%vkKWl;QzNuq(QJ~K-K#!O9 z!K_!%zl+ZLdc>sSaLBCZk3lJ0)!B4#usxtZ!MoW;X1(q?!K|-FOzF@1+N5!|mvQpV z`bOcGO=Z^I`h1C5;h=w;jBz*DqUJ$zL`x;A(!+5k(Xf=A(+WV(oKNCPk{~~T8<3xa~E>vzWHK-Pvmd- zZTQ>0zG2H1tkP|~RTgX9%D!^m$Aq7Eu#`Q~DEPy@2?7-dl|C!jXL?AbJhd*IC}q9a zRAiTOUK-hpiHew2^xiG6U=D$a$|yKQno{w>s|2z;R7Uz z_gs!$K7k~|rB_f)W87P>2BycKfgHE3sbp8PxxNY*@h{xrC7IF|v|#95k9}VKUz99F zHgz{aJ?B$EksW+DQcVNXU*qb_f$0_qDcXW=INy&{DejYO6i_FiA=Fqcz>$Pa<)wUA z2|Ny&2mCF4*2vxkdndDlB6JLWiuNIj%!_4wQBl;z1By;ms!qT;6k-GCQ2GzGY^(0` z9=m+9VCqG^ylc8%jWmfE!u?Ap;RpB{l87&DBlu!FAc^>5+a!ElN4WQ;_RbEzt{}eF zcvFx8rXjxM4k&}NUq;F6;2=XRu}QalYuaOgzfO_A1xdSr^MF(tnE44ZUQ_5y#j*M; zq{_ysy!kxdG+2Ec>$TpRqEK}Fu?vVigV?~Ai6s4PMpJ*0@lX>uq1jApTTUzD=-ms0 zMnQ4BooaA2RD^E%i}p(wQjWdXLW+dL-k=7^%C-c&CAd2K8&;tuaCOApiWA4d?4#p0 zh2t`BICl9I>|hcq*<=w7};*~ zKj%-?L{HT0n*Sgo>`u$Z9m_+AxzseY=dXB;_^p?++-QaTESINqsYcVL$^15(Hlk^2 znYqlRNW>yhCi}=Z7oF%st@#S5JF%cWE<^J)J(p<)ov&RZz{w2t*ooiv55_`;7GY35 z+Sa9TF&DnhsPo}9K!b~wHt0;K;5xCd+8BufIRDaFf#n3+38 zs$Ei6T!7KisWCvu7!xeQb4*>`|_s-j|kt_8^$+OnzBZy#cXuK zTaueK2R*)}?FMq;CuIlIp7?C_6)nXea%-SQq(qz{;B@khD9g5 zj(7T{rs~ZT5muOLP`x3Qx5El^slshIV7`?szmEZtA`o>cT-Q_=zEWBvlM=Z94|{JP zUsZ7}?jKN;T9nkHrAqBlfd)kk2ow-82*IEMBQJ^qhL9H$2}wv!cu`aow21+mYtd4r zO0C+|QcW$|)S^bEB(>D0_fm^WB}J|t4YjGIN)^cOS!-s$opW*k+wbqM>!uw8XrdFg^o>=HzL4Oa5#`P+{zx%5Gjbiv776ViJaB)e@a{WF{?LNbI z=Z|FwGo!|9e~BBNdqlq~(%-&yBW4nOlabZBG;gNdGRwFz#v86mc+jndeoGgcaVUhG^ou7jSdQccicQDv;{u}w?k`xR1tFtde{nc4Md zGd!Kjy!wvT`yQL_@h9Y_jFsKqUo1*rJL0G2da1pBmt0r9*d*q-kLABqULYvJ<_6T2XW1{Ui&*feuRfK?gQ-P-Z8BNmv8srF*JT-1u32hmV26ehESNgyCrH4wfkhQ0MVYgsJDFsmXtG1KJ5u%hcB}P9&l=7oT^`1Io zh75ogYiH0X#j|)>K+vwd%r>nzL=M!6z0*4wBHep8iQS_{`wLPjLb~rgLnc0RG;bDl$m-?ApX06Pw`{CdiX(cCkyv zE0OQvy~IT`r^{Sxf(!HOcrTjRLT5B<^_#D5O!BB+@(%)ckc)j5tntUo^4v48C99^* zJP7>c?{1i4KB4TXtXW1j3GCkgZ$4#&E_A4HW4nMeKA*=jC~%_}T=-gWqiw-wHL~)@ znfZ5+%r!@8LFe8N#83k*7?mh*6|tKJ7jE`maPdv`^4wj zTJva*F8Om@oZwBS;)0Ex(mpojSW?cR6%Fpm>jQ2_L_yE@9Q5)g_E=;TEgWbXB%V+DCJe-*S<4h6ZQ;CJw~jeVIf> zYsHH;N8f#&zipv!O1sEA)-2WNtvd5tEsri-K`>AC1sB(tPg34%v=MznwekI3 zs*NE8UX;OfK^DlwjY#>L?mHR;hO9S#tW<7f)}woWP>-hE4ei5{CC{q8PSag|zMc*nMXR$tbDpgN&C>gJ{QB+H%9-H14)+^N4y+^Wkx?}eu~TYFSDy~q@F1TK zxm0}P+S7XSH-l)6pti@fYh_Gsj^6P-l;|x=usOAhUoF%pJw0Y$Bbm6n)DrHMzz;~y z)zWOWJ>G~(;gu!spy!J{vEb-fnPJo8?Jvv|PF}C|?2#v_MNd73>R%d!ob?COkoLnV zVVNz(>pChzR4$iqfbQp;Ao8p!I=9|GBrO^R2O{a)o5v@apkLp$@*j$Rwm=K`2a=&yJ(fTC4pV6|DoS5{JyJZ zX6$Etu;Dg5DqTwb`yre*PqoLIQ~aKkMBDm}SzSu%;Hf)VD+e8M8^<9KA?7oe7I) zy{JkFtiFBNq3)wU`Dz#sx>3M>+mn-G>=}aWKTZG@#qXEWO?3X zuDAYNf=sQh9&Q#}68*DOu8H-D*f^Qn@TFxZzq7PhX(cMCk1b)y*68Z{xmmAB^^Z~w z+wXQ0?KJL2b9C+AlvR_R@_siHZ)xM4F7=dp-7lNwiXHK8V_n01SwQdF7;MoW%$BR$ zxLIrbG!eq&8u@&V(Z^)^{ho)kZz z6X`uxbWlDqr&f zZ@z+fzJ4=HgM8N$jz)#Peskk0$B4Dxe8!03^_xXlyJxXwm`;*@lP=P9)o*@ygRLMP zY$@}WBlnxX&obG?_M5Ltb-L0|*Zn3K9XNr64o$z9uQ|qiOYGko%h!oJPB4~l$Af!; z^)Zd*f4@n)pF5WSWjRH3J(l15T3{?ccVic0`IuOV;>Pl#Z)-tR#~I7-XN2_c7?p@I zRf2oQ={ECfk3^036(2MHpDKM&gLw>`Ni*Jbck~isY3^$yZ_;XGjCmMA&vK@JdcHYY zb&Vt@A-|W_!k-B?+9RDWe(=SC9q#*4@$X=3qAucBUplS2UGsQT$`@70u>_>vD{;ML zxk4J<_%h#Y4gG{&k@S@|sc&=i&#y?ma3q9bs$LyqF26PqDv5TBNL`P1Hzs_KD5IWf z*zVz;Jz#i@mC66xiA9i^L0-H=)<^$1ZH3#x50d^yEG^oV3AzeY*P zd|zLRm4Q&o3gi)QftooQ0=so%=6hc zCox>@6bewu`raXlMLCY|l|S2L3c~3|?~PJwtv}dJm3Y^SYrb7PJ)(dAF4?TwS`>T> zFr28_?qDBe(m#ulm}3!>=UgBewg9-|KC$?OrO2IcD4x6*%qwU4{rw z$F=+KiGQ>j$=dx18DL}DeMyldbld$G3ptEy_r?DuH(_P|b{`u|d5fluZTD|FDU-MR zO7#67X!n2JM>a9-Zs*lrm=-gqruUo4F*K7?^Enhc!E;lj+7erEO=F9wwbM4~ta_SG zZ05;aVxCNoG3x=hYFZiTxxn-AR>{h&pSS_+8HF|%ZWt$PZV6t#W&#~EKf?u?aY4LP z;C<5vrUsmq^>mb<;L|1nN1scmC7md+XtQZuV0?7fC+tENU3`!}X#ajKXTN)_kT96jeDs~mZ4N@tfXa{4VZ`HMQgZOiXf&sT++yyTO+ z#%t`kg~i1~%qs0r75$P-V$2wImyG`LKeLgFh^iPfM4C3c@oGutw%Kewlr}?3XKneK zrJ7&hldN$gN+6_J&FJqNe-aGPrt~=yvw5XY(ZiV1cWyXx>4`rtGj4WVYkH;?zity} zTMykDsKwFSRu9Z5&bG2y=kdi}|3TYYf}DQa()D0nFACV!Q;%^Qebck8zvk$fSlfE{ zd2H(k3=)oQnf=9Tx97H+RsET*S22N!v$Ct&Ba=FV%s{0-z6WMj{iAnURo4 z(m1GYb&{F9cgB~2eVR%*HpRLoHML8{qdBY1+sp%ipEO8ueP!*}Z!%&8nW(k}4@_J& zv(di*YMvT$iT$s1Ok)#z__ZH#IEe3n$foZmXP13MUaf1C+b?kD()|5*nJcI{Y0^D< zh?g|)K_IZksBei|M-TjforzHO=4a&Q(HU(Or;@Xt`T2LunI-bujC$q=oN1Q@A3{j~ z&z$#&e&rtSH9u(O%oU`*eO-61jmYn-ES9qSY8D8bt%pgzb8N9SUVRP(&RHErfk7(7}Kxc59*!8iAIw=}zj((AOp)0d(<7P3v&)>#7OC`Bjy6Sb% zWDb+G@nn{joO=&t29kcLhjDHEi3~__ZTynl9BAXKL=et+ZF~u561VY1Yq<9)+W3pP zQnuU12PI-g6scdrZ`pjlHhxTX5^UrDstvSpte3Rbaq|QOs)?OZ{&RD{2Z}WmVZHUq zXOyG(LWDb`{7963b6{qn5=xt#eTG9mUj7>;z0U14rDk*`NjxiTu}+))$Pnf074iVsjE*>y(A6>eTvV35@i#seSA?#5m@A@ixf^I))PwnI%oK zW_z@fKH*>|QRjx#8{Zeko+O!tO;`hagR4MfjN*lraoY9M4tBIy%P1|)ru zLthZIhC@%%QrSm~;=E6ZY><`Z1fLQ~bH|#lzvUNZ0Q=l+_>Qc$CTznmT+NmLL>rd> zAfTBaydWjIZFnz-o@VwJI^sBOcuqEF5@@EYHr#`if8I8{>ym(^OF8sOTEw9z>3K5B z2imYmf|dl59!(o=r^5u=&{w^LYmVp3f{!jVr+@=IJ4jk-j6RGhabHyjVH*9_1QqL+ zm|6rEB45MjlU*Q+IEryYeqzY=J}3QHaz5gU)DIAYGqpv{7CUReJom z@jKQOj1?zYiJT=NY|OZ^SAB-NA~itf!ExipZt-TZD^gDcQ?^S#A|ME^L_X%EOg?UW z_>?66w^t$;$a>u|jTj<{FklQ+%Fury9!f0E2ss5}aoB13sEXDBab zqUj0iD%pNX_E-6mD$(~3eRIlfXo;5Y@D|&Qhr?~?B&(3MG+z{TRjs%bUk97 zLOA8iC(bvCe_}qXZK(0YAzW`DLC`(Yz+ZeYx#;b|WrNQx`agIY7my@qNVRdIKQcZh zfnRj$m}07FfCJ%6GXAfwyhGn3r5u(eWhIV5VXa=N#OVY#1`*d zD5b^Q;-7UmbH@)Id89RCz7Z+9Q(Lj$hqQBDwYBGN_UkQC;KkNxr`8v+!sZJT&yoc< zFU07rht1sF%F>G&!lPp(R*xO6k5EkEh(Iyda_B22Qx9W``MC^(vBeyN7tr_NV_m>0 zH9; zApXm&w|Bg&tSj)FWlo< za(rR6J?oAhC_x|OA_teSVc zH!7rJu5b-AthlS1RQxt(ZCZ6(z0hs$u~t7UOFUL_TjiR0Sn~#Jkg}XgsDp zJ^BuEwcX^B-9KK}_e`OB;W)C-U_`;X1 z_54j^A2ESjz9C{_0)L8r8gYJjG<&5KYiga(Oym8x&6@`;UK9Ur8n?(gqHW?s#vbKZ zE%<9VBzs+(_>#NrF?Ie6*Xp<^!IAyNs2eeaaZ64R~CrZjtj1(r{C zW}e;3OIB%e*`8mUJu7pqIb+n<{(VHOxpS279;=(&3*`!{=(-?o?l5MybgnV8heh<- z`~3OacuI|fm9mnha z72>zwHLY4;X1ZkT)HanrO~!h$T`aCyzfOR*P#5po^9%NpdXPSOKHq$~pvJrsSi9C} zQ;O$1D=a&X*H#xX!|H#$a&%2n3z;OVJknJ1s9N_dnpftmP{zZpow?Gtx9H0ci_rIF zYc;5+H5+m8MiRbI^sVn0iIK%fEJ1^LA+UC>zPu|k)|zjd@HLn5y6aj*ty07Hi;8S2 z>Vu^ss=hp!Sz})ObG{B^UZ8UnnTet5_b_NE^P({6e;@trP?4M0hHB@^Zkc}0Eaes4 z0585tZ}R^ppBd}Q>6Fz>jaK@@74sBnrF_I{^=z|wxH@xfoeov{1qPM(t*Ig}uQN%R z?og4TA(_)R zl76E|FP$Qb(y~ax12!&0Su2p+V%4yNc{vH|MUuWMi7d~eQ?2*F3ek4XbY1;0y41xg*LE>H%t&0%*S8%obGzp?zSz&m13G- zTPs`E0iBPt$eVrV%R^0Ru_^qMPy4W)kJshAZulXSN}G{1I~JlDeMKY6U7&S) zRGn?%vCL12Xztv>D;kK!>E`Hq4H13ry2<#xxSmkwE*k!GxRz*ou^sFGVjc!kTdlEE z;;Zyg`ta#~;XW?V3xxn-mQMVHFIsh?#}f`M5*F4_$|fQS5amf31=aS z$*f9l))&d1PnUuRxn(ol~ zZYAXBF!kK~oXf1`H3xhAh90MCs zQ%04Z3s%w=f4yD`Y5(xrl|r=M{jTiNnZ6|5Godq;UPU+6II03OVBG(G^fK1G+PlSh z#?Qnho7YR7y#aKtF5mNfhR=FLudOA`+S9`FiI24-9ud^`xbbVo0X~qUD?CnSUBBn) zJid~tHBI&M8Ses<=r>g>J= z^V4LSq{H-Y`E+EAT(v#kIwROlzWKgsqaMFZaz{+t54=U%o_R}8J7C?1kLsf3#a8Yo z{tUEtY<>*)$*xagLYnOQ{mjghUB3o5lkED(^tEKy-z`ag{ThPhWa%IOQS$2_pPl^r zKo*md<^OSR^6T%i9VuD*(;4cLU4MLT^6Mwa;4F5yHaktJYeL$2eiriX)Q%xrMJ{CrjT`#%Q7DZlZ z-%WmRhRN;-81LuE9mot7 zX-0n-;AM2*IL*j$;;~x#7%e$v-YubyQ945&`}uB9F2fbfpp(h?0O=hWO|)Na9u%pk z-}m+8sY_&N5^0bmMJ7!?$(mLsPn*vD$t3I0*^A6W$lk1frFo@Y#)?(4)aM+*a{hZo z^7PZ#;o{WnE}j)Se?P&ZXH2FPdcr@Pxu)RPBwAk@U zdI_(U#>WoMQ$ih4RrKGiNjla%hp+be3upWnBhY7Fs1RZ0hp>Dy_`QmBu;%DLX#TE7 zG;_z<%-9m$5ijv~*TvqMhEjK;)Wq-YyYCRZKwU5g8ZTfB&G#oZhTZx|z^}t}%4wmT zra0>&)|H{|To7!# zgijjgMkxp7@ciJSz+C~C|usm}UU#JcGG_z#CA38!?95*b` z7JDc8+hUt+E>F}JX-8-0@5+Oscst*129h|Z_sP7(nzJn#t(42IPdqjcpzVaBya#Px zHLEQ#Z|eqZXVTe$5*=@1+Pp&++H3re?wWZhAT4op-sAnTfy}$C&V4vEf8X~cH2!LF zz_rsbIK<#Ptd87Pk4^M9zlAtSyjSPl>5sr9 zzaVkhzPw}ea?M*uiSIDvBD^a}Ox^b18J~YZQq&VT_C3?q2a1s^sk07%gT5!mD`5u1 zqn7faneT^I39~OMl zk;*^0aLh7MZxc&*zZ&F^(udPoV)bih_qd}JQ}rXf-<>de`iVMYcNc1w467_1xeuoK z<_mA>B^ z*NH*6Et6okqDFLx!21Harmpdud~Y0~X?N;FDBb&6=~dBcR?V<`|GQpHLbB0rBie0< z)2=?^KnpV)OlGRnyTq>abu5<9YNyyss)2Q=tvs1qJ+z>C6OPNMsyTXfsuU@nrb!O< z_eJFSptQB~0^m=pR4nvo8e$Q{iZM^eAlmQHJ7_%_pXHy-1gA&}ZfTBQyqaa|i$xeY zT04Nd&F2)H3?5+49nC2B-Yvb3iL>1b^BRe1>`ea?S*nxnR41;rnXzcO+8`TcCB`tDs2#z_zPdsYj#f*Pi!>W!wxmmx6pt#ir)KYB29`g`iXH8i?0dT zJ#vSUJ8mA2n{mgdTSdLjXB$uK9LwTomgoLfCq7eb~r)zc}XU9**p^Z?XHzeb|<1 zdS@eI)|;J2(scaE1jhHRke5t5(FL^|lDjt0O-RpettN^4&4>wod@ZF$VV-0?DC0?u zL}&E9uD%s=znEmq4*h!xAa$deak=kEhmG5z|7Z1QxL5x-BW49t{#a9nurnIQ{Z_~i zos?R0ynXGoMBfTII9Q4m<@{UU3i-8}EEoB|6>{{Rrgh&0{sx(~AFubvya{}=Y^4akeDCyq ze@8Rlkd4Ny@uh3v6TfEIcy)sQ9~EDM2sDqx#Ks3myofIBvDUDbFg!S~3McY<8o_1v zMKug=n(%3f=o7@G=fz5sO@56?6F=7L7|$im?=l%`Tyru*-VJ7O-$}8w3hO7{==~2d zTGLoGpytpfrzHHTCDzeFQ^$@kj8;J@r#X7EWGG#4kgn;+IMTB0;0QpEVh-kWI1Te96R>5Lt#%h6Tzp~Xgasf9vW(@g#GZPSF-T_k% zUqQFq9w;GhEDWgPDb~Q~-Crf2C`Mzoq(<}TDdg2Hke8!;MsTl59G|okaa*wpY;j6p zaaMfOs#{t^ky-eVoz#T8nxlUdZ>YgrleiKI=66TSPjo(}c!s$ThvnRF+*51&&z&>J z;Wz0nt?fTRdOIc1_WeMVBMOVzFn8}1&KuX`S&n1QI&VfdJn;AG=Gv7SPjj>b|Kp2X z9jmif?@jOWs#&k=8c%Jac8hb2(AmdK6aYA6U#t$?4|k(~^~JU$Gxyn}4$&H*U*#O6H{vGH;!( zf9T7X@{19iB;n+xa-)oYb|WchJ^x^*CY-^oD%j{Oy_N7zXR^+XFVn~jvS)Le^vnTuEYQySo;rRBA!qJvzr|$U z9R1m)1+DjmbS->mOvV>b94J5#Kg(vZal!PpjGbxTvI z>I#x`XLEG>J)x|?ZnNr)KasJv zv5voL81**!3r>hAMX$~5+51_!oAzVUNtu`8@Kh?(LB;HlP&Ts$g0L3j(RRjqv?kH} zs;$|&snpYMoXF>{Z$<%&=3h^yMRvDNhl7I-^ZO1<8jl(hp3N~ zwI_X>YKpZ=jRm#FcdG>CpuZ>m1z$+gUTh##@& zH20&;u2qXiUxs$-(auJUVm-}UM_;dw?dvkPv}?ajX2LD53MZ<;M1AJkWnl!YXO+CK zA(QmlDb;xLXb5{k-kMgi9P~+VZKbK};;=WeUnPX@ALBMfMxb=_z%-OK*36CViAsQ@HG=_=coVq0^axWQ(65GpT_7 zkYS_sHWHt3sWIp$F7sBiUOZK9knAPBn({t`kIMTkrUhS#*6L(mTO(A#wb>I%6Td&x z>FrYEQRn;nVSGer=Z$}nSiD980po{n7hBQwJhaXC=GhLeolCej&ecJ`ZkgSXwJ7|*8eJU5MVa!EoW|X}6D^?8YGTMr7 z9Q;hYzRYg9oAG~|9oY5gX2t|5r2WSUM#c|Kcc7=rT6plISlwmSDEPr7>-gf9+%#@u zKi4=T+S-gyobiPgBHzH~g(oiaw}u%d(_{TgY_HJ~asoAMksAIz8X*mJVSr3%bT;j-hIC@+v^DZD zKG(|7B9}A-+k3fQ!nk#TmP?Kq^Zp3q0mkq_kR~Kg4!uOL;GFgK zBHuU@7XwKzNJZEv!lE=WGP6cgCF?%Uo+kSxg`pBg#J|1n^dwr|j61QmVU6Mu37fRL zs*SwE1U}hU$$s8fBrx{$3%S=zTCax)8vOEHuFgARzdU!d+&tR<)U4NA8R zEjOm{1qxzXUTZPX__eLeUhxr2ruea#rzl(T1)gO{oko!})@Wzp)(#diq+xol&po-f z9LreMD)s#ETsdSaRZI;=C2_ZN_sP8#&LW60HO z3DuDsF{YRV$@k~c4j&V&`SOU1;Act$I)aSOlbhYhC|F{md6PG1qP8jS{jyi(*8ZBY zEDqF=lve}qlV1>#FOy;1d|&<3&pO*>Q-6oX9sYY{Jk)!BjCr>2^7&V>#v+PYk2Q8R z`|fX^nKY0Iw>3xe@xe4^BT;z+@3!^U_Y=K2_hV|eVl8!^wWR-3#~k@2WrO6=F-!Ak zXn&3YW4&|sMosFUuJ<(QzGm)_c*&$MH}OKl`%C2|Q9X%qYVaUKmE7drJi*?~c?_7} z-!g@BzeIErEaFXV<$jc_g4Nq=6jo7SgD7qRo|_6#zWb%T0q zj>oBs(o-AlsgA3-KR%n^NJQ7g^!yfPUUZ2>&rxJ-4p>e$O^WM}A0*6bJ(pNt)*wz@ zJMP^b$S0knlC^OPW$BpwKmXFk8;@j~@_%GZ7f%)38rsXG#QadbN%=jKG7xIm_)K&9 z9&?L!UL#Z8IK^#ceCDPBB7yt(@;2^N_a`E#RDZsKv|DMw;9Kw4j85J^rEl`S@`ZuB z?D)WfqT5$v$DTGxcHJR@e)Y*wIlNIPZu)~U-|kD?Li6+ute*J$uM0AS z=Et4BZzf2L^CGS`+4Tenj+pL1>x|8!pH)q11P~qOhjxvN^Xyu5lG&AEo4Eug)EwQ5 z&%ivT;1)TKq=`(j339iU<0CVS+xOe=ExkddPZDHxU)ly)P;Z@P4Yp-gE{x z*@u9)v1Ai{b0o(xPXb*ZUSmgw+7#6YOFQfYkAGXrziy^!(Tk^pWMd#z{lZwMs?N0I z6KS8AVelfv^p8qO##Nbmd(Y865U-VBT-`Gw(3|4MMAPSb=@OdMp0UU%0g(%3G*O*! zR+qJ0%W!&+BtTu<{Jk#M<+~`er$2YAcq>7mtl0j>Al?^2wbO6zFG|(<3O)&&o_j`jW2AJQR#&5`24CZ$@vD&i)Xe zXpzQ{aKiay$68q^);5)EC|c(?t&Dp+O=->1cZgze$S25c?eqA5Z4PGT5|Pkq&C#Jc zrILZ$>FO_v;)(o&E3iDll4y}MNMY0d(}WWT+E>>d!@Az@>D0bB#*LP2rfGv>EYh@y zMtkPrnRy%2%m_~m*;t#FW zqw-iog;DqYE5@Jp%gpLibM$LwVSv~nJFN8);t@eoiP%8tEIZs?>f&1V4|>Rglknoe%3ml_OM@J6S`UHWmgNWMY+!5%_!hff zz4D{a{PyU@FhQRl(Dxd2;Im;b4em72Tu=n`Y7-az&>DelDkcKLwS>ry-+kJ1oO~jI zBUT5iW|3Kg9AInb>fcoPW_34WU^pY~;a@NgsJ+LpXkPL$ldb8_eL79zUUfp>XSKQg zYG1xgd(8dkI`_Nz{*36>hGcF0u4z+*eGK9SMN~UDi5QLMX%~V@>u@A!5OZqP3?zo5 z#Z}ZN#iGpCTSxlerphVFH=U?OAEI-K2u8YI%xcqx^2j8(+7xJK%G@Q<4m_Uf#I&23 zj=Oi?_tve~TyADv3_0JX^Z7E34PUzWIP0Lo&(TZwPEoZqNnIKlQhdGidtVCl(wh!D z(;Kwh9Q~|YOfP-nJ(d{U4Y&PRaG(WRRlU}J=-1I57>48dqcZh1o_{VLA#pGDdI<(W zxv$7PC(-VUxbaFnQtX(eJ6NV!e~Pbaj(%BYA~j-Q=s@GRpY7EdZQRP$S7nYZ>r}tw zW=(rJE=SM&2#H#-KRmD6-w?4#-MbQ8#(e!e&c{3f!5G3g(gO^$9WoM;4mL+WU<#`h zm3nvQY*fzf@ShF-o{7mP;R;wVQm<2QCP7wk1#A@3`OMCdN$puW8wrlM0@g={YgH!| zQe6!DhgbrA!tNM;`JzCW;Z{AanArEmL`%_{s=fX;bv77^jzm5)*rKO3(uF0)QPss& zb&7_7?tET_K!5~Y)JRU*yuXke&`%)Fz(>oSrph@~IT(M6?7kM3Ru05o7cfBjQ^?(k z13zguC zvWjrPzlQNe#$kUA9qp;snxliUYoF~oSGBexO>^{9X6|I*wV88|&Ed_j#};rTrDrnz z^@|j*boY$ERKEtaPp-sKS`D;`q#JpnrhPzWOQIBAK)8`0N|fTOrJH@iSuHs&`I=uV z;^9)I6R7&#tr14;Z+$@ub=QAZyeqB3w|A$s(hYV}aNFdX7Tq_@BK_OZUE%}G>n#4d z`RqVil%1+UJN=_Bpm`~z22L9tMVykr*J<09b|9^+?R-~Ot$Re@mSfG2PdPP7s6t>X zs=Y6!$s8L9$e?bg76;&KTqkCoh5i~5xdB&O&j_)R8CT%lfSv4f)W_=0(E}ff6g=Rk z;C*;4X;W|G?%l+pS!)-eUbFIjhV163olC}Vn=iuFFn1-&<(Id!YTk>!)r*MeYtqm% z#MZnm=e|jw$f}_^o1=4>?2*$>Z$+s`v|iR0R&AYSLsw7haXs9O7)59Jtp#1DQ(4+v zA~S+MFOqwBm{dl5r37L8KH_dM7Kwv6wy!0+uN}xoZhNo(mkIfH%I))T3fzvzlU$uw zZq&%1F_*;d4|v}>rX&+>^3`1PqvqzTB;=N&Wr~H`FABw_Vqi717VbnsjZ$T7W^vAa z*^>Na1Sb<2C=Vc9Tf5B663L+9qb)h%b)B1$xI9^L zJi?Dn9I>M>O5*PE%w$fOn(#FY zzul$iIH4+Tjur*>i|7^UhZ4g`h`Q>QXee?Lij=6nBeL9_IA5EEW|ZzN^N_W@#0Ib! zIpxLs-JG^uo9Vl5lGv%sOI5Nhlax}R&X{d3?p^lxhXccU8iLZzf*&<$PrG^@ z>R?>nBVP?VvKi9ek`waKirz4wGS3*|1npdEj%IM^8x= z3Hjp3feC(06@Gw{{c#Fy9*{rwxC?Lo0Q&RQC#H~IQ+?|*ohJuMOuT-DJ#2e19XB5B zF!8_I%YKrYr5LrB`}R03i*@n->lC{;NFQU>M~T4VEG40(ODFRz9dE!ei}N=9D$E?v z%n@if9jtGdY|rqkvf3?lFc}qG8;{Yz2cINsgndMzp7xy4KNYd4c5rXSA9MlIU4Hs+ zYH2j03R*>EZYIU8qLJf!ilp1JYW^1er3NV>J$f&CKFh8q>}`>{NzAu9db$i4-r|$z zNku!2GtGZJJRc#wt}im0uFZ=NXhYlOk`pBN=4gS=ICYx(f%pY7mxqczLk+u{Upw<} zG11$%-i)1>q_}sp?BC6Ow5DGXbw6{aO~TW+_CD&gwSlzR79^O;akUG1WL&U5?=_~_ zexLWDM5{6J-tki3$Ha;F+A+EF`>upw>X)>wqIgo*wzrv}Y(Eddyx*+aCPCuSS+voV zmH$beReDmX4&4#NS9@vYSbe0q&FgvYWBjFvbkl$R!HfB&2-m(6c>vu;Z93qKy5=Dk zuX<`?1{bTR?bOiI(}&j9VtzO-QT+mXDkQLs(NnBG%>5d_#|JtC_XpC)drI|EX*%^G zs0mD|A27!5>cq8lV~0eq5?5f7@oO6K^@=|e;8yvlkdK1V~wnnWvU z-0Vhbs`YAloF)YKqtPknPbX;cb^q!eDMAldL%&eV&y-e?$Su>sI{Y zU(U*1H7)0}+(fnW(jfH#X2|jP)8PV56-Y8d%|F;@rDr4x^KMnk!2L4Y2N2>vTra=G8sHiBb2u~=>F9?Ui#iikS_Iz<^ zVR313MPzzGWMaXR;=F>?Mch+z!MJ=q8XWcxbBgmTXBTsR?x4`TP`I$7Y*9FrS5c4~ zDY$xRLFq+TUNa##uc#mt&W(gal@)mxYSPLJ7cb?1T+dm!BqvYq7?M|B9vTzMuPQGo z&WlYQBiEFIMS11RQ%eh$y2&pXx6p_&yCipcLB(8T&5h)y4jLSG3h4Td%7Vo?B?YD7 z^TMSImlT0~{l7>~W${e~gTs@jPMSP*^0XN_Q!=JZ$i6OT;;@N>j6jZ_63VmWi1KuG z$LEnJzoM)>pztLHd6BY;F=KM3u6yd;HMhU_ z!0LDJYj}6Ft@5C5=vmaG3`-QRvF0=GZ^C7LZqOyvaG_hjmq3b zB~#= ziYm%VDc0zC&jTC7)5cG|JR^MV_-RuoPrZChctB;iD7P|Puq;whu&5wRsZ|A)V@?W% z!&8b&xr}{I$z3LYoD|~Z)ZD4zyt1m&2xs!CLTRCQ7*3^&n;M793c~|#sv1NZ`Ogu{ z+Ht(zM1MKuk&40LNLeJeE8diS6{Hg^xg+oV%&JE8{`yGR~PMwrJbY!@&tYQ(idJ>nW z7v+{0aAjFYj%G+Rg$qisnbJrlr_T*fDagfONSRkwx&%K^kZ;n_`}MMzliViOWE<$* z@I)FuoLX31Q5gvjO-plg3}cf;1o7tkT7< zu{}qh-Z^sw(t_|T&T$20`vR`~2owCcE^@W6rLbG!>WvZ*q|g}KEg1^MTYZj$~b4QDSGguN8= zs|rz^TDUz)ckiMAxssrSuJe}!wt&Ss{1IBH|Gz~4e{oJJ{neg!|6Cx0h}>FLt|-(0 zFRv(@uYXh&R5IG>p9zZ9YJ_Z&SR6w<)ho^`rrU=X8QUSmv*uID*dZzQZ1_%id`XGp z0gb?wPEDl~RpsRsR8|&ZZOgfz!NzBXRh795u+srUh82dVTxLduaA`q7enGzDI*nLb z&O~uvVdGKhAJWalpiS(Ne<7ZzxUje&UsFuXxNIh4-X-!M!*BM?8R3j6mt{Q|4 z#SBTRlq9u4VEHd7=lDwE1OR$rG_L2_3wiDY5fO?%?g*7IOOYgtGGGU7M>k4-ywk06@;t zF{f)tBi%HdYRr$nH9QY9d{;^^pWw=L4Hvl2^f+HUW$fSR%^mox){4I*; zYj4WehID@#Ggjb zW;`5|z)n)_P^rlRy**Y3X^DGxT>7MDj08SQw!PPcfq58!m~xo4l||*b6=Doaiz7wh zvWnsb#ibINyTgG=Xbk@{M=C>H*bU-<&GGtN2Bu6H1l#s&N1OI zMd^RReiP7#+c}cx!tKp|MI`Ub3Cq=ns*Wg(`GRT1g^_Agg^ zI!ZYHqO5#^t~*VeGJPs>79A(AqPU!BBh?r?Ya00_%$En|b45e3!CqL1n4A?y?gfS3 zIonxsBau=~I6F-im}~+KAPUMWi%ZH%)s0yDFDt*iAYz5Nwya`dWjRw{>9nan0R}F0 zuIQK_OmFIQdBLJ1RYnC{JUFZ|rzhEwsz*A1lkwn3&v&vFKZn<1~{rrk?P%TFym*_Dp&?7-3oGrj1R8Ig#A? zB?X>s==nTivcUNm|8Z>faGFuvK~h1FaQQ!94{1*tBY_``X@Fq&bnX5K>Ai|~k^b1m zfiiatD=zgKOqUhBLw_eO>%uMhY+npg1VUV zPUeGoAVx5;#4DK!1@(3m7Mz_fgo#qLT6)SzjC+c2*fZ`%4lWI>|o2UvP2= zdIm=yM{O^Xjm!ki(&$mk&g}^UTc`Z4?Z+uFsC*~gG0B}@wzM>%Cys9l>x`vvNtIZ5 zpWP;x-xTbxK~VnbON%ehnt236uUuNZpoHbecwytu1Y5`F$B@$$mcZbh0=ky-=;B0O zM`wB2RW*NKFs_nK1O=&SgTwr9=%B=g)73q;HONOpP%m*@Wb5dadM2Oe<*0!Es)`$m z@k2+^pG78FupAtAo)37^`TMiKYVsV(rvYE>Bqt~RQTkU;0Dom1J(&J6jd^z2qUri* z{YX}(Hx0JWHRa|pWAbzopBh+apFDE-;PB12$Uk3I%+t_7DtD=hTk~14lO#jEG{!Bv zdY5+d37g1Dt$@MW%4$SWaejV5X-;JX_iC3R+I#AXtR zWET@+x8DpMBE3-l&k@hF`&fCHm?x3PStK}0dF&#=zatM5?j-U!lmDZX$4>r_ERSg# z9Ry?*Go;6+Hr}zxO^>ymc)ZOsk%F)o20K%3GS8P)Wr;8O{ZG(dVw?qI{Xf99{W3 zU?iBKx>HtGdje0uWVRZYwPU9~{~9xqOVa+_87QI4qK~RY(EiXN;S0l+iz_0j`9&PY zj0Jvs4NQ`3Tf|R{jzRX=y7)VZFaGlBNjrUThMJ^2b2l5}0vd}gHC9mxO-B_KKU9i( zy0lX(tlL7)oV3g{io_Bgr5;$=Nm6+maQZ_OXOYQTq(5S@i%hYuSHlC@Kg`aKa9a3M ztR*eX^RGa_VC*?BuY^r2(+cvd2<@H6H2%^P2qe#a%|lCHWN1Rm2@|q%rq7&W2dccX zMT@);MYYCLu;TnG8rMcqzIHHAn5>1$0C6;m3Per`WVPCoSzyx9(zzjCE}tFCO)^tU zDwuC7XIsoaiE#XiPcBWMQ|pB77;H69UompYj6R87xzov`lgXLvA6X(tCmLFXQ!|8$|hVIfK$S9`4{QW&fZ)RejM2-YT6<@HgIlaY( zgvvWYx70MA5&HBOjN`g>cbiB$F}y*Kp8L_e@HAi`Qb zG8YyN>1;QLJqNOryS&^kv-9Ud=TG)b6|g(Cs-nQ2aSYXM--?LcU*&8QD6evN7sSa# zq4IQ?ZS@%W$l-Q34nif`M#BF3&^eFu0;BKW2;blDKgRD~`Hubf9pswPXGLUkX>o)d z^=22WY_3ynJN}4qgl!p%&WXF0xIPg><(frJQE-oo+*R&f3+3kja>fva?a% zX9*XS%G0rYjB#l(`zmGocG%yk?M}}Az)}%graP^yD7kte73yJoY&p_qzJ9JQO z+@^ANdr!nf!n)s2T8UjK&R!xml(VwIzUtyUdEMaLaMntpBf0EY)wdSd&_^1Z(b;B8U(Efo zgIzX3+m|6qmYdgA!l~nqkiV5~e!=oG_N(h|^vc}A0<-U*qGHRDiL|}~GdgX6Q?*>q z-6Z?H*@?zR%ZSvv3M;8#*UVD8-<^EeS$mDXqY=(6FK1dKyS>D6`dG(;f4B zKfafU>}xkUQmxYbknai=3%*_ynx9))oX66bY<$u;kaUlCMO9uzcX4A#5m9Tg+4UO= zna}Zb-e#Loj@`fI9EZN+wDFFrVHf30FVE38X$tb?%@W0}Hgs4{R#_gaHD#1PEw#k`0#dFu3> zQ90~BWG8dQ!gLP@&mR8#+u6rWGv4G9p8V7wa2-CRt9TxVSib=tbPYTYuI&W=xL^{bof085ES} z733_fkakC?;W-)8N99Z(pELc+oXOKK<48CWPRb#h>6tkh`d@-td#A{Iq+h?-OTpx0 z&?4zaOZr8*3we2~h_0E>K7<8s0bC!NGlkbdmgO8(*3LU;30^7FJdL*S%WCMU11s57 z#vb?!#^o&1eeq_D^SvBq*08;o9?j0|3+WM>(7p|3=XDi4e0AQl^kvQ1!?=_q&Y7oZ zUXk{^yd4vi%k1~`-RT$3o7))K^k(~d%_Lsh6myQrFO%2jx~{9qM(%NwO{yp;&_q@( zgF?4XH<}9ElImV|8!YAz+HZv4KIngwv%hKcItn{EbM{3;7mcr!w;g4Zt$8tI%$c*( z7Ky*)P1Hi(6yZ;n-iu55gRS`Pdx!Fls%i_!SZ zQL-_{X!gyDXO;vq=BEyl`b3I~D>3Kbg^&@8;akg}O>_{NE6ezjn4`J8B|$-#xUFmB zLZdg$&h`^yiSk@Sgf`tJK1$K|_RQ;5Mt>q-c&tx4o1K-LP@pikq_QAnq*ozLb!bGM z0gBQH1I<+x)pyk%Ye7ykOu!{a3Z^-$krRDQ71*mn@T~e0y-KCDICH>;RW|=hk0NI$ zUpGhH`RS&X@Z3%99C*9Yz4Qoz=UDD$;}c9u<@E{U6fSZ8V6Uv#D#pwEcVQ0-Y3v%? zZh7TZu0wMTBCZ8vub9S&Z_SCrB%Ig(j9+E1n!HhBI-k72v9y9s@cN1gufj+GBPTE( zeSv`8#eu6Ikh5Yn`R-|B z1!KY{Em5gS>AS-wG$w2P!EC*K2-(b)Qp}Ml_>G2UM7pAl9p&x7V4l2QJlk}H?20lLXCv}_sPGP02@QEY|?(+?Lfk9ku)h{71BIl=?OY;kOTSHv5 zIi%I)pjbyt-zl%9lg^4hzkeBDQoNv4-sE=QWX|OkqNTiWjHV+?%MwH&X6UJiNKD2W z&;l&O7ZBXaNX=WA8uUp#Wc4JmkA}Fpgl!j^K|9AiVWU>(?I`;ip*m{{@dgtcBxBo7 zTx_!_m|C!Onu?Ni`nd8^{~Sw18+y@rnPG}lmoZfcbvyI8i;qudUU*_CH1D%dJH!0F z@(#YL1nb~DcpX0-&*2Go5xxm`!cXA=c-9%DNA7X39xj6W;0D-hVkop1romIsq&{#QEQdG1 zI@km^!=C(f;ZB$i55Re_FO$@ja14L8C^VH5lq?uFy{ICsxUOYQY=8Ek6qTz@}Hof`j2Mm<3PgSA$Am z7F+{Ee8H~)hG8q520P$*6R9Fu);p?yn_8Nn|!!%eAC&CtZBaFfou;+|W=s}nYpM#mO+r^}Z!(cU>2PNJJ=_6Xp!`y92fP=iToVd?0Mp>t$6@cV z8kWNcV4Ylt4RG3}v^%^J?uXC8zSp9+bo35yg|pzV;W9XQJbHuI!!585?t&k{gYb)& zq4!z%EjSk50O!CZa0RS}8{lEs1XK8htTvbdL)V2uvtbx6g6Z%rm=8aNtKq2=4|YBZ z`@^m90T_i_VbAL+Hv>NbAB0)(Ff4&7lkgL;1a5-2!R@dK?t^<_FXE$5;b1t1AIQss z`LG0Tf@|Ola1%TTx5HET`0qX#hP~!s*KjbbfLU+@EP>nL8hGkt(!()uJFJ5p@DZ58 z_;$(_v7w2mFOFm+yrk7 zg?8p1>dab*9nC+~SpttQIMlfwj)RS`V!@%#@tRYxDD)ec20w+_aNbhV!!@uDHo^v2w~X|# z19rgf%Sk^k6zU7p;1w_%-U!R#YFGz1!3MYow!;25kseNkDY>E0V=xW=8fL@2upFk` zOnP`BY=Bc?E4%}Cz{g<9{7~o*Fb#gR;!x)t*yq+moz<}aZHGEH!@N5Vb?$|CuHybY z;-b}uI>*A7s%b~~)|x||8{j|d4s~vaXWw_I^B|mX|Dn#*eDwdoq0U+G8xN5${Q4ub zA8dGxeBot3CSUBT`xCS;JO|E!>2L+ifg4~M+yYm^&DdEJ+yOs;`{6m84t4gy-s+wt z9c+Rd;F_lnb#8^T8V+^tgZ+QTeMPhb91Bl*n)2XaxB_Ou4e%zo1=hn|@MU-qz6<*m zGw%JI^l%cK1IytGSO+)22Dk;b!dE1Us4U<9VnU)I4ixEW@{9daF>|19l5|H+2S;PY@je6I0OXA^t_ zw!wB7Dhq{v{>ww1{bA_2L!A?0_*cjQi{Lu=ZMYfs+(J9RN8v%Z74|D9E_(h@XF9wA z=EIxdYPbzPD%W42UEnovuUv=6FGhc`KWu;#VJpms*)P(s;ddw3nzaoClY|O>jNj4jbVR3gyE7a5elUd=y^rN9-KF z2KU30UdNsy^jkO_J_cvNcVPsc^e6NMhr`YAD!2ny!2R%E*rzHKdKwOgLw90NZ~|Nb zXTl9|0o($Y!(DIMAT9_p-z`LG2(3Zrrz_FNhY{TZgh zzrjq{<1PFs90sf5ELaa$!WOs@Mqzjt`d&u-0Qrbls03!iyJ0ze0@lG-VFSG4&*&T8 z3lGAlVZY^}(0{|R@UFL!8@>js;YY9@4%tn)@Ol`9cfp=FF`mG|a7`QfhEKo<+yU3Y zUVlN~a17i5=fnN*F4%{0<4HIi?t(MmIeSPCr^5B{A$S133j40W&%jY|>^rm%Tn$&k z2DlNn!X`LuFZK*?f<12ug&v2ga2w2ossD@LgypafZh#GN3v7j_yo;Rh8rb(%;%_(# zM&K-X&-;`MpM)FXkPpz0oQF|44}0Fm`F7d`-UG8?-+iBwN0_Vdu@M*XSZiU<7dH+N|@DA8#74u#=9Hx9sJHjP!1+0b} z;InWGY=OJrr|=*==@aB!js3#0@ESM=u7)e%<8T9Pgj-+>+y(yx55hSI4t0j_B+iEE za1Shk!~ccca2ebL?}po9BisjHgT3xzUH}Kfzrrk7@hR;C?}c^n0BnSrpHVJc0}sN7 zV83snZ#Wk2fpcK@gUAifh8y5~xCQQlyWr!8@UPX#0sF&Yhw*dpepmz_ht=>ISP#c^ zA~(Dh?t}NhUNw{t2g4U&7L39Yctz-NXC0ge8{p%x74Cu^@PTfJJNw*?zk4sLBg+zKbcz3>Ux zYYpYVG??DwaA!802Fu|}SO@E11KbE(VH50tAHbBglz#$p!})LqtcMZ!He3fcoOrmi z5k3ib!d>tH{1EoNhjFv#;m%R;0XPfpfXm=vxE}WT0_ovYxD&312jHla4tMtZHued} z!q68Fcg}&s;R?75Zh#-bE%5kWhdXz{>)}DT2KK8%J~$S>4(GrqTmer%ne;FdZh^Dl zE_g3I2p@(0{tG`1$HL=JAwBE^SHP)o1Dpr9z_#`|C+hD(Y@e6P)Jg+zD;aIo= z-Uv6q6>tk&2Y12e;X!y(3hBRty~45ZZa4?-fh*vhU*Z`Kd>yvHfu|nsjKUeP=Q`wv zgW;Pn3%(CaVCk1>cen#K!oGdbJ3J2_fcs(izEJ46)37&K4fElLa19)G`r*z^a3b6e zPx%VZi{MDu=YI6^RpfvP<0H1Fvca?q%!3=@de{VCgKaSN0@6Q(zkp%53Z}zn;5@h+ zu7sIGNDnJu6Rd}Au=|Cif0%VQ7=}eK9o`G)!Gmxm9GFIWI0rVtXJH%s8w_ovJ%*AV zPJ-$18aNLY!j7Ys zf!~2q*aCb02z|g*_!-QEXIw&hI0{z7>tQ{tfGw~ZM&Wa?XFd9csqpl1q=)IS2v)#q zxB=F~pTHLQXBdS&FD3nt(Hk5LSHdj#2rPkbz%{TPZi3gP)9>Ia*a7c@DNo=B$20!H zQE&!)1V-R?xDIx|jBy+0z#VW6+z+?HK0m>4!r^ey1k%IjU=1;3qHwCr%){SqeL4OHz6ksN6#t&d_zUlVv*0GU4DNyJ z;n>NH1F#D2gipW&a0l%BBz_K#g2!J$dU!fq28YA-a1v~U`EVz^3m$-*Vc(}H2abX% zSCSr%gUetETo2d7Mz|I3gdOkzJTr^*4U7wL6wHRR;7xEDd=##Shv62OI)#1ZSCJm3!gM$n&Vvubm2f-U2)k#K z9$pLE;2IeEInN=k#%^I~8g>i&z#=#jR>MMA57)pJ_%MvZ7huoL#2qjd9zUJ*@M2g5 z7r|=y0IY{Ug)MM9jKVh9^BMfl4AR5NFcaPki(oyhhFf4gOr6O%4yVC=@H?>Av-n{+ z7`DJHIQ|;?4V(wpz)H9YZiL(6t8gDY2z&j4cDt7J@VhVz{u-9RPvIIkau(^~47eQ@ z!F}*y*sGEHz`^hXm<12R5}0-!>EQyn34RZ5huh#j_$KW2OZ0X<>ETG21&d$_`~zGA zJK-icd^YLf~>z*Z}v!R(Mhleg{s1DO;HL!Zdgv%!c)_96kr@;NM^a>@kn@a3t)2 zGhoW|)EB0~hha9{4$I-&unzXjB|V%BTVV<8fOo@`7Z@L48vG;7hTZ3r9`=ECa2#xa zx4~BU2<(8p@<{(8<1b8uSHf(#9G1hUU>$r7Ho!fw752#|J-ia8G=)O9z%=+Y%!a>) z;VpkE8qtPdI z3ESZ7Ftn9?O0YYa3$tJZmcTV|4crSi!2@tRJZlkl2gkx*zvezT7!E2$Zg??_!0X{U zSOGV~jc^A%0QbX`GU~mJc7?-XKAZu!zzA%G>)^+5Gwf4NdYBLQ!~0>MX2xGQ9G<|uzo8b(&4MyPGa2@PbNqRU7?tnMK{qPRh z=Qp$?91ah{8Ss<{>EU3w4laS4;q!0@>|e!r1INO?uaX}e1?%7}_!3+O{|Byz{|g)8 ziAxx-;2H4$BkX+O`kwO#{_We>WV*>@GM#-*CX>lzG9l9@lL?tjH<|3Z$#is+=_Z8i zGMP+Vx65R@>Bz2|&UL%E(2)soA7%z#;S;bOw!uEw3nO;0U){m+ z0uIAWcx(yl4W`0MxDGbL&9Dxj3^`hJFFL&0GGoIcrPq~&9DN# z0UKcIz0@1t4+r4;FzQ{-Yhfb%9%jNr@1xyt7OaFDU=w^9cEQi!Alwh5-($bJpZNy= z1he5suo(8kYIs~Z^93fuZg?{sf_K4~F4h%HhPz=l9D>F0f(J+sm%wJY0(QgO;Sj8c zG4Hc~!ekh|mh^B2EQWJoHOzy}a1-o?t#Ak)`YY0Zz^3EzUQZu${=;Wsc9hOcA3!xLaB zOop|v7Pi41uov!uo(~!ShnRP8DolgPFb|f(GPoYr!S7%jyrh!!Fc*40;ye|`!A6({ z6CP&V;kmF9X2M3e40giD-~jwRjQW`40ZfGNz)U!4J?UXGtb_|;Bg}!Fa3vgol`yJ@ zafV6o;78b};8a)yXTvJE4K~4c*af4jn0N3@82t&yZI}eFfm!e-SOhC!75p!3hKD@L zxWQxK5IhCOd`kOZGJFx{z`w&1_%W=3U&9u7*ar3~I2{hdwb1(+_W@ukd>!V(4`C_% zH>`z|9%J0$D%cC3gr1!|UxRV*RhS0<0rTKKSO%9>vroYXVLSXS?1L}Bh|hUE1>@mv zm=0$=&Ur1o43@*|VLiMHw!Tuq=$20IV^_tupYL~e@(hC zC=bTNi8Z8y@h~4=4$EO7tcQ2QcGw8};6WQ1uU?-2!z6eP%!1d$BKQETg3rSy_&)4{ zKfpnF+;2($PtHeR61)j!!AD>b{1dE#Z`IP@@ZWGJ+z(y58TTg{FL(+}g%`tI_!C$P z?|`-NLD&kPg*)M2pzBNK1N6eLVJbXy6Y1fPVJSQt*1}6*E4&fzgbzShANx7;tbu!B3ygV@^$8Q9cMs25VJiF?%!PNsQur*ag>S%CIH{R&f|tYm0s0kI!bLB! zZ@}q)!NZE}%>QrLhhRMHXl1{HOaI1k0&avg#={o47xuu|Hu?!JfU$d7A20mSl{p#7(qXk!+5w6ro&bkNxeK@F#q6e*hBdla2OW!68|60FJL-+3FgD^U^z_q zC-VO*q@i28a>EUBAAASofVA7YY3z!W%;7ZsJUxShVW&i7=UT`MN zfT^$m=D`YB1{>g0umhg*FXkoO0HcSvK88uq^A+ZOH+=3J$~8ZS4KQji<-!H9AN~cvt7iPc=SO8bS3RnRf z;B&A8z76}~0F3;B^Oix>cEG)`A5Q%b>4zB?m;g7!4A=$> zp!Yl054;36!n6-Mo6|AUFJe294nJ7A&tIjn?pzb8E`gPpJn4!}AXb%1pW z6X8yn2~+lw9^MTr;bX87cEC<}01m*)AK2e~j6Y0WUTg!Qlyegr#VKOBIraMFi| zxn{ydm<2Q8EwB*21}ouF9@4{^uoGs&0k|4QdBR*T!9+OyAkxDmSO^!xO4tM&VF&Dl z{cr$AP9Xh3VXg$22s2KNI zmMmw!%!<4GZCSuo8MFk{-^6op3Qsj0ki60%pR8U?F@4R>F5+ zBOHL8FzHargT=7mkT9NG`g~O|>2RN~6;@34`TAimjGP$e$~l600E>?D`Lf_%SOhz# z`Fu5Sk=N&IfwPYH`Fdd;>^VBj^$i?`6XR$XqO&!LPPjDzpO zG&p&N&zA@1z%rNt>tGpdgPs#9&-@#DBEwu?!8lkFPkQ(Q%!94440ggg_|?fiUkCgi z_QQw-+W8~uajMUk1T$e4+y;wadLsP;*T5#YANIhg(`X-@4Pz&TxzbM$EX;M$9G@>9=E8LN0L+IUz;bv1*25>yWc=Vaa1c&Q_W7ct z!dy4QWLO8YVGAsV|Af_W_E|n(Gh70@;eO~kJj`{;T>1;%0n^}{Fdw?krd&7^*25Io z4tK+T^Yaw?YcfBF39uAqz`}Ewcd!Xo!XKT>yn|lY3G?6p%sP*H9}(uNhDq=nm<4C0 z(%-P}eCiDs&7J1((AD*uI$lf<@Vk6P&e#`2{O}!g!^Jxjux$a4+;u<@mFd@q&wBE?fsoVLhyc zJ76muhCAWpYbpO2<}dWZyJ0Ha0&`&pEQNbuEj%`t^zdA`6E20WSk^7{!i_K$Zil(> zQ&0hb)<*q!kus_bp4q0&0u&lhN-X{X2K!39mbwWzAyz&zM1s~<6sFag0-*>w!+-8LNDxxsc_QInKy9iEzCPu0PA4!t;`#^7WTp=MeK7YQQj}u=U@)ZfNNj@d;?a% zh*dsc1AOH+#vOLS0a#Z|yJvDdzMXo*H((BoSWUfQHmrfscd$>u<6#fH5)Q*s7<)4N z8ccyt!yMQGOW;0O0}m-7J)8!6;2by%i(qU5?SU!q1DFG+|C01D8P>q>U@MGXL%+e< z&~*y^552GiroxRd7Y@Nv7K=9=-z8;p8%ot1tzY!yH%-OJF;!y_fX@6Yi(oiL6(c0vlm2 zytkZv0(QeX_{syMgZ;1%PFzd5r!j6Y9&Ur_a39QvU;T>lhO;UdZ+HXjfZJg|jC+uE zKAYnbOoT%)3#PB*xB^$gDp&!V;7hOzcEdrq8%Ceb`PoCPb2#x~)(P~&VwkX=bpkJk z&2TyFhFjne?1iyO>=%#FZnzBQ!eUqocfdM01l!#57)uSImE#PxE*G|E?5AgA0s{d3O1O3SCbyjhW#)XMxM!a2~2=x zFavIY1+X4gz?jF`U*SU7347rXbU(p5OQwHdGE9TnF!|TabC?HfV0;bZ4GUo(Tn{78 zB7YbUUxVqe7v{t7U^&d*NWI}2*a4r0{qQ{)IhS$zE%k;wU^d(bi{aK<$~FInEpXwJ ztWQ`7hv5s*dp7+KQ{kLVj2BFYrEmqTg;lT>z6N)~Zsc0s z%=(7Cuo6yuhVg>;)H7c2Q8)mffl=pi9tsoTw=ff4^gGrG%z;&KHEe>@wy;iMIUIyf z!06O4*XJ+^raw!+!;P>6Ho_XX1Gd0jum?so(C;u7#-7i915@CYFbCcWOW=A~12@4I zxE1!mZa54*zbE}X<||BvWiS_h1WVx%tc6paqu=3X*ax@6hzmF_!+4nRJo62fzP2iPI!TJ0t;XzTmu_n4eW%^!2#F7suGf$73 zZjzexY5bQs?+0I$`9bi1=I1&57jJ%U(g*)%@+{)Nm=AvNO))>4KPl=y_XR(ia0vzT zPvfijZxa3p81f4(-^7m}|N9TVX~v%)OuxkPUHtg-_kZwBmHtZ05AySm@iEe0Vfkp< z{~ms(^fy?ZLxSrQyjFe|z6WpmB~-sVt@K5B7wyx06+RAss!2c6Kg+H3P59IC?@8}| z!yjP!uM58$@2}rT`o&iI!Et=F+c|ww&};R}!hh#be$hC-Y8>Ab^vBEg%kjtWb0~i> z=(Y6G;oKW_NNLc4A&<{0`zC_=^XHg~l-B!SD*t zJqLwPJ;*;8UgAnA>zIxoe82PXkCCyuup;~(Px#{=qRe>K5wni^*!8!`qYYo$`Gap4 z;~L6i{ypL0U16a)(MQbf4taEwv>cxfJt?BjV%{cD#|C0%5fds$tcwi%OZbzeUuMm_ z0{m-uyLss)O%C-pei@%VEgq}i+*e!uMU3Yk!Ftw+7*WqI%84f?S=Qrbs~&^+RJ>h1 zOn*l+ZkmrXbz}Y*{|ECqR@8w$cRy{FlS4io#F;dq{<_BUCHPLf)?YRF9;KdMvaFyU z#$U!~&y(SzztThf)kw_5?qGl2C1ON>c_#2q3o%-M#0CBFloM))SZ`_g(?~yA`irb~ z=HcV;XGmXQ`7-=m{GHOfSNQ|{^B?~N{uJr+tn_X8dc4-pz4!+Fu`<0l9(WGs8K6V{ zarkD3{L}C)`0Hf;?(9(g^Y91ocKz!mO&0Ufc>g$l9`--uX!@X@JQ649lkn)6bQj{lH9f(|`H+ zO{BM5173VNz8P=MZ9@HP_LI$484bjwf25C5N6>3y)Q?|6daaKlC-N*7uZ>XxJ`b;r zQ3if7UK^u=V0z70;BQg#H+|F)O#i#Fee`Rqk2;C@7kQi`V~Q+4fS>U3557$4-52@; z%>J>HbutHE!oKObhKHjgbHN->SGxUiW;~M*<$b45^gSgTUyj$>T8!U-pD*iCV%4)6 z--4eb{Y{o{#=nc##;P0tZ~RP|Ud-nq{C9Y5F2qFgJ}X|EOUd~EI^>^?cYms%;}_$H zNIyx|Z@$&OYW!}z-Jb0wO*6h1fA6UDte4xYJbH*(KpoDMF)J-UjQhkil27D);S7@w!%pSSR%A=E*n$LgmWyl!!1^xgpKO4Ze zgq`3rq-jy~MwIt#Du3rYAz zc&&a}_$<6uzao4ge$*Ug4z8hwW_+sfHF#|uG~qYm2W0!ktb=Z1R`iZ3Be)Js8AJFD z__Q&3n6qxNH{rGV)#0n~TK(FBe$@D|9##bBaxcDt^jiOTm_%FgC(8OQ zwdxm#{}Wyt-!%N|cx`<1@Sozf@h!uDj?a|k2M87TH9m3*^8}wg%3m7t3HYt}Go%;S@frBn@!Fa# zz(@7jx5JcQfp_EW`oW8Dz|Y42RF?0*SLVLTRMOhbB)ZyeR-%KlJM{2wRUFVkNubZdhn8G5q>`2sU6kCloNA`tb@4c(Tv}K zpCi4vKIq0b;Fn164xC(?ej38>#@iiZjE|uMzr<@}os1vGYkNXAejna$98CVjh`o5b zHSfh&<9qR|_#B$6W}L*n(L&6euk7(@#o-~5>^@FIYjKy+-Fzv_mU(9 ze*%7`Odpt!{@ugWA(xn&iP8G06u%NrRjfJ$=c`Fyi(l?ghgN(x-l?B@iP=Jo-Lck7 z86GY`n(!rj9@W?4S}C5G$bai=A{~Dr{(6~*J8+ZGv?CwC7H>D+#+Tzu9P+QnH{-Qq zU_1UJvH6Q@&2;>3yj>sr z+m9c>r_21Q-|Arf%JDJ%djHhpC*rkhy>@)OL;61aR0nT5DH5M5>nGY5k56z&pN>Bk zueKj=>Zkc~{A|2-PF#;)iZ|zWq4B#muxESuQ9FJ)ezEjszX;q4_a7VjiK!+=n}d-* zW{%r-hCS8Tb?b^n;J(X4NrpS?Djn0G~Z&zwZX=-8rH3 z75Lxavqt%gL%spOnELt8gKVCAcHmzi{i!DXh&Lxe{L|FGAK&WWBiW?h#(Rkk^|AlR z8c3fI^xAn?27U+W&3Q;Dy|`CUfR8>(_Z9dh`211n1A}bp-+u<`=*;%^JUPn*WXcAY!F-E#<~v# z+FuK(!Fntb@gsE>_2?n? zCT}oyjvj035#i-~y2tJJP4!TZV@0g2N0@nn&zvJZnfOai2*$^bAD>VBnI{F~r;Z|iVGpr0)AsxR%4biPh&AUCF9hZwUE%VuPF^97>EiPIRm6$CF8c(&7e@?B#_BKe z%()of63ipCHbnYr{GajMsl`2dy`M{@%-NZ zZ+`Ht;QqVk?;@W-c~?y$(&SS~KA-sXYqmyw+Xei7Sz<%|CeBkk@x6F$j~~GA!Sh62iguX(D2@#EM>hGqcj&P1 z4Eilp&tQj{a*FYh7Y5JOM)tf}24a8~lI~%{n!58D}@n#K$`bWZH}xOFZ^tt&t@6b?Eiost#vSUH9Q3I&y_kpD_+6ye=3z0u-y#3%aeVVQ zzIz-$G>(s<;--FD|0LrFl=}D6C-Kbhh^S%TjbAcZVz)@D*VJ zetvd<=ePsBA^7tVPCO{^^WNakAM(udn)Lq&3vk-RfS(W^;G7^A2jjiLe~-=o{a|_d zK~@IiN76kWO!t%E&sPTHc6meVJ^f_PVZIyoooD95$hDWbxHo;lf64scQ~0mhS3DX2 zvif7>nAPpyV*J->RpgcT{jhJY%8F{V2pKkJ+Ci1#O^zHRwvr?~_<#PWv zkj~8ce)7#>575?JBnNLNpD>Qk7{?ck<15DT4deKZaeV(cKJt`t+CPrZ7{?dj_c@Gf z4r?lU->~m*95XyOuy96>3-UbcfrDlxx}Rp!m-vraY}c+T()~7!-*wJ0ZPfX5;62g; z6X+j*%J1P=VRwCF^64R;_uTx>f0IwDD4%0a`0~K{y2&SrO?R4SzwZJ*hw2!d>&9o{ z-=v%qqz~NY@Xzled?)@U>D_^w+om3Q%+nt;PxqU8yt&&d!yIG#t$MVO&t7E@GWF=e ze}lKX-uL2%@n7KE`5bD8IljJRtr;&jQ|@25-}au5S=Uj_sV2PtIpgV~t|RBs!RL%# zVpE8%znH&yz-Q0#B38VMSxI>v#GFV>s7~TpR3rX9{O!^QpY5BztEG&D41V7|pFN#p zj^FF7b`6nFEpc{xftR#>((ldr5_kDNbSKv+Gg!)t5_jPfz zXC+>{=i$ZY;qCg#l%MM0O__|{_&&VW)_#0H9$4wc^OQ(#?Ce*($KZ;j0W<-JSX*~EqB5HaR9ni-oYPB8Z1ua({% zI59Uq5kG60o<0+Q5k6I>C;yT4ScqSN*ZQLpzY@>ssa3zgt2_Svq7h$=*Y=Z6{H^%u zGJW9d1OD^__*HmqJx86Pdw-podnUbqZe97gRY%j$!MSDnDVu!WB5ty*hj{0r82?uX zUyXmk!8hZda`4^wM;-hSz8tUhLku@rHE;SMf_Y^8ANcI~$KfI560K9l@HbJlGg1KRLa_*Ch|oa@Cm;k7yEIg|V5_#;;yH32{!u)KRIC2iTWhimzX{K+3CoN3-Rki5c@UQvyB`ukD2y_-MR#PoMxl8E;olvt}wBylF=XbJnEy zw_~DxJG{iz5x0Uk|GvKW6RW*q&J9u4I)}Q%oTYpJ{AW&?^#1w3UX)?_RL+0mrrr>& z%WWb~)TNBF%-?*}=4Ty#9$wqC+we>AoHko+U1rU%Ui=EhoBryej;rvO@!9inkCjL8 zT+h^Nh?qKJ{O$d-h!O2IW0=WBYvOW!4iw^N;kneX%DK*}Yb8Dz&+u6OX3IC?Gx3~y zT0U^;;NRan@k{aA938+f$7_8ZMJKPoYilhLe~Ux;nfOxt0$IOb2I}wSXQ?!7Gv17q zr}Sg14Pv~i$m2O;E+xj=?*d<@F@4;G-^INI|Ml0%y@oYbA9dj;{MY_ogh@Y$-x#bz zD1T9YbPD(G@OJCeOL`CU$$0;KY5hplG4L?a#KaTRO^kmH)`}Q0?@K6W*r6UZ_=unC z^{Ai>ft$uF(*b_KXLd~4)uth8}R;mw0lPhiZaACTPrb>^GEta#E5!CpUZPlVh*EU zL-Rv?mn#W>g5u4(%ECwEr-qC47q1I{&{J{HJrlwwn5Q#l9+nU{;YR!MHTe{nd5rh( z_r={-UB(=@+Q?@qao3V2R9|t<+Kb=lP>xwI#{0|h>g6Py$9HBN%E`dbDcJ8@JE|Uc z`4^^XZvnm*KTUdZ&%46GoBlBC#iYNV&z|ysSZy`^5q!?=#nln#x+yrug(6Oj@epOj z5u@#sF{#`~aPZ0aB)qndX5*9b+Bsb@J`b<$v(@+;9nv@BA9L{C_%^&&zaf0P;>~$l zA8V}xf0>)UPW-^?Yk4m&*}7+&em>V#-{^aCK7Pl|_G54An9UqD{!Tu7y56_S8#C7% z$!7+C|Jklzy`=5LUydJju59v{7dSqc@iyya4Keh*MG54%Yy#|QA<>&l;O+I__ zM6AivT-5P%Gq=;{@q2cOTz=K=h3 zK6^U4tU8!_ig%y~$Y<6q_V++dJEAzKF2`r{+1g72FJu~@h~J3k6w~s-cc4u@ys7N- zcz-=I^?DSO&tBrR{;09DcMEa;<8qdW3m%uvyeg-xT4Jb{ zmCqg4`l`pb;kEhFj^BpY`nxZfeu~UrT%$&C!uk^F?fT2qFCM=|@m_p7{&D=hGXF9C zSwPG_VzmCuVNM#qoIE^(3#4IbQ3}m<#zX z1YYaUWV}hQ^=CGI4c@M;raz0v@zwZWl0H@TrK70(g$xqnYmd{!-w#1@Y&<) z9Nm{K*1U3YV*JD}^nUQ-EAV#Zc*#48d9@p#XUeI0+bYK$xXEtXnnxZpR_SxA3_nHj zri?oLL3pRR)lSTEVzlw@!>?7!Fm;T$i1Bvt@%R;ZZ9LP1-tM~Av>}=KRz&(`eD)OY z5d9RmMlf}#AdjuY9LeX@bEAhAB?e-zlhDQ8#cx^72xnIZHc!ikd zeD*}WHM*_h7&A;>u3~-8#PR?m4zI286#P{Dv1AkKPw|dd4n78N*A_EgCHM%%d+{~+ z|2nkQd zSTVu-2_}!kOW5;n*Vkw!ega-wqlNecytYOw@$>LbYqW`&9%8h9?!xcDYyCVpPWtFY zf%JBJzUhM`eEe$tTt5px7q6W=7U8q-cH?03ufi|IYx`hRF#TL&L*vMNHmBhxeHVTu z>9sx{#Fycl#-x|OYbEAV?qS@aukB>~Wc=@B9`0KIMr!JijjzFT>)ooOco(}E--xf5 zKKQbZNnedmF46a%X8bI?HV3-#Ui{)w`3Fw;O#Vao#rWCMi|e%*F2wTjhe#soL;aHRCcRd_ zY%U@rC0^^lYWx%uUuE{+kLboXlK)B4i|5Eg_&?#<3a$0&P7AHym@7E1#h)m> zcn>ETzXyMn^y0fB+4vd1w7-vFj?3xHzoq!6`0RP8&8mYrE&4j2p=62 zyq{y%avL!d){Kb}*FwGcIe5Fd;w8N&i|_D{O>gQU%7`bXkQlo?)zl#!e;?j%zMC@g z@xQ=3owwyMCqE#@f3EWDzm0CMIB#nrujo7V^N=olGJZ1UhK>)xXYZ!(+bCxyo+G@= zbKaP8#I;fEm7G%(r?n{s|7W~Yn-ZC4v8BN_{o@s@KH~h#^j|r7Eg?=jx2ngl!E0?Y z>!lFyZ_^!P$`NfEAfL&1+0Sd!H#^B^D*h-peRJmMe8LY2-0$=f7r~sGcu(-UPsWMs zz8nrd3yHDoKhx$C{9MJGGHUQA;kEuXd9)Ix(jju=i8t$f_6p*Nc)K}h>Yss6a_|NC84kVzALHN~@Dm+;2Y#6G z(dyTa-{asTuV(%`_yqjBijOz#$G73lef7}3Bj3qBB+NfoOl&c+&9vnu87rSHJZZ(& z5xd~N{rs&g`+VKs1&$Xc-%es@+|RqSe738D`%5c!nAjm=cWbeoR&4wd?oE{M_x(eQ zec6i5B6jHm`+Y51?B+nMS)ZlE?jd&6d35MmVcuXhsH zP24YJT;O*y_|GoA{Ctp@nZMdU`uBU7{aCEEm}~fZqr@c3e25wOoifSz)%a7S559+B z#vvPjKi+&-(rz5W4-Jg37hOhdNyYxrzpEoqXaAdkCZ?X4t;GC{&!PNGoyEJX9mEU~ zW4GU$Jo@op;9l$!~lK{w2KDM>Y6A+y*Z^dif@V&_u{qt7&-Wvc&&XU_#%h&HTX4n zt$i){Dm>e`m|vNp`PGBpf!EqMjQ$qmY zYuAV=_)QM=%fVOS*>*+#mxsou1mB6DFMWx%erxdm!E5uU1%K{(y?s6Sbi6jc!}!H` zZGOcrtHIxn*N#_a4zv>U z5iw`VJjCxU+KK-LukE3(>$!G*MDIf{J_@gmK`MTUL;75N9$xF?Qv7PXwnx?CoABBg zw&GvGA1T`}o*C`LzlhiB@5&=#l|%dS6Y*O8Q}JHBR{vc561+BtO7XYiwf5EGOYzzo zZpE*~YxUcSuXQNjbpyv>yf%lu_(r@ohg0zz@!A~D#s40EuN)t9-X@;qmJu`QQGE{A z;ZMM8bGQvZ7q9hEFFp-#cdljDkcW%oIe2ZKio@T6pD)`W_Ng@dV|eX6AP>Im zz80_baUFgg{#aSQc+S>_ufS`2Krj9&yw-o7pK>jR*Va%R{`U^$r{SM<@OgOCey#px z_)K8l~arh>O^lA9l@LKvje5XVD zGW;%w^mX`tymp?^hX2?h|6cspO8#E@w2SpIglDU8d4Bc0I96NlFvPQgCU3C+o|aid z>G%Y^Hir55S$J)ZmgDE+?bfKNV?F*%ywmki2QhaMqm4s9{#OoVM6TeSlX1#OB<6Ku zw6&Cp@4-`)=*Pf~chgUW_!*Dsb*#iE;rHjZvD|AlPiN6Q09Zv@Em3Zx( zG3rM88b4F!f1P#Cl8E1m*UlL;@rPII=ZuB;bMadLRN~9=TKSFm^?0p+I`Q>*GncJ% zT5BE;;QR4f|3npVO@!C_ClNpSalL+-_*lHwKZW=^@Ry9LpY`0O5?_y>A-(u6R3rX5 zyj}ghr0K-B;0I)SbG+SUjo~0M-_fZi(0k4f= z6@DIGJEv;Gm*BN?sxJJacx{ai;JQz%Wiku_jJ4Dep2g=uc5`Gi;Yd#C#=-`X++Z}urz5{=< ztlu(g{x;#e9efwQ*TE0s2ONBKq099HULAk@9=z7SS$LDbHouDS`xNh`%~g0$&6xRb z)^PB;#>A8}hvyPw%J6*rjJ4N^y}E~dQi;*#=`j9W2Os-$?(yNZen`R3!kz<4DuNIw-^Q zr$b$yUq5Y)A#*MKjKKYQQ_r|tIUa1(?}Mh{8}UcTdIZ12Z_?-C>+o7Xmf<(!wRu#B z--g$YiEa26hw^*z&*Qaq;3?vm>X1GTzX?BEwog1)Ov5+fwR77%{F@Hx%kZDzwRu{H z{|bMEEMN3*TQI%WzFvHfLwe6IIB#&MUmSiHUTa?({(Fb?dH6$rtGBNVAA`4RpO-e* z;Un>!wv3viZt*@!J26?rXnRQ?ekuNZnMbj8Oo>?Ka@~nPUV8C+EaLHZ<0nfWe5b<9 z@pOC<{&wlb-<-+EzlGPvs2tycx9bCwe?9)Mcx{ff2h*$V$A3V2Z4O7=#=ENyJ|4dh zukGdO_zPT~^7HX)@ml|v<7@C*{p#_5a7f>de-*FQzYqT|UhAI-Chlu^ zZT`gLI~~eT$M16R`S?M+Hh;?T=RB#;pL+bocy0c)<8$#^`}*)#;I;J|v6`&~PnfmG zi1izf--y?aLFxE9{4AMXyw91B--_4jUygqrukAne_&4wu$^6B;H|_Xu@rOwtd|%be z&-(DY@DUvAL+5JdoN}MF2S(n(9N46{F#+F!*ZMdE-|CRQAejGovW}~)@+rvRUb*XCdaemY*8BMta7 z@l#~~iRVon_*3wwNiUvP_v6pSYkO$qF9Yq<>Yso&_1E^O4E%+7Z4WBIXW`ZM;}_ty zJ*WYniPw(d9r&N)wKdd_FT$(sU&DQVyxM;Ja=hAp{4eop`|)e>YWwjec(wib`|xV} z@z3DZ_T#tUwLLhJn^uE(t$qpk-FU5i8TfDU+WINLht=uVa~1fb@W;yeE1q*S;HTiV z@#(-%#Lt%L#d&%^J{GTyUt}qN@5Uj00)9STTb~*DC3vlV1^7$wTK`ty&vWn%_|p{c zb+I?&&~Q-Y{w$72QTIKTcvu^y+_)O`| z@m{>sK1|F4Vq#=W;H&=rxe|L1#|yl+o>TA{csOZ=o$O(OPKQO@re%Qx8vjR+F11AXX3SW9dR#h#E+UU!FLJcD1V4` zla2okpFQtAV)d1oQx93^aGB)u7I9iX6ypDipGH2R^HA|SU@Gz3@mfDM;#G(8F{7Q%NqaNTK*P;AGe33)>nfO|V@(c0n@!I^Y z#8=`k8r44QJ0^|zm+*G|?9Txrj@N}hQAzJL?Z;2WkLn|{kBdBF ze#QM#Vzl$HWPAZ$Tc_Fh8}ZsW7vq2C;H&ZZcx|0D<8$$L`@CsmH~umFsC{tk`42IF zBSyQPj;-K3lK4sF7dlpnbFdWrGkCjmFf-3`@K56H#@UN6!9Rju$>&i1W8U$vCFWCN z?B=^EqZNNp!+zf)nTI=2N#l3olkf|q4}M?7c-MnmvpA&p;>+>J%k<*Dcq+aEug!y8 z{IBrm%k;tf?p}VD$y{u}pT}pU?lR-Xpzt`(ofZy!kEAaLBMY0`YEF172;O*9g zsb>w{w*$Yz)N{$0dgcV`8INltF6z1czMq@8b4A?9_ld5x+T?wRcch8c=3grQmw2tt zx%dW$^riUs9Mad~-Ot;%*-QOf@%u=xt#cF8OUw-N&|=Iwh^XXRg_u^Z1w!K+{-XaO zsPXamy^Z^Q3;7)K?(_W}Wqdk5_66=Oj7o3)U8Q{d>09;lmU4VHK66z5*6(<($Cu;h zN-uujb~}C(UYp~6L4S)(?+)A?@wfkB=D9=t;_*}eU_TE2_Twkxwes`vhvQF|<(n^) z@o$s993O|*_R4zv6#Ocg-mE`yf3t&_d}6e@+K+z-ug%rS^)A<6lscF?Cg7iU@EQ0g z9ee@)5xm_#<|Y3M{06+!ak-J0kBB*0_JepgrW5~9yms6jz<=T3qqwyD9B~;#=`r|3`0dx&DpU>Ys%F#KC9b-^OeEP!aw+ z{1jP#@f@cL-|vur6aG`YR{t*iA%E2KAH@5V{JqpEnv3^u@GE8g#+;WX6LU6Y*!7Xg zBO8AT-YJg~Vs0bm8d;9p`hHjqeydW3nddF|I=t2oJ@~irQ)K?4pNH|UI^-W)&3A_! zd8Mt;BQ@bBb)Ec$af0eivRlF1sENct*}I{rpz^%x(79fu?>t@z>$)=Be?n zCwL}_*RJKg_$7F|y~vzPM=^);@Fjfq99UzW8;+byi*uX;@@jIZO9g(jgKxlBJNOR# zTKpNNT_e{h;vJiQ{BQBvdWifr?@c@8pMbyL!Drxa$7}m?0e%DCt}o44RN!kJ@^8Se zckmtfG6&y}Uxl|@k6y};tl_wfcUo_W#Iz8j9eXqJ&pOnh5Wf*0LwTVw63<*J@urWs zU2l1J;1>(T@}oxlK8O4}@%{J=nV$1xbBofXAHc`Fs9(=VafAODytZBv@jr6#nRuT= z{R;8l;I%QR#1G<~#-NFqGac&Cg+Ix`58|EL7`-vjMy-xX_$cyss$(`WS2)z67(d^^ zSL2=9*o-%Aw5y{T=WhJTO8#E_5dK)aQyXG`%kKao=49eSb5&exq~LE-$}n}v!C&j( zOYoOD_!|5Lc)K=w$-f1^0Poa>oy4pqMr(tsmOam*4qp8I4n7rsyMxcg--y@RP>NrP zcWOf&Fq>+pZTA1m`0?}fDCUvNm@i?73L=SZGS>XuCjDu$d~x3)4R6|as`TP} z5PA6TD~HwZe3juhlm2{}Uc8%Jhi`YtzYX7wzd)v^|3`k0L@$0A&t}L8Jucc4H|I;CT4*m;=^d<;c@Dl0zs$i${Ep|T4n7{g)WN6Y za~ym=eu;xG$7ehEdi-Js-;U36@O}8p9el(Vo<%zNc>E&#k&Juj97p_~nRI-bgU`p$ zRlJuh%kk4681@}4^AG$gEAtcov8jQWNb;CwelYTP@5FJq17Ff^f1HjZiCHiG%-8L1 z?p0jNow<>F1e_~`H@f{BtBLC+?(V~p;BzRi)z%L^W`2gujv4P@CM#{B<$u6Y*)ZNjqj_;^*OS z9+f_DGHB8l;@3N*uf$j4XUg>A-ghJZ3A}diyc53;Z#Pd(U$(Li%JA#??0LSxY9oE_ zF0;li=J)gyaoRa)GCu0>`Z;knemY(|CoaaXcvb&B2-Wz>q(58s19iNKDw=+5#?Qjf zkY2ol(~VEUA0_>wK>NJ>YzTiQ<+tG7_%vC*_#1aU_!$m<82|c!{(f%k^JM(Ze&3Jf z@hv>se^4`Hkb>Wj&oJeW>?Pt|v>g1*z54ye61*30H&;#mHTWrtH*>jyxgLpM%V*CI zKeOs!=CXJvu$O#R5*JNesGb*F;%Z5Mk= z37!DtA zJNUkb9KVsjad(wXZCG zo4ER2`+YYZOkARfTUZf(k0*S|!NKt?C2r5Z_WS-y-kzBv&g^+F5X8?-`?{&y zqCNKSCdT92$tN8@&D85~kkp`RkI0Ux-hU zUVOK^48I(IhVtIFs38-Q0wufc&(4y z@n!fUWqNUa(uZG-&z9b8eJ>&6Med*rl0WId1oSi z1O7CbUfc`L#Mk4eOE2E5F2rxa+g(3+NmCh2UoO)J=8FH0u$e1O#F+CQySZTU=)y*HRiFG&V9Ft1S*8ecC#+J3vDl3ne-QVdk{`pVL)FFM1)#Df9 zC(C{j?~JwMQ}Aa;({8#u%(u;S+ z^YPty?U+`M--Vwd(~Eai>VxTzlU}@2*N*>)^qTL(f9&8R{>=HJgOA65g4g;t9lsZ^ zjZZ$l7e7nZUtH^y2h*$Te;nVA|AzeS&QHwz>cbzvYsZj?7Jd)%_j>)~@k69vE$bir zd!r`*bo}L{*Vc7DJ`z@Y->@AK!#OLe@dNcNh5= z_Ab0#8)8Y5fZvSYBG-C&;16{epMg*OVa(q|VE;9zy~Y>dv+&v+slYG5=g9iG0~gFD zeFOe({6(X@_1&Njd;|V4>BW1I{rDOOAGw|L1iaRN3HS|ot^OJKa{M*2eDOY80lp1C zUwUzDs=$AR*N)*0_|Nd__~QrhcE=`jtY~2kOc)-y_d_p_9B0ijQr!EA=H#-FIIWx{ z{C)TN+0;c115bH-m~9+9(eIh_*DF;^D)zgdDb~y zH!%gooG;4=K2tS)(cZ!}H2!4fvggl?*hpXa*JSuH=ICYW;%()A-GTkSESZlxaPz?U zRQzjryS94qx%g)MN>KD@TSWZ=6Ud;$J-ymkz%2$nxrmM^~J z(SRRts9y)(?K)sP7tK8D#}ATT-T(ggFUQZ2>BT+aetb4wtAAu0_W|){orLC~SR)DeyYbxJuzc{B z(3$?tz`udF+Y5{@z_;M1%k<*fpaQ=gzf^j6paV_%2K+ESMtbq?UudTuUaeU;f z&gm1z@fks%BZ{Tulj z_iFKW$4}!E@C)##%JkwrpA7urAL`FE3h;ACf4)pF-t(!zXFKHIfX~NYAk(u)j=bm7 zfv>nj7Fk3UNK;4>Hh{Kx0w zwf&(2zZ9>n&xTOO#{2K9c|HB`J_pjd(FP;Zg;19$5+dtyPHBtlqXovDU@G(VR ztr?Y%S)qYv*{RIIRrpU2CZCr7$>&XTpV#D5O+E=n>iLwD&szLvZt{6l|?6T9Bz8QatYw)ZYgdM|zvUhB71{C)UTVncmG9Y*e3=HfTw zO&^B5xHnvi--6e!d1~>s_+*)${LR-YO#NE%&3LBo`Wy7`h+oPnbV#3!cj2{b{cQZ% z_;Y0b!TW_?epZan#*g~@r)KRoSbJ6tF?-_lJX-LPCphP^lbB62?EAsg!}T`T+xXv^ z&m;58^yRPpQ{3N|al{;bqJCUY!$;vyl6i>db$P+`mq_nU3$39t`~uRSD1Gqxj2ZJH z=I$c=Q+)PZN-naVj|S@L#dVO+--&a&FEBvNe~6hyd?=sbG1lY})yX>-@y>Z95pyar zc5~G18Cm!=ymlT`gipb1{Z)mZkJtLE34aS->#sWI_G-Mpzs}3B+G)mB^w%)?{D-(q z>Ky6|6C>`Uc;9ikjyp;3uT=c0c&9w_h`ELst-s3fh4`7W9@km(uMWQ&KUwKF?)$glX(QbR^^}L5%02p&a^MXjA=YR0zWD}ePu2WO&!wl3HSk-zv;_e zRyzua8B+2vWmMo}Pd=djzPtwfmH5fB4Dk%41D}r9>e!Djz&q73>OIcY672IZbxg$P z;IEc-5Wmwe6aOe)+xH6bTk%f&UKKIvv+R!{Udk=&xevb5k({cYkREaNgNZ*KG?U23`Ux~Nd15Et~@J;xi$@1m% z?|<-4A6`2zO2i+)&y(qI7wyM;Pq&|oru;(u5%@D?dhwjL5+9F0O?vTsy%B#FUMs&7 ze~m-=1NcP_yLfbSRPUgH(4(vj=Rz%$?*%$2y8xIyBkv*-4j zxKt4r{;~gJJ(igCPkDcb{m^c%nmqFHd-2!F`#9zu(9FO)sirr-P2Gi5cASJPQz&SSz~@YltZ)rDVlHqVh{SzL>Zlx6Z5GI_8!oy_M@yO&vOKjt&` zPrTjQ@{%NixnulgeD zti{h##@LImqzvQF;j^cW{~sAc@orWxd89j(;n~SMnhs?QQik#VGV<)pNGFdqN*<;i z`3~NT&;BHk{!Bi5Jfe)?_XA8G4Ldm|QwCS`Ry*An_&dw^4*YifsB3ZYo@zgS4?dOH zP(Lv*M()){e$G9w`TD)u1pH+Dkutq_$20>Uf!B_Q1;O;k%Jky-NCiHM^xE;R0Y3$Q zoJ=q7*>&J!9DF~1szd!EcX6*3KkD2`oUT!rt)YxATD{}o#%uK}!h7-B7*yep!)w>MP55a1G+94!-qscLv!xgBu?*s8kp3{~ z#d~tm|748uXGtG;JuA?De4<1BvhdUKcK29K{}lz(YxS=RdaZs8zYWqE%;&lbn*-Br{aB@9{d3ZKa6*!>-%|ZAID{Ufz021edu_bg0I7CeUO9S zj6XC?)Gyo}I!`OXufOPk?{b;H>4U(dk-+&6F|EWzj>;q4pNAQzR(!)k`+knYH!{~Z z;@{@8=fxzepG-L~TYFl>zqr=A_<-+4Sr_+np|-~3KgBN`<^6{(e?8Oj5gGbe@9&E@+1*FZ6-8foDL(ou?lCVq z;Nw===$Lk^zmkcWbLp7*E9Pc4emS1AUMr8lYXJWKDmG;}q_4&|;r}etySIkMqZyx> zsqcZ^`1A1EF<=OP8GfeB|2nIEG5_Xz#3BD={Ka@}56#A(@8FB^$#}c*HuJg~zu2Mt zX8e4I{JZgIJNO~|Y=`n=`UCx|wJ#aJfbBhU@Y-=M7ry|1l&nMG{=&hg7G|!L5i>VSzo%S>UxwG}(1y>&+ttD3-;2N8!F&FX z`*3)3?HSq=#dT*KJ_pa4g5||Ci!}T#c)M{h<>%oy;kCV{4Bvp)&bR6ud_4JAF^^69 z69<~x&cjWySDNOTr+-T zaxwkeM$BGfwEpeI5977@hB)y==kH<+JB_R?WzHp!lvu%piM( zLmm;#r{0k~zBtJ?j~w#YM~vP5GEmuFSA8Lu2ey#o5Hv3s6^+>?~ zk~H^f>@Rq={rCdB zwl350EAUff{^D6%KHij{D!r&*IldJCzVz-6{{Wi$)#K+}XFm^&Z^x(Nwfgnp&&6}Z zv)Uhg@6F%-|8oBlueKj=(rZ2)pXlK8@zWLWrAp=aXndipU*I_Bf1+yY*g#A!F|%b% z@V=H=3pLE~Tkz&M=J_c*?8v${YoWqg3xnj-OPtnM(L+4T!E49UBz!;qC|Qq1f&9Jv zEDLY)XB)EW5NOL6Cd#y>n3$$z_Rn9;IZZzEwH4pSXU`u_u<{xCz4C!e!NBpKyt1#? z*G@0K6tCSU@qEwsn($L7E7WIV?#AJZ@mhbS;a55MJbV$}Zti#~vJAf*zijL=CCfUd z)Du&erympA@mugG$~K7K@zIBW0k6*gef(Wg{2ZA+c#dnvV~F+h0p33z=P}zx=9bNP zWRcg(8|?QgFL@Q=SL0ulWtlnpJ8Of z50`R!1HaeI-p}5ZPEtvdDG5C=BpIjpCXCQQ2jeVKQ!42=DoF>8B%PazB*{pn z95ND;WRN7uAeE%v`DeAzc1w_Z4cc1TcmvJ5$ZoE<>@2juXRwKsc?k+&82+T5kB8Z%J)in zTWQNb{d39R{nJ;<16@mBuO>@*Z7G-0`RDURes8o~%IivbY3EH_q`Z!lm)1WArTlCu zFKv8k{UyKmIYN1JDX$^r?TWU~|Ihc4ous_Q5kB8n%9}`eN!yixLTXOw9UyFP{dxn&cl=70cp#k|wJ1L(e!rNZ<(Kk}Qhs{T{M`Q%5Ld(RGZ`uGU9?=s`S1I~n@dcCBaC4u zDX%5v#pZ!u53JlV=qu&dNO@^}G+D~SrIfqJyUV4#y_AnG+D5xwdF-F<$~K8xB5^f} z#ufS9cz&MG%$M>|(Q;XD?R_EjqsG`He@&-&)Gs zN_nyE0{4!%%g{~A$4L2+mSKd%Y&=38bESNZl$W*)?@RfrBYb|BlrNF;(#Em8qo2R9 zXg?RZkLnC5|3*Gv(!S0egLYEB>j?Gtl=7`dC?6r^X(=yl`Q}Rb`Xki;zLak|!smBM z`KMCers(qd_4z^W`m{yX6VvO^;f?Zl{xF#d_88yYhpqW%eLAPWXXZ*=X>;W|DW5Lo zp`!H^IoH2m%2!Ew>!RgFu62x)@?BD1+CJ)KDc>sPrR}3ON_jzVIiD(;-(3bhG|!zg z?w6Pr5>wi-ca1`yX({EUwV|n$HPJb5kCL0l$Vq8 zlE%?|*mm@0apVr2M#|{afVz_rr&M=bq`vJsf^# zOBpj%?sWaLtQl3{j45jt6gUUUm?sLIm1WJ$LT7eav!n3D&1KE+WnPl?M?P5ezq0o^ zkC?GO$yBsJ-c35$SM!37#(eH`w&eK+P?$iub zFBf)kKK7ZHeaXT+@q6V+|oK5-WXP>jkF`xLI;g0#y@9fPt!}Fbw9rIScv&S*NUdeMt7Me-<&h`ScFyHy8z^rqe`Gsb3f%9RZdB1?H|F1@8{LbnE zvpCP0T4TC05^YFRjEG;lY3!Kjj%*Fy8#nMYEUF1CNGgmvY++_HBR#v}9F6O=AcRnp= ze(}#Q{IQ&gAEmRYR%a<0>~~@n&0Hz3V7~W%QtrnJ=Glr`-g}zQnOV!c>2qGMX}z8m>MzBA-_^LDSjSX_wTjkoQ>7Z*rVLP=N{$ks&4k>{*IS-1|4fYEbsa; z6^0c|KGytRL)%yGK4-koyv{oRfBx>}ilhC`m??Q~KOdN!XV&?hZv@-@&bxvkdCnri zj67$eU`?Jg&@Ibzc21JAd}o7TLcX(1@VfJXbkCb*oVk+gi!#!vcv)w~b9v^avd#>_ zs~E$3Z%_d+dakzmIO&WIQC%&^+d9(niSiOza? z_sfaS62Xc(PJDWv$<%T7%e%x$&IWn6@+4=OU{5_~+>3eU*^`}v^6r2wyy5kN8N#n_ao!RFgA_ zw%1F?xK)oH=yUdno+sQQeEV^q^U}OL^R;laaNi)Gvsw7`V4t&6IA@5@8986ZT==`Z zpZJ9Djk3Gm%QFMV_?(m!B!n}CZ^>TnH{n;pFN71u`4W{@OW^S7K4;Vd8E@e%;bGw_ z;cGAYoUeqx2@eS8&G0!-C-cmgGkwll;UB`JaQ1AU^QUl)@LS=&IX-9pLg~!8K4-je zne6Gl7H$%*6CRem+_*(~X37Geb5P!|O8T64(h%{} zgzJQ#y&)eQC?A~pmTVngk(REMmI=2A4@$s-w{=ZqCTwMjZl3U+Vuy_PHZ zC?kz~O#*~(3P1S9=RCVC&-^SLAbe|^43}`5@NMDi-}%N@edS$Q%0GPPPkBAh%s=33 zI9NF4Pv7)>x8UQ$zDt*iSR3=t$#ca3***NQTpA->DI7V(@5~lX7mg9mdcyDgC?7~Y z;XnCf;eaRk;JA@~XV4q6G^6~^R^cY$tHJ}qslx9^`#P=Z1;g1sVt?*;vd$JeWCH^nrYT?v`-3&#o9Pxmh`bSvB@8|kk^jGX6p-VuH#9J4ad#N>eS zfV_WWzQ3LuH*|qCK>RY{YT*{)JmIROG~n$#vsXA(IANhQK=|S!|BJZ_K7L(VB4U^D zYvB~xJI`GuEfS6se!oH*AdJZ#_8sBEl;0V>I?uc-+$Zk`z2zTY{Us3}Zt{Pa%lM-l z`pjP=nZA~)gl`IemVgD{(97$-^E*S|$uoZnzjgy<*RxJI@_YY}j=M0cf0A921nd*O zE_`;k-+4he;Ag+He60+R@M+;E!oS4N%aZw>-~7(_buyoYKTE)*J$~mi;iBLDBaU)2 z4UV zW~blTRKdI>t*Bt87F0CjDmc$nG_xy77WrtcurtVS zM)_)_{YJJmulwEn)BI+N|LRFwb*j0c7W$bEO^A zyYcRS#ZDnA)RRNLcBegL;+}Jg@~2C|wN6KKpJsJGS-1b3j0dj$7Q9rp*L9EJ=GAqN;GVAQ z9>Iy1Be*Z?nx6|lEI0Bbv!Tk*6@IU4(kK1XoZP~MoqNr2T2tDRds|to2iz>)0Pp*M zb(<^ffB(v63reFm`@i0wDP4b))7d=w&$3oto_C^r`<3TR&X@DLk~_cD#XnW@UFY0n z?$^&1J-I9GZ@2A5%Sw80m#=7C(Z6%$a@Kix(h0sQr%T<}I5+zqEjQR_o^c1P*g0vD z^UZ%(CFi!sFYsL^6W#Bm^L(-{%Mb7J%I%l^SJ9JO#>CqM?|VI0{?G05k2(6B$Nj!t zZezy#%O(6~k-Osm(=PX%`u`+O*28O^wSIT4c-C)ZGw3dJHD84m&P~1t%Jnx-{j<2V z@u-(%*Y7j3lg+J0WiE0qH@EQjrA`}j8-HKvTxWX9-)^5&K25$D9Oc`Q7jw+Bjs{e2 zC~myCp^kgLsBx}a<(rtl!!ZL2aQfK|zF%eA$n?=Rp6p!iyW6?n$l;h>bg8TSdGfBk zbDQa>pD$lW*vz@wv^_=ZxZYXpH>3T|M}G6ZU$=|Z?~u)>bm@LMx%Y4Nk9W*dY%!}| zA)oH;$l=6V|0QzF_zPRAN^UJzf1kHG-|YRnQHOnXnu&hW=Uyd{L#$Wx&Et8_`}t;L zp7V9S8O>6jN+isy7=gC}F^(0t--bl?2KSe?-MgEI0jq?Q}vd!dr~zI>%s7 ztHo%a^HaWAC%GN-y8rxyBNu>9o#mJza#icUb+lVvv~x?{bbCj{`@U$W7yVoM|G#YO z{#*G!`<0T$;pinkXP?h(@SP+*n>$RAC32Ula%<_Dr)4bsG8SLT)W!dcY>%t9kofD) z8|m}OSKer73t-M+Txe`3rciekE-&S88X~R?QFiXAUuhv7}j`v-%+VA_y zU+xFL`Ljs*Kg(G0VhQW!43`W1ZUdy*vfV7-#JSLSyVK1)@^?z{$N4HfFo?=SGYZ|iw~o#%aL-j}!i5Xv1UnhLy6$=B%aEBC#-F#Xz5x%bEC z{x82PgmMKW_bRhY%iZtn-IuP7^4#7-`cK(WozYt4;@&m7WOl9Gbd7$loWdOQi!Q%r zmHXWAa>A&ul$L0eK8M~ds7kSJ73AMB;QiV=^K*Y$KeVI1x9tp3kuLQ=^&L8_?eF#< zjX%tLedqo=H(Fl4g_=KjZ}U+7|M34&4 z3r#EPhXELbAsB`c7=2IT)DHtN2tzOoBQOeMFb)$i2~#i)GcXH{ zdx_fZe;9y47=mFKfl(NPahQNfn1X4TfmtXw`)d6#0D~|D!!QD)Fb3l=0h2HV(=Y?G z&|FFVFaU!v1j8@_qc8^JFaeV=1=BDCvrxWl(B+2#7=$4hh7lNrF&Kvln1m^qh8dWJ z<|^uk0T_fK7={rTg)tb137CW_n1&gcg(gh>FaU!v1j8@_qc8^JFaeV=1=BDCvrul~ z)BcA67=$4hh7lNrF&Kvln1m^qh8dWJ^4*};4+Ag=Lof^@FbZQZ4ihj5Q!ouPFbn0Y zZmk~%U=W627)D?e#$X&KU=pTa8fIV?%J;KcKMcSi48bsrz$lEtI84AKOu;nFz$`S^ zQ$GyAAPm7UjKC<2!8lC7Buv3H%)l%(9jG4$U=W627)D?e#$X&KU=pTa8fIV?nj5Ge z24E0|U>HVV6vkj2CSVe#U>as%7MdHW9|m9$hF};*U=+q+9425AreGRoU>2JHP(KX7 zAPm7UjKC<2!8lC7Buv3H%)l(n-E1yvKxOx)b9n_|5QbnFMqm`iU>qi35~g4pW?&Y| z{T8`;E5iT`!VnC@2#mrQjKc&>!W2xy49r5)iTYsx24M(>VFX5D48~ysCSeMuVFqTQ zxtaQ500v4@?IC1;`Hu@h1VF-p{1V&*D#$f^`VG5>U24vUNOu`gQ z!wk$qb366J01Uzq48sVF!WfLh1WdvdOv4P!LURZ8!vGAz5DdczjKUa!W2xy49r5)h5BIt24M(>VFX5D48~ysCSeMuVFqTQ z=}P@D0D~|D!!QD)Fb3l=0h2HV(=Y?GP;P?AZTBj}01Uzq48sVF!WfLh1WdvdOv4P! zLeri4VE_hU2!>$U24W2XsgdrG)5g3Ir7>5a%gejPY8JLCUUh0Pd7=$4hh7lNrF&Kvln1m^qh8dWJx!c^` zqi35~g4p zW?&YYho~P0U=W627)D?e#$X&KU=pTa8fIV?nun<$24E0|U>HVV6vkj2CSVe#U>as% z7MecP4+Ag=Lof^@FbZQZ4ihj5Q!ouPFbi|{db;!P5p72R24M(>VFX5D48~ysCSeMu zVFqSl?za|h{f|;V48jl$!w8JR7>vUNOu`gQ!wk$q)0h5-0T_fK7={rTg)tb137CW_ zn1&gch2}BphXELbAsB`c7=3(Y|4hXELbAsB`c7=W2XsgdrG)5g3Ir7>5a%gejPY8JLA;5cR_V48jl$!w8JR7>vUNOu`gQ!wk$q zGno2e00v$U24Jk5QbnFMqm`iU>qi35~g4pW?&Y|FS2y~ zg8>+XAsB`c7=!W2xy49r6LU7;>N z48R}^!7z-#D2%~4Ou!^e!8FXkEHrWIhXELbAsB`c7=2Hj)DHtN2tzOoBQOeMFb)$i2~#i)GcXIy)6@?GFbG31 z3?ncKV=xXAFbPvI4Kpwc&3NjE0T_fK7={rTg)tb137CW_n1&gch2|ORhXELbAsB`c z7=!W2xy49r6LMY*p3FaU!v1j8@_ zqc8^JFaeV=1=BDCv(QYWei(p37=mFKfl(NPahQNfn1X4TfmvuKQ9lg8APm7UjKC<2 z!8lC7Buv3H%)l%(lc^sDU=W627)D?e#$X&KU=pTa8fIV?nkm!|1270fFbpFw3S%%1 z6EF!=Fby*>3(a%X4+Ag=Lof^@FbZQZ4ihj5Q!ouPFbmC8>W2XsgdrG)5g3Ir7>5a% zgejPY8JLA8LH#fQgD?ccFao162IDXRlQ0F-Faxtto*tp=KMcSi48bsrz$lEtI84AK zOu;nFz$`S=s2>Jk5QbnFMqm`iU>qi35~g4pW?&Y|BmZ>yVE_hU2!>$HVV6vkj2CSVe#U>as%7MfYq4+Ag=Lof^@FbZQZ4ihj5Q!ouP zFbmCW>W2XsgdrG)5g3Ir7>5a%gejPY8JLA;4)wzT48jl$!w8JR7>vUNOu`gQ!wk$q zGne{d00v$T7=S?-f?*he zQ5b`9n1D%`f@zq6S!iCRei(p37=mFKfl(NPahQNfn1X4TfmvvlQa=p9APm7UjKC<2 z!8lC7Buv3H%)l%(uTeh?z#t64FpR(`jKMfez$8q;G|a#(G|Q+T24E0|U>HVV6vkj2 zCSVe#U>as%7Rpnu_4p44U=W627)D?e#$X&KU=pTa8fIV?%EQLBei(p37=mFKfl(NP zahQNfn1X4TfmvwYpne#DK^THz7=ck3gK?OENtl9Zn1NYnR!~0-z#t64FpR(`jKMfe zz$8q;G|a#(G;dNr48R}^!7z-#D2%~4Ou!^e!8FXkEHo+VhXELbAsB`c7=W2XsgdrG)5g3Ir7>5a%gejPY8JLA;CH2Dq48jl$!w8JR7>vUNOu`gQ z!wk$q^EUOv01Uzq48sVF!WfLh1WdvdOv4P!LbHncVE_hU2!>$|98LjB9DE(n!L@u z`pEx%CG`IU@g=o?S}A_FtZya7zmm)U&v@DAF>M#`_QP>J?@b=koqM?*=K29V|II#s z>J(W(*H6OR=S8)_FUB_@&qMex@izZd{6TyR;y=WPI_VRutp7v2+s|(??(#PGDkID5 zGykj?_IXif;`3QwZ2XmY$K!k83p{=_zR=?rhjDWo z(zw)MJ@Ssr`cazS_M1;0%;nbG<~4o7)|u4X5dW~p+kRVFO219^>-NFhZwb7&-=6n) z+i%l6-uBxI9)FXxQ~sqilk~JFO8Y0w(B-2ZXWwQL_5CM{g?3#ysq0l@AwYdsm6}4 z9j~`N{bT2Omw zD0y6J;JxE=oX6X7sp;`{T#onn+oYe}?V-unw4m6vVPflHB`L6gyZ2+IWJ@7ISwC}%k%H!&ne_zN`AL>?#tbM_}r&& zdX7Ke6kk{RXZC(=iJgbv+`o@1z9)YF z81+qZ(elsZ;IAE^{#^X0_$FJ_pN>C>Z^`!49uHhBUE{WA;`Pij6@RK%cylumGc;oIkE{|pYy6($HIIFM?ala5TlG?1?llrWGpfELeg(c$S8Znl{5SZjf5-|bFWdggvVd+o zS4`17jfwvczQ=a;wjYxC{8;gR_yXT~kjA^`b?&P`{y}>;|I8NZJF7NZ&-9l#zK;?g z_AKul{HvcAZ_hsbW7ny-+vgMH0lRL!@77d*EA?KD@BO)si|wE8_&%)fc7DE&ulR(< z-$9;j_;HN;S@;iRBDwXBx<~7^<8lZ;@D}xU9M6&sf*U{XVa>mcc{>uHX1%cO*<$@@ z&C`&2>&uGf=2`oqdOHvM<5z~%+i`ye->sjHV%~?0@^WC~=D(&^@pc}E-#A~rd_4EM93LL1`E7gV z;G>@H;Zpn=uWFuFY_UB3X znfPrhbiK3lZ3Vu=TFrAW@f+~dUr=8kUtTWFxb4}sRhL(WDfiliKeV`beB%=|z83ot zJB}aYoB1`q9eJ|&p0v|$=ksL4?$&$GTUxJ7zuc>;OeEKzFaH*h`?B+A6#kHBes0D` zKGt|UZ;uYR+xd#-M9=x-N#fo7of>GoowrT#8^^1+>&1Qe#-|i-=Lqsw`Ag&b8hw3C zeC2%gt?|3?+vaN?yZ>z<6}j#C=05ed{ax_)ey85%*^DofRo{X9z3OV77R|L@+x|KD z*Ew#n`_s?xH@>8KY(HEt6U=Q-ZO?w>9{hUx&9-Np^&@nC+WC;dx4%>U)wI*d3hU;- zai_M^?jIZB*SAz}w`14fz4PZz@$UF`EUWFmnEdw>e?b1}YxiZ>i)ZltXy=8*`%c#O zj5?rsWVOw`n&F3>rtRsB?`q>uQZH$9ukrZi4K&Xw_{HMgcJ8@e{f#~?SeNqz6~Xbc zV!Hh^p{2&#_2OCl_SIT%TR!szzGV;fP4UM_1Kd21+@bwo=T#Wr>RpX*PkantPX3vA z_hr}Lgm~G-|LgS*@h5$&d88S+*RoT!-l5DxJC0lNk)-Cg^Rvcj8ecF^y`3j>@M~LX z{ubo_9KUCOVb0ij`v-o<4;pXh`KdAi+$M)MU{5tk8c0Ij7CWhPoFYvv|^DKTC=LNQ%Gw_#keqp!I@8WB+-M0PE zSOVPq?XK2#THgU*JEHw&*RKTr{hk`%k@mcCrshdA?y`Bwy$<5rUt7FCj}B`5sD0WG zZHS*Q-Yz>ThrfuAbkX+MzBT*^-8zrUO(VBd$z-4 z8*BV6^s_xqcnjZchvu>Kx!lk;B#n5X&e_H!%#(%03?^w9WoHGkWS)Z6(p3BRvU z{obQBzDyI1Z!k!g%dQvo@R7C}Z^tWwZ^-ey?B;T>ukn*_)%dINRnF7=3kPXG*!@Bn zzdNm7-sWEYdf@=6R!^dON<;@pHb=_SkX!3V-L4;_a*? z1LyXqZ<2c3&I|EFnCI70?@Rc`Z2#?i{s#X|8*Pv6hiVsT{#V8pueULN%8AAOGV%7a zRGGs&#IIdZe0ghKtoaABAFR!XCgbD1G{4<`rto8~QE$iBgf!0x&daKkry<^Zo$6Zr zj6bwJc00UGd`a#9n)o|Yx_z+Qw_ou~PSZTHI_F+pFVXh&^Bhk;hhH^K+kXu4AK-uA zs`=&LmCd~l;l0Oc=UuA#U*@>N_CqiH3Ej0lmy`cV{QU#f{|CPmzn}ATyT1R1e`>0> z({8UCUZ(Y4&|mY|dD{lR=q+u3U7uF)D1Q3F;>)`iKYNSjvGeeF*>St`x%Lyf-K#-6 z?-pOuxC|iveb4#w^Y~93t@jG@{ETn&q}JOTe}iK75;w57)X z>bcJL1it?L+WtDk|A_B!nYPEy=Srr-{ zXbx`^zvEerZ$RN;{BE}AcD-oOUh_BSdY~Q0_V~*8>$uqdABJDQRP)<$OyTEoJ<$5a zS84uQyETvX$IFW5E^isOZ!(Q@uQvDxCu)AXj8EVzx?dc~>xSI_jY;66b2Q$z^9%fn zi?!Y{#MhGp2Djcn8JEWR8}LW5-RO;f1%H6!VcVZot|__y?-1|iAIEmg9uM4$-@HWI zBX4uBt@wVNU)cU>Bs(}a&rFWD8xemMzGXdKUTI$LbvHh~K=U`mZ^4)OUVUBsG1qDS zy?r&099HFCH{lOHqP{!+as2W&>g{;FBfbhTMe!x^0nhdDo%kB;@8obL_d4r(ZO=1X zG{0?U5B!x}f3tNxjURD>#xEvMpo8YQ_#^c;{zCi>lhxb)T!BwvZ*Yp4^j*C#VQUk;i+T zWD0pMV!Pd+&%BOr$MJw|e`DEEy7jKPxgZy5>+Ov{$a#~Ef6sc(Yg&_k8-C_Pnx`@T z*p8ZiG4~gJj6V~9Q-6)O^ZaJ~bq934w(WTW-+}F_9hc|v^|=0H+p`59*{uC(eds1_ zPxol?b`Hm%FjV91^1g(B@^x*`S+sv8e%x(3zBayQr;_LY!{Xh3Zou`@n&cUQ53yd@ zd7Hv}&$Bw+ta(x#pV;lsQ}`~I7H`j1{2hFm#9K6fd%g#>zN_^w zXdc`C;rRLwY92Z4$-TDVM--^HSg$Hubc48A5t&fo_jry zuiHd@fBc*Hop%+F_eV5;Pqq&y65koWZK%fE{_Ka}!Fj>C#LvQ4=Xl9(m#W^T`8(aJ z^QtxR4e{RNpU3crW@x>a5dQ(bqUZaIKk(O2*Z5hfnY=J=f=J-=TTBu^((kp0VQH?OQ=vohLTWJmOFPUhAz*eBw^c z^Ah*b1@N=+l`qviwts%Z-{M*C&g-IiZlOQza@~#h9-r^TPhowx<6BJ*wA}UdmZsXy z811~)aFTT?}cKcALo95s0fO^|M=i`@iUN0Zdy~=ji__~*9 z{3H0b`2C*a-u?J#p7A>OF3q!=^I3a-cLjbN$JH&!KN>%GzxIC(eAynFXUA#Up4Rv? z@yDO7-X6E#iT{G{@ohhJ>#2Dz;rmeQ=isYy-Nv@3OfQX(F4FeMV&-0x@S%X#YuD?U z_-0S3xBWao4&-cIq{v~Gcz4`i8Ljy*rSLC&8OGh_sc^T(j~J) zH~v!d+cs>+hcX&3(=zw!5Y_y*{GndLa<3`)Z|bW*9lrwK`U3U#c;+ko2JX8!jrcD2 zX#QOsx3|H+h9A33^VoLA@74I(T(`9S{5pPxr=OksG=6#?%~O&5Gx675r2S9{zX5;K zu!5YC%}(xh>iwE$$a9*}^KUqy_1f)uCwvRfd>DeC#d&UZ+kX6_WraD@4gU%LHqIApn-Aj; zjMDbl_4-=bkh<%8XZGWEo(vT4_Rr8V#oIrD_{bOP?f&cfM>WrzwA0qx5C662d$fD{ zYW%017c`~)bMaa3udwrZBmQnro_8M8Jl^eFqM!OpJ;zrc;G1>V{+~s?XY|+jjV#wE z`1Ubl>d1_aH7%Dj=irBkcl)jRI_-y(d|GhL0PVMPDytvLyc#CnjbF}jaildBVTxJUo+nd*Q2}tMT@G;TQ1}xR1!zTVb%~Npn5$T=F!y2?iS;l4%7TH zjdL&mlbYu`zDKkDc@ci0=X%}m;-8m!KI<}_=QPt)8m4)crL;Y^p9kPqoub}uw_n8n zdaAakHSO7mZ$4MO?S~77YyL~PKk6#tpU3BOo^JE(!|$j4U5Ot$Li1chznz5Nfe&-M zWZPeLq{iRR^^H@B58`_~uI+4$?}gtmOPANqp9e>kygrT-@2-!%yX$21)@nN&(EbDXovqc^$M=Y9e(!yflf>I~lh7P? z65pEjRCdR?SBs}K&!t~$d+hQKz(4h=dV5^)9{$@mwLP|F1?`plR@XmN` zPyT~Cj`sRS2l4K4$T?Hg+xt8pCH}6@)Hfr~F6+78Yv=iC&uISg59@q5hWH!t%Q#-L z+mqqq-FkoG`(L|1ol1PurCM(l^5jp@{LMHYJRN^A{w}U(*1>ncr#2$mC{ ze4{XDkEZM^@wN>`!cD_xwLN!mKGKoG<0h*AjO~E^Yz)7f^G(~&o5kChqWO0dKmAAT z2iwn&PSX5ee6HhbKf4_t<9OH}KO8$*SfJ^Vc6P@Z)HB$#$Sj3mE)+I_yoQi*OBVsU&Alw zdWD@E2gJMMb!t}IUq4SjaLx>^cUnlj9rt#4?{WJ~e8)$1d|OcOi7#oMGX`lMyZyWa z--!EQZT>O%77eu@tY3lO{ITY-{ZM|U=J&IIvGK>@@4ZFiE76{I`1em%Z};QAStYM8 z^~JmMYCiY>JVKtH_$qBRkKMn&fbaOE_Jdu2-^KUYrtwvMn&`FJnm@U?Y);f6|3Q39 z?$fh*2F}s=S{G^_ySzz!%kwpl^>5=Z|3Lk<uFp3ltl=7ku~%60}tU}8KBRt#T6P9RuYA)9J8su+-&+t8L2tP!;dmSms^HM6~ClcSH zhxSi@{2TbaH);N3{aWw~ehA-B+41W8seQc^dY`2-&-9+;aq&pdfI-w zU*3!#!2KjPXT@b&@3x`39ea>G9q_AO()QT)_rZ60OY`4F{O5S@@1eHi3pnqy`>)2Y zYrQu=qZiF6}v4RuFf+KEQg~27fm`!+LG|;WfNr zzh=Gv4bAf@*GVoWPY?XSeY(9m2|ph{jqexg;|o@3o^kAdo8r&J_q$B{`8xate2n|y zTjMLfsd+Bqdk5Q}1M%(8)c!e<_=$MK^|~AIOYs}~YQI&(S4e68J#Dl-_BikweEd<( z--P%<_?bgAz6(BqzkHp>pN(HD-aYPWm#6W~@!N>^-cNY(TiTuzpU^yZJspX^;|BG1 z+*jjQa6W7M$G1}R4B~p`_0&5EALTlKcl?|9`8&0pc00To@BMwj&*JTzAw>=~-`09( zpQ7!w^R^k@dmqA$_mtsTQG9N{ z%yVq)JS@LP^Xz5&a6WlL_yHdj-!I>TKjXCG{WAqWgZnnSkjHsP^LOSr$M$Cf{DyZl zj~%c3@H;%$E6T0aJUuzzvGeds{9dlR+VhU@#JlrtQ76sMZN%o7b(&}OH##nM><-~) z%_{Csc~|3i->SY1%he6v?gaIAy<3Za`djt&iGTe)@-$L^EB+_^gN$Qc{H5=gJg+*4 zceigHf7az{Mf_lV@22YO;8)@Ij#O{^t;z@5Kl8>FvE0BGaB#x zonVFa8t?tx$_4lWey3x{{dWAj-HNyKA$&f^|8`tf;cuL$@u$(A)*n(Y*S&2&OvYcz zJhAmI$6v#7i=CfeCEGv4@Z*PRy&>{^kM~~p=)FPnPvE%7_RlZ)%AW7#mwZfn zUM)V)KgK5}sBcWY1sgTqd!4E_eiFwUc75-R_g;s896#k1&3_j8-@#w@lP<6Ahhsj` zdTacy6IrHx?loAvyWLyLcCQBhGvd2_sd=LK7e3WI-rtR`5pU}vG>7fP58?WlJs#-w znda%h_4W(+P#o|5J@^v*Ue9{B42mx}+=o_M!%fq#bv@eQS&yRP-TdF|*7eTD zKPleb{|@DNFW;|;W)VM%>k^g7{{jADj$12ea*| z^PT#x>$KnOb&qSr+c`iJbC{g-70vV;biO@IVVXQIKd9}o?RU0o{u+G$UXA!m@WC}2 zf1XbZZoywPL-X6?nIG`e-_&^PTYRti_j7&H&d=HS`#s<9e}KP^`@8IM=CwOCPm=Yq z4(%L?zvcJh>**o z7HRzbp6ycAA2rW)%tL#g(oVcP{}1u}!?XRmVC#t=Jzn$M{aTA%n&;OO)z_yVZWC|Y zfXZPi@kg;-Hs)LWqFK6JcD%0mN%MC+px*Y|!}#9UYQ2rA9Ac#d0A_)c7Zv-yw8YW@aX zZ?yZVOYy6oE55vs;}hIhb`uM+8sCWRr=34P;s_OsLQANNPk0vols%f?`+Kz8@y+>OpbPOs z@!snd%kcL;q4WP3;t%6Hao>Vn@0$Kz@_1b<-re85+h3R0&i{w-RsU3P+rJE7hVx>Z z{}+6i>sK9VPqV$6-}_v<&iL`%@7IL*Pi*|af}A-QfAc=g(}4B83jRa`v%P)n&)fgLv`}h!EaotE5m1=@dbG~fH>jr!owpTXKQ}{I;Khz}uyW-vJ z*ctahfAR|D{%=ghgIaH`b-KK^|D*Wm586)K4=eF;o^yLAc{bx`Kd$5672o(T&3}mN z|2EHfd|^WK+rHX}uj<*a`44HH=<6D9+u05u^xWS#8vn+78sDDw%*StgO}*U?|A61h zaZ7*VFFRcFIEKZ$>z(&L?c4BWJ@-+3gFi^UcHW*SPm^@>c&}sk#|IwOaXF3lOvW$a zc;E(nMW5#J9&enBe{PTF=|cQF_@{2sei)1|^lP47kEyrYle6(lf7Lv-iQkLw!uK_H z9M8$qJl@}9_QlWOdi9CqN#M=Tg*jvWdVG-I#oP12)AKcd@A}$KyPdxp@4fyoU%YJ_ zRp#J0nrATE;ocPX!Z#eM`LD#kh#$!Jtv2T`;_cr>^EWEc{MUP~4@?zbQv6TEukd^y z`e>o%d5-T9I#J1V{HRSjE_S>&m+_VOe5CSaHP5JsE|=ZzosU1mbG&+w_1tgSntF$c zFRA^TiLcmR^V{wJ3FS(TZzbNH|Ea4q{v7i6&Ba$V-uq16C%yygwLPw`CSPc}dA!dj zJcPfC?-#n0|Jw2zziEWlYyAlP!G-F>#8;?La{C*Jck8{B<2t)ucg4^3Y(HPchYo1| zG2}mp|MdkO$E)zQDr)}oIq$Rk*$2g!)c+%hukAT+`q_HVbB)fgr1>Ya9<`+2oAKV~ zo(;fz@83@0hw(kjgXB4YAGlnX%eMK<%3ALTuGiV)pBwPr=Om25|HygV$>d*#ueVeC z({7h`iMMqTnuG6Xt@qUx8h;6eL-GCRskiIJBK-Ic>g{~^0DnK%A#GjdtCU=C?Na=? z;@$bvjq^>re;g`xV zr+K{3IjoEC?75%(4*XBYX*(0NXDYrE z=W+J>#`>Hms0gyeZ{vo(i}SQFP}7$?Zj5o<*oM#a67Tkp_c>Ry@UKO6c`FhB4SpH@ za}qvqyylt7d4b)2_Qc=l*&mI?U&HTePxon}{92kP#CWwP|GoHfoTu1%xCH;H=REf? z-g`bCIzjU{I9~ht0`hmoJA98|w;Mz8-s71i_!Ylt|Jda{zqaPD&2g-q53})x`}Xbj z&Bz3H``>%L=yv>IjyG)nukfSTKG@@>G6Bs~aGtj3ZkFp>y!ZKy!|{8#&TqHdYw+{A zU)rvZU*danKR|!-m#L%mWkPv)K7O9>2Az=4ni0w~H^S{do;a_8rB${ke+o`|N(|5xn=gplk3O zuhe>*QST0XMV_l>kI(nun~&D{W48m>%Ldu4H^}zZ?#G|TFIcYm?RAO$_zCRS?08*! zhUPigU-QqRJ&)tP&wGCfUyu8{Y+wC|uew(AoJXF;4K=^_e%&?rB;#wx%hyQbzw}%; zJ^^2Kzb=>EuQkM<>X}y)@%M6HhwZmb_+R-Ra}e!pcBa-lW0dB%^Rp*@JM+ipnTq#5 zXL~WeTwMFXo|kRKPvm-es`2*w=-&7Wp6&K#{2`Ce zJ6H2>;roSIwErr+_rCo{#kHM+%b05Ce039#8K0b$U=eb{K3*I|F?~yMG-R+F`{*We3)ju*{ z^Z$qb8Gx@;OTFD5&cq+&dzQ%zJgohEA$}}=L|gSXegQt7SA4lXz^~sm9{n_7C(tA7f}Po}@*PAkXfu z+Rg_ldxtz%8O`4ZAH7`j-^%%jUEWvl*>^OL?enTFHNIa9&12W!k@#L*2eaGB*YS7N z(Re!#x8S|^fnVNA^H1k`LSx$57e8RD=CSQrg73?HT=w`sgYV6K-FE(*(^~Tn>#ce0 zb@u$e7w9>0yi=MUHX^{Mw&eEA+azV`U`TYO}(#@|4Eqjp;F z2)4tv{jKq99@cm}A8y1~5H&|AD zxjw;P>iNA=UVF`R)o5K_d)(LvznMJty4+Xz4ErzJ{_0n0p07Edwd>17{P-(0zg@1+ z@fG|k@Vp#LEU8?aP;#c5@j@9L|`_sMnd7IVSd2-3snx}O~^U-k9 z#ZUBH-z&UE^Bm&3hh0w_;G5i9Jbwh=yPbNwyrc1-4%6kb{q{C~0rTG;_g-AMM%B z{4Cxs3svS&wu9#1s z?f-NCqxrqpm#)XB#%q3iTsH_mm-DHy48pf~@9#BQ%7LQ0-`wW8Z+R*Hxr;QvZRZyJ z^@EEqSDBkMPhHMa?0NLn_*I_!6=V42J2a17-u9g|PagMI*s*>BUz_cT-CtC^S>uB| zm&5jdcf9xb`2+lRes5xz>sNe7wqthtf7vaX|6a}``qQ7c;lrP6`|bAdJ-qk2ciCI% zAC4dF^4^A@{dw{2z%+cao%%DW_gDOG_Jg+Gv5}I$|6M5F9p6`nX*=!rI2rM_Z%Lg) z%iA>GyC0v5@A9bTvHPzDc<;Zjw;z8t=P!2qd~#>af6*#kuB)iLC%#Eq^V{`zB)+=m zyx`>9HP0g)H`;OUCEh*%9rc&y>CbWOOX34X7nHx3ljoKh+8&#y)g7AO`+Y=j{54#+ zwDC{idlqQD){nz`e{V7qKcrvr<^2r5lj~r1jLP1r%hmZOjbFoYZ+-FZ_|&2tw15c_?*-nGJe ze=pM)f6iEKk6j;k;%9DFe+KP8;V!NBz3ti_JHEH!7jYfQ9w!XMkK{U-jsF0@ab;mH zb7kiFUwH3*={Xo3;xm&N`z0WP&h9B{w)@zrm(LI{K9@o9?`9&Z6cFsra`ZXQj)^pyv z8ejc@=C{WuJ?_=~{hlfA$Ko4tTx5^i8{DVy5AyFgjA7D^!7pe3(imU!evLo)j^?-h z(-_~F``PUB4#j)7CmFo=`LRbmp!vPOQy7o;{yRP&;rsFZZ-{p8!`Jkz*L5G%{4>7P z_S^XD@%M6^Z;wx2#5cXKAeY&0Z$HO-x1am)Lu+e(d!E((A+2`>&$G4tGX(zz_sQ7v z>y3Etzn`)nzk0gnx7P=nKdkwutygcy>kWM6^VQpW>-N$3U4N*z>-*z)@BQja@a_3N zy%qiJe?<42z2|8EJXu}~9uV)Yr*m14?EHKI?|nYTF7mXyLAO^olK;d8Ki^Q9T+D^ND`?{3)a~>davVp zf?XdM;=T8!{D$}5ht{E==J8&4is1)vU!UF1Z@_#1eY(DKAm_Hn`|ndfDc(KaTU17u z*Upp4InN*!!Pn$z&3(&uzFi#C{6k*VakS@m{qYsKKhVyT(*~5>Z_UNK?c6j?^Vs!t zApT;$kF)E=8vMpB#rK2P57ayrxxdb~=Sh6M(VE|0XRG_T#y{sdj!ocObKj%={^b+= zn7g#z5PcLJqgZh@XJ(NPpVx!)|;7uBX}K|KMQFU(IuVcNhL(XI-vk z9M5zcqVc=ApX65ReG2dWy~9Sl_xC&h7kh64UsqM_k5@zmMZ^I_5FaQHiekyl)C3Va zUD~vz4M~BD!{sKqX>Z9*LUL0IBJflYkUf4KmhxM<%LVWqBwI8!>UX=C zSKmwgRN}jSUFT)z_b%cOmG!&k?_KXW?I(OE=uO0*qkqQih=s)G(O+cw562VVobfWl z?ul#}TjHOhUS#9^DDfkh zUtx6hDdM63>YK#h&-jb2!=1!GcCLQc?44as(($wge&8zcI}XW#Weh^6QW<>UhHUVlE;c%E3Lpr0w^;GPy5h z;KB#4Vtsg%r$?}TsJ~qdT=?W)E>^vbppY>_&-}H?~h@a1V50g9Fenr3gYsyVqpSPc;__@5FZgkZ_{FH6A z{~z$XpCEn#>Bstc!|B>jnE$g2@y!`uv-tQC#J_Nk_HXC?C)N-BGqd>J7l?=Vm9{!V zzk5tzZ+?S#c;Ea6;;Zl0@7ft(|Et-`k^95U&oQ~&M0{0;wzvF{GV!o(|FOgyKdJ34 zUUN0^Cs*qFo1J?P@#h&gEOR_>|BjC5yxX+>JmLotKPPzS@u$RhKTg{lKYz*kIZ(%A zYqjFL+J9KDxJtYs@W-A^eE12KgD1XP2lPYYZ=v36`onLCAIbM;&3=5%_w>8l&~ISl z*^_vf|G$cOSXc9O;P$t+h4S;qY~RJWsUgi4Dr67X#2a^&%z&QKVe+>cH)hU+ZvtUPrQ%$pVt2l7im8qo2TQn_U9WuOL1F= zn~9$`Tk#t?p7;Gw`{`@caqdh0*$dp(icQ^*VYdG(--$4}d;{^zXpflPxA~8>|FHh~ zuEaw<@EsRx`>+mVJMr)x&kFHPsQ*vrV2>t#4cEce=WE1|XWn}U+wXdb%y;8y$G%3_ z;l%ay7as*K{AV%aIBjhIHR4w?U&8Fl`-s2gH@eOy?=HMl``I30{f{c_^_uz%z8 zM-boX6WYJ|seVFyBjzpG`mBGI_J3`s_HX=QI`M4>6*qph2l4Rzp!)JK)kIpB44gDd;_kKnE>>ytGIPoz4_|9weyN5il-!*wOkN9Rh z7dD=gh@W$~uA9lbtB8m9)c!(z@AI_1@$F52s^{y8ujssf$9YW$F7&yEeB0<`G4W?v zPs#MC&$xd0$KH?AflL1FQJ25Lp(gLX%YNRwEV<4P5pOMZV7}kJ9P8o4FKE*5?nyqg>^f~9=5LP>|Jy^_emnM4x?bDg z$a$H*bR_Zcy_;VWuTl@Oex4_OEb{|x9o}?<_W$r)-AB{&E5wV(B+t<)#Lu}%ar1ZG zMtu1JI{sFUf4!e+|KWRNdlNt6K5cL7|2^Vi{r|rZzm0J~unS z{%iO(z=i*BT+;DBzCZ9lzN-o7|%P zta^|3eo_eAcrEcAneT1%u>S4Z&u=K-XL0;Xfxnsm?*FK;{cU499;5TiiT^S3$hMSw=j(pE68**`vDt zjoh67A$|e#0mg_o|5E!8>&72QyqSE)PS@w{?B|{L zrq62_xb)xrKJCZer9F;#SU>zi;$c2~@2|9hdQ1C;&1&m+tY4Z z!*L!%{P`t19^+$WJJ4`ph!mY5UmA_UqH$@8@{VFr0n^vuplL zJb#1k%ey(wtsl_uu4JCJ(QOCu=daN5?8g2tCLX>o@C@;=9?n}IRDQlepoi=OT=uIy zh$DZF`1+JnHuo2ZZ^!$mb2*;D-)KMA@_xr*#7`vNN`Kva;@1-o-(`N1_%B(<&-nAQ zhqV9UK+pLk@mCG$`j|fOAn~1!QaUkwqu**jVg7tO@iUn>X!9K)zA^X3*8h0om)xQK z+j)GD_@{SP-0Drucv!zXi*>qYaUI5ppYn5UzcKM=h==uBm;6rq3GYXIn)n^8t88}4 z36E&|gC5oO`A^bY^-;x7dbi>~;CC+}{t(Zd$(N?bw0)SjJePQwuiHsHtdH>(;!6U% z@-2^R|C_QNrrBTbBYqn1eb~ONAs)Vydk^vO{@PAYX#b5r(D|DD`Y7?zm3khxwtXQU zzO(Zq;upL(x$n0V--Py^jq?TKUzn%uXS4s0|6a$l1?z&^I)8!q$4}JuW(QqE{Qj$S z{HDjg^GWTeGdLG5#6SOA?Z?j1QN+W#L*D_Ork`BP_Mc@Qwb`eC&1k>HA5`wNEq$$< z=}xcF3BMn>@R?8ZK9t$B2e|fJ6z|99GVpJ(pWVKx>tOoHGsOGzirYAk{-b_(3F9*V z!Ec;Ld>8r|OfGEmCmnyohPrM>hi$;q)~5no=CwxxdsgI=6MWjrdMCYkM2#lf>6B zZ_4mZp4M@O?@7I%_y*KtZT*)LKlxX>p0hZfcRr*29I{%!YyPVq;x7mOsQ)H@-FLK~ zZP-u4v)a$rtcP!M^BCg&G>;xOaI$mr2Q}Fysjs{IFS1f6Muwt4EovrW#VC6`-9JG|Lf7dw*5MU_zlza zyQi_AYl!d7Iso&DKSlh2M|8fnzq`Dk{eR;!?Z^1xTV7N>XFBzq2KMtF;4GB=d@GJ$ql$_RoG$+nYUgBJtxN*75Af{%<6H<1Z99em?K7>HB*;a2e0Ay;S)*@o?mBHf z*H3s4;&R|N7dCZ2Ze#mR7$>!L*z{G}PfuCb;V}MDf%tZxRs1rJ|5DI5&t6j-UeKsq1R~t%NP%`^K~k4$rn1D zd4msg-Tv#f+CI$dzJz$_$Nn|(Q$MZrU}N6#buxa`6Fx=Pv-oQLz&zkG-%#KFEb#|d z>bx2`p1X*L_0Qk5f%bFl@yYKlCH{WqiyB`o0hjSCc}ClBO@G~Cz|-`H&$FM+CWPsID@>UeBkZ+(OIAJ&sxK)mnk+I~0o|4HI==zlOd{t)qt zf_u#-4iO*x)#N__>i62k?VC(jdH)}s({h5o1 zhxad@B0kuv^D=q8(_6HkM;}tT^E~-L4{$k;Yu>8uA0>V|+s|VC1Y3tkiHG?l+rL%& zzxH|UXBOA($%CM>HD}ZaM{Pdk`I_XI&L#vpIc8;`dmc^#EZbC{hh)2YQe*-g717qa{Z4Y ze$B}`&b>I!-vgKRzj2wi-+}UC+UDBN!JpLjc3w6o{%+<=Z_0lDi+F+kn_Yaf^%LY# zZ~PAJ{{ytkY<-p!KZ^eRgV_JHEwud^k0tl(55&WF7~i+0wh!OyoJ;(!r**!zU)#Mi zeP8wgF6((4>DlzCtBHqoJ737?XQQpOp9lV^-z{@qGl{?FdOa8BZ|o<22L0Q%KGza| zk#Tdw*L#=tU-OOt@h94K-S**kR{^)bwI!CHhp~NFN9&u!Ur)Qq}= z`;YYR&H*m#{P*2;eN5lJn0Q!6;%~(7WZtRC-vw< zy>xxd&iV}T2Ii01TK$ChO+VK5=AVBYxV7=u^98ow`YYPr^u2A~!*K?AnIY#E-vL_vLhs z=Q-lfF+Mn(?H6sY{mi33|03e25`Ug?OpB*%Fv#v^&BlA{?;9||9y7T@!R=r1)jEFtJ(e>-a9bjiHG_2A8FL@wqKy@ zY3KS>;!p0N_;k+eN5DmI+?sl08}SYBAex-(7ynb&-`a0Wyz8Bc8$Zth7k%KJ%eDWv zudf~L!S?qauI%5Mkp0g|a*_U|u-u;(=%lbS;d)(;rVYdJ3h<>-sel}{+{x1mXv^No7&3jVD z2UZjRWKsJuy1I$@>%XnzH~QbGRr`PK3EJND&)LB3Z#fkAUZDE{C4}+p2Wj?Usd2j5AWJQ$8UDuQEdMt&yn4OxY6PM$o1c!VEgdBv*~R* zo`plo-)tR@B|eXO+TQFk->&V$eC87IPVyO}=Mmx$pQYp9f&Cl@+~)3&_GfG#-mCgu zM*H=5*71bzYwtn)k{{^)8oeDy{7m}m7VzU26F;D$<23u?uf)TAjHVChcel7u+uM0O zhWHcRNjkiq_*JvDz1d}(?xOuv13ThF#4lt0e;fZ-i2sFgBGYI7O8lH4POsTV_9<#Fqu*FB3oQP+bR;%l}RMuE4JO zGjO5v!#C3JUi51H!1UepyJs+e={)K`ml8iakVm%?|HymwyGDl(6TgZ1Ip$ZKJ4^e2 z7yU?fAM<45&B44bA|BTFdxZGbuhs8bK1TCwuKz7MFFRkS5fAGV{)+g|C ze_O$E?le#PZwv0l59PvEdHwEr^uzRUJd21=uj+h_ zKlc(}!nm9De>L$w#u?2HdfwU}qWzd0*>!jQZdfPhyTlKr-e}*wg7_^bCC76M@vvUp zGsMID!XKWm-#wasdmH~b#KZSW-nKy7pT@d*Hm`k%pG-Z+&P$1Sm|uN9@&9E0lJUvk z5fA;w4IO%}x21n+2ky%p;G!38NIT-~#Jh=4KV8?y@)8dvzNA6%W7z&2;y(-GChZHg z|Gk6z)T@Yx_dw4ee!{kTPCvx{e@J`|-zhMC<7(ix=4|SIJi_*GTaY~On=jJwe3SZH zBmZa-@nzKiZGBb(PwW4v4E)O(_*Lxx^*8BwO8oBri*-DUFH`-{`1}#T)4uz4wjVfF z^#p74H{yS#zt`xee-G`yOnb`O-%PwYh$FvdiM9{pIMa#$DTvE789u1=VEuOzZ{@o- zhM!D4%vZmJ_}!!{8~>feI~QpGM&}zX)o~ubh2kbJW)KhK+Pe{d@K{}cTjviGKR2)| zk0-uKhqgDqaW(N5=jwOOuK5#io1;&M8+Gb9yZ9c#a{h7_@ju+7{g_<1fcWz_DQ@TP z_r&j@9%A3xYESKd3(9djzg@(4-&XtCliwX79_G*Q-lhFK{9|pu58IzeJiI@3AMwYi zzfEWR4ZG9VVGH2GXLbqlW_Kn2>|y#{o7Z{7AO5MXzwv?pT$bMdSO$J0aQW`_pV$6P zzuIyy#lw1ldE%GsrtM8XECCn2=58n%TlK8<#={U=zt8G4{-#zj% zotN?dbBSNWxRveKBg6-e(|%gn&kp-)KW8yt!p=np@sFLO{YU(Vc=*231|QabHak{)l}6}ZrC+nCac>B)x@e>sS=pGCYtd2Q=|D{!IP zkI`;5KK5I-e~|aK=5l>L_)-0CKkZpthfd%|OKj?XoWu4juGe{)KkZ)P2h+Zu&p+DW zW7_|*)X%N|O^Gl6u+G=)xUGo4_eJgB^u1oaM}0QouKP^&TI4ozhL|Doq{Dj+D}+d>?Y!$q}^onv$R*+ zhxs<2B7WLxotNdMJU~2r&*&|E+Ru?c)%CISy9@BN^*@sB7yL>4w>Cc^{*E)XAB&fj zO4|Qi`n4Ll&ZiLHgz@wW@o)5N`>^ ztR()_nq>cH5)bPizG|iR-%9_0or@X7pS(-gVLOh$Nc^r>Y5xm|A51*FUwpD%5An;GSKQBWuD43R`>unP zZl5APm-v^@(0+^#_anZNaW1o?pC*1H_hkwDdGr3-|BYK}KRXcLo%pt&()rqVhlq#u z#&0Bk8~s{l$89vA{olmAU!%{3#IHF=ziaqsf!n`xDDKDOZ2uoT7dHNPRkZ&w|8pMk zu)gI8@$ml6*NOM!b-u=j?*N`Qo=4e!T5wNr&Y+GbeCOp*;;(x+x$g%IY5Qk`Jf$BH z5AQkt&hXD_KejKM4{JYTfnRq|;A!KpuzmRM?cazWPW{B_>OT+A{&!xb>tp)ZQsQr9 zp1JM&S;RjU=&SD=(SB|W{5C_x&te>T4mbak#2=ZV-!*!^op_k1y2+^a|JZ}NJ~sYN z;@egA9GTvG5%CMDpV+ycR@HvOx`OS*8<^KRo8$Ze@#oLgahm-#du+XFUqc?4$GQRQqqZPRC<*%}&H$X5Q?^9Ow6m zzdonyyf5vb$APDfXWC)f{}oNzemUE>5f9(5IhS}?Px6n%cYjZE-8Ma3`+u1FPbbHD zD)9x|>vv7xxQO^Mmnm-h{sQsNS&G~GeCWS8e%ezO7wh??;$gj}1Bi!q;_<}qVE(Gv zzmE`qn)N8Q=lEauDeXV>7wty;THa$d{iKKZA6s>OY#lBjz9s!Lwnwu*t^IE|tnK&W zch4ps*2TJv_!IMWeT+Z2KZEyLmyB?|i zyr2G}L5_13@zIyne%XWgCB(yc?n}TWUOb)g;x@Mbz-P4o%?kQm(;p5d9=@M)3h^B; z)^*s3^nBJ)?1yq>0sFs#c=!(ZPDg9|qju5n+Bxke{-cW&H@SHn@g12bVC#P=aNARU z(*4-#814U?ft`3K@lxR5p7mL6AKnu=lK6Q8I^UD{kv|Z>{Wcw^?aTJZ%6K-O_RDj0 zU;aS<(oEnYH}5-N``Lm0TtfUZ+9_sVJV^WuzN4yKSn&P2Y3?k zBTmzC+PQwq@!C&V2eOm+0_q!2@w=xHU;M1j>n7rFIzjvS82x58&IQCj!Mc>V9^(G* zYT#-6^*gqI;&7dpwb}A>+J9FNU;6}bYr~;K)lK1Tn~ z6W`){ire{}exi=&h==vHK5(-3^C0tq zjSklkpGSXMAII|};A#7HGu!tuj{gz1e~$Pa-_r3IUH$zF`rRt;rvyu39#E;oV@xzF}NPPGv{q8S_AM+*c|GK{@{m&wPJ8*SemKPmU44fO)NDe)l2b`|_Nc|MtzNYyWpK zeqsD@8S$|0T#@)M$cHU{aUSu$6LtOVzQz~N(Eh{xjqeeEdXBcYbMZ9s578fD>(lgA z?dKDuA3LYpovHW*!Tsyqh@Wz^uG+{@AXFqF*hxZXKB>vM6X+L(Z|42N1pKhzK zYyWS1udajjzcX;7Re!!cZ2xNVTho(2Mf@U+u?9Z$tds`)fbjP%h6o zN85+;TW-&C@#)xZLkM zbLZE)8CksGI`$L#L*8<(_P^Pyb-vqj9TpQ`$@@4nh#y5fj2Hiy_>EuJe(bwX6Mrq^ z1=i2j-_q}1#5@q=t9ug<@7H|I+LO+0-EJe^5Y%mWg7|lTuH)H)s{DDW5eypDAtG})Nhw^tN@tcDFk0O2@?-|)RuO_|;*JlTg|9$6c z|Fd4uc^Th3l=#QUf8NdZUj%Mvhmia61lx!80G}s*8vP7c@Q+Uaj(&IKce*|ch`;%} ziXTru&o;y>#KZgDClU|+ke3n<-?zG*_^*DX^R<2Xv-R_guD|Kqz2DREJV*bV&HV=8 z!nZg1kjlr0c&_hZ`_Mo3<_ol+lY@M``NZ#DuH&)(ze@Zu$~)VaCy4L*HErL)_1XIS z+W+UKD{lUxy@|h<{!*idRm4wwT-U$Aeh$7cJ+68taN!#d->&VAZ~rgxF2)fIU-AQO zANr}zAs*hdxRQ9NA3g(I);X-}zQIM>e^~c*Iq@*Q{3YVQyCb>I_Y%J)sGq;`huZ%Q ztoL;w*Jr&SDgIOTW9Q;qz-{l?)cyFS!#AGx+?RCScH%F8&wk!Pd1U;&(^^0Z;4yOtv4UK51&<=V4T@5e9M z{^=n9e8bDMpG_D?uyOVgU%Zvh%g+0?#J@%R)byMUe@uBdPvzYk`Ry%$JNgguz*~ra zgm&TrwqHU#eD80F_=(KNGkxF;;$M78*VFcKgUj{1Cj|FOwgzr(yycnJ!uHn%{^0|N zhxN6-O8l-pb$uGy*<-}t#QQ-jh;Q~2-N$=3)Nw-naR0XgxUA3hztnzgY(FBtJM+ZM zu6&;O>x1?uT%r98kq(XT-9$XB8?)1u+I}(pLiXJP@zb``@$Am=e~S14zfO+xKH~R% zPv_Od_8nJg|NHV@yq(iw;xAH;*mqAR{zdvR)=)0I;cD&YcUr79Z##8LPe4O}&%x{^_ zaekTjj-S)^}pq{Nt{5CnB6R+3ypQ|Wt^ZFt2&9_#3dk*Rj;$gj@2Z)F7(r$Bu_W!Y`beuNF zZxUbpeeK`kvriKb>k#huGwtWrMs06?l9P#7gS@qyh#$Fwwl_WG1>&LovdN9w|2_1t zn!Z;gzHmS7$K=B2iHCKT?j|1A0eivPe_hwX&fUBJN532Tm&(MqeNfw9%k|&D!bLO%Zk;&0=zA|G$^$c$N`gO#C$Z^KJYG z6YsoC`^mBWorVYgoIem>NxR$T)pLh__X6h28U0^Od^`F@P5-$IxcCDv3F1Cm{5+k{ z%m6NY%^P>oNRwD@zXkWzNUx2?oRFh3f3PmKJWqJ&oNG6 z-z^hAoOYkZyU!=SWspbxGvc9L^LOH5UAy=Fg7alO+WZ6afN#w?`aina{>ks^yo?Y0 z7x6!SMe#EK=w;%Y(JuZN@y@&SyBjedYzN|x5TC`kznzQs{ZiY9?+Jd5czFNi`^1l> zp7{gzzsud)&)xJ3+4)^c{0jO@O}>m0Kl_)uzqbAx-=qB;N4anN^{RUnKP%wFyAprb zdb&Paay*9;zc%p0TuuCC%FW%_e*Ir*|6x6%t%zR`+`k(l-ba6+otI;Q8||@c_v13Q zAEunLdA&gVJIt#$eC2&Qo*Q`&$ojm1_z-UulVu1>iU@c zJps7&>5ucA4E(>@&mi+}%};&{@zWURwYd4;iHG-k8XnN`{0H-?ERHab_%)B}yo`RX z2fnpGK>qu0*nXs4=Vf!*^g-=Ee79&3@f&}u{hPddo_Lrq{m$QLKR>3PY5ng-{IA{m zT|1|T1GhH*aBpY(hTkN|`NoH||Bd$7@mO51o%q)DR~TLWn0Sffx9>hdd~@Qq?;HPC z`#T=@YJ(|0TUr#w=b2S=#O$8#wA*$KGFg|6Lnefo(nBVIa5@iOs4h<}awy(@@+hxpdP{mN^IAHqD49oYT> z;t%s4f$6s|5xch)1)o&Bk>8y~yp3@L8~+8wKX-F- z{I?U|_We2@vyc8jd^74}-(vrVKcV0K$fvX)8|V4NyB}8E==03qYx{do~_Y2~WXA=+Wyxl-NtlKvIkLmuy-GIw@ z!gs4b_$Q9@tGW(luG=W^H}l{9AE&VWwbX~rpYT)SM^fI|zTZ#$na}8Y8a+>cO22zB z*TMWz-NgU=P92Y}{~5$T$$W$DIjEllw>JJbxBIj9AI3X+i1(6yj^Q7DiFkOw=3?SK z^L0GN2ks-@N&ktRi+!Hf?}mE#w}^-J&HhGw7uK7#xs;yKe!8e<8r^=Ac=&GJmB59M zJ@hJF=Q8E*FWLUCpX&H+U!EcUF21j1_WS0~>UYC>iSvoCnWOz2%ym1C_)dY`Jcan7 zfgkc#;++AVzeN0}^v4=KZ1=x9o(=z)JdfuQKlN{~b0e{JyPx=LpU`pI`P%q7?WcH% z_G9bN4?K;|j|Ogi`gCM_MlE)NmkI=H9#2=UYC&o?`ITJK18 zv^v(`KcjbAUumSYqC8qHjTEYtLhr!PU}cg-5tXK79E%HrUPQlU7Aj{ExZoqLzdVY|jcp;#TNlzX#UG`189 zVNZq8^5BYrQoggcKa3&nFS}ji%#>dH%A>=@YVXRFej9U%euI|X)x}uD@{p{3FhTjl zzEXd2Y@k{g80y`>Fg!F+?mZ|k%~}hE;nC_qdudIncdS|-8f;vhTii5SI-t&Phl?ZG zzir2F3uA+IduWM%-;AM+R*O}vy$rlu94H@*{~8)}8`f7Et&WU=9Pr-_vd;4tFP$@c zG5%z%r_fP2;K0&gQ%FmTeWIz1RZDB!oI7(Z!=x_-Wsw#)V(X3U6doIt6-sZNYgvJH z>=|2u15oT4DD|O>eRD14!T!?7o-&p>Zs(RP+G}8S9w#opT2{EbnrrJF8tlXDFpbs4 zk+S@4^MS>3wIk)P-QVW(tGn=bxv|0W0b?aMfsXEKiyLBoS;2j`aiBOh*t>FGX?3}` z)U~MHjbM*bwb;2R*XaIZpHg|n%Iaumu3Z~Bu&XpWG&X|2$Tf7xz`L>f_>V^SA6+GU zZnT#Q8Yz-}gW#96E-MX=4vn-P z=*TPIxjNTGa;Y9PTq;z%7R75(!CF+X7D~FH+g!u&NO`bY=qpx>9gC{@)t$=HB=iv!>dMM$vZu&Zt?Y2&Vr6bDE9hej%CzpRuhLn8+jiUZO!Hci_U zcJG?KN4}6>GOvI$R+ztJuR?wS!3FcWL@Z;9#mM}FwF_eX^y2$E6n@QlG%H7%*9_X+heYE&RHzoefMr0=w-9#EY8cZ7H-in zy0SP@>MIOaM>=M9V{lmf{M^Ql?uNqZd>-p%?FtKuqbsYzO*%SPcUPN-i{+7yMQGMt zZEPxZ^%SZjrBVUBXS7-GXM9Kp> z^ivuDN#{nZdEj6TP#UU3G7kKA^GbZ&{ri&c_5$Sh9z%TvEIDMY{071iy|ov>jptGH zxL@|U-`cM_N6JtkMoRJ*CY4-wGvm2o7GRa59gBM!^7{Wd3^R`{j-JImXxG)<$mO!rNt!Bl=RspRTg%pn66|wW5uZcigO%bM2=jwWi9c*C z?Cnl{p$9{;qo0e=G8C)1jq(=_@+oW#Bo4T`$vSLv@6a#^v7>X*K6PTRx$=hOpS$nKDi9@a`=KWilO0o zBH4!y>e{1w2~H^ft9PV44DuAx2K&gh#w#>Z7=RWF@u&2i@l8-J)G4Taxm>OhgM&hW zFMtX=dhr)IXn|0Jm5_R6fW1OP`DVADDyDG)s=4^s3_=EE9G1B<6FsZB!m=@_&IQ@$ zj>VPc{6cgx6MO;6_h4_SaNtOB7`!1#N}z>~q`=LG(j-mdVz2W{yW5Op)0Kl(RbV*4 zVn@IHl^boW8RxK4gIIPPmxwKNW~^(>%0ZFBC7Q6Y;6_5y&2A{zuF0-^TW@i=*b82n zUp;Hq{KXw}<}P1eXqwSHqY=i3@W3|XUL6(N_eJ=(Npm$hH4zi*=x%k4YoIhb3ZW#E z0T8VZ79dI8RkWyvRE*WM!z|;hK&LKOio-I_NsbN^^G0IayP^x+R2uAyh-|87I@b&y zxuOJN3qFn$EhDeCJ0Xp8IHXJ6s##MDA*WA$9IJ2o?W{Y&0YjgjTYiw?6YHVr3{^r~&4T|uJ*L(t`;24_lE90xNBT>MX)3GAjJhBt(*ss1Jh!?@{Q5x1(Bum{JX|q9-O`Li__yhY%7;c1nXIMr8r! zWeuS-*QS=7lfsm{LK|)sXPQ1EeK;+>mL)932*#6f^|(f}k+b78Z3QX=B7c3|*VGAT zBcv{lODq%+ z`-Fa0LLG#H+oXzR4n8Wk1_gK-M5pY)os-2@d} z8eNdXPNhZ%fu<-D+PXKpZ&|_DI#cux98;*gc`CWVpJPr`^5IZ}(*^E3?k=Skc()T< zs(&r%F0>Uo23HU54~+}%lB%d?9Sb{mU;WCvN3c98M+8fnhX&U@yU@L;(9t~y|0O@? z;pgrJh=lG1SPJ}~MxU;`!C% z>}KbH<0A8hX^?osN!fcis(h{)Qx7^ZhD24<5xH=_sU8@ypM|hvODVaqsXj6pgb*i| zn^s=L6+!&q;9|OoCqpzV4pGd;Sms8&0YguQn+)iFx>void2g;Y>#cf_x?+cGZ6*MqeZ-$0L-1c%4Y4`Qc%ev0^TQ zDxjhKyShRgEy6e3=Tx+OMVubWC&hhc(rJ;m#({XMdd>}1|FNPTVN}H5>;3uKA`CH4 zZFpYy!RX2!lknBDOM0*B>ILQ6A+5t)K73Eshu&R%hd!NE`8 znK=+Z>hL%jXw9WNb2bG8@9_g@Q}YH@nc<0T9zC!)3_0!;d*9_``hE5+qEnyGHS{UN z#@O?nn3mgr9F0hehb(nke6MKn>e}&qh9XBNtHMLvf=nXwI`#=dk(A%1y9Bd5f=;c0 z62FI2h{bK$hneU%BJ-mI!qYWWEiZdzQkC3rmGqQEOcYx^HFr*KYVMo~u<5yK?S*Pp zOAjKugJxoojzC#Um5Ba-jwsII#2RG@U(kGa!e5$cIfJO{8;S4`*9hoZjin_^#ZLs_ zoTD4UG1OU-Jg*S}s${6oT*8s6Gg_*;h}Jv_Tf#$Ma;yg`h3N~vstF%ZycUi{)St1j z?nSwVSfsZ=)WadBA;(}K@c}u_G?Gp*UMoXzHk!*%JbQ?4Bw`OuV{lqaR!d(&Jg!slI$>Q_ zLQix`eN9A54uuy&gQeyZn%PMBDvbjAMLLeZ?Z*1gEglvA2PN*0n0q&7)~SA)$ofsO_5>bMjOy8TLPF|M+Ct zQdkJjfzo7ak|rmu$iy2NtgOFg(UwO3QQP%YRYRJKKdrn*H5F{TwWs~!g`O61Ia@}> zg_wu7t;ZclYEl8MR*5x zmHA|oTB_~sNmEbh4-~XqLw;e6V*{!^obIyvH1@s42WQS8R@pIe5XetC``kI~n;OU+ zo#BTn@(lz>o7CDxs5U9AVpielYs)%>VI*QM`CzXa(GB(rXA}{|MojQu$bW?yo!+*(7I6=GTI{y`tCseI%a?ZB`0#6k^M;+zFA8dOx|Pc&kl0Fe z)D{WciKI3RlhdukVBO3TeB}~tS=fb@5sxewld$1vpk^$(C5yacF;7?ps#}%MBhs=* zl&Kd-z$2Ho??$npQ>4IQY0?U~uvA1YAhJwM2zmTGDPzVtGcn&3 z|ClpGTJ$jBFikM1PIshysW4YkR#g932T_0cGKFGYSRqxF2_#{>O?8&RsS74avMISN zQq?KA?vVSjk4T~PciDCZOk=XrK!hWdHTRun7ZB*Ih&K~Q4@|2ukc1=V(fuc2@F`2Uu!x^|#wuD4VxS{4y5x z>V%xYI<0n_;&OiQUm{7em&U0$o~?#$%d}e|(-w;tr6%JiiXUfxCmL_9);Xa=A&gMu zsc33mxIv{KjFxUF^7~5!6kaRH#s4eO8{z#8Ox=|eg_$wELVnY1so0v8(b3(7)$zj; z{$VsOU;i{lYb)%Q%7v%C{{vxqs8SGWkQob|J+?5v5Qh*M4HjF2B**B*F6>x{e@8=^ zVk4q(eMr1!0B-px<$ZStn#q5Oq9rsALT-L{uREAE-Nh#ZIVqFNkjnRTEs|rG!iJFU zDSQ!8KR4vb`duyV;rVvk3bXr)l0Jvo|7#@GAngot4(qH_j1>guRC#g~QY=JNl;D<)#P2XG)-=k(I-w|t9+?^aV87gtaBd~TbqUm^)b%^b5? z!OalHlZ}((hTLJ~amm3&K5L!mNlRb3-%=zsWC&q^f2xhGl7$5?T(P$oX?l@Mtxzql zASoq%pbhzdig^?{TIXZ+m3b#*dVjwY=s8W|Re_$HIa);eoMx=+MG`nOk(-dgsDyz` zUh6Bv5_7vjoo`V^bdfl#IQ2lLD{hJ78OfNbMoFfa^y+20M5(XQz#KOT?Q&u8B8a>Z zr@Ee?PkJt;%0j1Kc0scEI%cir6Pxueb1?GxYs8IGrro)6I$#Y6ABjX#%{t3Cv*;S1 z4jctab&}X+dDL@V4%yt|*l2N}d#JyP9CeAS$H~;7t%|A9U~)S*|2^v;KEP0(b#XVPc)oI%P|`tEB^E3$mk|k=wTdIgyrv2wf#k zyH1@)NP;!vX~cLmD2uNl64Kuh-`5iUoZ_)v1jf$nu393jAtcc#5XIJf-(|y=4mUKS zHX0Tn^0hM(?6M`+RV3KEEdmn9oYn7+e;1`_a*<(Gqa+5bLr%XB;!?#9FpD&%X2&8j zIIO9HDfvB6UzUP`906ry2f!eejhRGxfMqeoThSUPuIQFY)}BmQGg_R&3#INK(BL&( z9{HAxT@`x9diO6?NAsdv6)ZU*R;l#ps0{Hf(G-0aH&59T6ADu3pP zHB4?7IC8TxqCM_!S%*kI!9`y9A4Z4zJJMapVBli3@KdQL7TIix3q4VY4IYJBqoF?c zk&qo^4vwH$M_@HD`*{MbFQ_k~s70=e+7mqviMrT}>Z7q$k`oni0Qww4Vm-L^jkHIz za%ACTnM^ecb2ZE2xsKFxC~k6jOfCz+=ZwIcBzmcLRiKQqVG>htN_By;2}7p8YmT#q zqWc3Ka1qJAIflh6Y^h8c`#Yh$aL@NHi~dgtBF1+)yqybtjPMIafu(?VrA49A8J07j zuyCff+XWWqRYsPg^Y!9J3yud(8`;=I26A+*tkJtxmOX4)(8zvwq33k@>FHTe^Th${ zeB_=t&lDPZ#j-UTZ2DknO|{q$`>?;?@#nV4=opt%r9`B0Q5Vxl<>IwwN_*tZn_TKa zJmhZNBDG8M$nlRT+{UPVRnW=p zarkb0WKwy2LP<)d=ft|0WXeNKfqAQtGRJtX)z*G?wK^!b6zeX(OrhR6SqI*IYyu9F zrHF;OYlw_?E$YODVKgJ7uaf~9GYw%0wE{2-i+WNswnDl}QWpeQf4Px<2j;;v2G?E1 zP07qVnb{)6dCi=IQSLw@>}<-}lzjH%Yb@%$_f?h}uM`FwYjSH_VGCSZrZ`Q=UIHG#e;T5sOp*QYnpBE=;I{mN;Q;#ORQ zC~N)m=F)q~g(k}oC94jHAjRp7Z{iowq`(<)`({8c_ToaJ*l(dzAEXi3?xGuZfyLBZ z2b9qGLdavf&+Q1shQYt#<^C4E*`Mh^PF4GjkLV0Nsnfb5TpUToY%s6Ko&RP3QVU!iL@)71BmW`35trU66W9al8gAZv`~ngZf3}ZgwvW|IPppfj{4lX zdRJXiv-w=6m%s5-z5vsvxneww-9rPT`DCgd*?-b!D}&6?$>fNuA{ZT}fkYZ55=RLq_q^ji+0|{I(&S4WJzF{l>DUltlbqd#q8Zv(3o2!yA ziNv?AZcTji7uoaZR7w>%bz<)ymU@!9OdePA_7JPz+@fZY(KApl-wgat%NKMe?lVsC zYN7aVr~tLCLNHgd?CFX2$S#e4mhp$o_Gg$l?-A7Gl2hTB!|3C?36kqmk1}z!Ac!L% z7}T$>(5dAHrOXDE{c3ZA>cm0(<)?~MUsOXUkt{surRD;~I!`WOMvRJ0tSi8C(%H}@m3`8** z6fi4R=q_^$Aazf?*km+R=PgRt;?*-R^+9CsE&hQ^k`25hTNBW3lkzOJAw5rVwe<+R zr*;X|YZbAA+$kHDW$XUOBl^_g!egm_H!l?y-9=uGDNl= z@wGSVq(g3wZHAE!^_J)Y#W-SPovdpliLiFE$f(zVnqiTX18WKEEWSTKP#IVU*QCK( zKoUtC&{7E9wSC_r)eK+pt2$RhTtJR-A-Kvb-NPtT=uG5jeq3zwnB=MQ9Kq2qT#A6I zi15e56JLRdua=V4yregNHa@b!_R>~&rv%HHi0$DltULEwr-TUIiNI%UP!rD+^NikT zQ5+BQI?Gnt{k+&ZLHoEbYC`cc1`OSl{7g@MOP^G_3+`TWyHig*l3}vqqb!AC)75NV zY9_Ja5Dz|ilHH;46Z}&-iiC*7wDBTDlHN&8BP%Hs;!89N`SnEPq{N6PbIUT#2CLB+ zn~>>Y7t6)d;-#lcA{Nu`QMG1?cmkUW9Ydpq-7D4pf#QnMZspKX?KLz=u|s|J2uCvf93a6SR~YCZ2v)vegVaK~T?ACa;i zRPH^o7#Q&bNn56K!1+wF+hYl%S^3g8B_-?Rx4Xj#}uAiN}#{>0Ut6NnzY%5eXJ85k~t;S0be&W z5t)*_fxk+vCgppEg5%xq$!A8o&`Z740^fm&?XRf-o93%f7c?(0AAuak8C6#X5l|%>NK9%DSLVSzNJ!>Z()0IQ;yW$SyR`a;jE@4UXB~+K%|e{*G(!6NJPI(10H#VskgGJ3l! zAigzutYDc`j8*rGoKeJ3!Q=i-oW1aG7kftK2k;=Jh3vJfZC)R<_KxGYYION)rR&l*cu9KB_(CIpRAz7iY!#f#O z_Y3*E-SdgKe(W<-bvAo_QoU7-iVQ5b356=_qzcSVrV0dh zQ`TT3u^_sxiUlwwIB-ybKyq_oz2YBbP^LDe(yHk**?wH%{n-b=DT`fvrNSkTnqn}u zx<#?Sg8EQ2dhewh8+7EqZ<5Ox!Ny=@DY6yz91~JJj*DYsajyRxrDyAS6ZZ53hPMPY z#z~zn;M0j^PD~n=vgBY!emL(EL7f}4wJ{R#X(U;gS1-!odIcU4h+iW~aXv}rnmbQY zc$WuLM|LP@+p*>2y7i2y!WT_YViHB^u^RDEJ_(x<`OS_jU*N~)!&fVeNYAeIH zC2m5=u1)xiuo})(;!7j)O2)W}HEHxAK0+s!xhnReGfo+C*sW>}%)0U4iGX_OyFMnCH6~u3;MJn^XG?A670KV8F!=+HICdjic4rPZ;Q7xvW=ZeEl!f z{L?iRSPZb%JEDi*#DoZ?--OhX2^23~r0f8>0uYk1*&KQcJ;xH>SCHi{2zA)`pJKJ=B%J!~+Mzs<$P zBBz?8Epf7di`2O-j|Z~oh(pL+uKd7n1(t!O)LUd2bA5La>Yb%L6rm0wwd=E|fX*ip zfbl#Y;-g@>qIln-&Bv}SKkP&W0l$|_-zad9>a&^XSp3W2kSi8}5G1d)>@q zd{u_OB7W!W^|c|4?edu3hL3Kez!{pj32*iQgHiqyaP>%e9k?+PS!SKkSe@6=C!<`8 z?(RP2>dK{3^>hM-bIsl6j2#;cKc3bn zV@`bp5SLk1U?bX)RG8xGEA`hUaPPaEH!Wnf=rf{28~54&Uvr(g%-%_4FE6$1Ww$!MboLcY^S8LG ztYKA(Y_98GH-l(T=7i&U#ebJojjV+K8}0#d%c^@IBW&-RGLaw>#X7KsCVs;MnZ*sz z;>x1(wj$XJOf8I*%UJ%sZ`HVPse3x6r-UF<$rT)dQjOITm1J2rRtW1C2dP}vSABV@ z;x<+~U%4757;1dUMQf*s8lUBDiOU#@ijO1|)H0T9u`rhqOBQ|KS+FM0+HfJGTPkBq zXgl&%TIyZ$Sp%n(dlC=V_)^|F7O$ZzeyJidn;rTwSkHqPdpOhX_Jes&s6vKw{|pQr zC{3(lg_0PJQ@yVDC+`&v!GnLfQb|t z_w2yQ5+lyHqM3RFbAWlqzG;;j|#ka6wGkxOtjIZh1LI+Uq@94AP6!|*Qsu#AiO zDR6@y)?|swNvCZ)$Tzt4v-S-bs(Q)C+;M#ZW^&5e0E9e8+U(BKNuT@LxbP$j zv0wKlH66*7#Dic#ou1}c`Cf}2RVkb5(%QTW_Si?<{${vq{}OU4hbqJ1CX%e>UNcSK zOwzva647I#+qI=$JmF@M>y(sqyQpS{$f$Ig%;M@$X$)dla8BV{D@DiK8D$;ZjPu8cE(9>5W>XL@WzcT#OazAF{sG`62TGW zPdNP|)qtqYZFdVUbyWEVV`N5~qo)o77vGv!>~$#=9?Tp^G3(9dCfFY7Ver&ou|I{p zu)iS0ytbU|&Y6sPwx`6CWGmy76b7mAN3mR43*mQ9Nc&RLwq2f-^;8AhqVy#N{eiL~ z+L^qkDLjijiil%oA4Q=S|2j{9rGFh<=K-~=Jf2FDgJk_eId4wAwh&lK{}91WFU(?gcxLa)S+nLYUp{+IMG3Iv>pW?>a{YE3D?p>QpVg;s?(sgAaNO9fX9ifz^{M6bs1MoVcnBJCfS2W_2LF z1-32UaLV4~gdra1YfCDfr+<&fQ2cU;@KL#4mN+9gOAu#F?MbS-C3eL9rST+h@K|ls zeIFdd1JA6FkCzVeJ@|;z@|WxPJgvsjXZ#qNkO66EY*z=lv%bH{}wWuZ^ z0glqO@XRChti`Zf8Z~_C^S#_mOKw|IITs@U_JFIiIcV-s z5wRsaD(zmQYnda)o8`kH^fj@|+%d19OYHxuw3N6hBG~SL=URqGkWlDC3x%HkTx;i| zD)fj*>!);cCP5;$88?w*oDn}3g(!X(BL0*5?(irymCLwA!3MboeKY_{r@Y2nuW)@B zhlqzlvqq0@a7IF&A93B9B5PV`AHq1)3P62@zM;r}lXwpR3zweEiMJ9so-8{);KOz< zr{Zi!Al-M=%tSH~6nlAF3crE3roF~nvo=3oTg1;+(8?Bw&7JbDZsb=;ijekv+0=1F zrZu%i(nC77zqx!xoBG4|Hc+v$bzC_;!0CvesQI>u^Io{)IL-GH;4aeY-yd@+}6~ZYr%iy1Zi$; zZ*Fg$+1M_Kv}_0kO>@9V}?GQ;c+YT+|6qgCEZ|(XO0=&3Zs?I8$M_ovX zsFs-Ci4$q`?kqN`e1|bQ`?Wl{dI)d4{!_A^GpZWMoRG>RcY={GiKvOoo0Q@^xD;G2 zmvPW|B@WRCH+;~}=-5CN9z}U0T0&ZC>MZTvL$WNpmpHS+qt7&FfF#111jGD*`)bSkPF);pld{it~a%8$%H#H0I%g)J?O-{Oh* zfa4BkQb%Ya&!`wN0XwO(^GpX zn>*`>#H?~cLCtZ;Wc_&~Nr`e~jYcF=M)X5@u-9!cK|A!^cF-r3pVYUJ;KhN5L2d3K z#6wXO8HY(u5|q(|4IS||-=$O6sE9Rtx`CZVY_{3Nv%b_%f1qr;T3ERcDVb<=9n zwXhKl$nGu@+Yj``$G`4;bTyry*}@cs(wg4Vu*j?+F~L=-r#5PbH1Sv-)IkJGqkutt zEYtiu!j+&F1OP$O}eK3ImbMgbH2fXd#4Ko<<(}JnjCa5n4=@R zyOwK1?qgqBmKpxXk+LW}xduF2UhY2#LBUe!?aAdpWXtxRpc8uBW9%r_Q=!vd^Z}0s zW#;PE*>tBiet1tFf7`pB0j&@3p;s$rXaP`BD)Gu_q*Th$%w`732ytyhFZ(Z36_2n0 zMY5v0-jkI9ImI`MdDc1bNSRJNBkrF}KLnp9x>G85__^}VAM5!^eDkEfI7vH6|H#R7 zzFrMcxHksIN}@N2JK2;6>~mUG2sG)TO)Uhbx1gu=L;+OqC~v?EVBo0$ICG-pVd#(l zfHEU~0Qv&*?)Be9;y@dc3!&kdqa`p$TC!?UOeMcn;>~i?9y|p1wA|a8U6Lgp-qV_h z1r^&Y;@(e?(?m6UbP*S6onk2p&sQ1Umpz$a9Rb0pk-O`Dlidfbi#E~7ZiS&1x%U=F<^rDk31LhQD>5Fx1 z)dVEg3`bE-)&V>nQ%|nNGJ7_IdvDOhcg5X}5XYY@b|!8ecOI!>@RfCtDcE09W3myu zO8u)UtHd2vlf&e`>_^vIvZCvau5?QmL^)mpqjy656-pCWDe)xt22oOqb{{m@i>Ge~ z=~BpEkVH;|9&9VH7aVMhJw?ae>E!A$^CFH;)@0Ruho%%o60w6RIg+Wa2k>Cw7800H z1srx^OOz4mNWr%$um@<3Z&?B#xYXKr7J?({pPwY_5YIz|4R%gNzQz5$d_-Q4{T znOla3#wm1I8Rclf{*keHql|XaF>WUJk#XY_X_;}pmY7DJ;*Cy?D@FVj28+xanrk~IZ|#*#Zl<Nz-hDL}1x#3dQ=KTO2I+c)G5-KL>gx>MPz& zVkKwzv5kcF@=BYSQUJ}#F*OQWMh2ym^-j3A6PdQT<@;FxBLb88bJpZ8Nn z$CF7}6OZ~N19uUUK0Te{-+)e|$5se*LYv#3%ydx_@p-@HiATJRQVOivf{!*)GDMhr z$UC4;(NOT8f1#R(d-sw+9~as3vwqvLgQY5RHG8AD1f*45lmJtUsvmNSLI~sf3}AoA zeSa_%yX-%%3YY48%6fGPy=A)4${3!Nlyq;-CoF1Iy(Gu7NxJbyO|5+-k-Zbot8rc%fu*Hj;q@9|E@0u zLvGqr!IIbQ!ryh%N#YX|Mn?Z(Q^q`0^*)iR%rUOzypiH0opk;|D#gcIyIZ9A0ywlp zz0#QUd@dT_^Xcda*pssmW9FOyPfzPwZ|3x6iZF7p4!nSHv<(Jx+scpVZ1@TPCQ z+YwRFF0f9ryhCxbI0R${>MJ@kDWK3dGBlih@q<(Im4k7gG=yPlm-gqAJW5j6=F*r1m*`fnDH~2?xoyxgN zE&Zw4qKyr`14sfYR7-=SLy?^TTP59+jLhQH{J>P7L|}wVmZ-#VXo%;U#&EL*6Zag$wl!Sj3ou%ze%3q|=8YbdQdjZ)L zFQc>k94Dz#%{2e=Fy2wu1(vG~O>p6ZqxSX=%pF2ZRq5WCa=bBqpa+W_Uk1kKBg5M@ z^4-hEBjP3r5Ask~=A)yzhGK8;7;5XoXEwE5V3%grgHCC1A?5vwxVVpo&m>`K#s}Pel>negIW|Q^}@)P_H#HNcrp0S$6C}M8y zdtT}VwkPC$iBx9gP@eNfF61`agDV^slJZO+-(?(dT1T!ac3JTkPZTDoP|ze0&3dZaF#R?xu?Q8o>vAw%`S6 zGWRPjq!47oTdI#OI17<928BqVmB1;JjDxyVeaG&>Jo!6UCqT-(bs|qjS%nKa>8?h% zH|xn0$pnYcvqVo7DCk5o@4?DVomdx~?`_xhNZ&XJ?P#Is_feS?owq6I8yZ-K+2P-RX?IT|Oum*X&Z>LIGIs5PIiwL= z#FRD6IyB4^(S2Zf{^GsF`#Vw|t`3d3<$xqjOuhD15!|;^>JVjQEv^apvR8yEF%!Fi zp(=VL(~}o*+Uz1uque>ri#VcsrSj2SqHsu^sxRmu$s-~*lA#hrmwD4dka$eS|FEk= zqz4p<-=?EQbcbjT>KjP9f0V}GD(U(7e@P^A84sl4@W~PU9*UOnb_g>Q)EP7)lW|^9 zmsW3gdF+pTmS@wh`#N3xy~KR&Z(WelB}^i`fXqf$wLr38CGIVG{opj&_2lZ6rm1JR zh5VsTqC};*rZ9^90}G4TLey@4?IkNrf+s0stz#d1t8)Vk;*GWnGeAHLcq=fX@FeM_ zZ9`0*BC9<6SNq42b7iOu{khkh(tW3aJXL^Z9 zY_F(F-6uP}JTLw8Yzwu<7>W~45RZiM#i014)D&!FeKp0U%3BeTNaIRLmrZrC_ki8V z{l1=1nG)HU`)=`$vxvQIU})&v!y4>vmd-0SA3 z{N|iJi+t{t%auow^-wLZ7#ib)IN7vSH7B^22_|`ulk-GOARCfbVq@DAhM1@|kz5aw zzjN{;bm+kHubqy^m=ao6QXEAdQ<-IJtZt1I25_k~@^NP9DmadlJ@(K{qY4bZ{Vg7q zH1ou*jOo4sM+z536iDz(xfxbtq1W7h4uj6A&i^R$QI-d!6GvPN@Y^B5X+pU@QRe@! zRk7hgM=i5|lPF6vz`f`A0dA=3kye45tUbj6311g*DO#%*WO(4Zx@&LdSyjT8(Sfq7 z$r7{W#K|!7QJp4M>Qm`b8cXV%n$NW<;Hd$alnk6cS`n$_ht3%yh@@UUw+zO#E^Ke7 z8pK1%j(vqD)VP^oErpKQED;9a7XmLMImu(i5@JoHG@?BqR9YSF1#ygYz%Bw)lppM2M@dt1UIO6B;73 zL*WO|fBdWUi_8Vnia_Sj$U#Q6fubGxU{I4Oa>7Jfk5o?IeRY@jB-O+o*X{BR->zir zJ*=~n7P7!{r6mY}nq;C-*s|~2Pf_c^K4kB#(HcDVg^UmvnGVJs1;nDaX~a(>3~b!^ zN(wIuqbPp%X)1pb8WyKaESHluL9qW&WI~b63vnv9)hm~494LkNw*S+HOpxl2l zu5_PCO_c4U@RB*-=SrC#w`H+hqOH{&d6__@u{i82q0~$T_nnQQr~Bw`Alh05&C1mp_D-!_Jv!*B{`J=+K%t-HqF5$P%AwX}=&tq- zIeO?A8ZGQzsrC;PSB%OI$A>+%{9JQZeyNvP5|d#}nl)!Jm6%fVmt?h^wHnRP7oOg5 zYu~B!7477+bKak&iK22WVz`pjC2UuMmwG18WvlQwFS0L$Q%@qB} zmT;?JrST+AEm3R4wpODwU5BuJja5Ux2MOWGLK)R>m?B?)c~Kt(8DNaqfHRsGc&&=~ zlwu30d{-mlYv_5L1;+8fh;6>{aytgW+9_D}I%!Il-jg6Q{0X{skths6Y|E6=Y*L3E zncrwA6;?8m{TKp(ys^y~XGV#;&V}Nr6s9ipqjFSu_7l#bgq`kOLbi*z)YhH2k~Gox z#i?#|SGAgbnl7DcWJP9Vup8sd8~sODOGXh+FRm!rMfwD*G^XwR@?H`@BzCmI6TXiJ zJ1|J#l;Z=F28M$eH(9%?DV(rUvhGp&+%%2Ph`Y=e01@MeCxL;00lK_NcUe~c73X_d z0SVZ+FYDdQ^7V;Gv6Bjn5Wz^y#d>gbZt23}G3O=m0%NX37+Ln7A7I#jfHg2|I-a2Md|)yT~} zr7X2Li`pgAM&{5faVbRZi;ee&FD_4HIx8u5aS+ypz8=LNTh=E$dA;Vnema( zrDC^*@m@&?3GzuWv#^$|(|VzOOBeyNPvh@$Ds;sYd#iRB?5aAO+2m~|zHhlh;zfs? zaJqApJ5KnllkqZ|FcjVYQMHgm#-gRogi^DpxJLzjp=|GpAg zGqBL3V*+wYDncTowvDCFTuOUT{veh|;kHKBdPW9^vqutYU#eB=ls2{Jz;I(J!9?0# z^WeJsN=h?oe`N;@QmqtMN5D_f#`@`HCh&U+65^WSP@k5il%gOj%Y!Qfho&&E)L%s1 zVEtX6`=w?kV^2|jWKDIX*sF2uaf1vg9@Tm=hSU{{7CvG-&ez}!h4)x zNHUi*8u39VU|s1TsX-E4W5L)bm96EXl=85emV3xolDcD~R29V+p>lc7n?~oo@sD}R zlHo`8VW6Hj9aoxTdloaF?2a?` zl(Nr|#nIef2_q}^?aRrRsJT;?;-ihJxyzX#H+U{u^xxrp>*-ts)u-8V=k-X4HvaEg zqr+w8qs|a=Ccy9UGD6Q-@BX;5(+QQeUFt(criu$Aid}&0Xi?O`EQ&ql)j7Da3b@!i zI_h#OoUMRl>BPI!etBopm=oolp#XHceNja(P!(G9Va&*4#A4)mS2v7{T3OEM^@NDh z+ra7ZUKUDsW40{d-HRid>%}97W?=NnX~+A#qG$%h-s+eQDjoa-MvC$)2ozS!8Cro6 zl}1K}M$(Jr)Gr{GYZxw$6!D4$3^nL~dTToI?hp#u*2$mp)ef!2`7MSuMl^KPP(7ZI>!h#VavvF{2V4yHOB$Y}=Iu=3B_dxg1rL|K@(HEtq zk89R0D#zr`KhB$7+dDLxlEyGJFq$`b`SL>3jAplH*`deCESUTO#u}+p2>Px~k>2%q zitYfLdwjw*D#g6ili&(C%7yc8T1M6vN`b-FbanA~(=qmPpn2y=xgu$%;&Q;YAR+r8 zDR|^6O-Aw1q>W~ldL3IVN_>iEj0t90c8-(PIjOKY*iQ7DNgqppn1-;)Nfp zPvsjltjQt9W2=)nv)O*Z;JM}&ig4b2zzb{tF6qABL9ENa6X~8%bYgiVt`MBHVqDz? zIs)n-m<(#Fln3!UIQRv|KaL3MGhce@d7#JPyAr1m>kHz=yM}MRXO6_Q*!ClzCfDF$jnt4U2f?s z^^C2+Iz!!zR=CkOfYaWL#{3`>UHs2U1Yf1gv^p{wv8U-`>MAHq!>Xgijdy{XQ^f#t zmE!*PoFgi&?BOmO$YYxe;Zq2RFc3n49!0V3+?xL7i*jY!uxfpU%2>6u1~pUU@n$*( z3XL-wXSCpK zgK2@w2%+!DM-)ewFyKkrc8%s_7usz`gd9hsOI%f%iVj$!6j)&zN2G2U2M9>gVxvUGMppu1;LdSYKMb_=WSJ(Wr z?Ui*=8uZhQ@H$o^&B(RzxtM8>sfWS#X-dgtP=@h_5=<8A!nxscskh{=QQ=88xEsfM ztA%QJ)zSixYau%^yU=0paq=v~y*gYhkBDCbc^t^7lzVNi3kOYUa84C~(6zLpHur8J zKbU$FC$2hsRD>_I(wfS4GuR;N-e__vLfsop*xHPNuiaR2g5|Uu%Za6asu3g3*oZaz zTyo-xLs9H5GEl>ehxCi!Ee0}_meoBh>%6j_~7}dc@2>t zmpYd3pfk20mI>&2l3*T6$=OhU74pJk~k|NJ*xRb6` z;B*)B)ae6&^GlUVP*!VlWngR?INjd6m#O|OR7J3@@pBV=Wh~8KofeU;I69HUVHf}q z2SREAy?E;sv0E2xkLy2j9pMQ`uT&@8uclI+F^~))Hzje}iPTw3&tG6KDBuIhv|r9h zKAcEaTNfij%zMK5Z+Us96>3TS{5SUyNo`Sw(-f%dS(gQ2_7AXbjITi&e?;7O}i?KN3^~78)C+lc9g*1630yhKd7MQ znOUiV7`^Kpb*VO@oCU|Br66q^lGIRd!`0hki9P==4*GV*8^=4UX}(BiPn4NG;gaet zd18F$X3ABTz7ltO9bW;ZH{Q1+)7D8Oj|e=ofjz^ftsnMk!@rz`0O9)uyKZ@bL%^| zk!DiI0|;({o+mxFPDa8yFY5Rb)=z~>JAV@u8#wx!X(mi66suM|tdWupZpPmII~F3tDc?2VNwApH zIIS!8`9lZfWj{-Z7Z$62E-QIBf%)nR*tYG448iGui(0IBFnv`)ka1ji@N4u=;+lW z`2V%{HZgK$U7C=G8HS#rVHzQ%Mo6Pif*oKv^3V3JU3OGe+EwN3O!wCXEP`-kq#dq^ zjQk=Z?Q-u%8o_7~3l>PgswFlI;-^_55JKN(G_Bdt%UL!%Hb@{OGzf{0=Q;PBd(XY^ zec$+TrtRvQ>dzf^W<e!5ho|!sI)ixlT2JG3-&q?@hH*{LZ3hP*W+qBiG$PaQswX=k}PzZ}yMYWE&{uG-8#T zzlBJ{b?UyeKGpOo>SMyOj?%0rFIVX}h1Q?I2t`wvhKoI)|}_)%;!nQxHWQB!;45a zxxbXzDq-O+QHOdtUIauPxe1KmB)4z6w|)ha2-o?Fkz6EmqEcVOMKm5apE;xOJ52RT zVAh(5mrz557}X7^)L01_a{TNNW^kTNgiI|wu`8RJEE=gOYp7rs@G{kAx)wrOTBO^iIus;Tj|uZimiVc~ZQC2sku z)uj4ECeu{BvZC?A)3ThT)?TyUk&A!iYSz7@P@2_RRzggKz^xX%>FrjAhm!rLaE91J z=XAol#4iLc_2Fdzz!1f#QcgX-lN7gYlkJR~KUIx)gZ6Qq>ZfquNCZiKHjZ?>IE>4K zSp8`tC4wuYqRC#piIn{|rE#@+j#MWbc#cS3B=m~_&z)*_4Iy7&k3n{#%(Rbkl9Af1 zt+6$MGXOGDC@(~B8e|HJuB9Ifjz82^b$nZm+Y>@UOf+|0_q;=lyD++LkRJJ#LVH8E z-aj;fSD_?v z=_1rsmD#8^+rOzRHYsr9Y=tu1)nr%})5#C)A&D`H=J;o`1diGU8epaQGdX#bot93RQiACXTi(ppp?BbMzmc=-Mcf-Xw`bSkW z1nMUU%D%0M94gy<=}4grZUblYNtqTJ9Cqt4QMU!FVz8uYptfv;B&i7VL`FeUm!%A- zgK~1_Yy-B*U8n*&eJDV4-QIF+iBEXTyUnyU8+BoHt&k|ybyP)MRzjO~=5lyadoZZgSy#fS_Q98v(hZR8hOOwSU1I zw|Xj{xW!sLnRR=OcrxBwP!^w~Du5V1k8q2f3a$p@m-E20Qx5 z5@)O_I^|D|%!O+78|wps>o!*aEKVj12kD_~RLGtwL4P$56Z5A?b~Fp?2VFp@mBFH&X&NOpHz5DA*v(Yjm6?lU#CkC?g1wh}}_T6MjkQv7?Q< zFcqwBei?lSH0htt?EN?jI8=E=QZl)a`R&YM)qAL=bDqplp(hFGlFH+{x=XStzA|ZM z8>zUFf+kE9sHig)*WDeVsI3H7UJsdKEi+e4@Nkod&|FBa;j|TfX6bN*xaN;F}M`GoVFu@aX>F*=UPs)yhCkn>0>ZJezz=@#z|7peB@Du(Ch| zKV1{j_X^eBBg__3qoQ{*lVh-kTiFy=+!VRe5eomxg~AujPQqyiGb4P*jJ}t z5H`TnHbLU6$?Wt+1<3 z)d4b9vrE+Gt5i2q(tu}c9yAukmmBhVxGC62DSv44a%xvjX81$Wd;|{Y1+9HjZ*`6oBgFKp|`f>1jaz0*|j`6DG2(8yCYrXwY zkMcxj*sQP|(cO#jZOG?qy!)I5+Jo@907IiArPWx{AG|_#Hc9|)AkZO@%tb(R4>Yzm z$_8h^2xcTu$~YVzO9N5$E+3A69F4D5NEJ1NN%g$0BY>}<)Ota%Nbn$fM;JTIv5YZd z#n?zdAI|~YxR~a=qsg{x%%kP|%Mnt!%->xngrtx|v6uOc=Cvm(4T|?HHpjzoEs%Fo zOd|ziK@1ivGzS}9y_d_&CbveplP_Ly^4oU3PG6Mo;fC3i!Zff63~@2yrG-Ll_#9=r zkHMnd@Rds`gy$@+6Yk?aTIZv9%|T2?TEosKB&Ar&hag%@o|7~Pron_BN=4?RqpdX9 z2Y}+i(9PKLkUN93VTyAPes`|T!yF|HB0VYZe-;#O%)*!HN2#Tt zZcRS0hhnBE*-Ga=`{m;y@^jiH=RPE-Hr5QArYZ;0hFwd&GEp}H(h(B$P%DTg7suZ( zF%F8Gi&UeRla(~-(oaab(F;_|YO>J@&tK4KB)a5(j#{T0tb|jl02H@DwddTq=Oq~L z^6cV$$alB^0KZpoMR8xb+S$YbF4;|K&XP4Z{BFKLfb*en@86DB!;*;SUSdgdud_TV zR1l0%#~8&Ud`h5_OCHQ0DuR5BH6@!hcpV~V5ygprXf+m!D=pg+PToYCPW;1 z-T{Pt%=xLpQG98m-%V^E6xAKfX6RbdP6}FDFt8!*Mx+I$;t=a<3sp~bQETJ1i9Axf zP4LpB#;SX>ElbuP-BPH#j&4TEIIpH*#O+BTOI6PX&nnwT<3Jqs9o^`X)+6fN1BI=tUr323p}xNn(mL zZ$0i79Xm+a(c8_Mr2K=jIYG2TRa>rvtz%Cmf}}dPeHHwHiOvOmhx`(q=W~awr0lZ3 z$^~jlBK20d#&D%*OWabe>wtzzB=86euF}xv0^~dGtk7nbVV?!JE!E0{cazm?YF#Sg z5|mu7mW%Ija}*Rr%m(A-#VWLr#=`*-7wOL@>wJO>jb_IDhysQ01BlkQY&|&uRJ0Q} z;6kT1yzL&QwO(UF3EcMGg31$WOBr!;HJyx?sVTZNIYXdH&qFoy&Ddx! zIYI*@0H>QRD!;I$oojS#JRMp9hjn{VUmV5}8e7sh1or2E!mu{b{&{=9R|Kr9OWnDx0BKMbiABjFGl0K&FylkOO~L* zfOHF}Su=>KGY!NUjMRkT1DvHvUn{`fbuk-!6N_$mS^e_EC8qb}m1G$ytzucZ5Ss z`4u`;vME=kLnTOz;K^)0z(IWlfd7b#wwkNCj}UiBh+tf>Ez-AAg%FqNww!Qj#P@G1;2{5+;%1y9E? zFPq4iLADZ}gUj=Y(XU8?gz+e*;K6xrq(6=j&VoB#o6OpDeiKcFFRHl`+#~kd;43rn zI~-!4-B3Wis>9Z7fkq(e+!>D#x`DyR^D33z1;aQ7jp_4UdJa__v`!xK?E8jPdTwQ{%!a6>V7#^S%BjX^^GSOu`SV*!~o2 zySXNX{%miXhUQubd}LFWB1sJah$tvWCGkgL2hNQMfUmOp5clYl;6!6bL;q~Q;qfce@ zJGD}l|Au!V7~=-bXS#;GaqrMHiH&OmR_Ps{PdJ@JipQDw)F-PTL0o(!>02!i`MkSy zGa(iYOcujYYH+z7V~_+l^vE|F&@kcQxvY^!b-HvITQ71H4VEydIp2EGpS)1meP9`W zPTf{az(yt{wF_uxay`1!v)?VmfzIWmU4a-z?G@pisJ2a?q~B|Xw=P|jtoE=Wvkzh! zP=l-vB3k5tBucHdJM%Uab~62TW;F;14d&Cwhy?LJ9X8Hmqa`QDjg+5MY|=f@aqIn2 zE%9Nk)=>s!avPTyq;$>hVSZAJ)Y-W|mqBUz~+9KuZretD*n2oTb zqoOpdgU`8k=YFN2@T}Q{Tc1MHp_DBgE;ue->H6rYjJOQ$(Slp){MYjP@XpN&xYGd< z%na{945Gu-CCk^TtrXT|mdeyq%aO<0LuR%u(PxsOcH3^PSPk^BXgE$mSPjv~sPvV$ z#;@3??#XtzUfdfTRD@n;(ufF*M#U;6TYRlMUrYO}3dmGwnpvk<) zk*KDSsR8d59#Slf`EV=ZGtp~y-yDPOPes0t>7hS`m2&z6USF*_kH?{*RcY{MV~8!! z82+SGp~!<$h0r!`xVvUGiBcx+29J)Xb)(=1tKqAvq*g&|fIYGV&@!-Q!{xXVwLPuk z&TyG_&ukk9rRJ2$(?!y7FhY*htNG$RH0rcni{6jafW1EgI~v_;8kBcp8@T~!4i7%b zXm#kCHmgnMwwh4*mY!Q1&YlP%gR=P5@=Xop0FAZd6!k|oJ7 zCk=z&Z`=rG#2N}!ChioyrsrVXfiuT+RTmRxBd9GtgmFS2=NU@CI7k&f$YS#Hnnkb2 z7id5PZ25WQEcE*qr>xWd4yF~8PKJvi23nD`56=-1=l+d8neJc)CN5Xo8v8 zqSPN?a@y z_Z78NN^&#&_8qdp;Ff+n$JmMC<*TvG9GzUgl3yS>;fYRfl3$+7Z|zqn3&w^PWA@D& zT+NYK_da~x{q$@W_o3|%f6$;W6>c@uuCd(}eV|n18Z^+NotzeN^JG%kq8hUET(bl} zQQ$y~X_d|ync??Yn;V8RQirm|W>FZl5uS0jI~34`=5koy6#YI$cN@XPK3iwQUL@>g z0>&)_UfjADOjXO0M@!QF$y;3n3#94%Xbd9-qrNOP&h(0x396)p>LiSAcrC* z;@DI&>@pTe>S;6`=l}#5Tx;l%P*j5LK}53An9b)AZwjk%yn6$5R)v=UA;Z_>YLKtU zF(>tNTO%!z(A!97aw9GAtXQM0r55-Bh?yoHS;8<|m{+^J^gG&u9A<4n$p4sK!MD{w z1own6nOuN5M!Do{R;$JS3Q}7I;#hS`?q9Xd4uH}k4kQ@j=R9NqnJ*+lKQy^$Kcko} z>2vVW8&ke2G&a>rcm-luXWBK_i}I*n=A;BCP#(zc%i*{3-J8QW?I8kwP?yN|MlDLy zRnhb2m=#MBciUClR<5dS2crpTTos&*usPPeY)do#GT`-Ur9Q>z^raB+T#{Hvn2Jo_ zjz`RO^;U@D$KBtc4ayN0&6~4z+M(ieV}c^Nck~&+g`pU-)z~EhN%v)T3SCw{H{@(v zjSi+3Ggao82tIl}T!2U)fn317X0V=AGHH#5FPP=fsiIB(qJ&?CA&S^BKjq%P>UIsI zePFjLOcWcv{$X%2oJ{pL2YYu9qMN~l`ESqI>Y$M`pLy`A`piK1*cNqI$zf+&;_wUf zG5UcTY^Ol4$GiNZ!K-Tyt~pzUR2QLXaFn>{xaIC%tKR$Q4U;})w=f$SmkzF-{!YXX zVlJN@9riEj!r27r6fHe?VGBD_CYwo4P!^zoZ4PXb9u4)|4R@FlpO&rrfknefXd^li z0-u4;#Gw>R?Ly%2>a^)3$gup zzI?$pC%)uJ1SM@xuK=R6GQ%7~RcRbG%WLy&CFC&F^!>&Pl$s-Nu3-Y%e$QTEbCDn> zpe+k3;j8g%R#FIs1a%8KGZcu5UR>EoBum#80r2~QAJi7XunYu`EsZQGAOg0uGg-0_eBAri}Z+Y{c_QA&=O?)>rvf4>I>J zK$xen>kv=4aHz$Tov?E{;o4br$`!m}o83B6!BnmsCKR1pE-rft+OI|vMu zXr&^Mnl?gd;a#{uY*uJf`1G3@(!!tO|1dNI*js{Q$4}A63a5cuTmDW@5zl~*DVOKt zAHzJ%!-g@H$UpCl&R39^irA&VRswy7cIuQzRFto@F=i4W6R% zh7~t_eT>(>RNj1rec?z7C6I3Bq&ufC%kDD)cbcle*teniy?0rtl+U6)br?pl=@gxy z^%pDn=b@zn4ollilUhBmjx!-uiDIbn>RrZBh`JE{Y=xdb!|;*L993;G9=yCB{V-lF z<&22hLDxXby|QqkRw(4z{wW)N~mT zPb(BPo}-1K7B{n&;ovkogH*9Yq0FrI=%YB(7!Y8NDz4fBwN;&ZaLtQYbF;91DFE>C zDv>e!$@_w#D4Ag^&3!kn1CLg7Y&W1WOghZym;I7#k0Vxv;mdlR_dJ0JK1Q5;`uWrh zt+T66ScnSxMj|>W_vr;@6^2@cwOK3aO>#GTW11eT($hdU8M?&H9SHN6!WNz*Pz0k~ zfZ>}G4t^5vlmRyD*%oc(U~*~M)`TVL=giub=5;k4o9hBgAl`~fF^OrBJqnIz^}1Wl z*6*ieLN3JBz&CU3N7PbwBb^v5_!N=E_NT?uf~?qeo^L$!?R9ncge>(b360`9%hj_R zRbZ`}{$hANU11j|tI2RW`Jsy}c0E5k zvA~uce3gg1{63ItM!m7X%__gU4LEr`i9&~I-`~KH?1du=MPXM|Ra3NE>FrBd%@6_O zFNVvpS_vqNo$^E~C4Bvpsbz=A1Wmy?^`5s=JcNfhrs+}&oK)6Sqms?K%i;i9Hd(Zs zLYPX?hJqH72{%(1Vd+i=H`%GFaAxaWCpG|}7D^}{uhW>P_ojyy(~kRs60fl;RYeJ9 zKusq^dV^!BI}1P_E)o}^0u30hoEkpa?9`1>t7Q@PZGNi%UGP%OeedoezMw~jNJ_ zgt}e{YzPdO1h3$Gv@dSSFKb>{$hm6&Sk}oXq8Y(B)thmMsL3R6CV)V2X zx{4?5ZO;=~_}^(5hL+*D?)QRL#EU>3&7Nfj(EdY*FdA__)nsSvr9e*q^%e4O^Z z#%JJdo3-ZQp;NEwSkU)($RNPO3abf}mK=6@2ZkD>aLzyN5I&lrcZ zo*6fDQVz!p@SB#FBut;?$)OhHa*2o%nu3uAl^DGumF?PhBFn0TS1=|r_v;jfqdmV8 z?sIe<#Gi!TX;Nbwnc+x_Wy9w?^q9-2_PIB?e(s(wUWOf^SgZmbV~oY6()HZpTQyct z#sa;fSNvMb;sjCI?JOn+@7cNup~ByRM~uPET8%WavLey9?by6hXrPz7(SskGU~2RX zY);*O<}%TW8o04Yu)l05?tyxR>BjE8#C{3!)&zr_V;p>kE9X0Xbbp&J**>J#bmlFf zw@WEmBer-m`b9_(g}=e>(zX^@mL@kG1=}_aOJs9Yr+iEMkp+s?36H6?+X0>YjOq7d zXZ`ePyZRc5K@jMS8T%F>09?Sqf=nDD;D-d{?jzRpgQ_u@?vT^Tvx#4d8eK9;u-?Mh zu^`!noN56rNvpB&l%&;!-AS-jT&9}RagI#B;^tp@O{ z8d#ROl=c4N_0D28Am#>XlF$pyR4YzKE)=C3PhAXMK#R+BX;$uM4iKHVLUY2eVX}6J zu)O^wIRb4-5!kx5Ndg3XjH2sP%vZ0+3l?i9q?hzomK12P2C{)~XciJM+9IxA&{auO z+&+OoP&r1iQ}E)4$8VI53T~pd$4f4Lt|2z<)EjH+l`-tHgg2kpV@XBCdm15T;jcu4 zHA~Z~M02?1fzc*qqPL)gSmuG~1W5);nL6+7T24th6}WD_-B55<`tD86*oGdAXRGnA z00tDb>`l)55Med6_jc;6l=Q428HE;JgB0m**%@OYwC*OW5@Xm$iqgj@OJ_7{qlMc5 ztm70nTmpuSU{A3M2qRzO9s7gh`Eu~(Y;`dmzFMAz-*k3+cIDh*G28FgL;a0*J8Soo z^ps*7ph-RCw)~FqYhp$udjc#13PUdOr1knvjp5<Txw2ZB~vkYCN6&^%L;KknQBnZy=&A0sb=iLEats8jaW4XYFBG^Ez+#o zR6L5aF`6$gf~qLYpN_w%{w~_)Xcn{r7`?^r&LGEj>k3 zcU_VL=w!T@qs214NPxa54&c!P3o0diige znu_PZE#%NKjh^i60aFdaguXCRlv7ph@LbK+!3NsXKSWIv+YF<(>x3KR_M~PkYn4L5 z9t7fQ-N&(eb~+T!Q*a#bW%t$FgL<8e^xlp~lub5iv(~W)r?Uw&Scd=(AZ1TgTPi%z0iucp1gNI`2suSA#oogYctVw)(BMNc zDSA9cgK23MjhJaZPWx`}0Ywcbs!A_G+ohNn2OfrnMS*KCFSS?wqRF^LkrS?f3$!F5 zS26geKJ{M=RRGogJSYJdzfB!eIKKnRj|oDuQJ7sdf6M0II>=Rq#iy?N&Nvimqqb=! zuT<+}**Z)~$asnJ8T(A;j1(Q6s!5t38P8$0Xs= zA!aZgU#!}axAXwz-3Oh5{;(mP;FrP67ehL6oc6CVFdUIgobB20O0t`_RA8{UY#%Sx zBmnFZuSD^6_&51j^(nIRK8|{}f9fvZ;U- z0y*YRSaHTF84;V&fJcv{r2_O}4 zasruvAt#+Dqv}K&D3e5 z?1TE4Px-up_$$JJfGGU>DbSDKsKRWn( z-Z2UlV@9puUL0|?O;(H)gXoDa9|`wgi^&*6RHLDdmcb{p?v1U}0od1!z$!$I9qv2z zd{!6e1MwL_A5Jn_mtdYY%HkMk)dUyIVC~%taka1K%N3SxxdLujG)#E=8Sxvi0k337 zD%lP`X)H)Wz$r`SH>l*4O^Ff=hT=qq=^&hq5|AmtQE2z;*RUnczH2!V`-scw<+Wj7 z+oQfNypPTKhE2-%2H&vvx=!dplL&v13KFBBA;iT#nVg6kqlr7;tR%!2e~_g^_Fixa%yd>= zT&}5>ajOqjSFA#Cus>&`e>fqFKKWXos}g<-YZ30%XaClugn?gs8mI z@=1K!2o9v1HrTHA3KN*{L%DPD*mVmo_ltts0Z!b_)XsP`~Q`N1hP1nOcN z^N;HE140{n(o_GGA-@THP5h5aylkjX>|v6xe-#s(B=Ah**iX+NI&KSL3iaT@pc_E$U9hT@4#TAp_8_13CI|KMp$U2E4S*O z^oPm&!*sWN+UqTc2q>(^m&^GAz!?pQ1{b{}D4)92NJ`(f+tsILt)dgJ#_JlvlqhcK zW-xn9Ti!wLKvBC41@SpjmI77`*70IE1!Ci@UOuoY;=?7Tw4!^e%e^BZ_d;9eVzYYU6pqG7bmMICic=x|66aw8}u zJ1RfToW~Ok_byuT*+f-MVJL>0X^|tusx*EQ$p4V2Y>gr~nOYKm9+mAxQPsRD?&w?1 z!$*7mN7L!)`22b#3Nm@;&}~YKkb6g9^GBq2qvdqO!LbkI?{88??(_t?IbmYWYZ-kiQvGl)%6!1UxG)n7{I0=TI|%k@>YK-*ZdcbU=5uYq z^`gW~QpEyqXI$RV^`!<^k;XQfT{XT?6q--qu1W!Smr5alUIKpO81NnNRxZ!cX&esF z+u>qDRocXPJsTBT%~@ge$p?|a6eHk!H64z|hI{okI`KL% zIFJtL3o4;@dF4c+$I>md4;cvLAXxWfI9}n+PY38KdW!dexA^@5{&$M1o*-+|;Y0wp z(|oFiJp4^ZQT$J*f!hn}JzZQF^oi>}Z2EqFrNeTG6`js3J}Tq$Ow(hZK!}`b(S9?j zeg{5)Ca5Ja6hN%%;Lp9YEbuM`rF&`>H!oB|AIl7~kfgqcMD>0AIl!Mo{9*F?ef)WV zKM(N-;YxDdSo_K>GKiqk%VupdzQimyRvDk~E~Ab`Wk?eGB&hcTj^XYtG>fB_cyw3Q zBqm~_oSa5?PRGN=VtDy#oRiD4OWx#GJty0iwm1w9+RgaXd(1VZN2Bf0kR1dX7NO5& zgm1W3yKVL!o8EM{fi_kkp)AYwqVUUt;a+TP2^xoT?kU25P@R~(c`;wS!$9mdR z%7=nsGc%Q1s~{-WUX@#KmJ*Uu+RdaK*xO-EiBd0H`T_c$fk&+#C3spNE6|WXmtx_{72b7t9KimF*jNocEd8nM&U_<~{ zT~_X~1*9`}27&9mX=@XGM(31Yl%k47)b9b~4`m(@REe11qec!qh`Nv^@Em$og>y|} zh$(kIO(iyhbG9|djr*g836=F7D7niHDtF4$)1F0?H$eT)U#fNrDzV-3U~ce~m6QHc z!(=W=N{XqxVJf_guBG=Htb6^{Qfl7a1rUyyDJsT`E7bX|E-DO3$}s+T6B;|}qt*${ zvx@Fa2|rl`BiZzp!^G8;{2Cq)%{jvP6EsvyzS6`Vkbst~Rq!r*sC@11Nmdz!`@^45q&2^fXigWa&N4a{=e*X;UFd6eHThGJ7My$CgFeNyBg@ znZwn7y8QFna>FC=%p^?h$+wtNR zqiOsS!n&rzIqkbzx#o7Z7#OEYWLJClF823iCSmVLQspZPBph{-LZ#_%2_NR*c1m}I z*mhsBEv(J({O#}(ETt^6Q6Ca5TF@XXQw`s?ZFK3N$-}aAp?!m$Wdg42 zcFyv>wE>#W;!Is}03UN) zYke89H7SQn!xJoSku$-yd&!fT5IG!NTqEcT?5lBR2eMZ%0cr+sv{eRqhH1*Oo$39x*waHZyvDfxU^VwdQu9 zv6+{tm(P`NVls_(_bSBHN<*kNOJ|er{P5JSLWhQdBs@#~S;Ou-8&rd4lk~I}B*Og~ z4T{@C%9-mMHEVY#56>r@1UiNak=%nmyIQa<)!q@vkn{OWXTC*c;wcYvC?VLtx*kp~ zEKSk1GC)W!mBTiQ2{+5|=9#1u^pKJV)DY8mc?f1=wLxWSp%xz;*>}&t6|mrn9DhrE zw7TbNu`EIivBi|wlv`PLfpx#ygjS%;G$;jJ!|Nf9p2agdzX~~`XYVFoee(wm0^uE0UEKaNbhI;0!}u)PqIXbf}5u^(r^Geh~8k0hUET zZU!Vtdj_(=XlOlHBGcB@h+ zDa9+xT)?WR3L6rfBL$B~@5Ea}-}~Nv2y;A<=Us6V?Gf}*W$$qJNKgh(f3g0QRcH=8 z*u8<43pY!G#aE>8Nzym=cdfCc$3MA*S#cgeH7HbtsSQ`qw66CM3%0-tghFaaL726E ziS)IHPo(XkIJl?d<@|av8i(FPB8O`I5Xwl^@+^r%-dpN!*^P-|mr+Y+Gh_4%qe_v3 znTt|IJ-M_jQIx%yAcEDVLioNGf=zrG$sT)-5>QA+r~`F&!;L7ok@@FE z|JB0L*bEaoH1VM*yI{Lfd4#ByxVml)s)h!OUDNUDY!MMcScQzay0Vbdz2#^^vw1NY zp~BL( za7XeX%vVZ2bdSFA$d2f>ukjb3Jq(RWkgh9kYFmRAw9iO2Y&;Kng-Z(+fur<1_~j0v zwQVq#nZ%dgdGoARfuL>Qfoy$wf+~cFBLZy-k63m(%ZealP#$bR^rk2Tx{rOzgnZM5t1- z7@KNrABJppxaP8SoJmLCY;YVQ5K{06wIe!Uh?sRiTK4Y=g$Alw{_V6QQv5ylR}bZmNoJV22sL zRhFBA8C?gQEUz`J0=YRu+Svp{ZKhQaS`<20tSwKu@*N+cwHpQ(4i=1Mbou)J*6XHT z7c+`Pgw@Kh;}_V68efj%ER1aL9^CcuV)bHiJ)R(i0Y7EiN~BIsmX|~A7?V%}Y6vGQ zUamooILT=uHr-OROZ}FhB8p5EJRx_d07XavnCd3kAyvGJ^bcfA;rlF)O3QS=J!8_i z<6w{-z0CPK`@~n{)#KsyayUJkU#yUZ{)h9$56dg05T6OjO3LJ8N`^R!p@&DP29hdq zW`V-7CNY#~xzO_oxkavHmBDP+hV)Hb9QE{1=O=x)G+Auk*0^>LE6x zi6>b8YOzW!#$8t(jOwE6Y5)wV0>{u=?m#Fu-8Fnicno)13oGte)LQRH)zR&ruYyl+ z_i8?iaO{Q~bv=!TOiyQF!T+_FJYW?NSv_B&6MGV|ZoTlJLYp zHAJ&6^AN-k?Z<-b9fA}=ux=(z`y>Lgn7q7ZWeCWZ-o2~gVmOm9NAG3Jw(Gtaiiu|u zIRdAjBK5yzdx(69wLx+qv;P>8kG{Nq!5=?iX8-d6{&$KQ!DX#Atk61!3GK-14AUgA zNfv&l!&p0nOsIgCg7=~=Nx+d^Y5xM{l94#eI!W3ai|3j+w}lfd$bU6|=M!b_)7NO% zB5*;1rfJvfo6}!gDEDOPKP2n`aH^#%B-H`I+KFxwCgOIr5oE1S1>(p6Zf}tokzuh= z8h^H$E^Blg3m7tFFTityR^P{Dc**=hKFg>$(*TCtC*RT z+-|A@O>0c-qDPSWB7b+4vW~l!ET?>MtL}i6a1DBpX?H-#rwfmFWLvj{$K{+6hiy)0 zLMGM#M-#yCtj(BP_s4^X2L_@eoCin0NCTxLW%QbOdX?M~&s!S*CX}$I+En*4J=BJh z+XQR10}V2?Qt8Wx3hQE{qfqI}3ft6353piqi#;hpQvAO%uNNyhuxSx z?l@9y1~Hu_Ib(8KLfouz&NP=Q?M7T8^`#`8y#i%#W32=hr(20@RRgS>?S*!>q<@p63~p@7y^I@y(1H7%oVFd^RiS5G#HaGt1jm zs=8I>gifI;{*;CcjK&Z%1eY-UQg@YB3&7mhddHr4HEU-oVv4)Pk~&f`=~rnkvfU zgd7fW)rYuzHw*o@YC(bx+f4PdAVHbO0sV&CJ4bW0#r*v-iKE&PL0%vBHu`<>@QE`W z-6-8V6I_qU3(s^9dDBI*c0It2`>7Z!uMvnycWU#t0o?zZp$yyt#!1Q~18Mj&L`!pY zW@#crn_@1wFr3M4PdzADvaAw%I~Zib7XW+K_HDC<7ce9_7!SuZ++eaYvXzRK+tZt% zCrr;MT7;I0ZJtGouq(tsrnmw^Y2Vjm_lpa*$eWpk^zLK?0}5(CQ&ps*p(-(;dp)p< zlXK>xX@e`N421=$3;%A-tu8_r^%4``hNa+YF=`aTan@5Sf&jFqiZyM)Pp1Qa8@As# zO2|RF`t3}?!?Ui-1T{fzLs%-gX@W^pxT~JlDJKf7_q0N#YxJ`^1omFi2)q`B=lXCD5-6X}8?`i2Z&t^K6xqb;fRDz-LF|(94-L4jJtJY*Ig|3-F_L()= zqJpXhlit*VNpCcmL>KC3Tof9EXS9W#5D*NagMcpG;`GtO0qUMRm=FN)ln>K>RxVr% zo(PzzV>M<~%}zxQh~O8o9C!p+<(%z_X?CQ1rZaT7QXuQt9g5DjRW!#Z#y1H&0W5<$ z!11dEvgxIWXNVySFU7~?A#O4;aD-iRduW{#+*lyQ+m{F=Ko%E00VXe^q8bKIURtuV zhUng|9U$smCUsTTh)2~ji^#-se)m|F!~!VmQMV%EI6OzytwP{iRPbet?ZZIdC%_)4`V#=$4n z`u5vcusG?5@q6H8JZ-$NgsLr%5KntCnN9~tZoQc4aAY`-NTW3_H||pRYB&Yu4^bj3 zTh>bU!HQRIU^0@1w_1%%J?7~&+9}vsIjT)7jKwPA#~sa>8WQXsCdv>z?2Z}+KEVD0 z6JRE@Fv6XK>nijQ{C&4mT-5&>a zq6SdM!}7=uHV6ufw=)~h=q~C4#Ia~tTHNGf3?wGSFfZS*yEnEZsibyc7nKY55^tAU zJD7tLo)|qa?N47_Xie8st4NjrJx7sv`co_Q-?9CYhbKiqyaT5oG@Ecr`fEqdoIl}m695D^PMnm=ya{mBAvgk1SuYu|cI1gBQ=9u3@kz%A2-^hjMKiQv4?{Qsz)W%o@ zcK$_z?;levJ=o$Q`*_K)E~Sxbm7JI5V9`3u%;MURMP?$e&{4gQqOaJ}tvwQG3?1aN zb?i7?1)188J&Qm=J*QS#l62ROanw%9@3irA_$~B8Q7UO{O{blskisTJ2g2FQtnR=2OJH@e zxFME-d)a!55WPRGpl4Q+1TPxq1AYz$4B}2jPi0XcW$N@Ps)iPP(?g-D7T2TI@eG5{ zg!h(Kdw%ShBm2j{f+M5dA$HHVcu8M#04VYuT41J1Wl^v%2Hz8o!7M*XTj1$xyFM4S zV%1@0fd;l3E7MYzlV5ZLs=^ayxmrz@yjGQFPej19r6Q1+P|HaOTW|MWetK|w(DL_G#lSl zqV_uHQMHkgY9&-}R`IXNs$>h*2I}T~Y^cNdWjcin`A#6LQ9m$TVF(cEidPtGMYSbO z8<26$wsr#qQimL}%_fgjLncv*=|#&Ot79B|Jx-QhmF6}7x;I;V9WmH$O4y|`qp1`V za5u_jPKS4U%%|Ij1-Eso`mrfs9&Omia5)QT5UkuYpF({fi6RrEix^C)htm*)@U{a6 zieocHZ<~vjlEc$^YwgMa@fuAfWLs80TK90vyeT7Qlj=aeOm-lLD|73~Cn(*8ZH2Wqu}1R3EV<0lbMK`xKQuUEb%{ z-urBSF7J?KE7%)I1MRG-GByk7-zBN-X_ngT1V~?!GC^hqf8)b@upp>+IA9|dRj<3+ zUp>2QBD%9T7H`4h@e*GVH8RGVbXwwHe6urENl;ckr!BhuaJq(9p&FIm&4;wX;~imi zRLH_=?}mqqyhiGY+&$lmPG$3Ron*jYOfFyXJAhlBXsaN}Wji16m-?HN1><;&F@I!p z=lNatQ&!7BA)j%$fi#l;kM<$#yBC((On}p^iLq(8pI*y0#@?eHyi557(aLJFQ((w; z6bAf2)Ub|Py6}{0)w9hY7p6Kjriu{{d+*UBAP`(6oq%v~vnqZ2+D`FYS~M{d!!Z*# z7c=Gk6h@}kR@~_;IW|nP<)CEOxT%lF+bN~ib<2>MhPdqnUu&)s4&_D-`q)8mI=)!b z24P$K7Z_6P*6`KUuxzp z$w+{%3VV4`JS#nS*Q>1eAa~ytKH74$JOHGATS1AM z77dj-q>VsE%e~rS^OYNm3s#e5Q~+AOWv3^G8fCuA=_73Pv(W}&3SWt7`_(NZSiRMJ z@FQK;%U5t8x8E0_VdPkRIboUw@(p3K0WAiYL^zmEmMeyn`m?hy27eSzq6YL2-pnVL zCATE$wLU@tpLWAMgCGZ42ef7mKz%lMM@zK*)Dlf{1(04;?GGJ8Rzh}U3d(jtaU9M? zI3W{JTq4YH&Gq;P zu{De-V=M6-!=jqrzZnc5s_7hUgu`U4TG$7-p(e2k`|1Qj4*w+g0LxM}&r@fRlVCSv zz%g1|vv0eoAu+--;~|Q^cVOD>VlmX>loQX2aXg}3yj}KC&`!7%LTJ08CFtNGCoRZY zlgyozWNvsyff4XS5r94ne}`RudBVXc@I^x&GbJ$jA%b1Ztpe8A7RSgs-KSgS3Q+Jp z^={(=cGSPZ-(`4c-NAa;?qSKV>Mqv94ScM0#CR@m@Z77L0-vf{yR0-R_Vr8?iRrNb zWM5{kZ=Kj#)v6^-JDYB)o;J!89dX53mg$o;1Q50E{|?bE5vI7=lCOiWbD?l1@L-Lz zse}RS%22|nu?pPf-M?;ai{~;|Lh5#RjWOeBIjrz}t89kCQEY7`ib)zzaSExm8*OCT zHOfZDLwp1mJVxf!(~v!tu~y91j^qmTK-@Zz{pdiVg0O=>L(}zQGz~eeXxoJ7LF5Bp z4Eh&WG;YrU+|*n|ttuEtW1%Xef+$`b_x3TNoJqU$#TUcnSUZ~pDY?O_h)DaS&AJYQ zA>6;e6vGfk5?ro^lS^ck2#Ff=9#8|$6K$|EDXI^lGqU1{BBIRGsKQkRF}QDcuyjsU zBPXN%{94!6n+0cS2zNIw%MY3&cX(7Nb8-~Eba8}gaBW<+pK>CL$azub5O1WpBukGD zfC;Dw7<>I-R7m1MgU0~x^K8X5eu6QECqMT=(a_6mDh4L%87Kkt6O-2`c>Pj_iX$|^ z|HKzLDL7vZgl)f?huB>i&^{Qh2FOU46&3x_Sxi{P50bWI9m?X&XgFd&wW0;-3%5p)MKcPB_%V*@{S z^h5jr^iL!xaUV6Z(VO^M>O(c1hao>9Uo2J)<7%i}YZe^i9XT>n<=#C3?lbF;)5zWpJ#zRqYlZP{L86>fW`w7Rz~7b$3sxb8oJ{~& z1xEwhtsiZFo|z9vK^AH~MLny7A__Rn;XM4oT?H;fOm^wH-KSx)!yi zvl#b;_;EJ;fg^F=&VLv$ezMIwnRr*l3^=fd=i|}vJ)4dN=3`rsr7NZo#J;8Ea>p&Vd%JQA%_;hAtr zPhPd?EMqi*UKrmA@oewbTPngCzTg9#cCW9C^rfVOJMIwx&CY%nS*x9$O7PK@uy37F zo`4~)F?U)*1l~9CF4=zq>BT&6(^FJEL}%p6>M}-=x>BWUfqIrQZ&)Q;;_Lg6Pd)> zwoPe2CPtIA2o=6}`(UJ53w(GF6_vzamSB6az|_dq>&Y@AH@DI?w18i0%%yjP5S9e5 z&nA}w>P}p)g2>U0bD&xu1W9%M8#7d6pfV{N;U_dOJwlNT%A)07a8hNTZKF;3Z_^tR%|$`tLcdpfH%9hPDu41rMt z&buXl z8Wmuwz|IsTDd1o=;+obJ$4;6nCKVjja|)u1S{<7WEWc4Yk~H>bJYEWvG)HB=z=>ge zQdFF#$nrVc*ShMFI`^u;Hr2ONz{wM+8iqNrv}E&TT0h1>%m(Sj7L*1qV&9}+aWsoD zUZv53%p6!Zr7*%W#CwW~E|K&HjW9nyo_#Yr+B@ifGwb0GtH|+6CkPW_M#JeSwg||f zUh8vlITX(&v${7q*CooM4u>iFcr%%uguyP9k>^uk$+1Js#N0il6H0-6erW1jP25CF zPf%{0+PswH&F-Pe5~ymFz*eC4%^V(z0$D&o_SA81u|7<>jiyveQCxFr@n$9i0ETLL zklO1PoIJe^t+AAcX#S!~T=Ls_-k==mfP}!OSE>Ak;@q2A=X7JG8gb;7X^-?4B50@G zU67|tCfz*XqrGLM_xO#edTB`1t5T@Ju8;%ck_UGgi*6O8X?<^>C=JN%i$ke6#o4b~ zbfZHXzRHdE$wem6E;Od=D`N}Tt2CRekZ$bul!eKj+>=w!pzo%=StjY+rb?X9nk8U} zB)K6!txEE$jjZ^I`1rR2_O2AKF{HZ7d?nR?u5nY0AlI4XuHz6xly}a-_)l0%S)hch zvF}&-z4wjElQ%O>Q4eiKS2FjbLmNN@O*msE%#VsSH~7Y8%?!V3ev*0%tz7LQJCr7@2wpXkMH*o!^@K8RpjKGppw#QNq&RT zals>T#xUkDA{wy;Tf}dmy@9E*>^EqL7ZM1TG=gttpCVYv+d={i5HuJOJ#vX z$NA3Xmh|p1fV>~{$!IivJcq&D*m3G;icLw|%Lw3(67pMptUyfig8UB`?qms`Kfw?$Te&&4OErnq+ zh92~cMK$rJ?W7z0i$8I+H@Hu!SKXqv50|aV9+rIToGq`oYf+RoT;Vh;6cV0;zG447 zW*BU_lVRoS@E=YJ_RV=(&p$~vOBaAWwy0xJ4Q6z=C@d;+st!nb;(wf2t3oTo{2-pD zesVS49wTu&IUe5|3>?Is+Dw&}(N? zF?9T%&mx@E!-i?)TC}7ccD~9ys7z)Ac^Vx1YP>poH~H$DKWI#YUZ-{{i4Ue*Y`1&8 z$r7W6R+Crr>)3DQV|rw+$5KqSBR2UI4Hd9gl_Y6WYRt`qD)}1t7X|h5ih$4AV@YW2 z^EA?aR(T>D(wL{(J=avbE=e;C4b5TaP_xDlWLCVTEg+hr_}M2LK!(->S!hCZ9#u@9 z+XpU2bcn1&Q}pb13Pp7(e`Kr0>}{-N*ZhQDDEFxkn4Q9Y3myZ`S51JBf|dB5Gp@kN zG=h?lfFQ9~)HF-Dp{4u{``@iL5D5qy00~}2#$oeZ^JD5vF}F&?^is=e znx6s}l}4bYDx7O1Xj2wsKnb!s@)dc@z9ehbDp8zN#KY&8C_+XX z#ZZLtl=zy+2Z$0C9gVJXER|tqs3D`uJ3fRBGF#exltG|WU8=lS9i=7UR|@@RhlMbS zdG_G}^Tuzhts1{oy#y~%yrUbfeaSHf(;Qj3Uq>2h5g-9ATnO%T^Qq=M1#%3MUAjaD zgHcah{ge+!r`*K#is4yG4fA)>(^{u_e6MvRzCe%Pw zclb7*xo86uO2ev(D{GjZR7n=;2px3UCw8NR0)Tg^E)(^@;y2oUHic-a2K?LgfFiKRDn{cVnC6*_6yNwH7Y=y__-9w%(+bw{sUqzg+f-g3(MM5cGQgzcjCb(1T z&a3wtoMGE3{p#N6^>ATri$?t;6t2Eny=E(yF2P}Dp>OtMIHWlMT$_>k;%G46I*oGUi79Ron2(@(t4HcFxDUPzWUdz__N;5!%r-1{>N;sh^ zY)w`(=F>x(+zPa`_VS0l+i)xMy;yxKIPuZ){iW=1dX}!Hku>|siO=0mCB?cgFJtZS z9fBNilvC#Xx>3}^q6NU=oCKrTv%y|qzVU7lWfa|TK>fL~&oHTjNk=cP3^Pm05k7??(j={btq#A-9Ufap4T1dpI+$Qr3|ENhkqoj%dTkRcbD>QV`LWxNwk#b8DgDR?Q zie6Cf$?V>K3P+T>;W0W|N#dhQ*T;|-oDXb5r587Z3D(@CCk)y>VBh|5o~TEfpPFWk zv*~%AhPdiS2=K$=vARztO{b#4viv#Z!bHA_ahIPM-?V#aXW>8X_v@J^Mm7sZR4}Nx zuIT}E+qco%mo2R9<|f){tP?!NEW!(Kr=7;dK|RcjOn)M{gY7-%wnW*{5Mzkg&!TP? zAan=yLdxtJwbJ2?zMc10{TZ;3AQptx9_$^!T41~JwA&vf`aS$>AaGw@l7|CtSx$Vv1f>u8Eq^su(Ed)Nn-*odK~ukzZ;&r^U3f|6=)WNjF+ z4A;NmK!Xt@s`AhbY4+`S!M=<2Xu}FO0%8Ln^S$fspy{60vqWf1eM>Gxe%^8;V4|7M z0u70cgf4>(wf@!nug5>4MO&O~9^7)fnnCK(bUX|wSHgfU2+I3}dNBevv-sZi=nLB( z`dUD5q)1^e1$NMm7oEU@C`Gu0C)hgMPU@!cWd_YDw)RNcI0Xv`Cl(`*Bw7uEEirfP zs0w<}_3XmRM$MXWsdPJXYu~H(JTBX1VHMZFIy0NaT|l+xH_-kFxU~X7)l5x{wtMml zAq6=p#@0iuCVeF9FbUi@V7rppQrg0##iD?~&mgTdK8uMyAAwS5DU-Tw$t4ZlWH`cJ zEG9NsS*myA;SVAxwgHh#f^5z(^K*Q8iOfy(w?n%lRK>kQ2Rc?m?(d-xGL%H_KfI5> z@3S&;@9^mU!Gryy-XZ?c<4+Fu9v(bAy1)03KjHiKj`oip?majwEyeG|^C1pH=eEIO^6K@PF$lc_7}d#WFuwrRHXhE7 zM>xDiyez0+D4bTzE9xKZ&&fW9CKk^Dc4x5JnNO$`NJDd3(4raCGrNY5RDHD;k|^Or zIb%Fv+i0t37j`M1u^gg~D^BBkJgNF2yPd^^NjoFU4Nr{+8I2c62ud)n_0@V`^^lquHU)|W4lfVqLEHGA ztW^d!b*GwLJh2prUQT>U}?C?ef+)ZdDlQnS|{)a=EeUo zeh-y3yBaRWOjFj5>;xesdmwDJ-j^pqr)6VAyh^KOM;VQHWAE1n?%F6u&e z51oD4A#(9ElGag4V9!qb2vZU9p|B)DY&_zqyus| z`wKL+;!HT;h`Moud2c#MLk%@09{@{%MQb2rgWAj3M#IzzCbW6F1a1reeilb zy+Q|1xTYxZDhe!NnLFkeA=Z4w)ML@>Mvy;{Y1ANcrH3dS0u`>fMIRpsX%PV^OU5xtZzQ%%gnM~I@d ztg1m1=R+tRo%ss3b{Ra4x5RV^IAi=F=~RQ{W8LS`5*( z@uh63Sc1+*^U*eF4kI*kIHQr=J~&3|37;bKNcg698$#3JP*`-hQalXv) zT`j_I-hy*QvOxBa0e*+h-a~#s|8_N9-MPzwNO$hyuXkU~^)Jiu=+52M_(%McjYaV8 zIStO8yW`h`i^UL->h(Fk&HjbUj7Ed;kE8JwX9cp2y?pCnxL6Ev`}_|3JHLmwfv=QX z9?mAX7Ou=Yz)$2dFPFD{tNw!z5YUfZZE%Of6>>w zc=ezEQ~kWX{%ghi|4Q-iNBH;u{hQy+?fU5JZx98yl_PO%<&+zMi7JuyX|L8Ql z-(G)l=P$_*{%QPw{>?uByWwqj&hYw|@y9;@k59wJVteS5X{>BkzM z|IdogpXlf9^>@}>|35E2|L=YmRwBOs*OadGUwJ=&55MyF|LgzvugRP3^`A>0mA{qi z_uuO0ZU67w`Rn?L@cQc~`4{{m|L2|m9{=X`?eo8WVxJ4If93b`50uyc`^D$~^lyjH z+3WBAHGX^g&*aB`^1t9O{`!CZ9|nQ7*Z)+{kG;RV{{M`_Ha#uYUu- z_n*J>>+rcde-d8*__6yi{%W7I*WWEZ|Hps%x8(zW{C|bd{|ouYF2NPjr#hF;@|(OUjJV4bNl?iUi_TT%K!Xpu|@LxKiBL3llxo0$-mvMYwzXy)IR^`zoAs#x$|!nA1MBL=l=(^3 +#include +#include + +#include +#include + +struct JiugeModel; + +typedef struct +{ + infiniDtype_t dt_logits; + size_t nlayer, d, nh, nkvh, dh, di, dctx, dvoc; + float epsilon, theta; + uint32_t end_token; +} JiugeMeta; + +typedef struct +{ + size_t nlayer; + infiniDtype_t dt_norm, dt_mat; + // 0 if linear weights are passed as W, any other value if passed as W^T (default format in pytorch) + int transpose_linear_weights; + // [dvoc, d] + const void *input_embd; + // [d] + const void *output_norm; + // [dvoc, d] + const void *output_embd; + // nlayer * [d] + const void *const *attn_norm; + // nlayer * [ndev, (nh + 2 * nkvh) / ndev * dh, d] + const void *const *attn_qkv; + // nlayer * [ndev, (nh + 2 * nkvh) / ndev * dh] + const void *const *attn_qkv_b; + // nlayer * [ndev, d, nkvh / ndev * dh] + const void *const *attn_o; + // nlayer * [d] + const void *const *ffn_norm; + // nlayer * [ndev, 2 * di / ndev, d] + const void *const *ffn_gate_up; + // nlayer * [ndev, d, di / ndev] + const void *const *ffn_down; +} JiugeWeights; + +//////////////////// APIs /////////////////////// +/// @brief 创建模型 +/// @param device 协处理器种类 +/// @param ndev 协处理器数量 +/// @param dev_ids 协处理器编号，长度为 ndev +__C __export struct JiugeModel * +createJiugeModel(const JiugeMeta *, + const JiugeWeights *, + infiniDevice_t device, + int ndev, + const int *dev_ids); + +/// @brief 销毁模型 +__C __export void +destroyJiugeModel(struct JiugeModel *); + +/// @brief 创建 KV Cache +__C __export struct KVCache * +createKVCache(const struct JiugeModel *); + +/// @brief 复制 KV Cache +__C __export struct KVCache * +duplicateKVCache(const struct JiugeModel *, + const struct KVCache *, uint32_t seq_len); + +/// @brief 销毁 KV Cache +__C __export void +dropKVCache(const struct JiugeModel *, + struct KVCache *); + +/// @brief 批次推理一轮 +/// @param tokens 输入 token 地址 +/// @param ntok 输入 token 数量 +/// @param nreq 请求数量 +/// @param req_lens 每个请求的 token 数量 +/// @param req_pos 每个请求的起始位置 +/// @param kv_caches 每个请求的 KV Cache +/// @param temperature 采样温度（0. 表示贪心采样） +/// @param topk 采样 topk（1 表示贪心采样） +/// @param topp 采样 topp +/// @param output 输出 token 数组，每个请求一个输出，长度至少为nreq +__C __export void +inferBatch(struct JiugeModel *, + const uint32_t *tokens, uint32_t ntok, + const uint32_t *req_lens, uint32_t nreq, const uint32_t *req_pos, + struct KVCache **kv_caches, + const float *temperature, const uint32_t *topk, const float *topp, + uint32_t *output); + +#endif diff --git a/infini-qwen3-moe/include/infinicore_infer/models/qwen3.h b/infini-qwen3-moe/include/infinicore_infer/models/qwen3.h new file mode 100755 index 00000000..39faaa84 --- /dev/null +++ b/infini-qwen3-moe/include/infinicore_infer/models/qwen3.h @@ -0,0 +1,120 @@ +#ifndef MODEL_QWEN3_H +#define MODEL_QWEN3_H + +#include +#include +#include + +#include +#include + +struct Qwen3Model; + +typedef struct +{ + /* ---- 数据类型 ---- */ + infiniDtype_t dt_logits; // 权重/激活的浮点格式 → torch_dtype = bfloat16 + + /* ---- 网络规模 ---- */ + size_t nlayer; // Transformer 层数 → num_hidden_layers = 28 + size_t d; // 隐藏维度 hidden_size = 2048 + size_t nh; // 查询头数 num_attention_heads = 16 + size_t nkvh; // 键/值头数 num_key_value_heads = 8 + size_t dh; // 单头维度 head_dim = 128 + size_t di; // mlp 中间维度 intermediate_size = 6144 + size_t dctx; // 最大上下文长度 max_position_embeddings = 40960 + size_t dvoc; // 词表大小 vocab_size = 151936 + + /* ---- 归一化与 RoPE ---- */ + float epsilon; // RMSNorm ε rms_norm_eps = 1e-6 + float theta; // RoPE 基值 rope_theta = 1e6 + + /* ---- 额外缺少的字段（新增） ---- */ + uint32_t bos_token; // 起始符 bos_token_id = 151643 + uint32_t end_token; // 结束符 eos_token_id = 151645 + float attn_dropout; // attention_dropout = 0.0 （推理时可忽略） + bool tie_embd; // 输出头与嵌入共享 tie_word_embeddings = true +} Qwen3Meta; + +typedef struct { + /* ---- 元信息 ---- */ + size_t nlayer; + infiniDtype_t dt_norm;/* 输入输出数据类型 */ + infiniDtype_t dt_mat;/* 矩阵数据类型 */ + int transpose_linear_weights; /* 0: [in, out] 非0: [out, in] */ + + /* ---- 全局共享 ---- */ + const void *input_embd; /* [dvoc, d] 或 [d, dvoc] */ + const void *output_embd; /* 若 tie_word_embeddings==true，可与 input_embd 同址 */ + const void *output_norm; /* [d] */ + + /* ---- 逐层权重（数组长度 = nlayer） ---- */ + const void **attn_norm; + const void **attn_q_norm;/* [dh] */ + const void **attn_k_norm;/* [dh] */ + + /* QKV 投影 —— 分开存放 */ + const void **attn_q_proj; // [d, d] + const void **attn_k_proj; // [d, nkvh * dh] (nkvh=8, dh=128 → 1024) + const void **attn_v_proj; // [d, nkvh * dh] + const void **attn_o_proj; // [d, d] + const void **mlp_norm; /* [nlayer] 每层: [d] */ + const void **mlp_gate_proj; /* [nlayer] 每层: [d, 2*di] 或转置 */ + const void **mlp_up_proj; // [d, di] + const void **mlp_down_proj; /* [nlayer] 每层: [di, d] 或转置 */ +} Qwen3Weights; + +//////////////////// APIs /////////////////////// +/// @brief 创建模型 +/// @param device 协处理器种类 +/// @param ndev 协处理器数量 +/// @param dev_ids 协处理器编号，长度为 ndev +__C __export struct Qwen3Model * +createQwen3Model(const Qwen3Meta *, + const Qwen3Weights *, + infiniDevice_t device, + int ndev, + const int *dev_ids); + +/// @brief 销毁模型 +__C __export void +destroyQwen3Model(struct Qwen3Model *); + +/// @brief 创建 KV Cache +__C __export struct Qwen3KVCache * +createQwen3KVCache(const struct Qwen3Model *); + +/// @brief 复制 KV Cache +__C __export struct Qwen3KVCache * +duplicateQwen3KVCache(const struct Qwen3Model *, + const struct Qwen3KVCache *, uint32_t seq_len); + +/// @brief 销毁 KV Cache +__C __export void +dropQwen3KVCache(const struct Qwen3Model *, + struct Qwen3KVCache *); + +/// @brief 批次推理一轮 +/// @param tokens 输入 token 地址 +/// @param ntok 输入 token 数量 +/// @param nreq 请求数量 +/// @param req_lens 每个请求的 token 数量 +/// @param req_pos 每个请求的起始位置 +/// @param kv_caches 每个请求的 KV Cache +/// @param temperature 采样温度（0. 表示贪心采样） +/// @param topk 采样 topk（1 表示贪心采样） +/// @param topp 采样 topp +/// @param output 输出 token 数组，每个请求一个输出，长度至少为nreq +__C __export void +inferQwen3Batch(struct Qwen3Model *, + const uint32_t *tokens, uint32_t ntok, + const uint32_t *req_lens, uint32_t nreq, const uint32_t *req_pos, + struct Qwen3KVCache **kv_caches, + const float *temperature, const uint32_t *topk, const float *topp, + uint32_t *output); + +/// @brief 启用或禁用调试模式（保存中间张量到文件） +__C __export void +setQwen3DebugMode(int enabled); + +#endif diff --git a/infini-qwen3-moe/include/infinicore_infer/models/qwen3_moe.h b/infini-qwen3-moe/include/infinicore_infer/models/qwen3_moe.h new file mode 100755 index 00000000..1c6b1241 --- /dev/null +++ b/infini-qwen3-moe/include/infinicore_infer/models/qwen3_moe.h @@ -0,0 +1,158 @@ +#ifndef MODEL_QWEN3_MOE_H +#define MODEL_QWEN3_MOE_H + +#include +#include +#include + +#include +#include + +struct Qwen3MoeModel; + +typedef struct +{ + /* ---- 数据类型 ---- */ + infiniDtype_t dt_logits; // 权重/激活的浮点格式 + + /* ---- 网络规模 ---- */ + size_t nlayer; // Transformer 层数 + size_t d; // 隐藏维度 hidden_size + size_t nh; // 查询头数 num_attention_heads + size_t nkvh; // 键/值头数 num_key_value_heads + size_t dh; // 单头维度 head_dim + size_t di; // mlp 中间维度 intermediate_size + size_t dctx; // 最大上下文长度 max_position_embeddings + size_t dvoc; // 词表大小 vocab_size + + /* ---- 归一化与 RoPE ---- */ + float epsilon; // RMSNorm ε rms_norm_eps + float theta; // RoPE 基值 rope_theta + + /* ---- 基础字段 ---- */ + uint32_t bos_token; // 起始符 bos_token_id + uint32_t end_token; // 结束符 eos_token_id + float attn_dropout; // attention_dropout + bool tie_embd; // 输出头与嵌入共享 tie_word_embeddings + + /* ---- MoE 特有参数 ---- */ + size_t num_experts; // 每个MoE层的专家数量 + size_t num_experts_per_tok; // 每个token选择的专家数量 (top-k) + size_t moe_intermediate_size; // MoE专家的中间维度 + size_t decoder_sparse_step; // MoE层的间隔步长 + size_t num_mlp_only_layers; // 纯MLP层的数量 + size_t *mlp_only_layers; // 纯MLP层的索引数组 + bool norm_topk_prob; // 是否对top-k概率进行归一化 + float router_aux_loss_coef; // 路由辅助损失系数 +} Qwen3MoeMeta; + +typedef struct { + /* ---- 元信息 ---- */ + size_t nlayer; + infiniDtype_t dt_norm; // 归一化权重数据类型 + infiniDtype_t dt_mat; // 矩阵权重数据类型 + int transpose_linear_weights; // 0: [in, out] 非0: [out, in] + + /* ---- 全局共享权重 ---- */ + const void *input_embd; // [dvoc, d] 或 [d, dvoc] + const void *output_embd; // 若 tie_word_embeddings==true，可与 input_embd 同址 + const void *output_norm; // [d] + + /* ---- 逐层权重（数组长度 = nlayer） ---- */ + const void **attn_norm; // [nlayer] 每层: [d] + const void **attn_q_norm; // [nlayer] 每层: [dh] - Qwen3特有的Q/K归一化 + const void **attn_k_norm; // [nlayer] 每层: [dh] - Qwen3特有的Q/K归一化 + + /* QKV 投影 - 分开存放 */ + const void **attn_q_proj; // [nlayer] 每层: [d, d] + const void **attn_k_proj; // [nlayer] 每层: [d, nkvh * dh] + const void **attn_v_proj; // [nlayer] 每层: [d, nkvh * dh] + const void **attn_o_proj; // [nlayer] 每层: [d, d] + + /* ---- MLP/MoE 权重 ---- */ + const void **mlp_norm; // [nlayer] 每层: [d] + + /* MLP 权重 (for non-MoE layers) */ + const void **mlp_gate_proj; // [nlayer] 每层: [d, di] - 非MoE层使用 + const void **mlp_up_proj; // [nlayer] 每层: [d, di] - 非MoE层使用 + const void **mlp_down_proj; // [nlayer] 每层: [di, d] - 非MoE层使用 + + /* MoE 权重 (for MoE layers) */ + const void **moe_gate; // [nlayer] 每层: [d, num_experts] - 路由门控网络 + const void ***moe_experts_gate_proj; // [nlayer][num_experts] 每专家: [d, moe_intermediate_size] + const void ***moe_experts_up_proj; // [nlayer][num_experts] 每专家: [d, moe_intermediate_size] + const void ***moe_experts_down_proj; // [nlayer][num_experts] 每专家: [moe_intermediate_size, d] + + /* ---- MoE 元信息 ---- */ + size_t num_experts; // 专家数量 + size_t num_experts_per_tok; // 每token选择的专家数 + size_t moe_intermediate_size; // MoE专家中间维度 + size_t decoder_sparse_step; // MoE层间隔 + size_t num_mlp_only_layers; // 纯MLP层数量 + size_t *mlp_only_layers; // 纯MLP层索引 + bool norm_topk_prob; // 是否归一化top-k概率 +} Qwen3MoeWeights; + +//////////////////// APIs /////////////////////// +/// @brief 创建Qwen3-MoE模型 +/// @param device 协处理器种类 +/// @param ndev 协处理器数量 +/// @param dev_ids 协处理器编号，长度为 ndev +__C __export struct Qwen3MoeModel * +createQwen3MoeModel(const Qwen3MoeMeta *, + const Qwen3MoeWeights *, + infiniDevice_t device, + int ndev, + const int *dev_ids); + +/// @brief 销毁Qwen3-MoE模型 +__C __export void +destroyQwen3MoeModel(struct Qwen3MoeModel *); + +/// @brief 创建Qwen3-MoE KV Cache +__C __export struct Qwen3MoeKVCache * +createQwen3MoeKVCache(const struct Qwen3MoeModel *); + +/// @brief 复制Qwen3-MoE KV Cache +__C __export struct Qwen3MoeKVCache * +duplicateQwen3MoeKVCache(const struct Qwen3MoeModel *, + const struct Qwen3MoeKVCache *, uint32_t seq_len); + +/// @brief 销毁Qwen3-MoE KV Cache +__C __export void +dropQwen3MoeKVCache(const struct Qwen3MoeModel *, + struct Qwen3MoeKVCache *); + +/// @brief Qwen3-MoE批次推理一轮 +/// @param tokens 输入 token 地址 +/// @param ntok 输入 token 数量 +/// @param nreq 请求数量 +/// @param req_lens 每个请求的 token 数量 +/// @param req_pos 每个请求的起始位置 +/// @param kv_caches 每个请求的 KV Cache +/// @param temperature 采样温度（0. 表示贪心采样） +/// @param topk 采样 topk（1 表示贪心采样） +/// @param topp 采样 topp +/// @param output 输出 token 数组，每个请求一个输出，长度至少为nreq +__C __export void +inferQwen3MoeBatch(struct Qwen3MoeModel *, + const uint32_t *tokens, uint32_t ntok, + const uint32_t *req_lens, uint32_t nreq, const uint32_t *req_pos, + struct Qwen3MoeKVCache **kv_caches, + const float *temperature, const uint32_t *topk, const float *topp, + uint32_t *output); + +/// @brief 启用或禁用Qwen3-MoE调试模式（保存中间张量到文件） +__C __export void +setQwen3MoeDebugMode(int enabled); + +/// @brief 获取MoE路由统计信息（用于负载均衡分析） +/// @param model Qwen3-MoE模型实例 +/// @param layer_idx 层索引 +/// @param expert_counts 输出每个专家的使用次数，长度为 num_experts +__C __export void +getQwen3MoeRouterStats(const struct Qwen3MoeModel *model, + size_t layer_idx, + uint32_t *expert_counts); + +#endif \ No newline at end of file diff --git a/infini-qwen3-moe/output/cpp_input_embeddings.txt b/infini-qwen3-moe/output/cpp_input_embeddings.txt new file mode 100755 index 00000000..e502e6e8 --- /dev/null +++ b/infini-qwen3-moe/output/cpp_input_embeddings.txt @@ -0,0 +1,108 @@ +# Tensor: input_embeddings (FP16 format) +# Shape: 1x2048 +# Total elements: 2048 +# Mean: -1.426835e-04 (converted) +# Std: 3.814535e-02 (converted) +# Min: -1.005859e-01 (converted) +# Max: 9.716797e-02 (converted) +# Data (first 100 elements as hex and converted): +0x2820 (3.222656e-02) +0xa770 (-2.905273e-02) +0xa1e8 (-1.153564e-02) +0xaa80 (-5.078125e-02) +0xa550 (-2.075195e-02) +0x2868 (3.442383e-02) +0xa188 (-1.080322e-02) +0xaa70 (-5.029297e-02) +0xa2f8 (-1.361084e-02) +0xa0d8 (-9.460449e-03) +0xa9f8 (-4.663086e-02) +0xa5a8 (-2.209473e-02) +0xaa20 (-4.785156e-02) +0x28e0 (3.808594e-02) +0xa8d8 (-3.784180e-02) +0xab18 (-5.541992e-02) +0xa970 (-4.248047e-02) +0xa340 (-1.416016e-02) +0x27e8 (3.088379e-02) +0x26a0 (2.587891e-02) +0xab98 (-5.932617e-02) +0x2a68 (5.004883e-02) +0x2b18 (5.541992e-02) +0x2b58 (5.737305e-02) +0xa100 (-9.765625e-03) +0x2410 (1.586914e-02) +0x2858 (3.393555e-02) +0x2498 (1.794434e-02) +0x2548 (2.062988e-02) +0x2ad8 (5.346680e-02) +0xab90 (-5.908203e-02) +0x16b8 (1.640320e-03) +0x2800 (3.125000e-02) +0x2938 (4.077148e-02) +0x1dc0 (5.615234e-03) +0x2ae0 (5.371094e-02) +0xab20 (-5.566406e-02) +0xa8f8 (-3.881836e-02) +0xa990 (-4.345703e-02) +0x2640 (2.441406e-02) +0xaa18 (-4.760742e-02) +0x2018 (7.995605e-03) +0xaa00 (-4.687500e-02) +0x2ab0 (5.224609e-02) +0xa3b8 (-1.507568e-02) +0x2c30 (6.542969e-02) +0xa480 (-1.757812e-02) +0x2988 (4.321289e-02) +0x2610 (2.368164e-02) +0xa030 (-8.178711e-03) +0x2668 (2.502441e-02) +0x25f8 (2.331543e-02) +0xa578 (-2.136230e-02) +0x28a0 (3.613281e-02) +0xa878 (-3.491211e-02) +0xa530 (-2.026367e-02) +0xa9f0 (-4.638672e-02) +0xa8e8 (-3.833008e-02) +0xa600 (-2.343750e-02) +0x2a40 (4.882812e-02) +0x1670 (1.571655e-03) +0x2d08 (7.861328e-02) +0xa858 (-3.393555e-02) +0x2678 (2.526855e-02) +0x2928 (4.028320e-02) +0xa830 (-3.271484e-02) +0x25d0 (2.270508e-02) +0x2120 (1.000977e-02) +0xa940 (-4.101562e-02) +0xabb8 (-6.030273e-02) +0x9c20 (-4.028320e-03) +0x2770 (2.905273e-02) +0x2af0 (5.419922e-02) +0xaca0 (-7.226562e-02) +0x2920 (4.003906e-02) +0x2ab8 (5.249023e-02) +0x2718 (2.770996e-02) +0x2740 (2.832031e-02) +0xa818 (-3.198242e-02) +0x2a30 (4.833984e-02) +0xa1d0 (-1.135254e-02) +0xa840 (-3.320312e-02) +0xa9c0 (-4.492188e-02) +0xab50 (-5.712891e-02) +0x28f8 (3.881836e-02) +0xa968 (-4.223633e-02) +0x2920 (4.003906e-02) +0x28b8 (3.686523e-02) +0xa538 (-2.038574e-02) +0x2a10 (4.736328e-02) +0x09e0 (1.792908e-04) +0xa868 (-3.442383e-02) +0xa938 (-4.077148e-02) +0x28f8 (3.881836e-02) +0x2698 (2.575684e-02) +0x2b50 (5.712891e-02) +0xa950 (-4.150391e-02) +0x2ab8 (5.249023e-02) +0x21c0 (1.123047e-02) +0x29e8 (4.614258e-02) diff --git a/infini-qwen3-moe/output/cpp_input_ids.txt b/infini-qwen3-moe/output/cpp_input_ids.txt new file mode 100755 index 00000000..8f25f89b --- /dev/null +++ b/infini-qwen3-moe/output/cpp_input_ids.txt @@ -0,0 +1,14 @@ +# Tensor: input_ids +# Shape: 1 +# Total elements: 1 +# Data type size: 4 bytes +# Statistics: +# Mean: 2.774571e-43 +# Std: 0.000000e+00 +# Min: 2.774571e-43 +# Max: 2.774571e-43 +# L1 norm (avg abs): 2.774571e-43 +# Special values - NaN: 0, Inf: 0, Zero: 0 +# ✓ Values appear to be in reasonable range +# Data: +2.774571e-43 diff --git a/infini-qwen3-moe/output/cpp_layer_0_attn_k_normed.txt b/infini-qwen3-moe/output/cpp_layer_0_attn_k_normed.txt new file mode 100755 index 00000000..40fb8634 --- /dev/null +++ b/infini-qwen3-moe/output/cpp_layer_0_attn_k_normed.txt @@ -0,0 +1,115 @@ +# Tensor: attn_k_normed +# Shape: 1x1024 +# Total elements: 1024 +# Data type size: 4 bytes +# Statistics: +# Mean: 1.320607e-02 +# Std: 3.177066e+00 +# Min: -1.000000e+01 +# Max: 1.000000e+01 +# L1 norm (avg abs): 1.092519e+00 +# Special values - NaN: 0, Inf: 0, Zero: 0 +# ✓ Values appear to be in reasonable range +# Data: +# First 50 elements: +1.310740e-01 +-3.676207e-04 +1.000000e+01 +2.511726e+00 +4.514882e-25 +2.540388e-08 +-1.000000e+01 +1.220372e-11 +2.637288e-05 +-8.579751e-02 +5.321441e-01 +1.171888e-07 +1.990574e-07 +1.000000e+01 +1.116853e-14 +1.264041e-09 +-8.071340e-03 +-1.178884e-06 +-2.275914e-10 +2.607788e-05 +-9.209678e-07 +1.840378e-03 +1.405151e-11 +1.367158e-23 +6.384474e-06 +-1.000000e+01 +4.157873e-07 +-1.112017e-02 +-1.998993e-05 +-4.940918e-09 +-2.536096e-05 +-1.122273e-06 +4.819952e-04 +-6.175636e-10 +-5.533800e-02 +1.000000e+01 +-7.685848e-12 +-9.917868e+00 +-1.000000e+01 +4.312511e-08 +1.383816e-02 +-1.069250e-02 +-1.790773e+00 +1.303584e-04 +-1.161586e-17 +2.681598e-02 +1.797062e-06 +-4.498989e-05 +3.523472e-04 +1.139505e-08 +# Last 50 elements: +6.930722e-15 +1.447315e+00 +-3.313673e-09 +-5.236782e-05 +2.840170e-05 +4.305873e-10 +-8.171557e-04 +-6.828826e-05 +-2.315294e-10 +2.994649e-12 +-5.143984e-05 +-6.152762e-17 +-7.462589e-09 +-4.170675e-19 +-4.277744e-02 +1.136325e-05 +-8.163680e-12 +1.000000e+01 +4.238631e-19 +-1.208621e-06 +-5.309336e-05 +-1.481468e-02 +-1.004391e-07 +1.712862e-19 +-3.857021e-12 +-8.286986e-15 +4.594741e-05 +4.773929e-05 +6.079393e-07 +-1.018217e-07 +3.284904e-12 +-7.163029e-05 +5.255237e-03 +2.893488e-06 +1.739355e-07 +-1.276804e-02 +-6.696644e-07 +5.253953e-04 +-8.101324e-10 +-1.000000e+01 +-2.771545e-03 +-4.914960e-11 +-5.330155e-04 +2.928715e-05 +-8.259592e-07 +1.000000e+01 +1.000000e+01 +8.319534e-13 +7.057679e-09 +-7.775052e-10 diff --git a/infini-qwen3-moe/output/cpp_layer_0_attn_k_proj_raw.txt b/infini-qwen3-moe/output/cpp_layer_0_attn_k_proj_raw.txt new file mode 100755 index 00000000..41603ff7 --- /dev/null +++ b/infini-qwen3-moe/output/cpp_layer_0_attn_k_proj_raw.txt @@ -0,0 +1,115 @@ +# Tensor: attn_k_proj_raw +# Shape: 1x1024 +# Total elements: 1024 +# Data type size: 4 bytes +# Statistics: +# Mean: 1.507176e-01 +# Std: 3.964079e+00 +# Min: -9.388967e+00 +# Max: 1.200643e+02 +# L1 norm (avg abs): 1.809617e-01 +# Special values - NaN: 0, Inf: 0, Zero: 0 +# ✓ Values appear to be in reasonable range +# Data: +# First 50 elements: +4.545788e-09 +-7.227968e-10 +4.831202e-07 +2.251037e-09 +2.630081e-31 +6.261424e-03 +-2.640204e-10 +4.733812e-18 +4.439191e-12 +-1.069506e-07 +1.521351e-06 +2.238383e-14 +1.100047e-13 +6.469512e-05 +3.559433e-22 +7.315505e-17 +-9.104132e-10 +-8.401670e-14 +-7.727718e-17 +7.239196e-12 +-2.062149e-13 +1.689252e-09 +2.909110e-17 +1.468697e-28 +6.389219e-15 +-4.447047e-03 +5.319699e-12 +-1.296380e-09 +-1.397281e-12 +-8.700138e-15 +-2.543183e-11 +-3.190253e-13 +1.333997e-11 +-1.725677e-15 +-2.830931e-10 +1.437183e-05 +-5.967951e-18 +-6.188257e-06 +-3.304688e-03 +2.577689e-14 +3.604265e-09 +-1.910571e-07 +-9.088827e-08 +4.221088e-11 +-5.820692e-24 +2.766438e-08 +6.446275e-14 +-2.140963e-12 +7.670605e-11 +2.822490e-16 +# Last 50 elements: +-9.406732e-20 +-5.132057e-14 +5.142243e-17 +-9.027856e-17 +4.488673e-14 +-1.590035e-18 +-2.596143e-21 +-2.326863e-15 +5.487493e-30 +2.545484e-12 +2.213128e-16 +9.224319e-15 +3.855626e-14 +9.688451e-14 +-1.199033e-21 +-1.328677e-22 +4.492044e-17 +-1.979558e-24 +2.359423e-18 +1.235909e-13 +2.462365e-17 +4.552621e-24 +-2.393840e-14 +2.708191e-15 +5.243916e-16 +2.086734e-16 +5.629214e-22 +-3.802460e-21 +-1.744704e-14 +-2.261808e-13 +2.486100e-15 +-8.626220e-16 +6.223572e-14 +-2.227758e-26 +-2.743256e-15 +6.388820e-16 +-9.500256e-15 +7.050857e-22 +3.556661e-16 +-1.101924e-20 +1.104224e-13 +-6.543918e-16 +1.337793e-21 +-4.119384e-19 +4.818329e-22 +-3.715171e-13 +4.572388e-27 +-7.525333e-15 +7.205990e-20 +8.030210e-15 diff --git a/infini-qwen3-moe/output/cpp_layer_0_attn_norm_output.txt b/infini-qwen3-moe/output/cpp_layer_0_attn_norm_output.txt new file mode 100755 index 00000000..f719a60a --- /dev/null +++ b/infini-qwen3-moe/output/cpp_layer_0_attn_norm_output.txt @@ -0,0 +1,116 @@ +# Tensor: attn_norm_output +# Shape: 1x2048 +# Total elements: 2048 +# Data type size: 4 bytes +# Statistics: +# Mean: -1.027799e+11 +# Std: 2.837654e+13 +# Min: -7.599175e+14 +# Max: 8.629949e+14 +# L1 norm (avg abs): 1.479860e+12 +# Special values - NaN: 0, Inf: 0, Zero: 0 +# ⚠ WARNING: Mean is extremely large (>1.000000e+06) +# ⚠ WARNING: Contains values beyond FP16 safe range +# Data: +# First 50 elements: +-3.100360e-12 +-2.883654e-11 +1.731343e-12 +-4.198440e-11 +-1.100027e-16 +-4.488482e-13 +8.053879e-12 +-6.100528e-10 +-6.624618e-15 +5.264157e-13 +1.030832e-10 +3.932434e-10 +7.918255e-15 +1.928898e-13 +4.414652e-11 +1.020276e-22 +1.223104e-11 +1.155678e-10 +-3.970114e-12 +4.364122e-13 +3.267266e-17 +7.852526e-11 +9.870220e-10 +1.784950e-07 +-4.568640e-10 +1.044339e-13 +2.979524e-12 +-3.656426e-14 +-3.136001e-12 +1.248668e-11 +2.053324e-09 +1.897818e-13 +-7.644810e-13 +1.215835e-16 +-5.809005e-10 +8.832943e-13 +-5.682219e-10 +1.069454e-10 +3.384527e-12 +2.042629e-11 +-1.826344e-11 +-1.425030e-10 +-5.604509e-12 +2.293524e-10 +9.263304e-07 +-1.726994e-12 +1.129342e-11 +4.194932e-08 +3.342272e-09 +3.391216e-11 +# Last 50 elements: +-3.559354e+07 +-8.115636e+01 +2.372335e+08 +1.779508e+06 +-4.082739e+07 +1.582318e+10 +9.793274e+07 +-2.075759e+05 +1.371022e+10 +-1.423770e+01 +2.835175e+02 +4.546840e+06 +-5.449829e+11 +-1.427806e+06 +1.606387e+03 +-1.099949e+08 +7.540005e+07 +-9.085531e+05 +-4.540288e+09 +2.641586e+10 +-1.272420e+04 +3.748120e+06 +6.142491e+07 +-8.052753e+06 +1.287670e+02 +1.398396e+03 +9.712456e+02 +-2.228687e+03 +-1.028828e+09 +2.285661e+05 +-5.317999e+06 +2.399118e+05 +2.051816e+13 +-5.816947e+01 +-9.160589e+06 +8.632875e+06 +1.039456e+06 +2.478465e+06 +-1.084403e+00 +7.801833e+03 +5.027351e+07 +4.967899e+03 +5.972693e+06 +1.585779e-04 +-8.182478e+05 +-1.820937e+06 +1.143794e+02 +1.691040e+08 +3.798064e+06 +2.041780e+07 diff --git a/infini-qwen3-moe/output/cpp_layer_0_attn_q_normed.txt b/infini-qwen3-moe/output/cpp_layer_0_attn_q_normed.txt new file mode 100755 index 00000000..87096685 --- /dev/null +++ b/infini-qwen3-moe/output/cpp_layer_0_attn_q_normed.txt @@ -0,0 +1,115 @@ +# Tensor: attn_q_normed +# Shape: 1x2048 +# Total elements: 2048 +# Data type size: 4 bytes +# Statistics: +# Mean: 8.596502e-03 +# Std: 2.723236e+00 +# Min: -1.000000e+01 +# Max: 1.000000e+01 +# L1 norm (avg abs): 8.320960e-01 +# Special values - NaN: 0, Inf: 0, Zero: 0 +# ✓ Values appear to be in reasonable range +# Data: +# First 50 elements: +-2.932223e-26 +-1.296833e-05 +-5.325266e-11 +1.193230e-10 +-2.035364e-05 +-1.653100e-09 +-2.904748e-10 +3.672283e-08 +1.298796e-03 +-8.323984e-04 +2.467656e-03 +-3.343207e-05 +5.813347e-11 +1.505783e-02 +3.404250e-13 +3.436869e-04 +3.930599e-17 +4.965085e-05 +1.134563e+00 +-5.711968e-04 +1.752410e-07 +1.508937e-02 +1.267860e-02 +-4.228256e-02 +5.086251e-09 +-1.184920e-02 +1.000000e+01 +4.008998e-01 +-4.156569e-02 +4.193974e-02 +3.713896e-06 +1.000000e+01 +-2.814211e-10 +1.116390e-04 +4.547528e-17 +-9.530256e-12 +-1.639004e-10 +-1.187140e-05 +-2.262199e-13 +-5.237858e-05 +9.163356e-04 +-2.223173e-03 +-2.002535e-03 +2.572868e-03 +2.132376e-02 +2.214066e-01 +4.834574e-06 +9.236157e-05 +-6.470990e-05 +-5.877244e-02 +# Last 50 elements: +6.930722e-15 +1.447315e+00 +-3.313673e-09 +-5.236782e-05 +2.840170e-05 +4.305873e-10 +-8.171557e-04 +-6.828826e-05 +-2.315294e-10 +2.994649e-12 +-5.143984e-05 +-6.152762e-17 +-7.462589e-09 +-4.170675e-19 +-4.277744e-02 +1.136325e-05 +-8.163680e-12 +1.000000e+01 +4.238631e-19 +-1.208621e-06 +-5.309336e-05 +-1.481468e-02 +-1.004391e-07 +1.712862e-19 +-3.857021e-12 +-8.286986e-15 +4.594741e-05 +4.773929e-05 +6.079393e-07 +-1.018217e-07 +3.284904e-12 +-7.163029e-05 +5.255237e-03 +2.893488e-06 +1.739355e-07 +-1.276804e-02 +-6.696644e-07 +5.253953e-04 +-8.101324e-10 +-1.000000e+01 +-2.771545e-03 +-4.914960e-11 +-5.330155e-04 +2.928715e-05 +-8.259592e-07 +1.000000e+01 +1.000000e+01 +8.319534e-13 +7.057679e-09 +-7.775052e-10 diff --git a/infini-qwen3-moe/output/cpp_layer_0_attn_q_proj_raw.txt b/infini-qwen3-moe/output/cpp_layer_0_attn_q_proj_raw.txt new file mode 100755 index 00000000..863f463c --- /dev/null +++ b/infini-qwen3-moe/output/cpp_layer_0_attn_q_proj_raw.txt @@ -0,0 +1,115 @@ +# Tensor: attn_q_proj_raw +# Shape: 1x2048 +# Total elements: 2048 +# Data type size: 4 bytes +# Statistics: +# Mean: -8.179451e+00 +# Std: 3.297291e+02 +# Min: -1.486108e+04 +# Max: 1.200643e+02 +# L1 norm (avg abs): 8.365680e+00 +# Special values - NaN: 0, Inf: 0, Zero: 0 +# ✓ Values appear to be in reasonable range +# Data: +# First 50 elements: +-2.984542e-30 +-6.481884e-12 +-6.319794e-16 +8.223318e-14 +-3.221370e-12 +9.854055e-04 +-4.083557e-12 +1.247232e-14 +6.700871e-10 +-3.113423e-10 +7.882484e-10 +-6.627551e-12 +6.739565e-18 +1.342635e-08 +7.575759e-19 +5.280515e-10 +1.169383e-23 +3.210699e-11 +1.883035e-07 +-9.081388e-11 +3.755240e-14 +1.260143e-09 +5.736301e-10 +-2.549246e-10 +1.163216e-12 +-4.336150e-08 +1.579976e-07 +1.678198e-07 +-3.101403e-08 +1.809832e-09 +1.138569e-13 +2.024194e-03 +-1.601783e-16 +5.800814e-12 +3.723537e-21 +-8.657548e-17 +-2.376141e-17 +-1.464625e-12 +-4.157527e-16 +-5.930162e-12 +4.295901e-10 +-5.335359e-11 +-5.942707e-09 +6.205521e-10 +2.426230e-09 +3.298649e-08 +8.650232e-12 +1.288357e-10 +-1.668049e-11 +-1.859898e-07 +# Last 50 elements: +-9.406732e-20 +-5.132057e-14 +5.142243e-17 +-9.027856e-17 +4.488673e-14 +-1.590035e-18 +-2.596143e-21 +-2.326863e-15 +5.487493e-30 +2.545484e-12 +2.213128e-16 +9.224319e-15 +3.855626e-14 +9.688451e-14 +-1.199033e-21 +-1.328677e-22 +4.492044e-17 +-1.979558e-24 +2.359423e-18 +1.235909e-13 +2.462365e-17 +4.552621e-24 +-2.393840e-14 +2.708191e-15 +5.243916e-16 +2.086734e-16 +5.629214e-22 +-3.802460e-21 +-1.744704e-14 +-2.261808e-13 +2.486100e-15 +-8.626220e-16 +6.223572e-14 +-2.227758e-26 +-2.743256e-15 +6.388820e-16 +-9.500256e-15 +7.050857e-22 +3.556661e-16 +-1.101924e-20 +1.104224e-13 +-6.543918e-16 +1.337793e-21 +-4.119384e-19 +4.818329e-22 +-3.715171e-13 +4.572388e-27 +-7.525333e-15 +7.205990e-20 +8.030210e-15 diff --git a/infini-qwen3-moe/output/cpp_layer_0_attn_residual_output.txt b/infini-qwen3-moe/output/cpp_layer_0_attn_residual_output.txt new file mode 100755 index 00000000..65a24cdf --- /dev/null +++ b/infini-qwen3-moe/output/cpp_layer_0_attn_residual_output.txt @@ -0,0 +1,115 @@ +# Tensor: attn_residual_output +# Shape: 1x2048 +# Total elements: 2048 +# Data type size: 4 bytes +# Statistics: +# Mean: 1.538105e-03 +# Std: 7.716089e-02 +# Min: -3.021551e-01 +# Max: 3.479472e+00 +# L1 norm (avg abs): 1.865306e-03 +# Special values - NaN: 0, Inf: 0, Zero: 0 +# ✓ Values appear to be in reasonable range +# Data: +# First 50 elements: +-1.047234e-14 +1.230577e-14 +-2.495599e-18 +-4.967651e-06 +-2.018149e-09 +8.735565e-10 +8.252172e-12 +-2.586568e-08 +2.702015e-09 +9.735753e-05 +-4.760062e-16 +-1.851240e-25 +8.876241e-11 +-3.579259e-14 +8.738076e-10 +-1.536230e-06 +-1.608290e-07 +4.744172e-07 +-1.558418e-09 +1.354471e-15 +-5.428957e-06 +6.663329e-19 +-1.064966e-10 +5.796012e-09 +4.189141e-14 +-4.733668e-11 +-1.291849e-07 +-6.719001e-10 +-3.159241e-08 +9.046591e-12 +2.960016e-10 +5.394291e-19 +-8.355678e-10 +7.317854e-10 +2.170812e-09 +1.089356e-08 +-5.415180e-11 +-5.462103e-09 +5.334016e-07 +4.460082e-11 +1.061215e-11 +3.636509e-24 +-5.135868e-10 +-5.626897e-08 +1.275711e-08 +3.103287e-19 +-3.857581e-11 +-1.252401e-10 +3.683057e-15 +-2.127536e-09 +# Last 50 elements: +-7.466156e-11 +-2.455471e-14 +-3.012736e-10 +-6.758286e-14 +2.602937e-12 +3.471575e-11 +6.081415e-11 +-1.205085e-13 +-4.639961e-15 +-1.632211e-16 +-8.941126e-13 +-1.011338e-14 +3.164465e-13 +-1.881740e-21 +-4.318848e-13 +-9.973109e-15 +4.312841e-11 +-1.219766e-12 +5.266368e-14 +4.713704e-13 +-2.106502e-13 +1.644674e-11 +-6.854101e-12 +-8.400270e-14 +4.084513e-11 +-7.921387e-04 +4.003689e-11 +3.050797e-11 +-2.734910e-29 +1.328185e-11 +-9.910929e-14 +1.802503e-10 +-4.322836e-14 +4.322952e-11 +-2.055736e-14 +4.487748e-22 +6.911179e-12 +-1.028365e-15 +-1.572474e-25 +3.082451e-15 +2.638666e-12 +-3.661477e-13 +-1.973131e-13 +-7.076717e-13 +-2.059727e-11 +2.921814e-12 +-8.937336e-10 +-1.077753e-14 +4.844420e-14 +1.062635e-10 diff --git a/infini-qwen3-moe/output/cpp_layer_0_attn_v_proj_raw.txt b/infini-qwen3-moe/output/cpp_layer_0_attn_v_proj_raw.txt new file mode 100755 index 00000000..4810836b --- /dev/null +++ b/infini-qwen3-moe/output/cpp_layer_0_attn_v_proj_raw.txt @@ -0,0 +1,115 @@ +# Tensor: attn_v_proj_raw +# Shape: 1x1024 +# Total elements: 1024 +# Data type size: 4 bytes +# Statistics: +# Mean: -2.750205e+03 +# Std: 9.669839e+04 +# Min: -2.330890e+06 +# Max: 1.034882e+06 +# L1 norm (avg abs): 6.491318e+03 +# Special values - NaN: 0, Inf: 0, Zero: 0 +# ⚠ WARNING: Contains values beyond FP16 safe range +# Data: +# First 50 elements: +7.855809e-17 +1.102023e-14 +-3.908537e-21 +-3.827382e-16 +-4.883356e-19 +-1.150644e-16 +2.621010e-19 +-3.307425e-28 +-2.188090e-20 +1.638687e-21 +-3.430574e-18 +9.325265e-21 +-3.723061e-15 +-1.673890e-19 +-1.767004e-14 +-1.405142e-14 +-3.133069e-16 +1.996685e-17 +2.734793e-16 +-9.472639e-15 +-1.985917e-24 +4.186079e-21 +-5.010239e-14 +-2.137530e-17 +-2.677309e-14 +-6.526431e-19 +-3.983571e-16 +2.772098e-19 +-2.856081e-16 +-1.158925e-21 +1.313655e-20 +1.731542e-16 +3.938714e-17 +7.542901e-19 +-8.577726e-14 +-1.149896e-17 +1.863244e-24 +-2.149917e-18 +-2.730639e-12 +-1.283110e-14 +3.563783e-13 +-1.538672e-19 +-5.290349e-24 +1.655807e-14 +1.428221e-16 +1.177590e-14 +-1.213426e-13 +1.034471e-18 +4.287490e-16 +5.893722e-21 +# Last 50 elements: +-2.905482e-05 +3.769574e+00 +1.428017e-06 +1.829156e-03 +-3.345160e-03 +-1.329744e+03 +-4.009689e-04 +5.477390e-05 +-2.659701e+00 +-2.463071e-01 +-7.458351e-12 +-4.180470e-04 +7.991121e+03 +-1.983787e-16 +-3.398012e-02 +7.927093e-05 +-5.616296e+00 +1.284941e-08 +-1.719176e-06 +-2.167707e+00 +-4.898409e+00 +8.957778e-10 +1.355272e-12 +-1.234176e-02 +-1.551404e+01 +1.202191e-01 +3.039030e-07 +4.313570e-20 +-2.887513e-03 +2.907754e-02 +-5.204398e-02 +4.171501e-04 +-1.871388e-04 +2.951900e-12 +-5.489370e-05 +1.290793e+00 +2.543847e-03 +3.091612e-16 +1.217426e-11 +-2.529018e-02 +-2.086215e-04 +-2.190959e-04 +-9.418075e-08 +3.497358e-06 +-1.069573e-08 +1.123944e-04 +-2.155048e-03 +-7.212871e-02 +3.011558e-02 +6.366351e+00 diff --git a/infini-qwen3-moe/output/cpp_layer_0_input_hidden_states.txt b/infini-qwen3-moe/output/cpp_layer_0_input_hidden_states.txt new file mode 100755 index 00000000..d05e4ee5 --- /dev/null +++ b/infini-qwen3-moe/output/cpp_layer_0_input_hidden_states.txt @@ -0,0 +1,108 @@ +# Tensor: input_hidden_states (FP16 format) +# Shape: 1x2048 +# Total elements: 2048 +# Mean: -1.426835e-04 (converted) +# Std: 3.814535e-02 (converted) +# Min: -1.005859e-01 (converted) +# Max: 9.716797e-02 (converted) +# Data (first 100 elements as hex and converted): +0x2820 (3.222656e-02) +0xa770 (-2.905273e-02) +0xa1e8 (-1.153564e-02) +0xaa80 (-5.078125e-02) +0xa550 (-2.075195e-02) +0x2868 (3.442383e-02) +0xa188 (-1.080322e-02) +0xaa70 (-5.029297e-02) +0xa2f8 (-1.361084e-02) +0xa0d8 (-9.460449e-03) +0xa9f8 (-4.663086e-02) +0xa5a8 (-2.209473e-02) +0xaa20 (-4.785156e-02) +0x28e0 (3.808594e-02) +0xa8d8 (-3.784180e-02) +0xab18 (-5.541992e-02) +0xa970 (-4.248047e-02) +0xa340 (-1.416016e-02) +0x27e8 (3.088379e-02) +0x26a0 (2.587891e-02) +0xab98 (-5.932617e-02) +0x2a68 (5.004883e-02) +0x2b18 (5.541992e-02) +0x2b58 (5.737305e-02) +0xa100 (-9.765625e-03) +0x2410 (1.586914e-02) +0x2858 (3.393555e-02) +0x2498 (1.794434e-02) +0x2548 (2.062988e-02) +0x2ad8 (5.346680e-02) +0xab90 (-5.908203e-02) +0x16b8 (1.640320e-03) +0x2800 (3.125000e-02) +0x2938 (4.077148e-02) +0x1dc0 (5.615234e-03) +0x2ae0 (5.371094e-02) +0xab20 (-5.566406e-02) +0xa8f8 (-3.881836e-02) +0xa990 (-4.345703e-02) +0x2640 (2.441406e-02) +0xaa18 (-4.760742e-02) +0x2018 (7.995605e-03) +0xaa00 (-4.687500e-02) +0x2ab0 (5.224609e-02) +0xa3b8 (-1.507568e-02) +0x2c30 (6.542969e-02) +0xa480 (-1.757812e-02) +0x2988 (4.321289e-02) +0x2610 (2.368164e-02) +0xa030 (-8.178711e-03) +0x2668 (2.502441e-02) +0x25f8 (2.331543e-02) +0xa578 (-2.136230e-02) +0x28a0 (3.613281e-02) +0xa878 (-3.491211e-02) +0xa530 (-2.026367e-02) +0xa9f0 (-4.638672e-02) +0xa8e8 (-3.833008e-02) +0xa600 (-2.343750e-02) +0x2a40 (4.882812e-02) +0x1670 (1.571655e-03) +0x2d08 (7.861328e-02) +0xa858 (-3.393555e-02) +0x2678 (2.526855e-02) +0x2928 (4.028320e-02) +0xa830 (-3.271484e-02) +0x25d0 (2.270508e-02) +0x2120 (1.000977e-02) +0xa940 (-4.101562e-02) +0xabb8 (-6.030273e-02) +0x9c20 (-4.028320e-03) +0x2770 (2.905273e-02) +0x2af0 (5.419922e-02) +0xaca0 (-7.226562e-02) +0x2920 (4.003906e-02) +0x2ab8 (5.249023e-02) +0x2718 (2.770996e-02) +0x2740 (2.832031e-02) +0xa818 (-3.198242e-02) +0x2a30 (4.833984e-02) +0xa1d0 (-1.135254e-02) +0xa840 (-3.320312e-02) +0xa9c0 (-4.492188e-02) +0xab50 (-5.712891e-02) +0x28f8 (3.881836e-02) +0xa968 (-4.223633e-02) +0x2920 (4.003906e-02) +0x28b8 (3.686523e-02) +0xa538 (-2.038574e-02) +0x2a10 (4.736328e-02) +0x09e0 (1.792908e-04) +0xa868 (-3.442383e-02) +0xa938 (-4.077148e-02) +0x28f8 (3.881836e-02) +0x2698 (2.575684e-02) +0x2b50 (5.712891e-02) +0xa950 (-4.150391e-02) +0x2ab8 (5.249023e-02) +0x21c0 (1.123047e-02) +0x29e8 (4.614258e-02) diff --git a/infini-qwen3-moe/output/cpp_layer_0_layer_output.txt b/infini-qwen3-moe/output/cpp_layer_0_layer_output.txt new file mode 100755 index 00000000..b0f9eb84 --- /dev/null +++ b/infini-qwen3-moe/output/cpp_layer_0_layer_output.txt @@ -0,0 +1,115 @@ +# Tensor: layer_output +# Shape: 1x2048 +# Total elements: 2048 +# Data type size: 4 bytes +# Statistics: +# Mean: 1.327144e+02 +# Std: 6.005331e+03 +# Min: -3.342569e+01 +# Max: 2.718390e+05 +# L1 norm (avg abs): 1.327572e+02 +# Special values - NaN: 0, Inf: 0, Zero: 0 +# ✓ Values appear to be in reasonable range +# Data: +# First 50 elements: +-6.254444e-17 +-1.864856e-08 +8.515404e-04 +-8.647280e-10 +3.445503e-06 +5.306644e-12 +2.511584e-05 +1.594691e-04 +4.427903e-05 +1.351308e-09 +-2.939051e-06 +3.586223e-11 +4.425444e-12 +2.375003e-09 +6.256033e-05 +-4.152077e-04 +2.155716e-13 +5.929780e-05 +-1.401656e-05 +-1.773127e-14 +2.531627e-10 +9.037456e-11 +-6.860614e-06 +6.492863e-04 +-1.164555e-04 +5.803131e-11 +-2.043446e-10 +-1.894818e-09 +-5.524510e-05 +-4.232009e-07 +3.999750e-07 +2.221357e-10 +-3.945735e-14 +7.887011e-07 +3.207259e-06 +-3.618960e-13 +-2.387178e-06 +3.757223e-07 +5.690768e-05 +3.050051e-06 +-5.594792e-09 +2.593135e-07 +2.691700e-08 +-1.729433e-09 +1.294411e-06 +-4.888456e-07 +3.764457e-08 +4.082262e-21 +-3.049619e-07 +-6.316031e-12 +# Last 50 elements: +9.877380e-09 +-3.251762e-08 +9.561470e-12 +-1.178374e-13 +7.987752e-10 +-1.434319e-15 +9.552248e-12 +-6.062709e-10 +-1.384145e-10 +-3.820872e-06 +-6.312822e-13 +1.194881e-09 +-1.447304e-10 +-1.501377e-07 +-2.804378e-06 +7.191108e-11 +-1.183669e-09 +5.452435e-08 +2.970397e-08 +-4.778482e-09 +-1.316007e-10 +-4.268899e-12 +9.178270e-14 +-1.402483e-10 +2.230734e-11 +-4.076159e-24 +2.283102e-07 +6.713006e-12 +3.476159e-19 +-2.331215e-21 +2.477314e-10 +1.184433e-12 +-2.350026e-06 +4.894605e-08 +3.012166e-14 +2.339637e-12 +2.111241e-07 +-4.709640e-12 +-8.852782e-11 +5.505982e-09 +-2.853353e-37 +-2.970810e-09 +-7.330150e-03 +8.047657e-10 +-9.283050e-13 +5.888642e-12 +-2.511064e-08 +-3.618962e-09 +1.136328e-09 +2.250455e-07 diff --git a/infini-qwen3-moe/output/cpp_layer_0_mlp_gate_proj.txt b/infini-qwen3-moe/output/cpp_layer_0_mlp_gate_proj.txt new file mode 100755 index 00000000..22be75a5 --- /dev/null +++ b/infini-qwen3-moe/output/cpp_layer_0_mlp_gate_proj.txt @@ -0,0 +1,115 @@ +# Tensor: mlp_gate_proj +# Shape: 1x6144 +# Total elements: 6144 +# Data type size: 4 bytes +# Statistics: +# Mean: -4.412920e-02 +# Std: 1.432461e+00 +# Min: -8.286533e+01 +# Max: 3.557556e+00 +# L1 norm (avg abs): 4.607023e-02 +# Special values - NaN: 0, Inf: 0, Zero: 0 +# ✓ Values appear to be in reasonable range +# Data: +# First 50 elements: +1.459077e-07 +-6.153776e-13 +5.244484e-09 +1.668780e-02 +-3.817330e-03 +4.777814e-08 +1.986190e-11 +-8.209082e-05 +-2.013044e-07 +-3.751819e-04 +-5.551227e-09 +-5.443841e-04 +-3.030578e-08 +-1.226716e-04 +2.930083e-18 +-2.833919e-03 +6.643252e-03 +3.571108e-07 +4.521011e-08 +-3.240723e-09 +-1.980692e-04 +1.384939e-07 +3.266606e-06 +7.043137e-05 +-3.274078e-06 +7.942479e-07 +-2.348568e-07 +-1.684799e-08 +3.108459e-11 +1.190127e-01 +-9.015595e-03 +-6.787063e-06 +1.184194e-05 +-2.521506e-04 +1.547889e-10 +-1.458855e-03 +-1.296510e-08 +-5.143061e-05 +-2.294775e-09 +1.051934e-06 +-3.939886e-06 +-2.076164e-05 +-2.910692e-05 +-5.334536e-07 +-3.706306e-06 +-9.921802e-15 +4.128962e-05 +-1.456301e-10 +-3.969389e-06 +3.757233e-07 +# Last 50 elements: +1.890290e-06 +3.612795e-14 +-8.740879e-09 +-2.625694e-06 +-4.034229e-05 +9.202035e-04 +6.132727e-09 +-2.721964e-04 +4.982315e-09 +-2.417762e-08 +6.423713e-09 +1.444640e-08 +-2.255045e-05 +-5.254951e-08 +-4.023037e-10 +-1.507271e-09 +-2.255752e-07 +3.300030e-12 +-6.831041e-06 +-3.288804e-06 +7.700954e-07 +-4.735051e-10 +5.462751e-09 +-3.795730e-05 +2.512895e-09 +2.403859e-09 +-1.053088e-12 +2.629348e-07 +1.181462e-07 +-2.389461e-14 +-9.339806e-06 +-1.055688e-15 +-2.070167e-12 +3.823019e-13 +-3.392350e-08 +1.209758e-04 +-2.616835e-14 +-1.502574e-06 +6.995173e-05 +-6.897188e-12 +-3.411495e-16 +-3.162384e-06 +-7.942563e-07 +-8.622288e-13 +3.284466e-09 +1.862197e-05 +4.950372e-11 +-7.281615e-10 +7.899640e-05 +3.761818e-04 diff --git a/infini-qwen3-moe/output/cpp_layer_0_mlp_intermediate.txt b/infini-qwen3-moe/output/cpp_layer_0_mlp_intermediate.txt new file mode 100755 index 00000000..c1ba7a72 --- /dev/null +++ b/infini-qwen3-moe/output/cpp_layer_0_mlp_intermediate.txt @@ -0,0 +1,115 @@ +# Tensor: mlp_intermediate +# Shape: 1x6144 +# Total elements: 6144 +# Data type size: 4 bytes +# Statistics: +# Mean: 1.129879e-03 +# Std: 1.119101e-01 +# Min: -1.685311e+00 +# Max: 8.602200e+00 +# L1 norm (avg abs): 1.877799e-03 +# Special values - NaN: 0, Inf: 0, Zero: 0 +# ✓ Values appear to be in reasonable range +# Data: +# First 50 elements: +-1.586659e-13 +9.603500e-18 +-6.008326e-16 +-1.192849e-04 +1.348094e-08 +2.489500e-12 +7.751732e-15 +-1.593225e-18 +-5.208976e-13 +2.480344e-11 +5.621742e-17 +1.117837e-11 +-1.475451e-16 +1.141982e-13 +-3.623443e-31 +2.480461e-11 +-3.949718e-10 +-1.245466e-25 +1.644820e-16 +-1.038320e-12 +-7.270235e-20 +-9.209835e-22 +5.895919e-11 +-6.216035e-16 +-5.574914e-16 +-4.421474e-14 +5.479108e-18 +-6.197144e-18 +-1.725525e-15 +1.069417e-07 +-2.130429e-16 +9.828403e-27 +-3.305517e-12 +1.624193e-12 +-1.685191e-20 +-1.234286e-11 +-9.845831e-14 +-1.263548e-20 +-1.816428e-14 +-4.541111e-13 +2.191360e-17 +-1.833049e-14 +1.995875e-19 +-5.288326e-14 +-7.626701e-18 +-2.148867e-19 +2.053084e-13 +6.393713e-21 +-8.645241e-10 +8.061957e-14 +# Last 50 elements: +1.890290e-06 +3.612795e-14 +-8.740879e-09 +-2.625694e-06 +-4.034229e-05 +9.202035e-04 +6.132727e-09 +-2.721964e-04 +4.982315e-09 +-2.417762e-08 +6.423713e-09 +1.444640e-08 +-2.255045e-05 +-5.254951e-08 +-4.023037e-10 +-1.507271e-09 +-2.255752e-07 +3.300030e-12 +-6.831041e-06 +-3.288804e-06 +7.700954e-07 +-4.735051e-10 +5.462751e-09 +-3.795730e-05 +2.512895e-09 +2.403859e-09 +-1.053088e-12 +2.629348e-07 +1.181462e-07 +-2.389461e-14 +-9.339806e-06 +-1.055688e-15 +-2.070167e-12 +3.823019e-13 +-3.392350e-08 +1.209758e-04 +-2.616835e-14 +-1.502574e-06 +6.995173e-05 +-6.897188e-12 +-3.411495e-16 +-3.162384e-06 +-7.942563e-07 +-8.622288e-13 +3.284466e-09 +1.862197e-05 +4.950372e-11 +-7.281615e-10 +7.899640e-05 +3.761818e-04 diff --git a/infini-qwen3-moe/output/cpp_layer_0_mlp_norm_output.txt b/infini-qwen3-moe/output/cpp_layer_0_mlp_norm_output.txt new file mode 100755 index 00000000..a0101a2d --- /dev/null +++ b/infini-qwen3-moe/output/cpp_layer_0_mlp_norm_output.txt @@ -0,0 +1,115 @@ +# Tensor: mlp_norm_output +# Shape: 1x2048 +# Total elements: 2048 +# Data type size: 4 bytes +# Statistics: +# Mean: -8.254847e+00 +# Std: 3.297172e+02 +# Min: -1.486108e+04 +# Max: 8.701296e+00 +# L1 norm (avg abs): 8.275272e+00 +# Special values - NaN: 0, Inf: 0, Zero: 0 +# ✓ Values appear to be in reasonable range +# Data: +# First 50 elements: +-2.777300e-14 +4.056357e-14 +1.047034e-17 +-3.998033e-05 +-8.307720e-11 +1.281475e-09 +3.755657e-11 +-8.297758e-08 +1.406864e-10 +1.006961e-04 +-8.507206e-16 +-9.575475e-25 +6.464428e-10 +-7.447419e-14 +3.131480e-09 +-4.058839e-06 +-2.027418e-07 +2.471733e-07 +-1.011012e-08 +5.223711e-15 +-3.570161e-07 +5.032583e-18 +-4.954622e-10 +2.417862e-09 +7.135063e-14 +-1.952634e-10 +-4.111069e-07 +-1.070976e-09 +-2.278866e-07 +4.414592e-11 +6.429082e-10 +2.469500e-18 +-2.643815e-09 +8.846160e-10 +9.207246e-09 +5.940744e-08 +-4.259403e-10 +-1.179751e-08 +6.712564e-07 +1.220162e-10 +1.872359e-11 +1.094378e-23 +-2.505527e-09 +-3.272218e-07 +8.313705e-07 +1.166805e-18 +-5.722127e-11 +-5.375596e-09 +2.272661e-14 +-4.575226e-09 +# Last 50 elements: +-7.779409e-15 +3.916387e-19 +1.479357e-12 +6.253735e-23 +4.899074e-10 +2.790914e-17 +-2.280294e-09 +7.400006e-26 +-6.291075e-19 +4.443924e-13 +-9.252887e-08 +5.901923e-10 +9.376228e-11 +-4.259743e-10 +-5.647300e-12 +2.607350e-09 +-4.848530e-21 +4.971812e-10 +8.443999e-13 +8.160070e-13 +-4.055025e-08 +2.148228e-16 +1.538932e-10 +-3.065757e-09 +-3.840586e-07 +9.386429e-12 +-6.443005e-11 +-4.575662e-07 +1.753985e-08 +1.294311e-11 +-4.155093e-11 +-1.469167e-06 +1.050206e-16 +1.534618e-15 +1.003936e-21 +4.207610e-08 +-2.837413e-12 +-5.628389e-25 +-1.738761e-10 +-2.439797e-11 +-5.077916e-17 +-3.525695e-19 +8.734733e-05 +-2.953367e-06 +-8.483660e-17 +-6.648759e-13 +-1.063960e-14 +6.202052e-14 +4.972706e-10 +-2.150803e-18 diff --git a/infini-qwen3-moe/output/cpp_layer_0_mlp_up_proj.txt b/infini-qwen3-moe/output/cpp_layer_0_mlp_up_proj.txt new file mode 100755 index 00000000..2bb49c26 --- /dev/null +++ b/infini-qwen3-moe/output/cpp_layer_0_mlp_up_proj.txt @@ -0,0 +1,115 @@ +# Tensor: mlp_up_proj +# Shape: 1x6144 +# Total elements: 6144 +# Data type size: 4 bytes +# Statistics: +# Mean: 2.074294e+05 +# Std: 1.227941e+06 +# Min: -7.821651e+00 +# Max: 5.774483e+07 +# L1 norm (avg abs): 2.074298e+05 +# Special values - NaN: 0, Inf: 0, Zero: 0 +# ⚠ WARNING: Contains values beyond FP16 safe range +# Data: +# First 50 elements: +-8.092209e-07 +-3.819099e-05 +-1.146579e-07 +-3.961656e-04 +-5.633445e-04 +3.772339e-05 +2.328764e-04 +5.265509e-13 +1.617906e-05 +-2.692974e-06 +-2.667749e-08 +-9.059810e-07 +1.841607e-08 +-2.644879e-08 +-1.511861e-13 +-1.182541e-06 +-3.745764e-09 +-2.140591e-19 +2.585977e-09 +6.472858e-04 +1.067056e-14 +-6.514823e-15 +6.487337e-06 +-2.710187e-12 +1.485911e-09 +-2.388459e-08 +-1.257068e-10 +1.267245e-09 +-5.584162e-05 +1.515537e-08 +8.024195e-12 +-1.059850e-20 +-1.104454e-07 +-1.868532e-07 +-1.175228e-10 +7.382305e-07 +3.605783e-05 +7.561608e-15 +2.707996e-05 +-1.925063e-07 +-2.866768e-11 +1.240946e-08 +-1.002368e-13 +5.670418e-07 +1.376861e-11 +5.202839e-05 +1.653260e-09 +-1.324900e-10 +1.561696e-03 +1.565363e-07 +# Last 50 elements: +5.305851e+04 +6.728382e+05 +1.730214e+03 +3.588532e+03 +1.013140e+03 +2.224854e+05 +3.221292e+05 +1.034252e+03 +2.243412e+04 +8.817059e+05 +1.835797e+05 +1.270339e+01 +2.511285e+06 +4.196182e+00 +8.196278e-03 +1.300938e+04 +4.122025e+05 +1.549100e+05 +8.370764e+02 +2.560412e+06 +3.145626e+02 +1.271411e+02 +7.507084e+05 +2.384405e+04 +7.465054e+03 +2.081531e+05 +2.892711e+05 +1.291653e+05 +2.166673e+04 +2.542155e+05 +3.896870e+05 +3.412010e+04 +3.075521e+04 +3.428946e+06 +5.909950e+05 +2.153050e+05 +1.281441e+05 +2.020209e+05 +8.858968e+05 +4.782037e+01 +5.971960e+04 +1.882177e+03 +1.307384e+04 +2.868514e+03 +5.416222e+03 +2.583354e+05 +2.666022e+04 +1.919240e+06 +6.360459e+05 +1.275417e+04 diff --git a/infini-qwen3-moe/reference/.DS_Store b/infini-qwen3-moe/reference/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..db35820c02eedab1eca70de30d44dc77eb88780d GIT binary patch literal 6148 zcmeHK%}N9@3{G@~6)$`ASZ`hzU!X4IRfK&2Wp)<#Fk@Fly!P!pdGvk!l2qBj3VIQd z1ez~xlKy-!?SzPU@z~9YW<)ea6J#+eBEwbJUI{(`vd(cv*ECR~b@x!2=r0b*-Vf=X zo@h;Xw6p(J-=@^#ST%QlJ72ESu+3yKhErH@hoe z=>}Xm1I~am;0!neKVd)~H^}tJ(6uw*3^)TH49NWu&;(<_Y^YZUl$HR%35Hc*OD!Qe zpXr2;LLONqgjj`2|9V!>=^>BOb@;Ii_2@xr<~<_{W991UGN1J1yhfqm=F z(sc=FZ;v>P-LiOXU@V2^$R(2;xOSe@n% aVk0gV%!V?HjB7g3e*}~eSI)pMFz^O{m^E(z literal 0 HcmV?d00001 diff --git a/infini-qwen3-moe/reference/qwen3_moe/__init__.py b/infini-qwen3-moe/reference/qwen3_moe/__init__.py new file mode 100755 index 00000000..715effd8 --- /dev/null +++ b/infini-qwen3-moe/reference/qwen3_moe/__init__.py @@ -0,0 +1,27 @@ +# Copyright 2024 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import TYPE_CHECKING + +from ...utils import _LazyModule +from ...utils.import_utils import define_import_structure + + +if TYPE_CHECKING: + from .configuration_qwen3_moe import * + from .modeling_qwen3_moe import * +else: + import sys + + _file = globals()["__file__"] + sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__) diff --git a/infini-qwen3-moe/reference/qwen3_moe/configuration_qwen3_moe.py b/infini-qwen3-moe/reference/qwen3_moe/configuration_qwen3_moe.py new file mode 100755 index 00000000..a23d19c1 --- /dev/null +++ b/infini-qwen3-moe/reference/qwen3_moe/configuration_qwen3_moe.py @@ -0,0 +1,240 @@ +# coding=utf-8 +# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""Qwen3MoE model configuration""" + +from ...configuration_utils import PretrainedConfig +from ...modeling_rope_utils import rope_config_validation +from ...utils import logging + + +logger = logging.get_logger(__name__) + + +class Qwen3MoeConfig(PretrainedConfig): + r""" + This is the configuration class to store the configuration of a [`Qwen3MoeModel`]. It is used to instantiate a + Qwen3MoE model according to the specified arguments, defining the model architecture. Instantiating a configuration + with the defaults will yield a similar configuration to that of [Qwen/Qwen3-15B-A2B](https://huggingface.co/Qwen/Qwen3-15B-A2B). + + Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the + documentation from [`PretrainedConfig`] for more information. + + + Args: + vocab_size (`int`, *optional*, defaults to 151936): + Vocabulary size of the Qwen3MoE model. Defines the number of different tokens that can be represented by the + `inputs_ids` passed when calling [`Qwen3MoeModel`] + hidden_size (`int`, *optional*, defaults to 2048): + Dimension of the hidden representations. + intermediate_size (`int`, *optional*, defaults to 6144): + Dimension of the MLP representations. + num_hidden_layers (`int`, *optional*, defaults to 24): + Number of hidden layers in the Transformer encoder. + num_attention_heads (`int`, *optional*, defaults to 32): + Number of attention heads for each attention layer in the Transformer encoder. + num_key_value_heads (`int`, *optional*, defaults to 4): + This is the number of key_value heads that should be used to implement Grouped Query Attention. If + `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if + `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When + converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed + by meanpooling all the original heads within that group. For more details, check out [this + paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to `32`. + + hidden_act (`str` or `function`, *optional*, defaults to `"silu"`): + The non-linear activation function (function or string) in the decoder. + max_position_embeddings (`int`, *optional*, defaults to 32768): + The maximum sequence length that this model might ever be used with. + initializer_range (`float`, *optional*, defaults to 0.02): + The standard deviation of the truncated_normal_initializer for initializing all weight matrices. + rms_norm_eps (`float`, *optional*, defaults to 1e-06): + The epsilon used by the rms normalization layers. + use_cache (`bool`, *optional*, defaults to `True`): + Whether or not the model should return the last key/values attentions (not used by all models). Only + relevant if `config.is_decoder=True`. + tie_word_embeddings (`bool`, *optional*, defaults to `False`): + Whether the model's input and output word embeddings should be tied. + rope_theta (`float`, *optional*, defaults to 10000.0): + The base period of the RoPE embeddings. + rope_scaling (`Dict`, *optional*): + Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type + and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value + accordingly. + Expected contents: + `rope_type` (`str`): + The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope', + 'llama3'], with 'default' being the original RoPE implementation. + `factor` (`float`, *optional*): + Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In + most scaling types, a `factor` of x will enable the model to handle sequences of length x * + original maximum pre-trained length. + `original_max_position_embeddings` (`int`, *optional*): + Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during + pretraining. + `attention_factor` (`float`, *optional*): + Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention + computation. If unspecified, it defaults to value recommended by the implementation, using the + `factor` field to infer the suggested value. + `beta_fast` (`float`, *optional*): + Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear + ramp function. If unspecified, it defaults to 32. + `beta_slow` (`float`, *optional*): + Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear + ramp function. If unspecified, it defaults to 1. + `short_factor` (`list[float]`, *optional*): + Only used with 'longrope'. The scaling factor to be applied to short contexts (< + `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden + size divided by the number of attention heads divided by 2 + `long_factor` (`list[float]`, *optional*): + Only used with 'longrope'. The scaling factor to be applied to long contexts (< + `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden + size divided by the number of attention heads divided by 2 + `low_freq_factor` (`float`, *optional*): + Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE + `high_freq_factor` (`float`, *optional*): + Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE + attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`): + Whether to use a bias in the query, key, value and output projection layers during self-attention. + use_sliding_window (`bool`, *optional*, defaults to `False`): + Whether to use sliding window attention. + sliding_window (`int`, *optional*, defaults to 4096): + Sliding window attention (SWA) window size. If not specified, will default to `4096`. + attention_dropout (`float`, *optional*, defaults to 0.0): + The dropout ratio for the attention probabilities. + decoder_sparse_step (`int`, *optional*, defaults to 1): + The frequency of the MoE layer. + moe_intermediate_size (`int`, *optional*, defaults to 768): + Intermediate size of the routed expert. + num_experts_per_tok (`int`, *optional*, defaults to 8): + Number of selected experts. + num_experts (`int`, *optional*, defaults to 128): + Number of routed experts. + norm_topk_prob (`bool`, *optional*, defaults to `False`): + Whether to normalize the topk probabilities. + output_router_logits (`bool`, *optional*, defaults to `False`): + Whether or not the router logits should be returned by the model. Enabling this will also + allow the model to output the auxiliary loss, including load balancing loss and router z-loss. + router_aux_loss_coef (`float`, *optional*, defaults to 0.001): + The aux loss factor for the total loss. + mlp_only_layers (`list[int]`, *optional*, defaults to `[]`): + Indicate which layers use Qwen3MoeMLP rather than Qwen3MoeSparseMoeBlock + The list contains layer index, from 0 to num_layers-1 if we have num_layers layers + If `mlp_only_layers` is empty, `decoder_sparse_step` is used to determine the sparsity. + + ```python + >>> from transformers import Qwen3MoeModel, Qwen3MoeConfig + + >>> # Initializing a Qwen3MoE style configuration + >>> configuration = Qwen3MoeConfig() + + >>> # Initializing a model from the Qwen3-15B-A2B" style configuration + >>> model = Qwen3MoeModel(configuration) + + >>> # Accessing the model configuration + >>> configuration = model.config + ```""" + + model_type = "qwen3_moe" + keys_to_ignore_at_inference = ["past_key_values"] + + # Default tensor parallel plan for base model `Qwen3Moe` + base_model_tp_plan = { + "layers.*.self_attn.q_proj": "colwise", + "layers.*.self_attn.k_proj": "colwise", + "layers.*.self_attn.v_proj": "colwise", + "layers.*.self_attn.o_proj": "rowwise", + "layers.*.mlp.experts.*.gate_proj": "colwise", + "layers.*.mlp.experts.*.up_proj": "colwise", + "layers.*.mlp.experts.*.down_proj": "rowwise", + "layers.*.mlp.gate_proj": "colwise", + "layers.*.mlp.up_proj": "colwise", + "layers.*.mlp.down_proj": "rowwise", + } + base_model_pp_plan = { + "embed_tokens": (["input_ids"], ["inputs_embeds"]), + "layers": (["hidden_states", "attention_mask"], ["hidden_states"]), + "norm": (["hidden_states"], ["hidden_states"]), + } + + def __init__( + self, + vocab_size=151936, + hidden_size=2048, + intermediate_size=6144, + num_hidden_layers=24, + num_attention_heads=32, + num_key_value_heads=4, + hidden_act="silu", + max_position_embeddings=32768, + initializer_range=0.02, + rms_norm_eps=1e-6, + use_cache=True, + tie_word_embeddings=False, + rope_theta=10000.0, + rope_scaling=None, + attention_bias=False, + use_sliding_window=False, + sliding_window=4096, + attention_dropout=0.0, + decoder_sparse_step=1, + moe_intermediate_size=768, + num_experts_per_tok=8, + num_experts=128, + norm_topk_prob=False, + output_router_logits=False, + router_aux_loss_coef=0.001, + mlp_only_layers=None, + **kwargs, + ): + self.vocab_size = vocab_size + self.max_position_embeddings = max_position_embeddings + self.hidden_size = hidden_size + self.intermediate_size = intermediate_size + self.num_hidden_layers = num_hidden_layers + self.num_attention_heads = num_attention_heads + self.use_sliding_window = use_sliding_window + self.sliding_window = sliding_window if use_sliding_window else None + + self.num_key_value_heads = num_key_value_heads + self.hidden_act = hidden_act + self.initializer_range = initializer_range + self.rms_norm_eps = rms_norm_eps + self.use_cache = use_cache + self.rope_theta = rope_theta + self.rope_scaling = rope_scaling + self.attention_bias = attention_bias + self.attention_dropout = attention_dropout + # Validate the correctness of rotary position embeddings parameters + # BC: if there is a 'type' field, move it to 'rope_type'. + if self.rope_scaling is not None and "type" in self.rope_scaling: + self.rope_scaling["rope_type"] = self.rope_scaling["type"] + rope_config_validation(self) + + # MoE arguments + self.decoder_sparse_step = decoder_sparse_step + self.moe_intermediate_size = moe_intermediate_size + self.num_experts_per_tok = num_experts_per_tok + self.num_experts = num_experts + self.norm_topk_prob = norm_topk_prob + self.output_router_logits = output_router_logits + self.router_aux_loss_coef = router_aux_loss_coef + self.mlp_only_layers = [] if mlp_only_layers is None else mlp_only_layers + + super().__init__( + tie_word_embeddings=tie_word_embeddings, + **kwargs, + ) + + +__all__ = ["Qwen3MoeConfig"] diff --git a/infini-qwen3-moe/reference/qwen3_moe/modeling_qwen3_moe.py b/infini-qwen3-moe/reference/qwen3_moe/modeling_qwen3_moe.py new file mode 100755 index 00000000..67f21d1b --- /dev/null +++ b/infini-qwen3-moe/reference/qwen3_moe/modeling_qwen3_moe.py @@ -0,0 +1,1052 @@ +# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 +# This file was automatically generated from src/transformers/models/qwen3_moe/modular_qwen3_moe.py. +# Do NOT edit this file manually as any edits will be overwritten by the generation of +# the file from the modular. If any change should be done, please apply the change to the +# modular_qwen3_moe.py file directly. One of our CI enforces this. +# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 +# coding=utf-8 +# Copyright 2025 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from functools import partial +from typing import Callable, Optional, Union + +import torch +import torch.nn.functional as F +from torch import nn + +from ...activations import ACT2FN +from ...cache_utils import Cache, DynamicCache +from ...generation import GenerationMixin +from ...integrations import use_kernel_forward_from_hub +from ...masking_utils import create_causal_mask, create_sliding_window_causal_mask +from ...modeling_flash_attention_utils import FlashAttentionKwargs +from ...modeling_outputs import ( + BaseModelOutputWithPast, + MoeCausalLMOutputWithPast, + MoeModelOutputWithPast, + QuestionAnsweringModelOutput, + SequenceClassifierOutputWithPast, + TokenClassifierOutput, +) +from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update +from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel +from ...processing_utils import Unpack +from ...utils import LossKwargs, auto_docstring, can_return_tuple, logging +from .configuration_qwen3_moe import Qwen3MoeConfig + + +logger = logging.get_logger(__name__) + + +def rotate_half(x): + """Rotates half the hidden dims of the input.""" + x1 = x[..., : x.shape[-1] // 2] + x2 = x[..., x.shape[-1] // 2 :] + return torch.cat((-x2, x1), dim=-1) + + +def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1): + """Applies Rotary Position Embedding to the query and key tensors. + + Args: + q (`torch.Tensor`): The query tensor. + k (`torch.Tensor`): The key tensor. + cos (`torch.Tensor`): The cosine part of the rotary embedding. + sin (`torch.Tensor`): The sine part of the rotary embedding. + position_ids (`torch.Tensor`, *optional*): + Deprecated and unused. + unsqueeze_dim (`int`, *optional*, defaults to 1): + The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and + sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note + that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and + k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes + cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have + the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. + Returns: + `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. + """ + cos = cos.unsqueeze(unsqueeze_dim) + sin = sin.unsqueeze(unsqueeze_dim) + q_embed = (q * cos) + (rotate_half(q) * sin) + k_embed = (k * cos) + (rotate_half(k) * sin) + return q_embed, k_embed + + +def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor: + """ + This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, + num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim) + """ + batch, num_key_value_heads, slen, head_dim = hidden_states.shape + if n_rep == 1: + return hidden_states + hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim) + return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim) + + +def eager_attention_forward( + module: nn.Module, + query: torch.Tensor, + key: torch.Tensor, + value: torch.Tensor, + attention_mask: Optional[torch.Tensor], + scaling: float, + dropout: float = 0.0, + **kwargs, +): + key_states = repeat_kv(key, module.num_key_value_groups) + value_states = repeat_kv(value, module.num_key_value_groups) + + attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling + if attention_mask is not None: + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] + attn_weights = attn_weights + causal_mask + + attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype) + attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training) + attn_output = torch.matmul(attn_weights, value_states) + attn_output = attn_output.transpose(1, 2).contiguous() + + return attn_output, attn_weights + + +class Qwen3MoeAttention(nn.Module): + """Multi-headed attention from 'Attention Is All You Need' paper""" + + def __init__(self, config: Qwen3MoeConfig, layer_idx: int): + super().__init__() + self.config = config + self.layer_idx = layer_idx + self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads) + self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads + self.scaling = self.head_dim**-0.5 + self.attention_dropout = config.attention_dropout + self.is_causal = True + + self.q_proj = nn.Linear( + config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias + ) + self.k_proj = nn.Linear( + config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias + ) + self.v_proj = nn.Linear( + config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias + ) + self.o_proj = nn.Linear( + config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias + ) + self.q_norm = Qwen3MoeRMSNorm(self.head_dim, eps=config.rms_norm_eps) # unlike olmo, only on the head dim! + self.k_norm = Qwen3MoeRMSNorm(self.head_dim, eps=config.rms_norm_eps) # thus post q_norm does not need reshape + self.sliding_window = getattr(config, "sliding_window", None) + + def forward( + self, + hidden_states: torch.Tensor, + position_embeddings: tuple[torch.Tensor, torch.Tensor], + attention_mask: Optional[torch.Tensor], + past_key_value: Optional[Cache] = None, + cache_position: Optional[torch.LongTensor] = None, + **kwargs: Unpack[FlashAttentionKwargs], + ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]: + input_shape = hidden_states.shape[:-1] + hidden_shape = (*input_shape, -1, self.head_dim) + + query_states = self.q_norm(self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2) + key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2) + value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2) + + cos, sin = position_embeddings + query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) + + if past_key_value is not None: + # sin and cos are specific to RoPE models; cache_position needed for the static cache + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} + key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) + + attention_interface: Callable = eager_attention_forward + if self.config._attn_implementation != "eager": + attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation] + + attn_output, attn_weights = attention_interface( + self, + query_states, + key_states, + value_states, + attention_mask, + dropout=0.0 if not self.training else self.attention_dropout, + scaling=self.scaling, + sliding_window=self.sliding_window, # diff with Llama + **kwargs, + ) + + attn_output = attn_output.reshape(*input_shape, -1).contiguous() + attn_output = self.o_proj(attn_output) + return attn_output, attn_weights + + +class Qwen3MoeMLP(nn.Module): + def __init__(self, config, intermediate_size=None): + super().__init__() + self.config = config + self.hidden_size = config.hidden_size + self.intermediate_size = intermediate_size if intermediate_size is not None else config.intermediate_size + self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False) + self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False) + self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False) + self.act_fn = ACT2FN[config.hidden_act] + + def forward(self, x): + down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) + return down_proj + + +class Qwen3MoeSparseMoeBlock(nn.Module): + def __init__(self, config): + super().__init__() + self.num_experts = config.num_experts + self.top_k = config.num_experts_per_tok + self.norm_topk_prob = config.norm_topk_prob + + # gating + self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False) + self.experts = nn.ModuleList( + [Qwen3MoeMLP(config, intermediate_size=config.moe_intermediate_size) for _ in range(self.num_experts)] + ) + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + """ """ + batch_size, sequence_length, hidden_dim = hidden_states.shape + hidden_states = hidden_states.view(-1, hidden_dim) + # router_logits: (batch * sequence_length, n_experts) + router_logits = self.gate(hidden_states) + + routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float) + routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1) + if self.norm_topk_prob: # only diff with mixtral sparse moe block! + routing_weights /= routing_weights.sum(dim=-1, keepdim=True) + # we cast back to the input dtype + routing_weights = routing_weights.to(hidden_states.dtype) + + final_hidden_states = torch.zeros( + (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device + ) + + # One hot encode the selected experts to create an expert mask + # this will be used to easily index which expert is going to be sollicitated + expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0) + + # Loop over all available experts in the model and perform the computation on each expert + for expert_idx in range(self.num_experts): + expert_layer = self.experts[expert_idx] + idx, top_x = torch.where(expert_mask[expert_idx]) + + # Index the correct hidden states and compute the expert hidden state for + # the current expert. We need to make sure to multiply the output hidden + # states by `routing_weights` on the corresponding tokens (top-1 and top-2) + current_state = hidden_states[None, top_x].reshape(-1, hidden_dim) + current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None] + + # However `index_add_` only support torch tensors for indexing so we'll use + # the `top_x` tensor here. + final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype)) + final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim) + return final_hidden_states, router_logits + + +@use_kernel_forward_from_hub("RMSNorm") +class Qwen3MoeRMSNorm(nn.Module): + def __init__(self, hidden_size, eps=1e-6): + """ + Qwen3MoeRMSNorm is equivalent to T5LayerNorm + """ + super().__init__() + self.weight = nn.Parameter(torch.ones(hidden_size)) + self.variance_epsilon = eps + + def forward(self, hidden_states): + input_dtype = hidden_states.dtype + hidden_states = hidden_states.to(torch.float32) + variance = hidden_states.pow(2).mean(-1, keepdim=True) + hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) + return self.weight * hidden_states.to(input_dtype) + + def extra_repr(self): + return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}" + + +class Qwen3MoeDecoderLayer(nn.Module): + def __init__(self, config: Qwen3MoeConfig, layer_idx: int): + super().__init__() + self.hidden_size = config.hidden_size + + self.self_attn = Qwen3MoeAttention(config, layer_idx) + + if (layer_idx not in config.mlp_only_layers) and ( + config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0 + ): + self.mlp = Qwen3MoeSparseMoeBlock(config) + else: + self.mlp = Qwen3MoeMLP(config, intermediate_size=config.intermediate_size) + + self.input_layernorm = Qwen3MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_attention_layernorm = Qwen3MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_value: Optional[tuple[torch.Tensor]] = None, + output_attentions: Optional[bool] = False, + output_router_logits: Optional[bool] = False, + use_cache: Optional[bool] = False, + cache_position: Optional[torch.LongTensor] = None, + position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC + **kwargs: Unpack[FlashAttentionKwargs], + ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]: + """ + Args: + hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` + attention_mask (`torch.FloatTensor`, *optional*): attention mask of size + `(batch, sequence_length)` where padding elements are indicated by 0. + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under + returned tensors for more detail. + output_router_logits (`bool`, *optional*): + Whether or not to return the logits of all the routers. They are useful for computing the router loss, + and should not be returned during inference. + use_cache (`bool`, *optional*): + If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding + (see `past_key_values`). + past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. + position_embeddings (`tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*): + Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`, + with `head_dim` being the embedding dimension of each attention head. + kwargs (`dict`, *optional*): + Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code + into the model + """ + + residual = hidden_states + + hidden_states = self.input_layernorm(hidden_states) + + # Self Attention + hidden_states, self_attn_weights = self.self_attn( + hidden_states=hidden_states, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_value=past_key_value, + output_attentions=output_attentions, + use_cache=use_cache, + cache_position=cache_position, + position_embeddings=position_embeddings, + **kwargs, + ) + hidden_states = residual + hidden_states + + # Fully Connected + residual = hidden_states + hidden_states = self.post_attention_layernorm(hidden_states) + + hidden_states = self.mlp(hidden_states) + if isinstance(hidden_states, tuple): + hidden_states, router_logits = hidden_states + else: + router_logits = None + + hidden_states = residual + hidden_states + + outputs = (hidden_states,) + + if output_attentions: + outputs += (self_attn_weights,) + + if output_router_logits: + outputs += (router_logits,) + + return outputs + + +class Qwen3MoeRotaryEmbedding(nn.Module): + def __init__(self, config: Qwen3MoeConfig, device=None): + super().__init__() + # BC: "rope_type" was originally "type" + if hasattr(config, "rope_scaling") and config.rope_scaling is not None: + self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type")) + else: + self.rope_type = "default" + self.max_seq_len_cached = config.max_position_embeddings + self.original_max_seq_len = config.max_position_embeddings + + self.config = config + self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type] + + inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self.original_inv_freq = self.inv_freq + + @torch.no_grad() + @dynamic_rope_update # power user: used with advanced RoPE types (e.g. dynamic rope) + def forward(self, x, position_ids): + inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device) + position_ids_expanded = position_ids[:, None, :].float() + + device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu" + with torch.autocast(device_type=device_type, enabled=False): # Force float32 + freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) + emb = torch.cat((freqs, freqs), dim=-1) + cos = emb.cos() * self.attention_scaling + sin = emb.sin() * self.attention_scaling + + return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) + + +@auto_docstring +class Qwen3MoePreTrainedModel(PreTrainedModel): + config_class = Qwen3MoeConfig + base_model_prefix = "model" + supports_gradient_checkpointing = True + _no_split_modules = ["Qwen3MoeDecoderLayer"] + _skip_keys_device_placement = ["past_key_values"] + _supports_flash_attn_2 = True + _supports_sdpa = True + _supports_flex_attn = True + _supports_cache_class = True + _supports_quantized_cache = True + _supports_static_cache = False # MoE models don't work with torch.compile (`torch.where(condition)` not supported) + _supports_attention_backend = True + + def _init_weights(self, module): + std = self.config.initializer_range + if isinstance(module, nn.Linear): + module.weight.data.normal_(mean=0.0, std=std) + if module.bias is not None: + module.bias.data.zero_() + elif isinstance(module, nn.Embedding): + module.weight.data.normal_(mean=0.0, std=std) + if module.padding_idx is not None: + module.weight.data[module.padding_idx].zero_() + elif isinstance(module, Qwen3MoeRMSNorm): + module.weight.data.fill_(1.0) + + +@auto_docstring +class Qwen3MoeModel(Qwen3MoePreTrainedModel): + def __init__(self, config: Qwen3MoeConfig): + super().__init__(config) + self.padding_idx = config.pad_token_id + self.vocab_size = config.vocab_size + + self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx) + self.layers = nn.ModuleList( + [Qwen3MoeDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] + ) + self.norm = Qwen3MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.rotary_emb = Qwen3MoeRotaryEmbedding(config=config) + self.gradient_checkpointing = False + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.embed_tokens + + def set_input_embeddings(self, value): + self.embed_tokens = value + + @can_return_tuple + @auto_docstring + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[list[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + output_router_logits: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, + **flash_attn_kwargs: Unpack[FlashAttentionKwargs], + ) -> MoeModelOutputWithPast: + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_router_logits = ( + output_router_logits if output_router_logits is not None else self.config.output_router_logits + ) + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + use_cache = use_cache if use_cache is not None else self.config.use_cache + + if (input_ids is None) ^ (inputs_embeds is not None): + raise ValueError("You must specify exactly one of input_ids or inputs_embeds") + + if self.gradient_checkpointing and self.training: + if use_cache: + logger.warning_once( + "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." + ) + use_cache = False + + if use_cache and past_key_values is None: + past_key_values = DynamicCache() + + if inputs_embeds is None: + inputs_embeds = self.embed_tokens(input_ids) + + if cache_position is None: + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + cache_position = torch.arange( + past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device + ) + if position_ids is None: + position_ids = cache_position.unsqueeze(0) + + mask_function = create_causal_mask if self.config.sliding_window is None else create_sliding_window_causal_mask + causal_mask = mask_function( + config=self.config, + input_embeds=inputs_embeds, + attention_mask=attention_mask, + cache_position=cache_position, + past_key_values=past_key_values, + ) + + hidden_states = inputs_embeds + + # create position embeddings to be shared across the decoder layers + position_embeddings = self.rotary_emb(hidden_states, position_ids) + + # decoder layers + all_hidden_states = () if output_hidden_states else None + all_self_attns = () if output_attentions else None + all_router_logits = () if output_router_logits else None + + for decoder_layer in self.layers: + if output_hidden_states: + all_hidden_states += (hidden_states,) + + if self.gradient_checkpointing and self.training: + layer_outputs = self._gradient_checkpointing_func( + partial(decoder_layer.__call__, **flash_attn_kwargs), + hidden_states, + causal_mask, + position_ids, + past_key_values, + output_attentions, + output_router_logits, + use_cache, + cache_position, + position_embeddings, + ) + else: + layer_outputs = decoder_layer( + hidden_states, + attention_mask=causal_mask, + position_ids=position_ids, + past_key_value=past_key_values, + output_attentions=output_attentions, + output_router_logits=output_router_logits, + use_cache=use_cache, + cache_position=cache_position, + position_embeddings=position_embeddings, + **flash_attn_kwargs, + ) + + hidden_states = layer_outputs[0] + + if output_attentions: + all_self_attns += (layer_outputs[1],) + + if output_router_logits: + all_router_logits += (layer_outputs[-1],) + + hidden_states = self.norm(hidden_states) + + # add hidden states from the last decoder layer + if output_hidden_states: + all_hidden_states += (hidden_states,) + + return MoeModelOutputWithPast( + last_hidden_state=hidden_states, + past_key_values=past_key_values, + hidden_states=all_hidden_states, + attentions=all_self_attns, + router_logits=all_router_logits, + ) + + +class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ... + + +def load_balancing_loss_func( + gate_logits: Union[torch.Tensor, tuple[torch.Tensor], None], + num_experts: Optional[int] = None, + top_k=2, + attention_mask: Optional[torch.Tensor] = None, +) -> Union[torch.Tensor, int]: + r""" + Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch. + + See Switch Transformer (https://huggingface.co/papers/2101.03961) for more details. This function implements the loss + function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between + experts is too unbalanced. + + Args: + gate_logits: + Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of + shape [batch_size X sequence_length, num_experts]. + num_experts: + Number of experts + top_k: + The number of experts to route per-token, can be also interpreted as the `top-k` routing + parameter. + attention_mask (`torch.Tensor`, *optional*): + The attention_mask used in forward function + shape [batch_size X sequence_length] if not None. + + Returns: + The auxiliary loss. + """ + if gate_logits is None or not isinstance(gate_logits, tuple): + return 0 + + if isinstance(gate_logits, tuple): + compute_device = gate_logits[0].device + concatenated_gate_logits = torch.cat([layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0) + + routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1) + + _, selected_experts = torch.topk(routing_weights, top_k, dim=-1) + + expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts) + + if attention_mask is None: + # Compute the percentage of tokens routed to each experts + tokens_per_expert = torch.mean(expert_mask.float(), dim=0) + + # Compute the average probability of routing to these experts + router_prob_per_expert = torch.mean(routing_weights, dim=0) + else: + batch_size, sequence_length = attention_mask.shape + num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length) + + # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask + expert_attention_mask = ( + attention_mask[None, :, :, None, None] + .expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts)) + .reshape(-1, top_k, num_experts) + .to(compute_device) + ) + + # Compute the percentage of tokens routed to each experts + tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum( + expert_attention_mask, dim=0 + ) + + # Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert + router_per_expert_attention_mask = ( + attention_mask[None, :, :, None] + .expand((num_hidden_layers, batch_size, sequence_length, num_experts)) + .reshape(-1, num_experts) + .to(compute_device) + ) + + # Compute the average probability of routing to these experts + router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum( + router_per_expert_attention_mask, dim=0 + ) + + overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0)) + return overall_loss * num_experts + + +@auto_docstring +class Qwen3MoeForCausalLM(Qwen3MoePreTrainedModel, GenerationMixin): + _tied_weights_keys = ["lm_head.weight"] + _tp_plan = {"lm_head": "colwise_rep"} + _pp_plan = {"lm_head": (["hidden_states"], ["logits"])} + + def __init__(self, config): + super().__init__(config) + self.model = Qwen3MoeModel(config) + self.vocab_size = config.vocab_size + self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) + self.router_aux_loss_coef = config.router_aux_loss_coef + self.num_experts = config.num_experts + self.num_experts_per_tok = config.num_experts_per_tok + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.model.embed_tokens + + def set_input_embeddings(self, value): + self.model.embed_tokens = value + + def get_output_embeddings(self): + return self.lm_head + + def set_output_embeddings(self, new_embeddings): + self.lm_head = new_embeddings + + def set_decoder(self, decoder): + self.model = decoder + + def get_decoder(self): + return self.model + + @can_return_tuple + @auto_docstring + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[list[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + output_router_logits: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, + logits_to_keep: Union[int, torch.Tensor] = 0, + **kwargs: Unpack[KwargsForCausalLM], + ) -> MoeCausalLMOutputWithPast: + r""" + labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): + Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., + config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored + (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. + + Example: + + ```python + >>> from transformers import AutoTokenizer, Qwen3MoeForCausalLM + + >>> model = Qwen3MoeForCausalLM.from_pretrained("Qwen/Qwen3-MoE-15B-A2B") + >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-MoE-15B-A2B") + + >>> prompt = "Hey, are you conscious? Can you talk to me?" + >>> inputs = tokenizer(prompt, return_tensors="pt") + + >>> # Generate + >>> generate_ids = model.generate(inputs.input_ids, max_length=30) + >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] + "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you." + ```""" + + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_router_logits = ( + output_router_logits if output_router_logits is not None else self.config.output_router_logits + ) + + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + + # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) + outputs: MoeModelOutputWithPast = self.model( + input_ids=input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + output_router_logits=output_router_logits, + cache_position=cache_position, + **kwargs, + ) + + hidden_states = outputs.last_hidden_state + # Only compute necessary logits, and do not upcast them to float if we are not computing the loss + slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep + logits = self.lm_head(hidden_states[:, slice_indices, :]) + + loss = None + if labels is not None: + loss = self.loss_function(logits, labels, self.vocab_size, **kwargs) + + aux_loss = None + if output_router_logits: + aux_loss = load_balancing_loss_func( + outputs.router_logits, + self.num_experts, + self.num_experts_per_tok, + attention_mask, + ) + if labels is not None: + loss += self.router_aux_loss_coef * aux_loss.to(loss.device) # make sure to reside in the same device + + return MoeCausalLMOutputWithPast( + loss=loss, + aux_loss=aux_loss, + logits=logits, + past_key_values=outputs.past_key_values, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + router_logits=outputs.router_logits, + ) + + +@auto_docstring( + custom_intro=""" + The Qwen3Moe Model transformer with a sequence classification head on top (linear layer). + + [`Qwen3MoeForSequenceClassification`] uses the last token in order to do the classification, as other causal models + (e.g. GPT-2) do. + + Since it does classification on the last token, it requires to know the position of the last token. If a + `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If + no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the + padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in + each row of the batch). + """ +) +class Qwen3MoeForSequenceClassification(Qwen3MoePreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.num_labels = config.num_labels + self.model = Qwen3MoeModel(config) + self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False) + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.model.embed_tokens + + def set_input_embeddings(self, value): + self.model.embed_tokens = value + + @can_return_tuple + @auto_docstring + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Cache] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + ) -> SequenceClassifierOutputWithPast: + r""" + labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., + config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If + `config.num_labels > 1` a classification loss is computed (Cross-Entropy). + """ + + transformer_outputs: BaseModelOutputWithPast = self.model( + input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + ) + hidden_states = transformer_outputs.last_hidden_state + logits = self.score(hidden_states) + + if input_ids is not None: + batch_size = input_ids.shape[0] + else: + batch_size = inputs_embeds.shape[0] + + if self.config.pad_token_id is None and batch_size != 1: + raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.") + if self.config.pad_token_id is None: + last_non_pad_token = -1 + elif input_ids is not None: + # To handle both left- and right- padding, we take the rightmost token that is not equal to pad_token_id + non_pad_mask = (input_ids != self.config.pad_token_id).to(logits.device, torch.int32) + token_indices = torch.arange(input_ids.shape[-1], device=logits.device, dtype=torch.int32) + last_non_pad_token = (token_indices * non_pad_mask).argmax(-1) + else: + last_non_pad_token = -1 + logger.warning_once( + f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be " + "unexpected if using padding tokens in conjunction with `inputs_embeds.`" + ) + + pooled_logits = logits[torch.arange(batch_size, device=logits.device), last_non_pad_token] + + loss = None + if labels is not None: + loss = self.loss_function(logits=logits, labels=labels, pooled_logits=pooled_logits, config=self.config) + + return SequenceClassifierOutputWithPast( + loss=loss, + logits=pooled_logits, + past_key_values=transformer_outputs.past_key_values, + hidden_states=transformer_outputs.hidden_states, + attentions=transformer_outputs.attentions, + ) + + +@auto_docstring +class Qwen3MoeForTokenClassification(Qwen3MoePreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.num_labels = config.num_labels + self.model = Qwen3MoeModel(config) + if getattr(config, "classifier_dropout", None) is not None: + classifier_dropout = config.classifier_dropout + elif getattr(config, "hidden_dropout", None) is not None: + classifier_dropout = config.hidden_dropout + else: + classifier_dropout = 0.1 + self.dropout = nn.Dropout(classifier_dropout) + self.score = nn.Linear(config.hidden_size, config.num_labels) + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.model.embed_tokens + + def set_input_embeddings(self, value): + self.model.embed_tokens = value + + @can_return_tuple + @auto_docstring + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Cache] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + ) -> TokenClassifierOutput: + r""" + labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., + config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If + `config.num_labels > 1` a classification loss is computed (Cross-Entropy). + """ + + outputs: BaseModelOutputWithPast = self.model( + input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + ) + sequence_output = outputs.last_hidden_state + sequence_output = self.dropout(sequence_output) + logits = self.score(sequence_output) + + loss = None + if labels is not None: + loss = self.loss_function(logits, labels, self.config) + + return TokenClassifierOutput( + loss=loss, + logits=logits, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) + + +@auto_docstring +class Qwen3MoeForQuestionAnswering(Qwen3MoePreTrainedModel): + base_model_prefix = "transformer" + + def __init__(self, config): + super().__init__(config) + self.transformer = Qwen3MoeModel(config) + self.qa_outputs = nn.Linear(config.hidden_size, 2) + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.transformer.embed_tokens + + def set_input_embeddings(self, value): + self.transformer.embed_tokens = value + + @can_return_tuple + @auto_docstring + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Cache] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + start_positions: Optional[torch.LongTensor] = None, + end_positions: Optional[torch.LongTensor] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + **kwargs, + ) -> QuestionAnsweringModelOutput: + outputs: BaseModelOutputWithPast = self.transformer( + input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + ) + + sequence_output = outputs.last_hidden_state + + logits = self.qa_outputs(sequence_output) + start_logits, end_logits = logits.split(1, dim=-1) + start_logits = start_logits.squeeze(-1).contiguous() + end_logits = end_logits.squeeze(-1).contiguous() + + loss = None + if start_positions is not None and end_positions is not None: + loss = self.loss_function(start_logits, end_logits, start_positions, end_positions, **kwargs) + + return QuestionAnsweringModelOutput( + loss=loss, + start_logits=start_logits, + end_logits=end_logits, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) + + +__all__ = [ + "Qwen3MoeForCausalLM", + "Qwen3MoeForQuestionAnswering", + "Qwen3MoeModel", + "Qwen3MoePreTrainedModel", + "Qwen3MoeForSequenceClassification", + "Qwen3MoeForTokenClassification", +] diff --git a/infini-qwen3-moe/reference/qwen3_moe/modular_qwen3_moe.py b/infini-qwen3-moe/reference/qwen3_moe/modular_qwen3_moe.py new file mode 100755 index 00000000..569b92cc --- /dev/null +++ b/infini-qwen3-moe/reference/qwen3_moe/modular_qwen3_moe.py @@ -0,0 +1,349 @@ +# coding=utf-8 +# Copyright 2025 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""PyTorch Qwen3 model.""" + +from typing import Optional, Union + +import torch +import torch.nn.functional as F +import torch.utils.checkpoint +from torch import nn + +from ...activations import ACT2FN +from ...modeling_flash_attention_utils import FlashAttentionKwargs +from ...modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast +from ...processing_utils import Unpack +from ...utils import LossKwargs, logging +from ..llama.modeling_llama import ( + LlamaForQuestionAnswering, + LlamaForSequenceClassification, + LlamaForTokenClassification, + LlamaRMSNorm, +) +from ..mixtral.modeling_mixtral import MixtralForCausalLM, MixtralModel, load_balancing_loss_func +from ..qwen2_moe.modeling_qwen2_moe import Qwen2MoeDecoderLayer +from ..qwen3.modeling_qwen3 import Qwen3Attention +from .configuration_qwen3_moe import Qwen3MoeConfig + + +logger = logging.get_logger(__name__) + + +class Qwen3MoeAttention(Qwen3Attention): # This is the main diff with qwen2Moe! + def __init__(self, config: Qwen3MoeConfig, layer_idx: int): + super().__init__(config, layer_idx) + self.sliding_window = getattr(config, "sliding_window", None) + + +class Qwen3MoeMLP(nn.Module): + def __init__(self, config, intermediate_size=None): + super().__init__() + self.config = config + self.hidden_size = config.hidden_size + self.intermediate_size = intermediate_size if intermediate_size is not None else config.intermediate_size + self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False) + self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False) + self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False) + self.act_fn = ACT2FN[config.hidden_act] + + def forward(self, x): + down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) + return down_proj + + +class Qwen3MoeSparseMoeBlock(nn.Module): + def __init__(self, config): + super().__init__() + self.num_experts = config.num_experts + self.top_k = config.num_experts_per_tok + self.norm_topk_prob = config.norm_topk_prob + + # gating + self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False) + self.experts = nn.ModuleList( + [Qwen3MoeMLP(config, intermediate_size=config.moe_intermediate_size) for _ in range(self.num_experts)] + ) + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + """ """ + batch_size, sequence_length, hidden_dim = hidden_states.shape + hidden_states = hidden_states.view(-1, hidden_dim) + # router_logits: (batch * sequence_length, n_experts) + router_logits = self.gate(hidden_states) + + routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float) + routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1) + if self.norm_topk_prob: # only diff with mixtral sparse moe block! + routing_weights /= routing_weights.sum(dim=-1, keepdim=True) + # we cast back to the input dtype + routing_weights = routing_weights.to(hidden_states.dtype) + + final_hidden_states = torch.zeros( + (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device + ) + + # One hot encode the selected experts to create an expert mask + # this will be used to easily index which expert is going to be sollicitated + expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0) + + # Loop over all available experts in the model and perform the computation on each expert + for expert_idx in range(self.num_experts): + expert_layer = self.experts[expert_idx] + idx, top_x = torch.where(expert_mask[expert_idx]) + + # Index the correct hidden states and compute the expert hidden state for + # the current expert. We need to make sure to multiply the output hidden + # states by `routing_weights` on the corresponding tokens (top-1 and top-2) + current_state = hidden_states[None, top_x].reshape(-1, hidden_dim) + current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None] + + # However `index_add_` only support torch tensors for indexing so we'll use + # the `top_x` tensor here. + final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype)) + final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim) + return final_hidden_states, router_logits + + +class Qwen3MoeRMSNorm(LlamaRMSNorm): + pass + + +class Qwen3MoeDecoderLayer(Qwen2MoeDecoderLayer, nn.Module): + def __init__(self, config: Qwen3MoeConfig, layer_idx: int): + nn.Module().__init__() + self.hidden_size = config.hidden_size + + self.self_attn = Qwen3MoeAttention(config, layer_idx) + + if (layer_idx not in config.mlp_only_layers) and ( + config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0 + ): + self.mlp = Qwen3MoeSparseMoeBlock(config) + else: + self.mlp = Qwen3MoeMLP(config, intermediate_size=config.intermediate_size) + + self.input_layernorm = Qwen3MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_attention_layernorm = Qwen3MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_value: Optional[tuple[torch.Tensor]] = None, + output_attentions: Optional[bool] = False, + output_router_logits: Optional[bool] = False, + use_cache: Optional[bool] = False, + cache_position: Optional[torch.LongTensor] = None, + position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC + **kwargs: Unpack[FlashAttentionKwargs], + ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]: + """ + Args: + hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` + attention_mask (`torch.FloatTensor`, *optional*): attention mask of size + `(batch, sequence_length)` where padding elements are indicated by 0. + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under + returned tensors for more detail. + output_router_logits (`bool`, *optional*): + Whether or not to return the logits of all the routers. They are useful for computing the router loss, + and should not be returned during inference. + use_cache (`bool`, *optional*): + If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding + (see `past_key_values`). + past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. + position_embeddings (`tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*): + Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`, + with `head_dim` being the embedding dimension of each attention head. + kwargs (`dict`, *optional*): + Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code + into the model + """ + + residual = hidden_states + + hidden_states = self.input_layernorm(hidden_states) + + # Self Attention + hidden_states, self_attn_weights = self.self_attn( + hidden_states=hidden_states, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_value=past_key_value, + output_attentions=output_attentions, + use_cache=use_cache, + cache_position=cache_position, + position_embeddings=position_embeddings, + **kwargs, + ) + hidden_states = residual + hidden_states + + # Fully Connected + residual = hidden_states + hidden_states = self.post_attention_layernorm(hidden_states) + + hidden_states = self.mlp(hidden_states) + if isinstance(hidden_states, tuple): + hidden_states, router_logits = hidden_states + else: + router_logits = None + + hidden_states = residual + hidden_states + + outputs = (hidden_states,) + + if output_attentions: + outputs += (self_attn_weights,) + + if output_router_logits: + outputs += (router_logits,) + + return outputs + + +class Qwen3MoeModel(MixtralModel): + def __init__(self, config: Qwen3MoeConfig): + super().__init__(config) + self.layers = nn.ModuleList( + [Qwen3MoeDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] + ) + + +class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ... + + +class Qwen3MoeForCausalLM(MixtralForCausalLM): + def __init__(self, config): + super().__init__(config) + self.model = Qwen3MoeModel(config) + self.num_experts = config.num_experts + + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[list[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + output_router_logits: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, + logits_to_keep: Union[int, torch.Tensor] = 0, + **kwargs: Unpack[KwargsForCausalLM], + ) -> MoeCausalLMOutputWithPast: + r""" + labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): + Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., + config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored + (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. + + Example: + + ```python + >>> from transformers import AutoTokenizer, Qwen3MoeForCausalLM + + >>> model = Qwen3MoeForCausalLM.from_pretrained("Qwen/Qwen3-MoE-15B-A2B") + >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-MoE-15B-A2B") + + >>> prompt = "Hey, are you conscious? Can you talk to me?" + >>> inputs = tokenizer(prompt, return_tensors="pt") + + >>> # Generate + >>> generate_ids = model.generate(inputs.input_ids, max_length=30) + >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] + "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you." + ```""" + + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_router_logits = ( + output_router_logits if output_router_logits is not None else self.config.output_router_logits + ) + + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + + # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) + outputs: MoeModelOutputWithPast = self.model( + input_ids=input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + output_router_logits=output_router_logits, + cache_position=cache_position, + **kwargs, + ) + + hidden_states = outputs.last_hidden_state + # Only compute necessary logits, and do not upcast them to float if we are not computing the loss + slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep + logits = self.lm_head(hidden_states[:, slice_indices, :]) + + loss = None + if labels is not None: + loss = self.loss_function(logits, labels, self.vocab_size, **kwargs) + + aux_loss = None + if output_router_logits: + aux_loss = load_balancing_loss_func( + outputs.router_logits, + self.num_experts, + self.num_experts_per_tok, + attention_mask, + ) + if labels is not None: + loss += self.router_aux_loss_coef * aux_loss.to(loss.device) # make sure to reside in the same device + + return MoeCausalLMOutputWithPast( + loss=loss, + aux_loss=aux_loss, + logits=logits, + past_key_values=outputs.past_key_values, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + router_logits=outputs.router_logits, + ) + + +class Qwen3MoeForSequenceClassification(LlamaForSequenceClassification): + pass + + +class Qwen3MoeForTokenClassification(LlamaForTokenClassification): + pass + + +class Qwen3MoeForQuestionAnswering(LlamaForQuestionAnswering): + pass + + +__all__ = [ + "Qwen3MoeForCausalLM", + "Qwen3MoeForQuestionAnswering", + "Qwen3MoeModel", + "Qwen3MoePreTrainedModel", # noqa: F822 + "Qwen3MoeForSequenceClassification", + "Qwen3MoeForTokenClassification", +] diff --git a/infini-qwen3-moe/scripts/__init__.py b/infini-qwen3-moe/scripts/__init__.py new file mode 100755 index 00000000..e69de29b diff --git a/infini-qwen3-moe/scripts/__pycache__/infer_task.cpython-310.pyc b/infini-qwen3-moe/scripts/__pycache__/infer_task.cpython-310.pyc new file mode 100755 index 0000000000000000000000000000000000000000..527fba345e2e63cccb8a38cdac6be80e86d4d6f6 GIT binary patch literal 2770 zcmZuz&u<$=6rPz~+iS;mel%^GmbRw+szFT(2!T*kDz$)6r3jR&he*|G@oW<(w%46m zm)6=`aKZs01S$uhLQ2G?IdDfR{{jEOUTFxuA#tmO;Cr*#_(!|ZetEO=X6DVC?|tK- zR4On$zq5O*4`Ka5jlsu4;}Tw70er|LlYA3zPb^bQNDD0sEiUaPR?XDlHS`FyF5%S| z06v2sTNaWpTgljETiP%GK+Y{on5G{qp_ZS6hGmytVhkgBmTGK4jA-F5}hb01@A2J1k-@ zmt2ZX7Ge`UeuecozF85WEg2TK7i%2;>61>hgb#WGu!2c$*c;f2&^~p%Gg{?sxfOm8&tdcextJzkVi_S<{wY=s7fq zl9A=z$)UBV_o;ObS#TCZXdN{U1};yBw9vuuM$_J-l@8jnO|ZHa|}_(g0r{FW+V(rPA2CE2mn4s>k!Eg9#U-W`{G z4#}FNxYTJQjlP>Of|8+&fFe%Tu_n09r$qO}P}YYAELI(2M$2pZt{WFz7mVsO$S%6> zt&Z0?=*hXR3~H{csHD_1!3@E31pSbVMMtkv%9v;zW^;!-8OJK-Oiv9iLl5Qo5?)O( zC_s5DFXW`mNoSKohVrstWN1tljSQ7!*~rk89G4UDaAXFWa3n69Et{IXhWhNB#I)v9 zUy}a(`oaD;Up#vFaPP|XDe z1qwi(aQMv4L>XW<%C=`EUm@Ab83H>X0dw<4!o*_z#>awH3$dunxI~hpS6zsJ4iZkp z1v>qMq8|(WX4_X@*in9Lhe3PYU^^~=C#h4uUjVZRw3;6vl><`M!#d`E1zn2NR#gdJ zC3qPS=SbAtx*RxKbEF*{!s9S>YS2zm`IwFYMUe)DM3Xo zF;mqLmDkW$6(;sLVIX!EK_xD9@+kS%>K}#J07}ZKht7#HlRth{oZ~t!LD{Bz7fV6V?gCG;0wz1Sk_{Atfu}YSl3* zPRf`RM-f}paRS4Jm&hVYq^FP%I$=Vx(UY)S{+;kGxSDG>E2hM>H3eZMDJojsBO|FH z?;&}m5+sQ=dASC|R1`!OJAq_M4pBW9n(W|5fI#b{*)euT0t+vT_88hyq+a*&t0xlo z*-v(le~!$Vp9`keyT`846yStc?LR$vesBBE*S+m5>LUN^CBsWd7m)X>*{{fraP;&qP literal 0 HcmV?d00001 diff --git a/infini-qwen3-moe/scripts/__pycache__/libinfinicore_infer.cpython-310.pyc b/infini-qwen3-moe/scripts/__pycache__/libinfinicore_infer.cpython-310.pyc new file mode 100755 index 0000000000000000000000000000000000000000..abba20102d24a00e614a276b1e6d54c579d78aba GIT binary patch literal 3917 zcma)9OK%&=5uO)^!}n9Nz0PA0Z{RS2EqFIDoFE7+QB=UJOloCG$u@!vM(h?f5;;Tn zG-XSWbBOjIfQ&w{~oxZ}SAP>t#Xjo{N5XhWUoLxUJYlbAz`Si=-a4Q*l%(QC)CP_{;bei6rVel%;N(Cj51b6m0w+7+6o8YXdEn$H zoFZ@vvAdH?8vBDpLh9|#R=d;o zHupZ;X?g9=!TNT4lT_xo8=vj97FpfyHYv?%wV#plJmbmJMb5_`EpndzbivI2qM7|A zGy6+s_Lt0TJ(KgR*?RJHk+WoG3(U;VYA(IJxiqT@v*sPveDcvVInUaF_kXGJ2M`(s z|7orrM?(e>2|uLgYwhqT2!m$C*&`qse>C)ia4qifU?k$TL2wu-g&rtCwu8ds(K*N5 zz;ua!4B%*%{%tNtSM$DkFZ>9MeY?E?d7Y8jO?DRa*fg^_CM(Be<(RA-la*t#a!l4v zBg?;wgYO}HA3&BiTL@j`iqU-cC*`=S6zSEY;_=e zXH{Rc_SS#9!0L8)Tckd#)!yDeSl?UUy$Mi4eTa+vApkBp#(g&LW$-0Rmwbtu&*6?w z0B(HA%p-#u%mOqOS_)GNZG~xtj>3$>EK7lYPGO$eKo=Ai6{p0~M@GK{)v1hSQkFk` z&BBjgFyS{}3O?=$nW4fPMEyX-(hdjyIpb2N(g=^G6`r0QOM~JaNQ?Hw8^E)uC(~>c z2ZJb-DRB&@Waco6JrSLja&lE?9IVZumgf46m}7CG-uod27XTy?zj%0D6|v(9kf1Q=x?=4=n2PAvjwz4 zSNb1y0jXqJ4DQPXy?4xeon;osF(WL&<0N50PQ3&XGsa+ZnV5dWJlb&X`+ughP!_2^C;aR|=aL7X8 zp7}iR4+qQ*BjLtOJa%^m%#Rt;ZXjF{xll;mn?v3W!dUo&fiHq6bZNj@PelCOjkwFl zq5JU7&_89@N)O$mfXCvo)JM{UR^VdwwtgUvE>^aKL+*naT*gr}4yo%4_u|v(a(dBl z6u>fZFdpwRdDkmx+?Pfa!!eQgK$?9fWVvm#GoQ z&fFUL0#2KgCi}>yB=~W&Jm6 zmGwMOSw-hf+pv|y)}615=_1JLpqp?}VuR-OIq3Y`DsbHDjqh;$8S9jQG4P;?-h0$_{V|`xQ~L`#NTrg zta5TCar;DYdgk@8ig7GSFbh1tD4{5+s0d6%hbjgSfWPQq=NHMJn_pt3j$Hsp`w-k! XHLG{3+3NkOTeTd^euVDBE35wxMH$XK literal 0 HcmV?d00001 diff --git a/infini-qwen3-moe/scripts/__pycache__/qwen3.cpython-310.pyc b/infini-qwen3-moe/scripts/__pycache__/qwen3.cpython-310.pyc new file mode 100755 index 0000000000000000000000000000000000000000..b0bc85f0650b6f90b67b3cfed66b4857811452d3 GIT binary patch literal 28544 zcmchA3v?XUdEU(I>=TQ{0$4l=l0#A=36TU}dRR2|p!l$0Nswexl(h`IS?&zL1sA)( zomo)W%*wJzM@nqT^doj+OEqBoq){B(andwyn^#XybDGnf_M9|nPA5HWjuJ~&lQc;X%-u>+Pzk5Yd{){5+AIiNS zU;7KHqL@lvF;!EmsB>Cgn~UTlD#areJ+Jd`G#};PSU$$T@qCXp=7 zSH6quMl0#L?tJ%LCZCz>$@k1<^VzxHeD7RezRz#FKi|)F)A_WOSk9TT6;)Ll2GZh{ zfw{r_;M`DtXl^(^JU5abnH$ZI&TYtVm>bKF&27wYoZFP&G`Bgw8EqxNfuPXU%Xl)QVDQnx8l~**%N34r~qszCLY3lDu= z4_HHp??Qag9722urFSDXY>pr{g4iB&)ZB1c$xm2&&9Mub_&sall7bOMPV#qrWATfa zoYaZ()0SyGe$lEPG$!ZfE7qJ#nJ_1wp^dFct{5(j0e$A1(nwVaM3EyoOkSSsHuY!Elk;l zZ4q>iWjy}SX(}^ol?Z!w>h=la(W(h>ENJj zsxVhGt%}!gTDBueHZzQS6lmj$Ms)+g3@xQco0y@?luW+B7)nIzCp73~XNn&Uo%C6wv-L=d|l zU;9%CEG4gk0BWY1kC>Vn!3(LIx*5f9)Qp*N{Km|LnZ$40OqpHyO_*u38^1|O7V;@c z5b|A;6y(#A1mwHrtDn!vmp@* zaPQ6ZPX$YasNh4-w4!+q(|9mdSN*bLYOi1;c{MWicDF=FrMQGGVWKhQ_N!utT?!g~SY!iR1Q#D(_(#NvAaV)4T_24eAj08x7{ zK-3<&F%Y$z1j3k@uh|Z*2{1d@Sqtl~ROUq+>c)u2wA3MCm||M z)SM`}y%PZ_xw5F#;-*HwO^9S1?=f|bYpz=BcQx=0(Z;@`=Phu|cO~g3V8^i0YcCIirp7Rog0&rjz3SO!}{BpfQ&*?&8 zpD9(v-9{60K+X&PX; z1kvbx%@hw%X2>yRE zwpLcMwqRs?%`q&#%w}qjf65EZ)Yo@ZB_%DYjm>Bbvmg!;Uts)6q@ylH%6u z_C!nbCvK*v>vE0EjtL?-5_#+Fs=WT6p%Tn)jjc1Ln`HaM#C7|jSM1|qKO@)b=LT$@ z*mvDd=oP+v<4$hCwuuAR?Sx3g^Ed9~25gzQ{kokH(fH{bcX9)^OWb+gPKd<47Z zVE%_lSMZ_daM#Zw(z)w{n1Uao&05?~0c{`SsyGrnOMLP!LfgyF;cLebBqFpWHip+3 zqrV~ciKEoWF$#`TaDoDZ0-|8=W8Y&@h@g-_wRSoN??KAI*ESGbZ=Fd?sDa+~#R4K0JXtNP0b2APg zYUs#AhFYo%KC;k@Se*0wTy?phbX+UoGYdJ6 zLDR*t0ggf6#j!z-K^e}mA&x;A&aq*RK^e}m5trH;H5K4LH|6zK>rjZ4P1Ev4Ios>< zqjtFgT_okVq&-rd*fQ~gwNzLvR_Ye(Q)~|kN@!r`ps6M{mn#p=6`w9ZUM#buh^@{t zP3X;RFSS@J70<~+X)$Lr`(Ci-?U@(<>erTk_`AP+ua_eAcmc{qr#OS@RFoI~?fqW* zT+NnRuR_@rgGlw#R$J1W^uBt%EqFK<(3c5sqqb76k0e- z5GM-#m;>%J>eK{utZ`Nl@4r*@Q|;{J)QQQd$-=QyPd#?LaANX%9)MDJ`LkBp)21Q}>m;65 z@Rf8J-29j(DDB$hjCi8ev7c_hH!^(3i;xRwImE)-}ejCTT!}tM?tFy6Y#)*TXCYYk?PI5K{YI)4* znoYl;7P_m$E~RCVHifhvPRnpw7HJTG`jhSy8l6;*t;aXM)4?B*u$K6}gZ4j#i! zw|50QnH&WfTakCGpEtfb?~1z6!;mD`Kyn+Wr`DwJ3L(6WA>19tcX51=GhueQ324I8 zuV4n*Jp7=pID4-j7@zw%+|S{Gp!c&=z$UGQg8{Bw%y=`$X}5Qz$u)CFN7|qpzoOcI z>*m_mMKk9Pbi@bT!H)PK=4jvY4man|*M2+?_|FIMJV?)&tAD|Bl3g<=cXE2~n)JIk zy>CtW2f{hKlQDdE7{80-hr;*=IQ~J#(cutBn|vJI(=qlTj9s$}?oh|rhA_6n7{d^c zL0fo`Vd`H4)4gGz_i~^2g>Mg>6jr@G_j4NqYtoN|Z5-h?9$>hSa(Ilx&u z&W`+;*6eMLxp?F1%9V)yxI0Fak@lM_5%HAsm^tG1;+-F*cm6r`N@U?ki0KbAeKOqM z5u!!={NMfl?*ae!u>X6%|9jN`J?8%&r{9@mb0c8d$S{oorcG@yeYpdsp9f4E7^X3I z6GImDA^RktD}VxRWGb~uQYq$`H%|!?3maRWsQTM8d@{)H5ZHcMU?J|Rvk0kT*A(HM zy+d?_HL#B9E0Aee!`a5mkf>}zP8f$?(&8fYrdD_H2 zu?tmSjexUc^&ZQlUU$IyF|*y`;Jnn~3S1>hwYm9whcLLe2!NQ9Q)5>fd)l?O#L0v_ z*~H-rSon&4FCK#0YlcFy7!QREz&w;BhN8@3AlQE(mvsz+`F3d`D!{sA*M#jQ2%m*{ zM~EdfAzIo)=I2Q7H42cKpNlCOfXpi1YXTykrB;Y5gl|GW^T<-E8FVO~qXO6JYbT+H z=4e!TFI`B5>CJU9jxNMybWtZ0EV*qwTr0x;=Q`ds(=}l<;)XOM*A1#h-WQ(cQu?BczBv0b|7O6RSqI4KF{ONgBonn;11l>Tkr~2%FO9r5M*dO2~yB0S= z^WrB_P&`ioSAHRU>TK&yLIP>e)&Lr;sC6fj-ou;2@*@G3-%AJ5`h0U8Oue5Dq-FW$ zI(YQ`bRez9H`l@0YjzM2C~+~EzHv=jTUfWwFAwX!J14bB4SwvVCp8PRu2c*7wBY)$ zrg=?p*IhCL)Ou(%zB!CE-dh*_)P-*uzPS#5;=OgiV+c*aH`j$5bkWEj6;R87ga_MF zV~YXr#3}N)13`hIZb6@xdl`5G&C8nK7TSH)-~f=YD~ONmMyLD9wzp9t*;i4GJ* zWwZeLW!Fn3%eJS{=F<4%;vB6VlLCu^X$nXT7o_YFH42`hfJA%IlJsde5oZy=_sc0j zUn>3yr4%UmaSDnEywqvZOC1-YCg2+;eTeeIjwn`bXrrL`C|9kbDEQ*IFsR~o3TOe7 zA}ery@{uZ+-Yv+?_qADw{M;~&G%S9amkfLJQc$co<(Yb|ZhKL{Fk^|2Q&SnU*nL|>mK85iuVgUsyYa~sHQ`C1 zWs3MTmFW%%1NV^#qJm8npP{xsOTpa~e2@at{pb5S-*`EO_E(fsxK!(%Vi35)kXCqI zZIvW2VR_QP)#ohoY4y?{sbkjX9mDQKpi2V(jiYqOSy?8sQ*lB(aVO* z$6uw_e_`@HVM9`>m~UUik_+ryraos#jH0Jn`6RvTybQerSYPJ8@`(%K(_;8k3!b_d zC4SK)2%Z1f#pv)OS>H|63|=-b7UaO^km-+K&-T<=VFR3TT_R``o*7Gw7pcG_YViB` z+7qB-u~bYA{<2zDrN5Y#ir|lxj;wgpQd)Z~2E}7KO3Fy>opfCOX5(o!r=``brmHzM zt!7}LNkBzOVKx%OZ(7yx9a9Z;SR3Yka_BDuRc1yVRTIx@Kk7@6I+Mzxw3b6V;Gcb% z($qHBL|smcMbINk62!yQ68AM60sJ@Fg0AZ-{@%&P2p;`vT8Co-)UNG+`VX_|VRaC> z`qiA?>513D%a12-pkHMg%aoYS^t`WgK=UPMW)V-nM>E5|A5g98aUEFG)WN?Olt;9^ z<3l{*{rK8s)NWU?nwq6su#Zeb#R_|o4oXkjG4k>yn^D47uRg#oG$wf!Hw-@lUv)Sd zTh)?fI7o+|xR5S3P?jTZcX$>NUqHP4Q38Ub2p3L01;v4r<|TT9Rb?ptW)lhZeXaPAqoZ)$1{qfGO6 z65qG_5U|J)Z31<(9WtGY*PVToL)pQRoIkIgP=DS8f4O1WT=Pnz{pGxGM8bkSR*Fq;};f7d`>_o-t3ya{Zat?@Y?>?g}uwM z=3oOWlBM;Ut$y>GV_3IMV0Qk(H5>cZdv9$1nvH#X-Np#25WAxRcA9;dLQ?8RfkQ7RP#P<5#G^%P28EAYR3f_%@Z+tJv}K-3SU;U^ZAy+sUUX zij^T_-NBg--f#20ZD{vPqzG$m&-3ybjT0=ya*J>C%umYcx_*80m} z8_q`h)IMxMGW55g_1^7{Hv(-c#0($BmmLZ~w}{>`MZ+(Qr(t){sL{fO5np4-XTM|7HLlh7O#cxy4>c2%9e*qaQ)H_W$xn=Ze=}|$cC44(R zw94dtb-P%3_>sm3d>0Nn`^4LDiE|(y zR%hx(oE>8)4m!cL$2c+h(c{OA)5kw@a`MqB)`?D)Xc-3|+F(-n@^bhRuxEmxgFB!z zL3%Ulz`90OBu87;V8e?!?6t8RBQJ{;&5T{rKAmmGo%nLXjl)ipa1*qxKB<_X3%YVC zqAJzIe#PE>5pB@lQ{*jzEq6JEEjzKIHIq)4i(`yV8urH{ZgNJ2pGFdGbi-dHGuv|^ z(o8v7Y~LwZsNlPCNnP$+QHV{rq>3|clv0{q)o!PMHbMR*3>kbl=&X{Nc@b7xC$~J{ zcCDzQjkRri&mkwvIfI;|J3~0;2tB$pjPdAZ-;3ncG2)B@@)0-f zcEOV*W%j>_SER7P8JpcWy9xNf`6n*78KpMQ8ti-WO@7`&VZrcm*p1!>T-bkyaY*vZ zx;Z2@;7A>IBhZwO`0)VaTir+wnteJn>1-p$g+~99(#$l5oLih*m$wsCE2{srVMT=x zO2o}vBItIY)iK`+xKUS}ov;EXVFm1g6)=mujc!k91>9sRmGNfQ%`V?Y7U<_Nhs~a6 zddRxh2KnZ-kfZ+{Y?n~QHfNV<5at@08O7Z4RkhiRUbd1QKpsY=m2na)gCTj?ha`hC z+ggzL7Dn_v596T>zi}`oz*Li%PGuPk=L*&e zwxz+Q=0uit7i%2rFXBd)qqL&V!Y2vOF|s#;S=A3K3nL5NE>4EQ6RpvG+k{=5#!*kH zIoIZ$@wlL^&x=8o4i${2&s1vXysq|ANs1KUX;bxdIP}0X2RI&Hib0)vZ{7f z*BbY&q9E&Veb;Tdv7PkfL{Qo(PvCF?ujcbOwlXm>(dd$HE)%mhd}vf5P5~P61cHv) zA{tFZ)uu184Ijr$on@wM7V~{UJB{_XfEfeswMM!f^K{ZQX4j{u%TN3Kr69qmjl}a7 zlRP|QRjsGz#l3$?NTG?+o*h;iyKan#ux6uoen|@c^0dIYa`lwxLQAl2r2VlMYSqe; zm&WN|`1{}tG0yDN??4e!+Q~!8jIbx=5!NYrbaChUXBWroq)H*{f~>jLh0)jSFZ>hyl!;ioHHGlwXq)7oppXkX_*G55Vu`; zkTc0@?$)*)GWo+;I&8aW?7oSa3~gt1V!&ul9^oxh9t`f(n$~F@zV`AL4OZA0T!e{1uIS$GJ%hUL7F9LT?@g)@Wbeb_wXX>NXY)>bu z>FKjjHY23>Q>VEV#MFpcS%XeHN1(Zs?=A6V`MzVvPfs2_USOZRBPWj@pE~A6iq$1a z{NiPd5Tr*ckw>Vh2Pk-ug3nPfNx=#Qq^R)X{)+>KnH0v*AwZ*Nd)-IsPVE#4vv7+P zUJsMBd10Xtw8h`xBxlR4>*dJT7hVbt?9k2s}uOcpv4&ucB?t z*ZwD>YNXlCL=wq_ItVpfLLGZIqvzoNmxyXtN8t;oLp4T9qS!mBIH|6v4F0rCOare? zIrNwKQ7nlK<7ygevxGJXC0YV%H%-%6%_qymew2=rN{;HI#|u7Lw}CGSg1g*b*;a0x9)h)#+Bc}++pB`E~L;N1ZvaJpOr$*Qs) z>N0|YNWxKkc?*}2fV{frfPiy1Av;EZh-WHIb7Hb9z?g zr2HIsh_q_BvDvH{y`YO_2qC)pFkZct_%?TGblqYmR|*X{~7fo%1d~IP(q3=8hVDyC(pvIJPS$W z^)gob++=`L60By2W`{3C#a|%5U&cbHl7K)FZy_f~BE!Oee3C>iw95?p2tXlFN47JP#A)hiP)oi5G6JJ39 z=l1~1f-r=m2xWNN`>I&+9V$mm4eMtlIewF?6-^Vjq}a7-+$uvGLKPnWk@MEmQc1-M zDbiGd)xomzBDhyXawV*^BGrdf5UtQtnS#HdfVN=qR|v+_E!CYDlk?5)U*N7_@(^#K zOg?>%j3roQuxr}^vw_u9AmH-ZX<7;eECqi%C#@j(oYrQ)ns5&U?a_@ia2O~E@+&$* z1l&Y_8R-3JC&w?$Tk!qQXet?^V)%|~qj*k5MvDpd+wHr1sXDG0mUX?<BSw;$N6Y;hZX~Yw=L=O34ER+E81sR2JzKP4Oh*UEP&4Yx-r7nxg zUx>o1HevRFIM$plGwY&F5LfcC?h9Hp`(MHK146FMrf0j|1c-jpO)h6(V(5Y1EQP*u zTyKEu4MKl3^kw}Oz>86(%we?sGbl5{x$MOWxU$Q=ZgPb+UiQ}z>kne$8%_@1l?1_< z^!QwO+!??qdT5)ab565yXOLkV!gkbSZhSGpYcaMQSqOh4a;Mw zF*>`Ex7D6z8YMP6o7{8@r?_e?*j9DZpwgY&s-L2W9|FEDgb%Z)*^O_;j5mAWkqW#+ z)AA?HtYa)2z)LSQGh5uQ8Jf{PKds;GHn-lT;7moIiSq)@9N^gIW--=Ve9f3{-srKHdQoc*pww=60B!A&FavXULF=^*o9M4% z6CDKJZ$p_0cQAmD+qh*JBLgl5+yP&s_7!t4!?f>}NOPc?f+zcSbH8~2Cmyz&2hH1| z!5Z-OWrNQQ;yO39{mdY3yl#&>04(<44H{^BgBYh03mDDs0sb7e?Hv-=v@YB%Myez4 zPuz^#@AeKWXe9&hcf5<{?Yv$wKV0H}xf%10oI>+8tZcvuJIplB`ptsRQtzzM3GPzfrI)Td=(umNlajd-Y20*k6q&tG8 zjxJz_ybRT=n(cy(Eo6wZ{+;sQ$*~s z`NCfX>_oqxdi>Fot)fBVY1(8BI@^csn1p?1a^P)6PF`}yZ|g$*0412yVZPxt?6&^4 z-Oh8I5r2nbjf}CgwS_~CqmDOH20jeknTB@uo<{F(BRsUnu7US7A`^E`&m8*HzSmw_ zI(~1Rj4D)z7=$N_q6&zhsNmr>HnB?^(70sSxFfI9YKm{otCRRZ5c+#V>@vAh@e@?-#A@EY+K|GsvSzes>VhGU8nJ~^)>0?@|l0%JQ z8p5i4$?AclTb+Y>*yTgMstQy*?7(-kHbSuSJp z7#0ISk@zTAC4*(K{~B_#hB0Xy;UWPT0|3KE{gIDMO-^CK`MAt=GHJYi48C@5Rr#q7 z77ZX6ICa=>UntKP_>Q$=#ph5OX~Q^Xkrog-r5$gpJKB1e>7mBy?dX{&rXDK1t)8uu z`1;xxKJKYTBX)S-)BBFx>#3>o_mSM_IR!O`0qZvo8(Z^Gru_lUqc&&%fE?in?X;#s zj=xl~iq-nOL7>gUxp61_x^RyO)M|F!8vg+?!q==5vY5NX&!U`ooq`yY4RA=GpRX(x zI_QGbl*QoFg2A$hS!%vo{2+0sl4X32o<5HNhGJ-~JRPd9Io5l5(Xw4FR>hxECZ4ev zScN3bkpF|~u`c89>52HE76&W-FGb&`;71gEk%B*_0D>~qSSC*cbEUYw#R-*QUARBd zfwLGTka-$Zt^_IvmA4~)i0U}ginF<1<_NqS%jfEj? z&Bt@+FW`=tHlubyvefiC2offO*yN;SOic%%g?qK8c$>UYMEmFN*(Z%MwwJz zEIs@ojP)GTkGkDgdt256EwU(Izj~YX}xQn0- zuAY``;%du~Fo?Hm#G&HqT24z&Xkin0pzbBx=YVj*uH=9?J+t&la6WO==|@fUyKD zr%5*fF1-n9a!$(nrB_^DXR?N*l?hN`h6X5(4=9U>vt~-QVzpBpY%?z;F<7R-( zS;#y%!~Bn1XKs(v<6@N@rggqeJ_Gh!SUS2@;xiaU1D*@5wFdtE$9!DEb&dYeg1_GU zJi_Mh`q<1McilBdW!vznnA;f-{Xu0bu@(acD-?~qF9YjHV(DkMaK6qVA|&n zF&>A5v7+zcuy zJL&O7l1aPK%Hh>La1OQq#Eqy8u2B!E4bRo;{pv1%-T3>J*#{qo57c*HkNf_&fBN+= zd=ocFzy8uMKpXtM-}sr=KljobU%LGLZ>_xXtuMa*&9A+&@~zjO{byn$_^QU?d;Sq0 zHitt{`X$!S9)0wY$DTNK zVF&wWL#Ybhc}be7;gSJnU8+G3i9bYr@ka<6aY&jtu3<>gD6;pU)X$42k@*I``25X% zqK`%el|3Y%PC=T}9&OWCLw?wd!ool>#qt^C2dKOsJ96^mAwDW4>JTR(xtsM8osEJg zjR7rjy#fimz7^e}JY>-mMl5fLbzl`1+qmkM)xxl*r*T93AuwEom>GQW`P2Kv z{WMmc0=)hTOZ|-`K}XHp(?}ii?GP|wBzGJ2YI&h<-BZ6CmD|yBs^xur&%TL$0e0w8 z`t|`ye!6~8)((-hy0(43R3b>O7yuuL9D`!ZZrtK7CK$ui}CLRhgvG1)vetZ8zTN!_suDV7OG&{7fqc;FCO zd<^3_C9ptc)QP8-ZPJ_Lzg;Lsnh4T2Nliq1F|384SRbu8`i}*yt}xE*+ii$mtj+p{ zR;ws39cfw~ECy?knCygP#vtPJ@bOc}$NQy$=Zk1U%FUu(No?hWp4aN+Zk8_=xo9=(e#ua099@SweOu!eFyiM_i@0?xrdqq<}=q ze3afZ8;8m80}m-K4f79Tewut|Z1Gc6IwN1ef}aR~Z*t_2L@}R{>p5tJC2F(}?bUuS zE^CAj2V};9yq|F}-$gU(!_6YNc$dKK^pf(1aYNWlmCj+7=I7xh5fbb;31d|Z3oauN zI_a8KauCU}t4AQ(hJtM>jZzt$z)Har1gE+n()y>du4Z)-W+B=}5)f8N0B-%Y>gy1n zAyh+jrV=FRrXdRRc`mq`jH!cU#De)s=QCL`RsRPaPAaXI+Jx*B)!d3!NR-a7Ku*pk z30*_ophm0bEI|!td5jEwte2m|sHk`UH!_BiY(>eIQ%tKh9sNNa0D*l})1S?120OaI zXoeozq0xTSIY#}p%oZ}zHFmBMd#^D;9cDA~VhVjOK95-g>jbw{uuiz(9;c;XAPWs* z5YpHnrhox%8N?zm2gO5kknXTQ|16ADFa{(b)W%&nZ9DJ?xP;meVIuHSNcjeuGY|n4 zcv-|*$W1{2#Fcpv?I1Ffu=EO@*Gj#JcMxJ5P6^DWVIq6N?K0yqW1T@M{(m!I81HIO zz6ygFOg!Y04bw;(rjQUEU=LASs2k)6gE8I{xJ00xAe;)d)H5;U~-T- znI)J*n%$^J41SIULN*jKi#pKkeg=j;I3)9v^ljiB@6W&c;-CNKZ@=;RUm)EgZ_8wJ z0O$6`x1Rmpi$C+m=fCp$3s)Kgsgq23jFxo)+Ep<^ga&;orvyxE(3Pe-`5)HsGj}8& zKtB98;k&Isd%tsX>hzI^Cy$Mv5OlGKphI?I1LbKucQmvevLVGEIB?s3?Q*eNSZ%}^+ z>4`6-xQm`zq!%o>q`#}_TTh_p!u^Is$jOt^{wUfS1To8iSOqV98v*GeLPz?zJ}pN& zI*>f_UL)t|L}U=efHZbl)P&Y8mi`gqk<7cKtpk~PcdaAWDhR-_F=xg+8Wa`Ec<;I= zK63!l>uGnVJo?xZ@|4Qo5-$EX1^U=UL-SwctNWu_n4_5B`f$Be7>sb+L>snE6>|%vQOAOP?gZYZaUuxz1czHcy`o*Y37QhXQ-rqjuH>hvV|Gs$}W4i({57cRna9k*AqPb@nW zcS$r9$lsQoV!bHmwyjWFoWwEN4iuh>O%fP6o8aU)RPe-O&^{T%9-G0A4Aprg2i`$E z4#4QfP-2=3h4FoSPcKfqF46zllCKC}eMS6N3Lc_&;*aU+zfo|8f{dxBK(F&{^5d9$ESzzg4=nyo2 zqMHKRu%)V;T}Fw)7VKcuPfv_$vf<_XAAj=r)WHJ&_gCS_X-J?)9(ee;oRqG~Ir_(B z^1=vWJ;ft5DkhOkoY{0l8gc1fJxNdGR4u)teuth2X_EM?Lm<4wVN4CK|EIZB>?s7g znoYBr0&_rrX&64}H-+o=_o;fCI5?W0K_Z3kWTTl#=AP`{>`u^;r!u>gkIBEp>BL63 H8(;l@e458r literal 0 HcmV?d00001 diff --git a/infini-qwen3-moe/scripts/infer_task.py b/infini-qwen3-moe/scripts/infer_task.py new file mode 100755 index 00000000..48873139 --- /dev/null +++ b/infini-qwen3-moe/scripts/infer_task.py @@ -0,0 +1,230 @@ +class Qwen3KVCache: + """Qwen3 专用的 KV 缓存包装器""" + def __init__(self, model_instance): + self._kvcache = model_instance.create_kv_cache() + self.tokens = [0 for _ in range(40960)] # max context length + self.model_instance = model_instance + + def data(self): + return self._kvcache + + def drop(self): + if self._kvcache: + # 修复：应该调用 model_instance 的 drop_kv_cache 方法 + self.model_instance.drop_kv_cache(self._kvcache) + self._kvcache = None + + + def update_tokens(self, tokens, pos): + end = pos + len(tokens) + max_len = len(self.tokens) + + # If overflow, truncate tokens to fit + if end > max_len: + tokens = tokens[: max_len - pos] + end = max_len + + self.tokens[pos:end] = tokens + + +class Qwen3InferTask: + """Qwen3 专用的推理任务""" + def __init__(self, tokens, position=0, temperature=1.0, topk=1, topp=1.0, + end_tokens=None, max_tokens=40960, task_id=0): + self.id = task_id + self.finish_reason = None + self.tokens = tokens + self.max_tokens = max_tokens + self.temperature = temperature + self.topk = topk + self.topp = topp + self.end_tokens = end_tokens or [151645] # Qwen3 eos_token_id + self._kv_cache = None + self.pos = position + + def release_kvcache(self): + cache = self._kv_cache + self._kv_cache = None + return cache + + def kvcache(self): + return self._kv_cache + + def next(self, out_token): + # 修复前的问题：KV缓存更新时机和方式不正确 + if self._kv_cache: + # 修复：应该在添加新token前更新KV缓存 + self._kv_cache.update_tokens(self.tokens, self.pos) + + # 修复：正确更新位置和tokens + self.pos += len(self.tokens) + + if out_token is None or out_token in self.end_tokens: + self.finish_reason = "stop" + elif self.pos >= self.max_tokens: + self.finish_reason = "length" + else: + # 修复：为下一轮推理准备新的token序列 + self.tokens = [out_token] + + def bind_kvcache(self, kv_cache, pos=0): + self._kv_cache = kv_cache + self.pos = pos + # 修复：如果pos > 0，需要正确处理token序列 + if pos > 0: + # 确保KV缓存中已有pos长度的历史tokens + if len(self.tokens) > pos: + self.tokens = self.tokens[pos:] + else: + # 如果tokens不足，这是一个错误状态 + raise ValueError(f"Token sequence length {len(self.tokens)} is less than position {pos}") + + +class Qwen3MoeKVCache: + """Qwen3-MoE 专用的 KV 缓存包装器""" + def __init__(self, model_instance): + self._kvcache = model_instance.create_kv_cache() + self.tokens = [0 for _ in range(40960)] # max context length + self.model_instance = model_instance + + def data(self): + return self._kvcache + + def drop(self): + if self._kvcache: + # 使用 Qwen3-MoE 特定的 drop 方法 + self.model_instance.drop_kv_cache(self._kvcache) + self._kvcache = None + + def update_tokens(self, tokens, pos): + end = pos + len(tokens) + max_len = len(self.tokens) + + # If overflow, truncate tokens to fit + if end > max_len: + tokens = tokens[: max_len - pos] + end = max_len + + self.tokens[pos:end] = tokens + + +class Qwen3MoeInferTask: + """Qwen3-MoE 专用的推理任务""" + def __init__(self, tokens, position=0, temperature=1.0, topk=1, topp=1.0, + end_tokens=None, max_tokens=40960, task_id=0): + self.id = task_id + self.finish_reason = None + self.tokens = tokens + self.max_tokens = max_tokens + self.temperature = temperature + self.topk = topk + self.topp = topp + self.end_tokens = end_tokens or [151645] # Qwen3-MoE 默认使用相同的 eos_token_id + self._kv_cache = None + self.pos = position + + def release_kvcache(self): + cache = self._kv_cache + self._kv_cache = None + return cache + + def kvcache(self): + return self._kv_cache + + def next(self, out_token): + # KV缓存更新逻辑与Qwen3相同 + if self._kv_cache: + self._kv_cache.update_tokens(self.tokens, self.pos) + + self.pos += len(self.tokens) + + if out_token is None or out_token in self.end_tokens: + self.finish_reason = "stop" + elif self.pos >= self.max_tokens: + self.finish_reason = "length" + else: + self.tokens = [out_token] + + def bind_kvcache(self, kv_cache, pos=0): + self._kv_cache = kv_cache + self.pos = pos + if pos > 0: + if len(self.tokens) > pos: + self.tokens = self.tokens[pos:] + else: + raise ValueError(f"Token sequence length {len(self.tokens)} is less than position {pos}") + + +# Legacy classes for backward compatibility (keeping existing InferTask and KVCache) +class KVCache: + def __init__(self, model_instance): + self._kvcache = model_instance.create_kv_cache() + self.tokens = [0 for _ in range(40960)] + self.model_instance = model_instance + + def data(self): + return self._kvcache + + def drop(self, model_instance=None): + if self._kvcache: + if model_instance: + model_instance.drop_kv_cache(self._kvcache) + elif self.model_instance: + self.model_instance.drop_kv_cache(self._kvcache) + self._kvcache = None + + def update_tokens(self, tokens, pos): + end = pos + len(tokens) + max_len = len(self.tokens) + + # If overflow, truncate tokens to fit + if end > max_len: + tokens = tokens[: max_len - pos] + end = max_len + + self.tokens[pos:end] = tokens + + +class InferTask: + def __init__(self, task_id, tokens, max_tokens, temperature=1.0, topk=1, topp=1.0, end_tokens=None): + self.id = task_id + self.finish_reason = None + self.tokens = tokens + self.max_tokens = max_tokens + self.temperature = temperature + self.topk = topk + self.topp = topp + self.end_tokens = end_tokens or [2] + self._kv_cache = None + self.pos = 0 + + def release_kvcache(self): + cache = self._kv_cache + self._kv_cache = None + return cache + + def kvcache(self): + return self._kv_cache + + def next(self, out_token): + if self._kv_cache: + self._kv_cache.update_tokens(self.tokens, self.pos) + + self.pos += len(self.tokens) + + if out_token is None or out_token in self.end_tokens: + self.finish_reason = "stop" + elif self.pos >= self.max_tokens: + self.finish_reason = "length" + else: + self.tokens = [out_token] + + def bind_kvcache(self, kv_cache, pos=0): + self._kv_cache = kv_cache + self.pos = pos + if pos > 0: + if len(self.tokens) > pos: + self.tokens = self.tokens[pos:] + else: + raise ValueError(f"Token sequence length {len(self.tokens)} is less than position {pos}") + diff --git a/infini-qwen3-moe/scripts/jiuge.py b/infini-qwen3-moe/scripts/jiuge.py new file mode 100755 index 00000000..23153a95 --- /dev/null +++ b/infini-qwen3-moe/scripts/jiuge.py @@ -0,0 +1,625 @@ +from typing import List +from libinfinicore_infer import ( + JiugeMetaCStruct, + JiugeWeightsCStruct, + KVCacheCStruct, + DataType, + DeviceType, + create_jiuge_model, + destroy_jiuge_model, + create_kv_cache, + drop_kv_cache, + infer_batch, +) +from infer_task import InferTask, KVCache + +from ctypes import POINTER, c_float, c_int, c_uint, c_void_p, byref +import os +from pathlib import Path +import safetensors +import sys +import time +import json +import math +import torch +import transformers + +torch.set_default_device("cpu") + + +class LlamaWeightsNaming: + def input_embd(self): + return "model.embed_tokens.weight" + + def output_norm(self): + return "model.norm.weight" + + def output_embd(self): + return "lm_head.weight" + + def attn_norm(self, i): + return f"model.layers.{i}.input_layernorm.weight" + + def attn_q(self, i): + return f"model.layers.{i}.self_attn.q_proj.weight" + + def attn_k(self, i): + return f"model.layers.{i}.self_attn.k_proj.weight" + + def attn_v(self, i): + return f"model.layers.{i}.self_attn.v_proj.weight" + + def attn_o(self, i): + return f"model.layers.{i}.self_attn.o_proj.weight" + + def attn_q_b(self, i): + return f"model.layers.{i}.self_attn.q_proj.bias" + + def attn_k_b(self, i): + return f"model.layers.{i}.self_attn.k_proj.bias" + + def attn_v_b(self, i): + return f"model.layers.{i}.self_attn.v_proj.bias" + + def ffn_norm(self, i): + return f"model.layers.{i}.post_attention_layernorm.weight" + + def gate(self, i): + return f"model.layers.{i}.mlp.gate_proj.weight" + + def up(self, i): + return f"model.layers.{i}.mlp.up_proj.weight" + + def down(self, i): + return f"model.layers.{i}.mlp.down_proj.weight" + + def match(state_dict): + return ( + "model.norm.weight" in state_dict + and "model.layers.0.self_attn.q_proj.weight" in state_dict + ) + + +class JiugeMetaFromLlama(JiugeMetaCStruct): + def __init__(self, config, dtype=torch.float16, max_tokens=None): + if dtype == torch.float16: + dt_ = DataType.INFINI_DTYPE_F16 + elif dtype == torch.float32: + dt_ = DataType.INFINI_DTYPE_F32 + elif dtype == torch.bfloat16: + dt_ = DataType.INFINI_DTYPE_BF16 + else: + dt_ = DataType.INFINI_DTYPE_F16 + + self.scale_input = 1.0 + self.scale_output = 1.0 + self.scale_o = 1.0 + self.scale_down = 1.0 + if ( + "fm9g" == config["model_type"] + and "scale_emb" in config + and "scale_depth" in config + and "dim_model_base" in config + ): + self.scale_input = config["scale_emb"] + self.scale_output = config["hidden_size"] // config["dim_model_base"] + self.scale_o = config["scale_depth"] / math.sqrt( + config["num_hidden_layers"] + ) + self.scale_down = config["scale_depth"] / math.sqrt( + config["num_hidden_layers"] + ) + + super().__init__( + dt_logits=dt_, + nlayer=config["num_hidden_layers"], + d=config["hidden_size"], + nh=config["num_attention_heads"], + nkvh=( + config["num_key_value_heads"] + if "num_key_value_heads" in config + else config["num_attention_heads"] + ), + dh=config["hidden_size"] // config["num_attention_heads"], + di=config["intermediate_size"], + dctx=( + config["max_position_embeddings"] if max_tokens is None else max_tokens + ), + dvoc=config["vocab_size"], + epsilon=config["rms_norm_eps"], + theta=(config["rope_theta"] if "rope_theta" in config else 100000.0), + end_token=2, + ) + self.torch_dtype_logits = dtype + + +class JiugeWeightsImpl(JiugeWeightsCStruct): + def __init__( + self, + meta, + naming, + state_dict, + torch_dt_mat=torch.float16, + torch_dt_norm=torch.float32, + ndev=1, + transpose_weight=True, + ): + nlayer = meta.nlayer + nh = meta.nh + nkvh = meta.nkvh + dh = meta.dh + d = meta.d + di = meta.di + scale_input = meta.scale_input + scale_output = meta.scale_output + scale_o = meta.scale_o + scale_down = meta.scale_down + assert nh % nkvh == 0 + assert nh % ndev == 0 + assert nkvh % ndev == 0 + assert di % ndev == 0 + torch_dt_logits = meta.torch_dtype_logits + if torch_dt_mat == torch.float16: + self.dt_mat = DataType.INFINI_DTYPE_F16 + elif torch_dt_mat == torch.float32: + self.dt_mat = DataType.INFINI_DTYPE_F32 + elif torch_dt_mat == torch.bfloat16: + self.dt_mat = DataType.INFINI_DTYPE_BF16 + else: + raise ValueError("Unsupported proj weight data type") + if torch_dt_norm == torch.float16: + self.dt_norm = DataType.INFINI_DTYPE_F16 + elif torch_dt_norm == torch.float32: + self.dt_norm = DataType.INFINI_DTYPE_F32 + elif torch_dt_norm == torch.bfloat16: + self.dt_norm = DataType.INFINI_DTYPE_BF16 + else: + raise ValueError("Unsupported norm weight data type") + + input_embd_naming = ( + naming.input_embd() + if naming.input_embd() in state_dict + else naming.output_embd() + ) + output_embd_naming = ( + naming.output_embd() + if naming.output_embd() in state_dict + else naming.input_embd() + ) + self.transpose_linear_weights = 1 if transpose_weight else 0 + self.nlayer = nlayer + self.input_embd_tensor = ( + state_dict[input_embd_naming].to(torch_dt_logits) * scale_input + ) + self.input_embd = self.input_embd_tensor.data_ptr() + self.output_norm_tensor = ( + state_dict[naming.output_norm()].to(torch_dt_norm) * scale_output + ) + self.output_norm = self.output_norm_tensor.data_ptr() + self.output_embd_tensor = state_dict[output_embd_naming].to(torch_dt_mat) + if not transpose_weight: + self.output_embd_tensor = self.output_embd_tensor.transpose( + 0, 1 + ).contiguous() + self.output_embd = self.output_embd_tensor.data_ptr() + + self.attn_norm_tensors = [ + state_dict[naming.attn_norm(i)].to(torch_dt_norm) for i in range(nlayer) + ] + self.attn_norm_ptrs = [ + self.attn_norm_tensors[i].data_ptr() for i in range(nlayer) + ] + self.attn_norm = (c_void_p * nlayer)(*self.attn_norm_ptrs) + + def qkv_slices(_i): + _Q = ( + state_dict[naming.attn_q(_i)] + .reshape([nh, 2, dh // 2, d]) + .transpose(1, 2) + ) + _K = ( + state_dict[naming.attn_k(_i)] + .reshape([nkvh, 2, dh // 2, d]) + .transpose(1, 2) + ) + _V = state_dict[naming.attn_v(_i)].reshape([nkvh, dh // 2, 2, d]) + _result = [] + _nh = nh // ndev + _nkvh = nkvh // ndev + for _idev in range(ndev): + _result.append(_Q[_idev * _nh : (_idev + 1) * _nh, :, :, :]) + _result.append(_K[_idev * _nkvh : (_idev + 1) * _nkvh, :, :, :]) + _result.append(_V[_idev * _nkvh : (_idev + 1) * _nkvh, :, :]) + return _result + + self.qkv_tensor = [ + torch.concat(qkv_slices(i)).to(torch_dt_mat) for i in range(nlayer) + ] + if not transpose_weight: + for i in range(nlayer): + self.qkv_tensor[i] = ( + self.qkv_tensor[i] + .reshape(ndev, (nh + 2 * nkvh) // ndev * dh, d) + .transpose(1, 2) + .contiguous() + ) + self.qkv_tensor_ptrs = [self.qkv_tensor[i].data_ptr() for i in range(nlayer)] + self.attn_qkv = (c_void_p * nlayer)(*self.qkv_tensor_ptrs) + + def qkv_b_slices(_i): + _QB = ( + state_dict[naming.attn_q_b(_i)] + .reshape([nh, 2, dh // 2]) + .transpose(1, 2) + ) + _KB = ( + state_dict[naming.attn_k_b(_i)] + .reshape([nkvh, 2, dh // 2]) + .transpose(1, 2) + ) + _VB = state_dict[naming.attn_v_b(_i)].reshape([nkvh, dh // 2, 2]) + _result = [] + _nh = nh // ndev + _nkvh = nkvh // ndev + for _idev in range(ndev): + _result.append(_QB[_idev * _nh : (_idev + 1) * _nh, :, :].flatten()) + _result.append(_KB[_idev * _nkvh : (_idev + 1) * _nkvh, :, :].flatten()) + _result.append(_VB[_idev * _nkvh : (_idev + 1) * _nkvh, :, :].flatten()) + return _result + + if naming.attn_q_b(0) in state_dict: + self.qkv_b_tensors = [ + torch.concat(qkv_b_slices(i)).to(torch_dt_logits) for i in range(nlayer) + ] + self.qkv_b_tensor_ptrs = [ + self.qkv_b_tensors[i].data_ptr() for i in range(nlayer) + ] + self.attn_qkv_b = (c_void_p * nlayer)(*self.qkv_b_tensor_ptrs) + else: + self.attn_qkv_b = None + + self.attn_o_tensor = [ + ( + state_dict[naming.attn_o(i)] + .to(torch_dt_mat) + .reshape([d, ndev, nh // ndev * dh]) + .transpose(0, 1) + .contiguous() + if transpose_weight + else state_dict[naming.attn_o(i)] + .transpose(0, 1) + .to(torch_dt_mat) + .contiguous() + ) + * scale_o + for i in range(nlayer) + ] + self.attn_o_ptrs = [self.attn_o_tensor[i].data_ptr() for i in range(nlayer)] + self.attn_o = (c_void_p * nlayer)(*self.attn_o_ptrs) + + self.ffn_norm_tensors = [ + state_dict[naming.ffn_norm(i)].to(torch_dt_norm) for i in range(nlayer) + ] + self.ffn_norm_ptrs = [ + self.ffn_norm_tensors[i].data_ptr() for i in range(nlayer) + ] + self.ffn_norm = (c_void_p * nlayer)(*self.ffn_norm_ptrs) + + def gate_up_slices(_i): + _result = [] + _di = di // ndev + for _idev in range(ndev): + _start = _idev * _di + _end = (_idev + 1) * _di + _result.append(state_dict[naming.gate(_i)][_start:_end, :]) + _result.append(state_dict[naming.up(_i)][_start:_end, :]) + return _result + + self.gate_up_tensors = [ + torch.concat(gate_up_slices(i)).to(torch_dt_mat) for i in range(nlayer) + ] + if not transpose_weight: + for i in range(nlayer): + self.gate_up_tensors[i] = ( + self.gate_up_tensors[i] + .reshape(ndev, 2 * di // ndev, d) + .transpose(1, 2) + .contiguous() + ) + self.gate_up_ptrs = [self.gate_up_tensors[i].data_ptr() for i in range(nlayer)] + self.ffn_gate_up = (c_void_p * nlayer)(*self.gate_up_ptrs) + + self.ffn_down_tensor = [ + ( + state_dict[naming.down(i)] + .to(torch_dt_mat) + .reshape([d, ndev, di // ndev]) + .transpose(0, 1) + .contiguous() + if transpose_weight + else state_dict[naming.down(i)] + .transpose(0, 1) + .to(torch_dt_mat) + .contiguous() + ) + * scale_down + for i in range(nlayer) + ] + self.ffn_down_ptrs = [self.ffn_down_tensor[i].data_ptr() for i in range(nlayer)] + self.ffn_down = (c_void_p * nlayer)(*self.ffn_down_ptrs) + + +class JiugeBatchedTask: + def __init__(self, tasks: List[InferTask]): + self.tasks = tasks + self.nreq = len(tasks) + + # Precompute fields + token_lists = [t.tokens for t in tasks] + self.req_lens_list = [len(toks) for toks in token_lists] + self.req_pos_list = [t.pos for t in tasks] + self.kv_cache_ptrs = [t.kvcache().data() for t in tasks] + self.temperaturas_list = [t.temperature for t in tasks] + self.topks_list = [t.topk for t in tasks] + self.topps_list = [t.topp for t in tasks] + + # Flatten token lists + flat_tokens = [tok for toks in token_lists for tok in toks] + self.ntok = len(flat_tokens) + + # Convert to ctypes arrays in one pass + self.tokens = (c_uint * self.ntok)(*flat_tokens) + self.req_lens = (c_uint * self.nreq)(*self.req_lens_list) + self.req_pos = (c_uint * self.nreq)(*self.req_pos_list) + self.kv_caches = (POINTER(KVCacheCStruct) * self.nreq)(*self.kv_cache_ptrs) + self.temperaturas = (c_float * self.nreq)(*self.temperaturas_list) + self.topks = (c_uint * self.nreq)(*self.topks_list) + self.topps = (c_float * self.nreq)(*self.topps_list) + + def input_args(self): + return ( + self.tokens, + self.ntok, + self.req_lens, + self.nreq, + self.req_pos, + self.kv_caches, + self.temperaturas, + self.topks, + self.topps, + ) + + +class JiugeForCauslLM: + def __init__( + self, model_dir_path, device=DeviceType.DEVICE_TYPE_CPU, ndev=1, max_tokens=None + ): + def load_all_safetensors_from_dir(dir_path_: str): + tensors_ = {} + dir_path_ = Path(dir_path_) + for file in sorted(dir_path_.glob("*.safetensors")): + data_ = safetensors.safe_open(file, "pt") + for name_ in data_.keys(): + tensors_[name_] = data_.get_tensor(name_) + return tensors_ + + print("Loading model weights to host...") + load_start_time = time.time() + + with open(os.path.join(model_dir_path, "config.json"), "r") as f: + config = json.load(f) + self.config = config + eos_token_id = self.config["eos_token_id"] + self.eos_token_id = ( + [eos_token_id] if type(eos_token_id) == int else eos_token_id + ) + transpose_weight = ( + device != DeviceType.DEVICE_TYPE_ASCEND + ) # y = xW is faster than y=xW^T on Ascend + if "llama" == config["model_type"]: + model = ( + transformers.LlamaForCausalLM.from_pretrained(model_dir_path) + .cpu() + .half() + ) + self.meta = JiugeMetaFromLlama(config, max_tokens=max_tokens) + self.tokenizer = transformers.AutoTokenizer.from_pretrained(model_dir_path) + self.weights = JiugeWeightsImpl( + self.meta, + LlamaWeightsNaming(), + model.state_dict(), + ndev=ndev, + transpose_weight=transpose_weight, + ) + elif "fm9g" == config["model_type"]: + if any( + file.suffix == ".safetensors" for file in Path(model_dir_path).iterdir() + ): + state_dict = load_all_safetensors_from_dir(model_dir_path) + else: + state_dict = torch.load( + os.path.join(model_dir_path, "pytorch_model.bin"), + weights_only=True, + map_location="cpu", + ) + if LlamaWeightsNaming.match(state_dict): + self.meta = JiugeMetaFromLlama(config, max_tokens=max_tokens) + self.weights = JiugeWeightsImpl( + self.meta, + LlamaWeightsNaming(), + state_dict, + ndev=ndev, + transpose_weight=transpose_weight, + ) + self.tokenizer = transformers.AutoTokenizer.from_pretrained( + model_dir_path, trust_remote_code=True + ) + else: + raise ValueError("Unsupported weight naming") + elif "fm9g7b" == config["model_type"]: + if any( + file.suffix == ".safetensors" for file in Path(model_dir_path).iterdir() + ): + state_dict = load_all_safetensors_from_dir(model_dir_path) + else: + state_dict = torch.load( + os.path.join(model_dir_path, "pytorch_model.bin"), + weights_only=True, + map_location="cpu", + ) + if LlamaWeightsNaming.match(state_dict): + self.meta = JiugeMetaFromLlama(config, max_tokens=max_tokens) + self.weights = JiugeWeightsImpl( + self.meta, + LlamaWeightsNaming(), + state_dict, + ndev=ndev, + transpose_weight=transpose_weight, + ) + self.tokenizer = transformers.AutoTokenizer.from_pretrained( + model_dir_path, trust_remote_code=True + ) + else: + raise ValueError("Unsupported weight naming") + elif "qwen2" == config["model_type"]: + state_dict = load_all_safetensors_from_dir(model_dir_path) + if LlamaWeightsNaming.match(state_dict): + self.meta = JiugeMetaFromLlama(config, max_tokens=max_tokens) + self.weights = JiugeWeightsImpl( + self.meta, + LlamaWeightsNaming(), + state_dict, + ndev=ndev, + transpose_weight=transpose_weight, + ) + self.tokenizer = transformers.AutoTokenizer.from_pretrained( + model_dir_path + ) + else: + raise ValueError("Unsupported model architecture") + + load_end_time = time.time() + print(f"Time used: {load_end_time - load_start_time:.3f}s") + + print(f"Creating model on {ndev} devices...") + load_start_time = time.time() + dev_ids = (c_int * ndev)(*[i for i in range(ndev)]) + self.model_instance = create_jiuge_model( + byref(self.meta), + byref(self.weights), + device, + ndev, + dev_ids, + ) + load_end_time = time.time() + print(f"Time used: {load_end_time - load_start_time:.3f}s") + + def max_context_len(self): + return self.meta.dctx + + def create_kv_cache(self): + return create_kv_cache(self.model_instance) + + def drop_kv_cache(self, kv_cache): + drop_kv_cache(self.model_instance, kv_cache) + + def batch_infer_one_round(self, tasks: List[InferTask]): + output = (c_uint * len(tasks))() + batch_inputs = JiugeBatchedTask(tasks) + infer_batch( + self.model_instance, + *(batch_inputs.input_args()), + output, + ) + return list(output) + + def generate(self, input_content, max_steps, topp_=1.0, topk_=1, temperature_=1.0): + input_content = self.tokenizer.apply_chat_template( + conversation=[{"role": "user", "content": input_content}], + add_generation_prompt=True, + tokenize=False, + ) + print(input_content, end="", flush=True) + tokens = self.tokenizer.encode(input_content) + infer_task = InferTask( + 0, + tokens, + self.max_context_len(), + temperature_, + topk_, + topp_, + self.eos_token_id, + ) + infer_task.bind_kvcache(KVCache(self)) + + steps = 0 + total_time = 0 + output_content = "" + + for step_i in range(max_steps): + start_time = time.time() + output_tokens = self.batch_infer_one_round([infer_task]) + end_time = time.time() + steps += 1 + output_str = ( + self.tokenizer._tokenizer.id_to_token(output_tokens[0]) + .replace("▁", " ") + .replace("<0x0A>", "\n") + ) + output_content += output_str + print(output_str, end="", flush=True) + if output_tokens[0] in self.eos_token_id: + break + infer_task.next(output_tokens[0]) + + if step_i > 0: + total_time += end_time - start_time + + print("\n") + avg_time = total_time * 1000 / (steps - 1) + print(f"Time per step: {avg_time:.3f}ms") + + infer_task._kv_cache.drop(self) + return output_content, avg_time + + def destroy_model_instance(self): + destroy_jiuge_model(self.model_instance) + print("Model destroyed") + + +def test(): + if len(sys.argv) < 3: + print( + "Usage: python jiuge.py [--cpu | --nvidia| --cambricon | --ascend | --metax | --moore] [n_device]" + ) + sys.exit(1) + model_path = sys.argv[2] + device_type = DeviceType.DEVICE_TYPE_CPU + if sys.argv[1] == "--cpu": + device_type = DeviceType.DEVICE_TYPE_CPU + elif sys.argv[1] == "--nvidia": + device_type = DeviceType.DEVICE_TYPE_NVIDIA + elif sys.argv[1] == "--cambricon": + device_type = DeviceType.DEVICE_TYPE_CAMBRICON + elif sys.argv[1] == "--ascend": + device_type = DeviceType.DEVICE_TYPE_ASCEND + elif sys.argv[1] == "--metax": + device_type = DeviceType.DEVICE_TYPE_METAX + elif sys.argv[1] == "--moore": + device_type = DeviceType.DEVICE_TYPE_MOORE + elif sys.argv[1] == "--iluvatar": + device_type = DeviceType.DEVICE_TYPE_ILUVATAR + else: + print( + "Usage: python jiuge.py [--cpu | --nvidia| --cambricon | --ascend | --metax | --moore] [n_device]" + ) + sys.exit(1) + + ndev = int(sys.argv[3]) if len(sys.argv) > 3 else 1 + model = JiugeForCauslLM(model_path, device_type, ndev) + model.generate("山东最高的山是？", 500) + model.destroy_model_instance() + + +if __name__ == "__main__": + test() diff --git a/infini-qwen3-moe/scripts/kvcache_pool.py b/infini-qwen3-moe/scripts/kvcache_pool.py new file mode 100755 index 00000000..81914535 --- /dev/null +++ b/infini-qwen3-moe/scripts/kvcache_pool.py @@ -0,0 +1,90 @@ +from infer_task import KVCache + +import asyncio +from typing import List +import threading + + +class KVCachePool: + def __init__(self, model, max_caches: int = 32): + self.max_caches = max_caches + self.model = model + self._available: List[KVCache] = [] + self.num_caches = len(self._available) + self._lock = threading.Lock() + self._not_empty = threading.Condition(self._lock) + self._shutdown = False + + def acquire_sync(self, infer_task): + with self._not_empty: + while True: + if self._shutdown: + raise RuntimeError( + "KVCachePool is shutting down; cannot acquire new cache." + ) + if len(self._available) == 0: + if self.num_caches < self.max_caches: + self.num_caches += 1 + print( + f"[INFO] Task {infer_task.id} created new KVCachePoolItem" + ) + return infer_task.bind_kvcache(KVCache(self.model), 0) + else: + self._not_empty.wait() + else: + max_match, max_match_index = self.find_most_matching_cache( + infer_task.tokens + ) + kvcache = self._available.pop(max_match_index) + print( + f"[INFO] Task {infer_task.id} reused KVCachePoolItem {max_match_index} with {max_match} matches" + ) + return infer_task.bind_kvcache(kvcache, max_match) + + def release_sync(self, infer_task): + with self._not_empty: + print(f"[INFO] Task {infer_task.id} returned KVCachePoolItem to pool") + self._available.append(infer_task.release_kvcache()) + self._not_empty.notify() + + async def acquire(self, infer_task): + loop = asyncio.get_event_loop() + return await loop.run_in_executor(None, self.acquire_sync, infer_task) + + async def release(self, infer_task): + loop = asyncio.get_event_loop() + return await loop.run_in_executor(None, self.release_sync, infer_task) + + def find_most_matching_cache(self, tokens: List[int]): + max_match = 0 + max_match_index = 0 + + def first_different_index(a_, b_): + for i_, (x_, y_) in enumerate(zip(a_, b_)): + if x_ != y_: + return i_ + return min(len(a_), len(b_)) + + for i, kvcache in enumerate(self._available): + common_elements = first_different_index(tokens, kvcache.tokens) + # print(f"{tokens}") + # print(f"{kvcache.tokens[:len(tokens)]}") + if common_elements > max_match: + max_match = common_elements + max_match_index = i + + return (min(max_match, len(tokens) - 1), max_match_index) + + def finalize(self): + with self._not_empty: + self._shutdown = True + while len(self._available) < self.num_caches: + self._not_empty.wait() + + for kvcache in self._available: + if kvcache is not None: + kvcache.drop(self.model) + + self._available.clear() + self.max_caches = 0 + self.num_caches = 0 diff --git a/infini-qwen3-moe/scripts/launch_server.py b/infini-qwen3-moe/scripts/launch_server.py new file mode 100755 index 00000000..4847a477 --- /dev/null +++ b/infini-qwen3-moe/scripts/launch_server.py @@ -0,0 +1,297 @@ +from jiuge import JiugeForCauslLM +from libinfinicore_infer import DeviceType +from infer_task import InferTask +from kvcache_pool import KVCachePool + +import argparse +import queue +from fastapi import FastAPI, Request +from fastapi.responses import StreamingResponse, JSONResponse +import contextlib +import uvicorn +import time +import uuid +import json +import threading +import janus + + +DEVICE_TYPE_MAP = { + "cpu": DeviceType.DEVICE_TYPE_CPU, + "nvidia": DeviceType.DEVICE_TYPE_NVIDIA, + "cambricon": DeviceType.DEVICE_TYPE_CAMBRICON, + "ascend": DeviceType.DEVICE_TYPE_ASCEND, + "metax": DeviceType.DEVICE_TYPE_METAX, + "moore": DeviceType.DEVICE_TYPE_MOORE, +} + +def parse_args(): + parser = argparse.ArgumentParser(description="Launch the LLM inference server.") + parser.add_argument( + "--model-path", + type=str, + help="Path to the model directory", + ) + parser.add_argument( + "--dev", + type=str, + choices=DEVICE_TYPE_MAP.keys(), + default="cpu", + help="Device type to run the model on (default: cpu)", + ) + parser.add_argument( + "--ndev", + type=int, + default=1, + help="Number of devices to use (default: 1)", + ) + parser.add_argument( + "--max-batch", + type=int, + default=3, + help="Maximum number of requests that can be batched together (default: 3)", + ) + parser.add_argument( + "--max-tokens", + type=int, + required=False, + default=None, + help="Max token sequence length that model will handle (follows model config if not provided)", + ) + return parser.parse_args() + +args = parse_args() +device_type = DEVICE_TYPE_MAP[args.dev] +model_path = args.model_path +ndev = args.ndev +max_tokens = args.max_tokens + +MAX_BATCH = args.max_batch +print( + f"Using MAX_BATCH={MAX_BATCH}. Try reduce this value if out of memory error occurs." +) + +def chunk_json(id_, content=None, role=None, finish_reason=None): + delta = {} + if content: + delta["content"] = content + if role: + delta["role"] = role + return { + "id": id_, + "object": "chat.completion.chunk", + "created": int(time.time()), + "model": "jiuge", + "system_fingerprint": None, + "choices": [ + { + "index": 0, + "delta": delta, + "logprobs": None, + "finish_reason": finish_reason, + } + ], + } + + +# A wrapper for InferTask that supports async output queue +class AsyncInferTask(InferTask): + def __init__(self, id, tokens, max_tokens, temperature, topk, topp, end_tokens): + super().__init__(id, tokens, max_tokens, temperature, topk, topp, end_tokens) + self.output_queue = janus.Queue() + print(f"[INFO] Create InferTask {self.id}") + + def output(self, out_token): + self.next(out_token) + self.output_queue.sync_q.put(out_token) + + +@contextlib.asynccontextmanager +async def lifespan(app: FastAPI): + # Startup + app.state.model = JiugeForCauslLM(model_path, device_type, ndev, max_tokens=max_tokens) + app.state.kv_cache_pool = KVCachePool(app.state.model, MAX_BATCH) + app.state.request_queue = janus.Queue() + worker_thread = threading.Thread(target=worker_loop, args=(app,), daemon=True) + worker_thread.start() + + try: + yield # The app runs here + finally: + # Shutdown + app.state.request_queue.sync_q.put(None) + worker_thread.join() + app.state.request_queue.shutdown() + + app.state.kv_cache_pool.finalize() + app.state.model.destroy_model_instance() + + +App = FastAPI(lifespan=lifespan) + + +# App loop: take requests from the queue, do inference, and put unfinished requests back into the queue. +def worker_loop(app): + while True: + try: + task = app.state.request_queue.sync_q.get(timeout=0.01) + except queue.Empty: + continue + + if task is None: + return + + batch = [task] + while len(batch) < MAX_BATCH: + try: + req = app.state.request_queue.sync_q.get_nowait() + if req is not None: + batch.append(req) + except queue.Empty: + break + output_tokens = app.state.model.batch_infer_one_round(batch) + for task, token in zip(batch, output_tokens): + task.output(token) + if task.finish_reason is None: + app.state.request_queue.sync_q.put(task) + else: + print(f"[INFO] Task {task.id} finished infer.") + app.state.kv_cache_pool.release_sync(task) + + +def build_task(id_, request_data, request: Request): + messages = request_data.get("messages", []) + input_content = request.app.state.model.tokenizer.apply_chat_template( + conversation=messages, + add_generation_prompt=True, + tokenize=False, + ) + tokens = request.app.state.model.tokenizer.encode(input_content) + return AsyncInferTask( + id_, + tokens, + request_data.get("max_tokens", request.app.state.model.max_context_len()), + request_data.get("temperature", 1.0), + request_data.get("top_k", 1), + request_data.get("top_p", 1.0), + request.app.state.model.eos_token_id, + ) + + +async def chat_stream(id_, request_data, request: Request): + try: + infer_task = build_task(id_, request_data, request) + await request.app.state.kv_cache_pool.acquire(infer_task) + + # Initial empty content + chunk = json.dumps( + chunk_json(id_, content="", role="assistant"), ensure_ascii=False + ) + yield f"data: {chunk}\n\n" + + request.app.state.request_queue.sync_q.put(infer_task) + + while True: + if await request.is_disconnected(): + print("Client disconnected. Aborting stream.") + break + if ( + infer_task.finish_reason is not None + and infer_task.output_queue.async_q.empty() + ): + chunk = json.dumps( + chunk_json(id_, finish_reason=infer_task.finish_reason), + ensure_ascii=False, + ) + yield f"data: {chunk}\n\n" + break + + token = await infer_task.output_queue.async_q.get() + content = ( + request.app.state.model.tokenizer._tokenizer.id_to_token(token) + .replace("▁", " ") + .replace("<0x0A>", "\n") + ) + chunk = json.dumps(chunk_json(id_, content=content), ensure_ascii=False) + yield f"data: {chunk}\n\n" + + except Exception as e: + print(f"[Error] ID : {id_} Exception: {e}") + finally: + if infer_task.finish_reason is None: + infer_task.finish_reason = "cancel" + + +async def chat(id_, request_data, request: Request): + try: + infer_task = build_task(id_, request_data, request) + await request.app.state.kv_cache_pool.acquire(infer_task) + request.app.state.request_queue.sync_q.put(infer_task) + output = [] + while True: + if ( + infer_task.finish_reason is not None + and infer_task.output_queue.async_q.empty() + ): + break + + token = await infer_task.output_queue.async_q.get() + content = ( + request.app.state.model.tokenizer._tokenizer.id_to_token(token) + .replace("▁", " ") + .replace("<0x0A>", "\n") + ) + output.append(content) + + output_text = "".join(output).strip() + response = chunk_json( + id_, + content=output_text, + role="assistant", + finish_reason=infer_task.finish_reason or "stop", + ) + return response + + except Exception as e: + print(f"[Error] ID: {id_} Exception: {e}") + return JSONResponse(content={"error": str(e)}, status_code=500) + finally: + if infer_task.finish_reason is None: + infer_task.finish_reason = "cancel" + + +@App.post("/chat/completions") +async def chat_completions(request: Request): + data = await request.json() + + if not data.get("messages"): + return JSONResponse(content={"error": "No message provided"}, status_code=400) + + stream = data.get("stream", False) + id_ = f"cmpl-{uuid.uuid4().hex}" + if stream: + return StreamingResponse( + chat_stream(id_, data, request), media_type="text/event-stream" + ) + else: + response = await chat(id_, data, request) + return JSONResponse(content=response) + +if __name__ == "__main__": + uvicorn.run(App, host="0.0.0.0", port=8000) + +""" +curl -N -H "Content-Type: application/json" \ + -X POST http://127.0.0.1:8000/chat/completions \ + -d '{ + "model": "jiuge", + "messages": [ + {"role": "user", "content": "山东最高的山是？"} + ], + "temperature": 1.0, + "top_k": 50, + "top_p": 0.8, + "max_tokens": 512, + "stream": true + }' +""" diff --git a/infini-qwen3-moe/scripts/libinfinicore_infer.py b/infini-qwen3-moe/scripts/libinfinicore_infer.py new file mode 100755 index 00000000..5587567f --- /dev/null +++ b/infini-qwen3-moe/scripts/libinfinicore_infer.py @@ -0,0 +1,157 @@ +import ctypes +from ctypes import c_size_t, c_uint, c_int, c_float, c_void_p, c_bool, POINTER +import os + + +class DataType(ctypes.c_int): + INFINI_DTYPE_INVALID = 0 + INFINI_DTYPE_BYTE = 1 + INFINI_DTYPE_BOOL = 2 + INFINI_DTYPE_I8 = 3 + INFINI_DTYPE_I16 = 4 + INFINI_DTYPE_I32 = 5 + INFINI_DTYPE_I64 = 6 + INFINI_DTYPE_U8 = 7 + INFINI_DTYPE_U16 = 8 + INFINI_DTYPE_U32 = 9 + INFINI_DTYPE_U64 = 10 + INFINI_DTYPE_F8 = 11 + INFINI_DTYPE_F16 = 12 + INFINI_DTYPE_F32 = 13 + INFINI_DTYPE_F64 = 14 + INFINI_DTYPE_C16 = 15 + INFINI_DTYPE_C32 = 16 + INFINI_DTYPE_C64 = 17 + INFINI_DTYPE_C128 = 18 + INFINI_DTYPE_BF16 = 19 + + +class DeviceType(ctypes.c_int): + DEVICE_TYPE_CPU = 0 + DEVICE_TYPE_NVIDIA = 1 + DEVICE_TYPE_CAMBRICON = 2 + DEVICE_TYPE_ASCEND = 3 + DEVICE_TYPE_METAX = 4 + DEVICE_TYPE_MOORE = 5 + DEVICE_TYPE_ILUVATAR = 6 + + +class Qwen3MetaCStruct(ctypes.Structure): + _fields_ = [ + # 数据类型 + ("dt_logits", DataType), # infiniDtype_t dt_logits + + # 网络规模 + ("nlayer", c_size_t), # size_t nlayer + ("d", c_size_t), # size_t d + ("nh", c_size_t), # size_t nh + ("nkvh", c_size_t), # size_t nkvh + ("dh", c_size_t), # size_t dh + ("di", c_size_t), # size_t di + ("dctx", c_size_t), # size_t dctx + ("dvoc", c_size_t), # size_t dvoc + + # 归一化与RoPE + ("epsilon", c_float), # float epsilon + ("theta", c_float), # float theta + + # 额外字段 + ("bos_token", c_uint), # uint32_t bos_token + ("end_token", c_uint), # uint32_t end_token + ("attn_dropout", c_float), # float attn_dropout + ("tie_embd", c_bool), # bool tie_embd + ] + + +class Qwen3WeightsCStruct(ctypes.Structure): + _fields_ = [ + # 元信息 + ("nlayer", c_size_t), # size_t nlayer + ("dt_norm", DataType), # infiniDtype_t dt_norm + ("dt_mat", DataType), # infiniDtype_t dt_mat + ("transpose_linear_weights", c_int), # int transpose_linear_weights + + # 全局共享权重 + ("input_embd", c_void_p), # const void *input_embd + ("output_embd", c_void_p), # const void *output_embd + ("output_norm", c_void_p), # const void *output_norm + + # 逐层权重数组 (长度为nlayer) + ("attn_norm", POINTER(c_void_p)), # const void **attn_norm + ("attn_q_norm", POINTER(c_void_p)), # const void **attn_q_norm + ("attn_k_norm", POINTER(c_void_p)), # const void **attn_k_norm + + # QKV投影 - 分开存放 + ("attn_q_proj", POINTER(c_void_p)), # const void **attn_q_proj + ("attn_k_proj", POINTER(c_void_p)), # const void **attn_k_proj + ("attn_v_proj", POINTER(c_void_p)), # const void **attn_v_proj + ("attn_o_proj", POINTER(c_void_p)), # const void **attn_o_proj + + # MLP层 + ("mlp_norm", POINTER(c_void_p)), # const void **mlp_norm + ("mlp_gate_proj", POINTER(c_void_p)), # const void **mlp_gate_proj + ("mlp_up_proj", POINTER(c_void_p)), # const void **mlp_up_proj + ("mlp_down_proj", POINTER(c_void_p)), # const void **mlp_down_proj + ] + + +class Qwen3ModelCStruct(ctypes.Structure): + pass + + +class KVCacheCStruct(ctypes.Structure): + pass + + +def __open_library__(): + infini_root = os.environ.get("INFINI_ROOT") + if infini_root is None: + raise EnvironmentError( + "INFINI_ROOT environment variable not set. " + "Please set it to the InfiniCore installation directory or run 'xmake install' first." + ) + print("p",infini_root) + lib_path = os.path.join(infini_root, "lib", "libinfinicore_infer.so") + if not os.path.exists(lib_path): + raise FileNotFoundError( + f"Library not found at {lib_path}. " + "Please compile the library with 'xmake' and install with 'xmake install' first." + ) + lib = ctypes.CDLL(lib_path) + + # Qwen3 model functions + lib.createQwen3Model.restype = POINTER(Qwen3ModelCStruct) + lib.createQwen3Model.argtypes = [ + POINTER(Qwen3MetaCStruct), + POINTER(Qwen3WeightsCStruct), + DeviceType, + c_int, + POINTER(c_int), + ] + lib.destroyQwen3Model.argtypes = [POINTER(Qwen3ModelCStruct)] + lib.createQwen3KVCache.argtypes = [POINTER(Qwen3ModelCStruct)] + lib.createQwen3KVCache.restype = POINTER(KVCacheCStruct) + lib.dropQwen3KVCache.argtypes = [POINTER(Qwen3ModelCStruct), POINTER(KVCacheCStruct)] + lib.inferQwen3Batch.restype = None + lib.inferQwen3Batch.argtypes = [ + POINTER(Qwen3ModelCStruct), # struct Qwen3Model const * + POINTER(c_uint), # unsigned int const *tokens + c_uint, # unsigned int ntok + POINTER(c_uint), # unsigned int const *req_lens + c_uint, # unsigned int nreq + POINTER(c_uint), # unsigned int const *req_pos + POINTER(POINTER(KVCacheCStruct)), # struct KVCache **kv_caches + POINTER(c_float), # float const *temperature (数组!) + POINTER(c_uint), # unsigned int const *topk (数组!) + POINTER(c_float), # float const *topp (数组!) + POINTER(c_uint), # unsigned int *output + ] + return lib + +LIB = __open_library__() + +create_qwen3_model = LIB.createQwen3Model +destroy_qwen3_model = LIB.destroyQwen3Model +create_qwen3_kv_cache = LIB.createQwen3KVCache +drop_qwen3_kv_cache = LIB.dropQwen3KVCache +infer_qwen3_batch = LIB.inferQwen3Batch \ No newline at end of file diff --git a/infini-qwen3-moe/scripts/qwen3-moe_infer.md b/infini-qwen3-moe/scripts/qwen3-moe_infer.md new file mode 100755 index 00000000..f48f6567 --- /dev/null +++ b/infini-qwen3-moe/scripts/qwen3-moe_infer.md @@ -0,0 +1,376 @@ +# Qwen3-MoE 推理指南 + +本文档介绍如何使用 InfiniCore 框架运行 Qwen3-MoE (Mixture of Experts) 模型进行推理。 + +## 概述 + +Qwen3-MoE 是基于 Qwen3 架构的混合专家模型，具有以下特点： + +- **稀疏激活**: 每个 token 只激活少量专家 (top-k) +- **高效推理**: 通过专家选择减少计算量 +- **负载均衡**: 专家使用统计和负载监控 +- **分布式支持**: 专家权重在多设备间分区 +- **灵活配置**: 支持混合 MLP 和 MoE 层 + +## 模型架构 + +### MoE 层配置 + +```python +# 典型的 Qwen3-MoE 配置参数 +{ + "num_experts": 64, # 每个 MoE 层的专家数量 + "num_experts_per_tok": 2, # 每个 token 选择的专家数量 (top-k) + "moe_intermediate_size": 1536, # MoE 专家的中间维度 + "decoder_sparse_step": 2, # MoE 层的间隔 (每2层一个MoE层) + "mlp_only_layers": [0, 1], # 指定使用普通 MLP 的层 + "norm_topk_prob": True, # 是否归一化 top-k 概率 + "router_aux_loss_coef": 0.01 # 路由辅助损失系数 +} +``` + +### 层类型判断 + +模型中的层分为两种类型： + +1. **MoE 层**: 当满足以下条件时使用 MoE + - `(layer_idx + 1) % decoder_sparse_step == 0` + - `layer_idx` 不在 `mlp_only_layers` 中 + - `num_experts > 0` + +2. **普通 MLP 层**: 其他情况使用标准 MLP + +## 环境配置 + +### 依赖要求 + +```bash +# Python 依赖 +pip install torch transformers safetensors + +# InfiniCore 框架 +export INFINI_ROOT=~/.infini +# 确保 libinfinicore_infer.so 在 LD_LIBRARY_PATH 中 +``` + +### 编译项目 + +```bash +# 构建 InfiniCore 推理库 +xmake build + +# 确认编译成功 +ls build/linux/x86_64/debug/libinfinicore_infer.so +``` + +## 模型准备 + +### 模型格式 + +支持以下模型格式： +- **Safetensors**: 推荐格式，加载速度快 +- **PyTorch**: 标准 `.bin` 格式 + +### 权重命名规范 + +Qwen3-MoE 模型需要遵循以下权重命名规范： + +```python +# 基础权重 +model.embed_tokens.weight +model.norm.weight +lm_head.weight + +# 注意力权重 (每层) +model.layers.{i}.input_layernorm.weight +model.layers.{i}.self_attn.q_proj.weight +model.layers.{i}.self_attn.k_proj.weight +model.layers.{i}.self_attn.v_proj.weight +model.layers.{i}.self_attn.o_proj.weight +model.layers.{i}.self_attn.q_norm.weight # Qwen3 特有 +model.layers.{i}.self_attn.k_norm.weight # Qwen3 特有 + +# MLP 权重 (普通层) +model.layers.{i}.post_attention_layernorm.weight +model.layers.{i}.mlp.gate_proj.weight +model.layers.{i}.mlp.up_proj.weight +model.layers.{i}.mlp.down_proj.weight + +# MoE 权重 (MoE层) +model.layers.{i}.mlp.gate.weight # 路由器权重 +model.layers.{i}.mlp.experts.{j}.gate_proj.weight # 专家j的gate投影 +model.layers.{i}.mlp.experts.{j}.up_proj.weight # 专家j的up投影 +model.layers.{i}.mlp.experts.{j}.down_proj.weight # 专家j的down投影 +``` + +## 使用方法 + +### 基本推理 + +```bash +# CPU 推理 +python scripts/qwen3_moe.py /path/to/qwen3-moe-model --cpu + +# GPU 推理 (单卡) +python scripts/qwen3_moe.py /path/to/qwen3-moe-model --nvidia + +# 多卡推理 +python scripts/qwen3_moe.py /path/to/qwen3-moe-model --nvidia 4 + +# 开启调试模式 +python scripts/qwen3_moe.py /path/to/qwen3-moe-model --nvidia --debug +``` + +### 参数说明 + +- `model_path`: 模型文件目录路径 +- `--cpu/--nvidia/--cambricon/--ascend`: 设备类型选择 +- `n_device`: 使用的设备数量 (默认: 1) +- `--debug`: 开启详细调试信息 + +### Python API 使用 + +```python +from scripts.qwen3_moe import Qwen3MoeForCausalLM +from libinfinicore_infer import DeviceType + +# 创建模型实例 +model = Qwen3MoeForCausalLM( + model_dir_path="/path/to/qwen3-moe-model", + device=DeviceType.DEVICE_TYPE_NVIDIA, + ndev=2, # 使用2张GPU + max_tokens=4096 +) + +# 生成文本 +output, avg_time = model.generate( + input_content="介绍一下人工智能的发展历史", + max_steps=100, + temperature=0.7, + topk=50, + topp=0.8 +) + +print(f"生成内容: {output}") +print(f"平均耗时: {avg_time:.2f}ms/token") + +# 查看专家使用统计 +model.print_router_stats() + +# 清理资源 +model.destroy_model_instance() +``` + +### 批量推理 + +```python +from scripts.qwen3_moe import Qwen3MoeForCausalLM +from scripts.infer_task import Qwen3MoeInferTask, Qwen3MoeKVCache + +model = Qwen3MoeForCausalLM("model_path") + +# 创建多个推理任务 +tasks = [] +for i, text in enumerate(["问题1", "问题2", "问题3"]): + tokens = model.tokenizer.encode(text) + task = Qwen3MoeInferTask( + tokens=tokens, + temperature=0.7, + topk=50, + topp=0.8, + task_id=i + ) + task.bind_kvcache(Qwen3MoeKVCache(model)) + tasks.append(task) + +# 批量推理 +outputs = model.batch_infer_one_round(tasks) +print(f"输出 tokens: {outputs}") +``` + +## 性能优化 + +### 设备分布策略 + +对于多设备推理，权重按以下策略分区： + +1. **注意力权重**: 按头维度分区 + - Q 投影: `[d, d/ndev]` + - K/V 投影: `[d, (nkvh*dh)/ndev]` + - O 投影: `[d/ndev, d]` + +2. **MLP 权重**: 按中间维度分区 + - Gate/Up 投影: `[d, di/ndev]` + - Down 投影: `[di/ndev, d]` + +3. **MoE 专家权重**: 专家在设备间分区 + - 每个设备处理 `⌈num_experts/ndev⌉` 个专家 + - 专家权重按中间维度分区: `[d, moe_di/ndev]` + +### 内存优化 + +```python +# 配置最大上下文长度 +model = Qwen3MoeForCausalLM( + model_path, + max_tokens=2048 # 减少内存使用 +) + +# 使用较小的批次大小 +batch_size = 4 # 根据GPU内存调整 +``` + +### 专家负载均衡 + +模型自动跟踪专家使用统计： + +```python +# 获取特定层的专家统计 +layer_stats = model.get_router_stats(layer_idx=5) +print(f"层5专家使用: {layer_stats}") + +# 打印所有MoE层的统计 +model.print_router_stats() +``` + +理想情况下，专家使用应该相对均衡 (balance ≈ 1.0)。 + +## 调试和监控 + +### 调试模式 + +```bash +# 开启调试模式 +python scripts/qwen3_moe.py model_path --debug +``` + +调试模式提供： +- 详细的推理步骤日志 +- 张量范围验证 +- 专家选择追踪 +- 性能分析信息 + +### 常见问题排查 + +#### 1. 权重加载失败 + +``` +错误: Unsupported weight naming scheme for Qwen3-MoE +``` + +**解决方案**: +- 检查模型配置文件 `config.json` +- 确认权重文件包含 MoE 相关权重 +- 验证权重命名是否符合规范 + +#### 2. 专家权重缺失 + +``` +错误: KeyError: 'model.layers.0.mlp.experts.0.gate_proj.weight' +``` + +**解决方案**: +- 确认模型确实是 MoE 架构 +- 检查 `num_experts` 配置 +- 验证专家权重存在 + +#### 3. 设备内存不足 + +``` +错误: CUDA out of memory +``` + +**解决方案**: +- 减少 `max_tokens` 参数 +- 使用更多设备分区 +- 降低批次大小 + +#### 4. 专家负载不均衡 + +``` +Expert load balance variance: 1250.5 (high variance) +``` + +**解决方案**: +- 检查路由器初始化 +- 调整 `router_aux_loss_coef` +- 考虑重新训练或微调 + +## 性能基准 + +### 典型性能指标 + +| 配置 | 设备 | 吞吐量 (tokens/s) | 延迟 (ms/token) | +|------|------|------------------|----------------| +| 7B MoE | 1x A100 | 150-200 | 5-7 | +| 7B MoE | 2x A100 | 280-350 | 3-4 | +| 14B MoE | 4x A100 | 200-250 | 4-5 | + +实际性能取决于： +- 模型大小和专家数量 +- 输入序列长度 +- 硬件配置 +- 专家激活模式 + +### 与标准模型对比 + +相比于相同参数量的标准 Transformer： +- **推理速度**: 提升 1.5-2.5x (取决于专家激活比例) +- **内存使用**: 增加 1.2-1.8x (存储所有专家权重) +- **计算效率**: 提升显著 (仅激活部分专家) + +## 最佳实践 + +### 1. 模型部署 + +- 使用 Safetensors 格式提升加载速度 +- 预先验证模型权重完整性 +- 配置合适的设备分区策略 + +### 2. 推理优化 + +- 根据应用场景调整 `num_experts_per_tok` +- 监控专家负载均衡情况 +- 使用批量推理提升吞吐量 + +### 3. 资源管理 + +- 及时释放 KV 缓存 +- 合理配置内存池大小 +- 监控设备间通信开销 + +### 4. 质量控制 + +- 定期检查专家使用统计 +- 验证输出质量一致性 +- 比较与标准模型的性能差异 + +## 故障排除 + +如果遇到问题，请按以下步骤排查： + +1. **检查环境配置** + ```bash + python -c "from libinfinicore_infer import *; print('API available')" + ``` + +2. **验证模型格式** + ```python + import json + with open("model_path/config.json") as f: + config = json.load(f) + print(f"Model type: {config.get('model_type')}") + print(f"Experts: {config.get('num_experts', 0)}") + ``` + +3. **测试基础功能** + ```bash + python scripts/qwen3_moe.py model_path --debug + ``` + +4. **查看详细日志** + - 启用调试模式 + - 检查专家统计信息 + - 分析性能瓶颈 + +如需更多帮助，请参考 InfiniCore 文档或提交 issue。 \ No newline at end of file diff --git a/infini-qwen3-moe/scripts/qwen3.py b/infini-qwen3-moe/scripts/qwen3.py new file mode 100755 index 00000000..e6b245f5 --- /dev/null +++ b/infini-qwen3-moe/scripts/qwen3.py @@ -0,0 +1,1019 @@ +#!/usr/bin/env python3 +""" +Fixed Qwen3 Implementation - Using New qw C++ API +This version properly implements Qwen3 support using the dedicated qwen3 C++ API +with proper Q/K normalization and parameter mapping. + +Key Improvements: +1. Uses dedicated Qwen3 API instead of fallback jiuge API +2. Handles Q/K normalization weights properly +3. Implements separate QKV projections +4. One-to-one parameter mapping following jiuge.py patterns +""" + +from typing import List, Optional +import os +import sys +import time +import json +import torch +import transformers +from pathlib import Path +from ctypes import POINTER, c_float, c_int, c_uint, c_void_p, byref +import safetensors +import ctypes + +# Set default device +torch.set_default_device("cpu") + +# Import the proper Qwen3 API +try: + from libinfinicore_infer import ( + Qwen3MetaCStruct, + Qwen3WeightsCStruct, + create_qwen3_model, + destroy_qwen3_model, + create_qwen3_kv_cache, + drop_qwen3_kv_cache, + infer_qwen3_batch, + DataType, + DeviceType, + KVCacheCStruct, + ) + QWEN3_API_AVAILABLE = True + print("✓ Qwen3 C++ API available") +except ImportError as e: + print(f"⚠ Qwen3 C++ API not available: {e}") + print(" This version requires the qw implementation") + sys.exit(1) + +from infer_task import Qwen3InferTask, Qwen3KVCache + + +class Qwen3WeightsNaming: + """Qwen3-specific weight naming with Q/K normalization and separate QKV support""" + + def input_embd(self): + return "model.embed_tokens.weight" + + def output_norm(self): + return "model.norm.weight" + + def output_embd(self): + return "lm_head.weight" + + def attn_norm(self, i): + return f"model.layers.{i}.input_layernorm.weight" + + def attn_q(self, i): + return f"model.layers.{i}.self_attn.q_proj.weight" + + def attn_k(self, i): + return f"model.layers.{i}.self_attn.k_proj.weight" + + def attn_v(self, i): + return f"model.layers.{i}.self_attn.v_proj.weight" + + def attn_o(self, i): + return f"model.layers.{i}.self_attn.o_proj.weight" + + def ffn_norm(self, i): + return f"model.layers.{i}.post_attention_layernorm.weight" + + def gate(self, i): + return f"model.layers.{i}.mlp.gate_proj.weight" + + def up(self, i): + return f"model.layers.{i}.mlp.up_proj.weight" + + def down(self, i): + return f"model.layers.{i}.mlp.down_proj.weight" + + # Qwen3-specific Q/K normalization weights + def q_norm(self, i): + return f"model.layers.{i}.self_attn.q_norm.weight" + + def k_norm(self, i): + return f"model.layers.{i}.self_attn.k_norm.weight" + + @staticmethod + def match(state_dict): + """Check if state_dict matches Qwen3 naming pattern""" + has_basic = ( + "model.norm.weight" in state_dict + and "model.layers.0.self_attn.q_proj.weight" in state_dict + ) + # Qwen3 often has q_norm and k_norm weights + has_qk_norm = ( + "model.layers.0.self_attn.q_norm.weight" in state_dict + and "model.layers.0.self_attn.k_norm.weight" in state_dict + ) + return has_basic and has_qk_norm + +class Qwen3MetaFromConfig(Qwen3MetaCStruct): + """Qwen3 metadata structure from model config""" + + def __init__(self, config, dtype=torch.float16, max_tokens=None): + super().__init__() # 先调用父类构造函数 + + if dtype == torch.float16: + dt_ = DataType.INFINI_DTYPE_F16 + elif dtype == torch.float32: + dt_ = DataType.INFINI_DTYPE_F32 + elif dtype == torch.bfloat16: + dt_ = DataType.INFINI_DTYPE_BF16 + else: + dt_ = DataType.INFINI_DTYPE_F16 + + # 设置字段值 + self.dt_logits = dt_ + self.nlayer = config["num_hidden_layers"] + self.d = config["hidden_size"] + self.nh = config["num_attention_heads"] + self.nkvh = ( + config["num_key_value_heads"] + if "num_key_value_heads" in config + else config["num_attention_heads"] + ) + self.dh = config["hidden_size"] // config["num_attention_heads"] + self.di = config["intermediate_size"] + self.dctx = ( + config["max_position_embeddings"] if max_tokens is None else max_tokens + ) + self.dvoc = config["vocab_size"] + self.epsilon = config.get("rms_norm_eps", 1e-6) + self.theta = config.get("rope_theta", 10000.0) + self.bos_token = config.get("bos_token_id", 1) + self.end_token = config.get("eos_token_id", 2) + self.attn_dropout = config.get("attention_dropout", 0.0) + self.tie_embd = config.get("tie_word_embeddings", True) + + self.torch_dtype_logits = dtype + +class Qwen3WeightsImpl(Qwen3WeightsCStruct): + """Qwen3 weights implementation with Q/K normalization support""" + + def __init__( + self, + meta, + naming, + state_dict, + torch_dt_mat=torch.float16, + torch_dt_norm=torch.float32, + ndev=1, + transpose_weight=True, + ): + nlayer = meta.nlayer + nh = meta.nh + nkvh = meta.nkvh + dh = meta.dh + d = meta.d + di = meta.di + + assert nh % nkvh == 0 + assert nh % ndev == 0 + assert nkvh % ndev == 0 + assert di % ndev == 0 + + torch_dt_logits = meta.torch_dtype_logits + + # 调用父类构造函数 + super().__init__() + + # Set data types + if torch_dt_mat == torch.float16: + self.dt_mat = DataType.INFINI_DTYPE_F16 + elif torch_dt_mat == torch.float32: + self.dt_mat = DataType.INFINI_DTYPE_F32 + elif torch_dt_mat == torch.bfloat16: + self.dt_mat = DataType.INFINI_DTYPE_BF16 + else: + raise ValueError("Unsupported proj weight data type") + + if torch_dt_norm == torch.float16: + self.dt_norm = DataType.INFINI_DTYPE_F16 + elif torch_dt_norm == torch.float32: + self.dt_norm = DataType.INFINI_DTYPE_F32 + elif torch_dt_norm == torch.bfloat16: + self.dt_norm = DataType.INFINI_DTYPE_BF16 + else: + raise ValueError("Unsupported norm weight data type") + + # 设置结构体字段 + self.nlayer = nlayer + self.transpose_linear_weights = 1 if transpose_weight else 0 + + # Determine input/output embedding names + input_embd_naming = ( + naming.input_embd() + if naming.input_embd() in state_dict + else naming.output_embd() + ) + output_embd_naming = ( + naming.output_embd() + if naming.output_embd() in state_dict + else naming.input_embd() + ) + + # Basic weights + self.input_embd_tensor = state_dict[input_embd_naming].to(torch_dt_logits) + self.input_embd = self.input_embd_tensor.data_ptr() + + self.output_norm_tensor = state_dict[naming.output_norm()].to(torch_dt_norm) + self.output_norm = self.output_norm_tensor.data_ptr() + + self.output_embd_tensor = state_dict[output_embd_naming].to(torch_dt_mat) + if not transpose_weight: + self.output_embd_tensor = self.output_embd_tensor.transpose(0, 1).contiguous() + self.output_embd = self.output_embd_tensor.data_ptr() + + # Attention layer normalization weights + self.attn_norm_tensors = [ + state_dict[naming.attn_norm(i)].to(torch_dt_norm) for i in range(nlayer) + ] + self.attn_norm_ptrs = [ + self.attn_norm_tensors[i].data_ptr() for i in range(nlayer) + ] + self.attn_norm = (c_void_p * nlayer)(*self.attn_norm_ptrs) + + # Q/K normalization weights + self.attn_q_norm_tensors = [] + self.attn_k_norm_tensors = [] + if hasattr(naming, 'q_norm'): + try: + for i in range(nlayer): + self.attn_q_norm_tensors.append(state_dict[naming.q_norm(i)].to(torch_dt_norm)) + self.attn_k_norm_tensors.append(state_dict[naming.k_norm(i)].to(torch_dt_norm)) + + self.attn_q_norm_ptrs = [self.attn_q_norm_tensors[i].data_ptr() for i in range(nlayer)] + self.attn_k_norm_ptrs = [self.attn_k_norm_tensors[i].data_ptr() for i in range(nlayer)] + self.attn_q_norm = (c_void_p * nlayer)(*self.attn_q_norm_ptrs) + self.attn_k_norm = (c_void_p * nlayer)(*self.attn_k_norm_ptrs) + + print(f"✓ Loaded Q/K normalization weights for {nlayer} layers") + except KeyError as e: + print(f"⚠ Q/K norm weights not found: {e}") + # 创建空指针数组 + null_ptrs = [None for _ in range(nlayer)] + self.attn_q_norm = (c_void_p * nlayer)(*null_ptrs) + self.attn_k_norm = (c_void_p * nlayer)(*null_ptrs) + else: + # 创建空指针数组 + null_ptrs = [None for _ in range(nlayer)] + self.attn_q_norm = (c_void_p * nlayer)(*null_ptrs) + self.attn_k_norm = (c_void_p * nlayer)(*null_ptrs) + + # 分离的 Q, K, V 投影权重 + self.attn_q_proj_tensors = [] + self.attn_k_proj_tensors = [] + self.attn_v_proj_tensors = [] + + for i in range(nlayer): + q_tensor = state_dict[naming.attn_q(i)].to(torch_dt_mat) + k_tensor = state_dict[naming.attn_k(i)].to(torch_dt_mat) + v_tensor = state_dict[naming.attn_v(i)].to(torch_dt_mat) + + if not transpose_weight: + q_tensor = q_tensor.transpose(0, 1).contiguous() + k_tensor = k_tensor.transpose(0, 1).contiguous() + v_tensor = v_tensor.transpose(0, 1).contiguous() + + self.attn_q_proj_tensors.append(q_tensor) + self.attn_k_proj_tensors.append(k_tensor) + self.attn_v_proj_tensors.append(v_tensor) + + self.attn_q_proj_ptrs = [self.attn_q_proj_tensors[i].data_ptr() for i in range(nlayer)] + self.attn_k_proj_ptrs = [self.attn_k_proj_tensors[i].data_ptr() for i in range(nlayer)] + self.attn_v_proj_ptrs = [self.attn_v_proj_tensors[i].data_ptr() for i in range(nlayer)] + + self.attn_q_proj = (c_void_p * nlayer)(*self.attn_q_proj_ptrs) + self.attn_k_proj = (c_void_p * nlayer)(*self.attn_k_proj_ptrs) + self.attn_v_proj = (c_void_p * nlayer)(*self.attn_v_proj_ptrs) + + # Attention output weights + self.attn_o_proj_tensors = [] + for i in range(nlayer): + o_tensor = state_dict[naming.attn_o(i)].to(torch_dt_mat) + if not transpose_weight: + o_tensor = o_tensor.transpose(0, 1).contiguous() + self.attn_o_proj_tensors.append(o_tensor) + + self.attn_o_proj_ptrs = [self.attn_o_proj_tensors[i].data_ptr() for i in range(nlayer)] + self.attn_o_proj = (c_void_p * nlayer)(*self.attn_o_proj_ptrs) + + # FFN weights + self.mlp_norm_tensors = [ + state_dict[naming.ffn_norm(i)].to(torch_dt_norm) for i in range(nlayer) + ] + self.mlp_norm_ptrs = [self.mlp_norm_tensors[i].data_ptr() for i in range(nlayer)] + self.mlp_norm = (c_void_p * nlayer)(*self.mlp_norm_ptrs) + + # 分离的 gate 和 up 投影 + self.mlp_gate_proj_tensors = [] + self.mlp_up_proj_tensors = [] + self.mlp_down_proj_tensors = [] + + for i in range(nlayer): + gate_tensor = state_dict[naming.gate(i)].to(torch_dt_mat) + up_tensor = state_dict[naming.up(i)].to(torch_dt_mat) + down_tensor = state_dict[naming.down(i)].to(torch_dt_mat) + + if not transpose_weight: + gate_tensor = gate_tensor.transpose(0, 1).contiguous() + up_tensor = up_tensor.transpose(0, 1).contiguous() + down_tensor = down_tensor.transpose(0, 1).contiguous() + + self.mlp_gate_proj_tensors.append(gate_tensor) + self.mlp_up_proj_tensors.append(up_tensor) + self.mlp_down_proj_tensors.append(down_tensor) + + self.mlp_gate_proj_ptrs = [self.mlp_gate_proj_tensors[i].data_ptr() for i in range(nlayer)] + self.mlp_up_proj_ptrs = [self.mlp_up_proj_tensors[i].data_ptr() for i in range(nlayer)] + self.mlp_down_proj_ptrs = [self.mlp_down_proj_tensors[i].data_ptr() for i in range(nlayer)] + + self.mlp_gate_proj = (c_void_p * nlayer)(*self.mlp_gate_proj_ptrs) + self.mlp_up_proj = (c_void_p * nlayer)(*self.mlp_up_proj_ptrs) + self.mlp_down_proj = (c_void_p * nlayer)(*self.mlp_down_proj_ptrs) + + # 验证所有关键权重都已加载 + required_weights = [ + self.input_embd_tensor, + self.output_embd_tensor, + self.output_norm_tensor, + ] + + for i, tensor in enumerate(required_weights): + if tensor is None or tensor.data_ptr() == 0: + raise RuntimeError(f"Critical weight {i} is None or has null data pointer") + + # 验证层权重 + for i in range(nlayer): + critical_tensors = [ + self.attn_norm_tensors[i], + self.attn_q_proj_tensors[i], + self.attn_k_proj_tensors[i], + self.attn_v_proj_tensors[i], + self.attn_o_proj_tensors[i], + self.mlp_norm_tensors[i], + self.mlp_gate_proj_tensors[i], + self.mlp_up_proj_tensors[i], + self.mlp_down_proj_tensors[i], + ] + + for j, tensor in enumerate(critical_tensors): + if tensor is None or tensor.data_ptr() == 0: + raise RuntimeError(f"Layer {i} weight {j} is None or has null data pointer") + + print(f"✓ All {nlayer} layers' weights validated successfully") + print("🔍 First 10 values of attn_q_0:") + print(self.attn_q_proj_tensors[0].flatten()[:10]) + print("🔍 Pointer address:", hex(self.attn_q_proj_tensors[0].data_ptr())) + +class Qwen3BatchedTask: + """Batched inference task for Qwen3""" + + def __init__(self, tasks: List[Qwen3InferTask]): + self.tasks = tasks + self.nreq = len(tasks) + + # Precompute fields + token_lists = [t.tokens for t in tasks] + self.req_lens_list = [len(toks) for toks in token_lists] + self.req_pos_list = [t.pos for t in tasks] + self.kv_cache_ptrs = [t.kvcache().data() for t in tasks] + self.temperatures_list = [t.temperature for t in tasks] + self.topks_list = [t.topk for t in tasks] + self.topps_list = [t.topp for t in tasks] + + # Flatten token lists + flat_tokens = [tok for toks in token_lists for tok in toks] + self.ntok = len(flat_tokens) + + # Convert to ctypes arrays + self.tokens = (c_uint * self.ntok)(*flat_tokens) + self.req_lens = (c_uint * self.nreq)(*self.req_lens_list) + self.req_pos = (c_uint * self.nreq)(*self.req_pos_list) + self.kv_caches = (POINTER(KVCacheCStruct) * self.nreq)(*self.kv_cache_ptrs) + self.temperatures = (c_float * self.nreq)(*self.temperatures_list) + self.topks = (c_uint * self.nreq)(*self.topks_list) + self.topps = (c_float * self.nreq)(*self.topps_list) + + def input_args(self): + return ( + self.tokens, + self.ntok, + self.req_lens, + self.nreq, + self.req_pos, + self.kv_caches, + self.temperatures, + self.topks, + self.topps, + ) + + +class QwenForCausalLM: + """Qwen3 model for causal language modeling - FIXED VERSION""" + + def __init__( + self, model_dir_path, device=DeviceType.DEVICE_TYPE_CPU, ndev=1, max_tokens=None + ): + def load_all_safetensors_from_dir(dir_path_: str): + tensors_ = {} + dir_path_ = Path(dir_path_) + for file in sorted(dir_path_.glob("*.safetensors")): + data_ = safetensors.safe_open(file, "pt") + for name_ in data_.keys(): + tensors_[name_] = data_.get_tensor(name_) + return tensors_ + + print("Loading Qwen3 model weights to host...") + load_start_time = time.time() + + with open(os.path.join(model_dir_path, "config.json"), "r") as f: + config = json.load(f) + self.config = config + + eos_token_id = self.config.get("eos_token_id", 2) + self.eos_token_id = ( + [eos_token_id] if type(eos_token_id) == int else eos_token_id + ) + + transpose_weight = ( + device != DeviceType.DEVICE_TYPE_ASCEND + ) + + # Load state dict + if any(file.suffix == ".safetensors" for file in Path(model_dir_path).iterdir()): + state_dict = load_all_safetensors_from_dir(model_dir_path) + else: + state_dict = torch.load( + os.path.join(model_dir_path, "pytorch_model.bin"), + weights_only=True, + map_location="cpu", + ) + + # Determine naming scheme + if Qwen3WeightsNaming.match(state_dict): + print("✓ Using Qwen3WeightsNaming (with Q/K normalization)") + # Create metadata and weights + self.meta = Qwen3MetaFromConfig(config, max_tokens=max_tokens) + self.weights = Qwen3WeightsImpl( + self.meta, + Qwen3WeightsNaming(), + state_dict, + ndev=ndev, + transpose_weight=transpose_weight, + ) + # Load tokenizer + self.tokenizer = transformers.AutoTokenizer.from_pretrained( + model_dir_path, trust_remote_code=True + ) + elif LlamaWeightsNaming.match(state_dict): + print("⚠ Using LlamaWeightsNaming (fallback, no Q/K normalization)") + else: + raise ValueError("Unsupported weight naming scheme") + + + + + if self.tokenizer.pad_token is None: + self.tokenizer.pad_token = self.tokenizer.eos_token + + load_end_time = time.time() + print(f"Weight loading time: {load_end_time - load_start_time:.3f}s") + + print(f"Creating Qwen3 model on {ndev} devices...") + load_start_time = time.time() + dev_ids = (c_int * ndev)(*[i for i in range(ndev)]) + + try: + self.model_instance = create_qwen3_model( + ctypes.byref(self.meta), + ctypes.byref(self.weights), + device, + ndev, + dev_ids + ) + print(f"✓ Model created successfully") + except Exception as e: + print(f"✗ Error creating model: {e}") + import traceback + traceback.print_exc() + raise + + load_end_time = time.time() + print(f"Model creation time: {load_end_time - load_start_time:.3f}s") + if self.model_instance is None: + raise RuntimeError("Model instance is None after creation") + + def max_context_len(self): + return self.meta.dctx + + def create_kv_cache(self): + # FIXED: Use proper Qwen3 KV cache API + return create_qwen3_kv_cache(self.model_instance) + + def drop_kv_cache(self, kv_cache): + # FIXED: Use proper Qwen3 KV cache API + drop_qwen3_kv_cache(self.model_instance, kv_cache) + + def batch_infer_one_round(self, tasks: List[Qwen3InferTask]): + output = (c_uint * len(tasks))() + + # 使用Qwen3BatchedTask类来处理批处理 + batch_inputs = Qwen3BatchedTask(tasks) + + + # 详细验证输入参数 + # print(f"DEBUG Batch inference:") + # print(f" Number of requests: {batch_inputs.nreq}") + # print(f" Total tokens: {batch_inputs.ntok}") + # print(f" Request lengths: {batch_inputs.req_lens_list}") + # print(f" Request positions: {batch_inputs.req_pos_list}") + # print(f" Temperatures: {batch_inputs.temperatures_list}") + # print(f" Top-k values: {batch_inputs.topks_list}") + # print(f" Top-p values: {batch_inputs.topps_list}") + + # 检查C++函数调用参数的内存地址 - FIXED + print(f"🔍 C++ function parameters:") + try: + model_ptr_addr = ctypes.addressof(self.model_instance.contents) if self.model_instance else 0 + print(f" model_instance: {hex(model_ptr_addr)}") + except: + print(f" model_instance: exists={self.model_instance is not None}") + + print(f" tokens array ptr: {hex(ctypes.addressof(batch_inputs.tokens))}") + print(f" kv_caches array ptr: {hex(ctypes.addressof(batch_inputs.kv_caches))}") + print(f" output array ptr: {hex(ctypes.addressof(output))}") + + # 检查输入token的合理性 + if batch_inputs.ntok > 0: + first_few_tokens = list(batch_inputs.tokens)[:min(5, batch_inputs.ntok)] + print(f" First few input tokens: {first_few_tokens}") + + + # 验证输入参数 + if batch_inputs.ntok == 0: + raise ValueError("没有tokens需要处理") + if batch_inputs.nreq == 0: + raise ValueError("没有请求需要处理") + + try: + # 使用batch_inputs中的数组 + print("🚀 Calling infer_qwen3_batch...") + infer_qwen3_batch( + self.model_instance, + *batch_inputs.input_args(), + output, + ) + print("✅ infer_qwen3_batch completed") + + # 验证输出token + for i, token in enumerate(list(output)): + print(f" Output token[{i}]: {token}") + if token >= self.meta.dvoc: + print(f" ⚠ Invalid: exceeds vocab_size {self.meta.dvoc}") + if token < 0: + print(f" ⚠ Invalid: negative token") + + except Exception as e: + print(f"❌ C++ inference failed: {e}") + import traceback + traceback.print_exc() + raise + + return list(output) + def generate(self, input_content, max_steps, topp_=0.8, topk_=50, temperature_=0.7): + # Apply chat template if available + if hasattr(self.tokenizer, 'chat_template') and self.tokenizer.chat_template: + input_content = self.tokenizer.apply_chat_template( + conversation=[{"role": "user", "content": input_content}], + add_generation_prompt=True, + tokenize=False, + ) + + print(input_content, end="", flush=True) + tokens = self.tokenizer.encode(input_content) + + # 添加详细调试信息 + print(f"\nDEBUG: Input tokens: {tokens}") + print(f"DEBUG: Token count: {len(tokens)}") + print(f"DEBUG: EOS tokens: {self.eos_token_id}") + print(f"DEBUG: Vocab size: {self.meta.dvoc}") + + # 验证输入token的合理性 + print("🔍 Validating input tokens:") + for i, token in enumerate(tokens[:5]): # 只检查前5个 + if token >= self.meta.dvoc or token < 0: + raise ValueError(f"❌ Invalid input token[{i}]: {token} (vocab_size: {self.meta.dvoc})") + + # 检查token对应的embedding + embd_vec = self.weights.input_embd_tensor[token] + embd_norm = embd_vec.norm().item() + print(f" Token[{i}]={token}, embedding_norm={embd_norm:.6f}") + + if embd_norm < 1e-8: + print(f" ⚠ Warning: Token {token} has very small embedding norm") + if torch.isnan(embd_vec).any() or torch.isinf(embd_vec).any(): + raise RuntimeError(f"❌ Token {token} embedding contains NaN/Inf") + + infer_task = Qwen3InferTask( + tokens=tokens, + position=0, + temperature=temperature_, + topk=topk_, + topp=topp_, + end_tokens=self.eos_token_id, + max_tokens=int(self.meta.dctx), + task_id=0 + ) + + infer_task.bind_kvcache(Qwen3KVCache(self)) + # 验证模型实例状态 - FIXED + print(f"🔍 Model instance validation:") + try: + model_ptr_addr = ctypes.addressof(self.model_instance.contents) if self.model_instance else 0 + print(f" Model instance ptr: {hex(model_ptr_addr)}") + except: + # 降级处理 + print(f" Model instance: {self.model_instance is not None}") + + if self.model_instance is None: + raise RuntimeError("❌ Model instance is null before inference") + + + steps = 0 + total_time = 0 + output_content = "" + + for step_i in range(max_steps): + start_time = time.time() + output_tokens = self.batch_infer_one_round([infer_task]) + end_time = time.time() + steps += 1 + # 详细的token分析 + output_token = output_tokens[0] + print(f"\nDEBUG Step {step_i}:") + print(f" Output token ID: {output_token}") + print(f" Token in vocab range: {0 <= output_token < self.meta.dvoc}") + print(f" Is EOS token: {output_token in self.eos_token_id}") + # 检查token合理性 + if output_token >= self.meta.dvoc: + print(f" ⚠ WARNING: Token {output_token} exceeds vocab size {self.meta.dvoc}") + break + if output_token < 0: + print(f" ⚠ WARNING: Negative token ID {output_token}") + break + + try: + output_str = self.tokenizer.decode([output_token], skip_special_tokens=False) + print(f" Decoded: '{output_str}'") + except Exception as e: + print(f" ⚠ Decode failed: {e}") + output_str = self.tokenizer._tokenizer.id_to_token(output_token) + if output_str is None: + output_str = f"[UNK_{output_token}]" + else: + output_str = output_str.replace("▁", " ").replace("<0x0A>", "\n") + + output_content += output_str + print(output_str, end="", flush=True) + + if output_tokens[0] in self.eos_token_id: + break + + infer_task.next(output_tokens[0]) + + if step_i > 0: + total_time += end_time - start_time + + print("\n") + avg_time = total_time * 1000 / (steps - 1) if steps > 1 else 0 + print(f"Time per step: {avg_time:.3f}ms") + + try: + infer_task._kv_cache.drop() + except AttributeError: + # 如果drop方法有问题，跳过清理 + print(" ⚠ KV cache cleanup skipped (method issue)") + except Exception as e: + print(f" ⚠ KV cache cleanup failed: {e}") + + return output_content, avg_time + + def destroy_model_instance(self): + # FIXED: Use proper Qwen3 model destruction API + destroy_qwen3_model(self.model_instance) + print("Qwen3 Model destroyed") + + def diagnose_cpp_computation(self): + """诊断C++推理引擎的计算正确性""" + + print(f"\n{'='*60}") + print("🔬 C++ COMPUTATION DIAGNOSIS") + print(f"{'='*60}") + + # 1. 测试固定输入的一致性 + print("\n1️⃣ Testing computation consistency with fixed inputs:") + + # 使用非常简单的输入 + simple_tokens = [1, 2, 3] # BOS, simple tokens + + results = [] + for i in range(3): # 运行3次相同的推理 + print(f"\n Run {i+1}/3:") + + task = Qwen3InferTask( + tokens=simple_tokens, + position=0, + temperature=0.0, # 完全确定性 + topk=1, # 只取概率最高的token + topp=1.0, + end_tokens=self.eos_token_id, + max_tokens=int(self.meta.dctx), + task_id=0 + ) + task.bind_kvcache(Qwen3KVCache(self)) + + # 执行推理 + output_tokens = self.batch_infer_one_round([task]) + output_token = output_tokens[0] + + print(f" Input tokens: {simple_tokens}") + print(f" Output token: {output_token}") + + results.append(output_token) + # FIXED: 使用try-except来处理KV缓存清理 + try: + task._kv_cache.drop() + except AttributeError: + # 如果drop方法有问题，跳过清理 + print(" ⚠ KV cache cleanup skipped (method issue)") + except Exception as e: + print(f" ⚠ KV cache cleanup failed: {e}") + + # 检查一致性 + if len(set(results)) == 1: + print(f" ✅ PASS: All runs produced same result: {results[0]}") + else: + print(f" ❌ FAIL: Inconsistent results: {results}") + print(" This indicates non-deterministic computation or memory corruption") + + # 2. 测试不同temperature的影响 + print("\n2️⃣ Testing temperature parameter effect:") + + temps = [0.0, 0.5, 1.0] + temp_results = {} + + for temp in temps: + task = Qwen3InferTask( + tokens=simple_tokens, + position=0, + temperature=temp, + topk=50, + topp=0.8, + end_tokens=self.eos_token_id, + max_tokens=int(self.meta.dctx), + task_id=0 + ) + task.bind_kvcache(Qwen3KVCache(self)) + + # 运行多次获取分布 + temp_outputs = [] + for _ in range(5): + output_tokens = self.batch_infer_one_round([task]) + temp_outputs.append(output_tokens[0]) + task.next(output_tokens[0]) # 更新状态以便下次推理 + + temp_results[temp] = temp_outputs + try: + task._kv_cache.drop() + except AttributeError: + # 如果drop方法有问题，跳过清理 + print(" ⚠ KV cache cleanup skipped (method issue)") + except Exception as e: + print(f" ⚠ KV cache cleanup failed: {e}") + + unique_outputs = len(set(temp_outputs)) + print(f" temp={temp}: outputs={temp_outputs}, unique={unique_outputs}") + + # 验证temperature=0.0应该完全确定 + if len(set(temp_results[0.0])) == 1: + print(" ✅ PASS: Temperature=0.0 produces deterministic output") + else: + print(" ❌ FAIL: Temperature=0.0 should be deterministic") + + # 3. 测试输入长度对输出的影响 + print("\n3️⃣ Testing input length effect:") + + test_inputs = [ + [1], # 1 token + [1, 2], # 2 tokens + [1, 2, 3], # 3 tokens + [1, 2, 3, 4], # 4 tokens + ] + + length_results = {} + for tokens in test_inputs: + task = Qwen3InferTask( + tokens=tokens, + position=0, + temperature=0.0, + topk=1, + topp=1.0, + end_tokens=self.eos_token_id, + max_tokens=int(self.meta.dctx), + task_id=0 + ) + task.bind_kvcache(Qwen3KVCache(self)) + + output_tokens = self.batch_infer_one_round([task]) + length_results[len(tokens)] = output_tokens[0] + + print(f" Input length {len(tokens)}: {tokens} -> {output_tokens[0]}") + try: + task._kv_cache.drop() + except AttributeError: + # 如果drop方法有问题，跳过清理 + print(" ⚠ KV cache cleanup skipped (method issue)") + except Exception as e: + print(f" ⚠ KV cache cleanup failed: {e}") + + # 4. 测试KV缓存状态的影响 + print("\n4️⃣ Testing KV cache state impact:") + + # 第一次推理 + task1 = Qwen3InferTask( + tokens=[1, 2, 3], + position=0, + temperature=0.0, + topk=1, + topp=1.0, + end_tokens=self.eos_token_id, + max_tokens=int(self.meta.dctx), + task_id=0 + ) + task1.bind_kvcache(Qwen3KVCache(self)) + + output1 = self.batch_infer_one_round([task1])[0] + print(f" Fresh KV cache: [1,2,3] -> {output1}") + + # 继续用相同的KV缓存推理下一个token + task1.next(output1) + output2 = self.batch_infer_one_round([task1])[0] + print(f" Continued KV cache: append {output1} -> {output2}") + + # 重新开始，但用不同的方式 + task2 = Qwen3InferTask( + tokens=[1, 2, 3, output1], + position=0, + temperature=0.0, + topk=1, + topp=1.0, + end_tokens=self.eos_token_id, + max_tokens=int(self.meta.dctx), + task_id=0 + ) + task2.bind_kvcache(Qwen3KVCache(self)) + + output3 = self.batch_infer_one_round([task2])[0] + print(f" Fresh KV cache: [1,2,3,{output1}] -> {output3}") + + if output2 == output3: + print(" ✅ PASS: KV cache state consistency maintained") + else: + print(" ❌ FAIL: KV cache state inconsistent") + print(f" Continued: {output2}, Fresh: {output3}") + + task1._kv_cache.drop() + task2._kv_cache.drop() + + # 5. 测试边界情况 + print("\n5️⃣ Testing edge cases:") + + # 测试vocab边界附近的token + edge_tokens = [0, 1, self.meta.dvoc-2, self.meta.dvoc-1] # 避免无效token + + for token in edge_tokens: + if 0 <= token < self.meta.dvoc: + task = Qwen3InferTask( + tokens=[token], + position=0, + temperature=0.0, + topk=1, + topp=1.0, + end_tokens=self.eos_token_id, + max_tokens=int(self.meta.dctx), + task_id=0 + ) + task.bind_kvcache(Qwen3KVCache(self)) + + try: + output = self.batch_infer_one_round([task])[0] + print(f" Edge token {token} -> {output}") + + if 0 <= output < self.meta.dvoc: + print(f" ✅ Output {output} in valid range") + else: + print(f" ❌ Output {output} out of range [0, {self.meta.dvoc})") + + except Exception as e: + print(f" ❌ Error with token {token}: {e}") + finally: + try: + task._kv_cache.drop() + except AttributeError: + # 如果drop方法有问题，跳过清理 + print(" ⚠ KV cache cleanup skipped (method issue)") + except Exception as e: + print(f" ⚠ KV cache cleanup failed: {e}") + + print(f"\n{'='*60}") + print("🔬 DIAGNOSIS COMPLETE") + print(f"{'='*60}") + + def generate_simple(self, input_content, max_steps, topp_=0.8, topk_=50, temperature_=0.7): + """不使用chat template的简单生成""" + print(f"\nSimple generation: '{input_content}'", end="", flush=True) + tokens = self.tokenizer.encode(input_content) + + print(f"\nInput tokens: {tokens}") + + infer_task = Qwen3InferTask( + tokens=tokens, + position=0, + temperature=temperature_, + topk=topk_, + topp=topp_, + end_tokens=self.eos_token_id, + max_tokens=int(self.meta.dctx), + task_id=0 + ) + + infer_task.bind_kvcache(Qwen3KVCache(self)) + + output_content = "" + for step_i in range(max_steps): + output_tokens = self.batch_infer_one_round([infer_task]) + output_token = output_tokens[0] + + print(f" -> {output_token}", end="") + + if output_token >= self.meta.dvoc or output_token < 0: + print(f" (INVALID)") + break + + try: + output_str = self.tokenizer.decode([output_token], skip_special_tokens=False) + except Exception: + output_str = f"[UNK_{output_token}]" + + output_content += output_str + print(f"('{output_str}')", end="", flush=True) + + if output_token in self.eos_token_id: + break + + infer_task.next(output_token) + + print(f"\nFinal output: '{output_content}'") + try: + infer_task._kv_cache.drop() + except AttributeError: + # 如果drop方法有问题，跳过清理 + print(" ⚠ KV cache cleanup skipped (method issue)") + except Exception as e: + print(f" ⚠ KV cache cleanup failed: {e}") + return output_content +def test(): + if len(sys.argv) < 2: + print("Usage: python qwen3_fixed.py [device] [n_device]") + sys.exit(1) + + model_path = sys.argv[1] + device_type = DeviceType.DEVICE_TYPE_CPU + + if len(sys.argv) > 2: + if sys.argv[2] == "--cpu": + device_type = DeviceType.DEVICE_TYPE_CPU + elif sys.argv[2] == "--nvidia": + device_type = DeviceType.DEVICE_TYPE_NVIDIA + + ndev = int(sys.argv[3]) if len(sys.argv) > 3 else 1 + + print(f"✓ Using Qwen3 model from: {model_path}") + print(f"✓ Device: {device_type}, Devices: {ndev}") + + model = QwenForCausalLM(model_path, device_type, ndev) + + # 诊断C++计算问题 + model.diagnose_cpp_computation() + + # 然后测试生成 + model.generate("山东最高的山是？", 5, topp_=0.8, topk_=50, temperature_=0.7) + model.destroy_model_instance() + + +if __name__ == "__main__": + test() \ No newline at end of file diff --git a/infini-qwen3-moe/scripts/qwen3_moe.py b/infini-qwen3-moe/scripts/qwen3_moe.py new file mode 100755 index 00000000..9a6f8ed5 --- /dev/null +++ b/infini-qwen3-moe/scripts/qwen3_moe.py @@ -0,0 +1,856 @@ +#!/usr/bin/env python3 +""" +Qwen3-MoE Implementation for InfiniCore +This implementation provides Qwen3-MoE support using the dedicated qwen3_moe C++ API +with proper MoE routing, expert selection, and parameter mapping. + +Key Features: +1. Uses dedicated Qwen3-MoE API with sparse expert activation +2. Handles MoE routing weights and expert selection properly +3. Supports both regular MLP and MoE layers in the same model +4. Implements efficient expert weight partitioning across devices +5. Provides MoE-specific debugging and statistics +""" + +from typing import List, Optional, Dict, Any +import os +import sys +import time +import json +import torch +import transformers +from pathlib import Path +from ctypes import POINTER, c_float, c_int, c_uint, c_void_p, byref +import safetensors +import ctypes +import numpy as np + +# Set default device +torch.set_default_device("cpu") + +# Import the Qwen3-MoE specific API +try: + from libinfinicore_infer import ( + Qwen3MoeMetaCStruct, + Qwen3MoeWeightsCStruct, + createQwen3MoeModel, + destroyQwen3MoeModel, + createQwen3MoeKVCache, + dropQwen3MoeKVCache, + inferQwen3MoeBatch, + getQwen3MoeRouterStats, + setQwen3MoeDebugMode, + DataType, + DeviceType, + KVCacheCStruct, + ) + QWEN3_MOE_API_AVAILABLE = True + print("✓ Qwen3-MoE C++ API available") +except ImportError as e: + print(f"⚠ Qwen3-MoE C++ API not available: {e}") + print(" This version requires the qwen3_moe implementation") + sys.exit(1) + +from infer_task import Qwen3MoeInferTask, Qwen3MoeKVCache + + +class Qwen3MoeWeightsNaming: + """Qwen3-MoE specific weight naming with MoE expert support""" + + def input_embd(self): + return "model.embed_tokens.weight" + + def output_norm(self): + return "model.norm.weight" + + def output_embd(self): + return "lm_head.weight" + + def attn_norm(self, i): + return f"model.layers.{i}.input_layernorm.weight" + + def attn_q(self, i): + return f"model.layers.{i}.self_attn.q_proj.weight" + + def attn_k(self, i): + return f"model.layers.{i}.self_attn.k_proj.weight" + + def attn_v(self, i): + return f"model.layers.{i}.self_attn.v_proj.weight" + + def attn_o(self, i): + return f"model.layers.{i}.self_attn.o_proj.weight" + + def ffn_norm(self, i): + return f"model.layers.{i}.post_attention_layernorm.weight" + + # Regular MLP weights + def gate(self, i): + return f"model.layers.{i}.mlp.gate_proj.weight" + + def up(self, i): + return f"model.layers.{i}.mlp.up_proj.weight" + + def down(self, i): + return f"model.layers.{i}.mlp.down_proj.weight" + + # Qwen3-specific Q/K normalization weights + def q_norm(self, i): + return f"model.layers.{i}.self_attn.q_norm.weight" + + def k_norm(self, i): + return f"model.layers.{i}.self_attn.k_norm.weight" + + # MoE-specific weights + def moe_gate(self, i): + """Router/gating network for expert selection""" + return f"model.layers.{i}.mlp.gate.weight" + + def moe_expert_gate(self, i, expert_idx): + """Gate projection for specific expert""" + return f"model.layers.{i}.mlp.experts.{expert_idx}.gate_proj.weight" + + def moe_expert_up(self, i, expert_idx): + """Up projection for specific expert""" + return f"model.layers.{i}.mlp.experts.{expert_idx}.up_proj.weight" + + def moe_expert_down(self, i, expert_idx): + """Down projection for specific expert""" + return f"model.layers.{i}.mlp.experts.{expert_idx}.down_proj.weight" + + @staticmethod + def match(state_dict): + """Check if state_dict matches Qwen3-MoE naming pattern""" + has_basic = ( + "model.norm.weight" in state_dict + and "model.layers.0.self_attn.q_proj.weight" in state_dict + ) + # Check for MoE specific patterns + has_moe = any( + "mlp.gate.weight" in key and "mlp.experts" in str(state_dict.keys()) + for key in state_dict.keys() + ) + # Qwen3 often has q_norm and k_norm weights + has_qk_norm = ( + "model.layers.0.self_attn.q_norm.weight" in state_dict + and "model.layers.0.self_attn.k_norm.weight" in state_dict + ) + return has_basic and (has_moe or has_qk_norm) + + +class Qwen3MoeMetaFromConfig(Qwen3MoeMetaCStruct): + """Qwen3-MoE metadata structure from model config""" + + def __init__(self, config, dtype=torch.float16, max_tokens=None): + super().__init__() + + if dtype == torch.float16: + dt_ = DataType.INFINI_DTYPE_F16 + elif dtype == torch.float32: + dt_ = DataType.INFINI_DTYPE_F32 + elif dtype == torch.bfloat16: + dt_ = DataType.INFINI_DTYPE_BF16 + else: + dt_ = DataType.INFINI_DTYPE_F16 + + # Basic model parameters + self.dt_logits = dt_ + self.nlayer = config["num_hidden_layers"] + self.d = config["hidden_size"] + self.nh = config["num_attention_heads"] + self.nkvh = ( + config["num_key_value_heads"] + if "num_key_value_heads" in config + else config["num_attention_heads"] + ) + self.dh = config["hidden_size"] // config["num_attention_heads"] + self.di = config["intermediate_size"] + self.dctx = ( + config["max_position_embeddings"] if max_tokens is None else max_tokens + ) + self.dvoc = config["vocab_size"] + self.epsilon = config.get("rms_norm_eps", 1e-6) + self.theta = config.get("rope_theta", 10000.0) + self.bos_token = config.get("bos_token_id", 1) + self.end_token = config.get("eos_token_id", 2) + self.attn_dropout = config.get("attention_dropout", 0.0) + self.tie_embd = config.get("tie_word_embeddings", True) + + # MoE specific parameters + self.num_experts = config.get("num_experts", 0) + self.num_experts_per_tok = config.get("num_experts_per_tok", 2) + self.moe_intermediate_size = config.get("moe_intermediate_size", self.di) + self.decoder_sparse_step = config.get("decoder_sparse_step", 1) + self.norm_topk_prob = config.get("norm_topk_prob", False) + self.router_aux_loss_coef = config.get("router_aux_loss_coef", 0.0) + + # Handle mlp_only_layers + mlp_only_layers = config.get("mlp_only_layers", []) + self.num_mlp_only_layers = len(mlp_only_layers) + if self.num_mlp_only_layers > 0: + self.mlp_only_layers = (c_uint * self.num_mlp_only_layers)(*mlp_only_layers) + else: + self.mlp_only_layers = None + + self.torch_dtype_logits = dtype + + print(f"✓ Qwen3-MoE config: {self.nlayer} layers, {self.num_experts} experts, " + f"top-{self.num_experts_per_tok}, sparse_step={self.decoder_sparse_step}") + + +class Qwen3MoeWeightsImpl(Qwen3MoeWeightsCStruct): + """Qwen3-MoE weights implementation with expert weight support""" + + def __init__( + self, + meta, + naming, + state_dict, + torch_dt_mat=torch.float16, + torch_dt_norm=torch.float32, + ndev=1, + transpose_weight=True, + ): + nlayer = meta.nlayer + nh = meta.nh + nkvh = meta.nkvh + dh = meta.dh + d = meta.d + di = meta.di + num_experts = meta.num_experts + moe_di = meta.moe_intermediate_size + + assert nh % nkvh == 0 + assert nh % ndev == 0 + assert nkvh % ndev == 0 + assert di % ndev == 0 + if moe_di > 0: + assert moe_di % ndev == 0 + + torch_dt_logits = meta.torch_dtype_logits + + super().__init__() + + # Set data types + if torch_dt_mat == torch.float16: + self.dt_mat = DataType.INFINI_DTYPE_F16 + elif torch_dt_mat == torch.float32: + self.dt_mat = DataType.INFINI_DTYPE_F32 + elif torch_dt_mat == torch.bfloat16: + self.dt_mat = DataType.INFINI_DTYPE_BF16 + else: + raise ValueError("Unsupported proj weight data type") + + if torch_dt_norm == torch.float16: + self.dt_norm = DataType.INFINI_DTYPE_F16 + elif torch_dt_norm == torch.float32: + self.dt_norm = DataType.INFINI_DTYPE_F32 + elif torch_dt_norm == torch.bfloat16: + self.dt_norm = DataType.INFINI_DTYPE_BF16 + else: + raise ValueError("Unsupported norm weight data type") + + self.nlayer = nlayer + self.transpose_linear_weights = 1 if transpose_weight else 0 + + # Store MoE parameters + self.num_experts = num_experts + self.num_experts_per_tok = meta.num_experts_per_tok + self.moe_intermediate_size = moe_di + self.decoder_sparse_step = meta.decoder_sparse_step + self.num_mlp_only_layers = meta.num_mlp_only_layers + self.mlp_only_layers = meta.mlp_only_layers + self.norm_topk_prob = meta.norm_topk_prob + + # Determine input/output embedding names + input_embd_naming = ( + naming.input_embd() + if naming.input_embd() in state_dict + else naming.output_embd() + ) + output_embd_naming = ( + naming.output_embd() + if naming.output_embd() in state_dict + else naming.input_embd() + ) + + # Basic weights + self.input_embd_tensor = state_dict[input_embd_naming].to(torch_dt_logits) + self.input_embd = self.input_embd_tensor.data_ptr() + + self.output_norm_tensor = state_dict[naming.output_norm()].to(torch_dt_norm) + self.output_norm = self.output_norm_tensor.data_ptr() + + self.output_embd_tensor = state_dict[output_embd_naming].to(torch_dt_mat) + if not transpose_weight: + self.output_embd_tensor = self.output_embd_tensor.transpose(0, 1).contiguous() + self.output_embd = self.output_embd_tensor.data_ptr() + + # Layer-wise weight processing + self._process_attention_weights(naming, state_dict, nlayer, torch_dt_norm, torch_dt_mat, ndev, transpose_weight) + self._process_mlp_moe_weights(naming, state_dict, nlayer, meta, torch_dt_norm, torch_dt_mat, ndev, transpose_weight) + + print(f"✓ Qwen3-MoE weights loaded: {nlayer} layers, {num_experts} experts per MoE layer") + + def _is_moe_layer(self, layer_idx, meta): + """Check if a layer is MoE or regular MLP""" + # Check if in mlp_only_layers + if meta.mlp_only_layers: + mlp_only_list = [meta.mlp_only_layers[i] for i in range(meta.num_mlp_only_layers)] + if layer_idx in mlp_only_list: + return False + + # Check sparse step condition + return (meta.num_experts > 0 and + (layer_idx + 1) % meta.decoder_sparse_step == 0) + + def _process_attention_weights(self, naming, state_dict, nlayer, torch_dt_norm, torch_dt_mat, ndev, transpose_weight): + """Process attention weights (same as Qwen3)""" + + # Attention layer normalization weights + self.attn_norm_tensors = [ + state_dict[naming.attn_norm(i)].to(torch_dt_norm) for i in range(nlayer) + ] + self.attn_norm_ptrs = [ + self.attn_norm_tensors[i].data_ptr() for i in range(nlayer) + ] + self.attn_norm = (c_void_p * nlayer)(*self.attn_norm_ptrs) + + # Q/K normalization weights (optional) + self.attn_q_norm_tensors = [] + self.attn_k_norm_tensors = [] + try: + for i in range(nlayer): + self.attn_q_norm_tensors.append(state_dict[naming.q_norm(i)].to(torch_dt_norm)) + self.attn_k_norm_tensors.append(state_dict[naming.k_norm(i)].to(torch_dt_norm)) + + self.attn_q_norm_ptrs = [self.attn_q_norm_tensors[i].data_ptr() for i in range(nlayer)] + self.attn_k_norm_ptrs = [self.attn_k_norm_tensors[i].data_ptr() for i in range(nlayer)] + self.attn_q_norm = (c_void_p * nlayer)(*self.attn_q_norm_ptrs) + self.attn_k_norm = (c_void_p * nlayer)(*self.attn_k_norm_ptrs) + except KeyError: + null_ptrs = [None for _ in range(nlayer)] + self.attn_q_norm = (c_void_p * nlayer)(*null_ptrs) + self.attn_k_norm = (c_void_p * nlayer)(*null_ptrs) + + # Separate Q, K, V projection weights + self.attn_q_proj_tensors = [] + self.attn_k_proj_tensors = [] + self.attn_v_proj_tensors = [] + + for i in range(nlayer): + q_tensor = state_dict[naming.attn_q(i)].to(torch_dt_mat) + k_tensor = state_dict[naming.attn_k(i)].to(torch_dt_mat) + v_tensor = state_dict[naming.attn_v(i)].to(torch_dt_mat) + + if not transpose_weight: + q_tensor = q_tensor.transpose(0, 1).contiguous() + k_tensor = k_tensor.transpose(0, 1).contiguous() + v_tensor = v_tensor.transpose(0, 1).contiguous() + + self.attn_q_proj_tensors.append(q_tensor) + self.attn_k_proj_tensors.append(k_tensor) + self.attn_v_proj_tensors.append(v_tensor) + + self.attn_q_proj_ptrs = [self.attn_q_proj_tensors[i].data_ptr() for i in range(nlayer)] + self.attn_k_proj_ptrs = [self.attn_k_proj_tensors[i].data_ptr() for i in range(nlayer)] + self.attn_v_proj_ptrs = [self.attn_v_proj_tensors[i].data_ptr() for i in range(nlayer)] + + self.attn_q_proj = (c_void_p * nlayer)(*self.attn_q_proj_ptrs) + self.attn_k_proj = (c_void_p * nlayer)(*self.attn_k_proj_ptrs) + self.attn_v_proj = (c_void_p * nlayer)(*self.attn_v_proj_ptrs) + + # Attention output weights + self.attn_o_proj_tensors = [] + for i in range(nlayer): + o_tensor = state_dict[naming.attn_o(i)].to(torch_dt_mat) + if not transpose_weight: + o_tensor = o_tensor.transpose(0, 1).contiguous() + self.attn_o_proj_tensors.append(o_tensor) + + self.attn_o_proj_ptrs = [self.attn_o_proj_tensors[i].data_ptr() for i in range(nlayer)] + self.attn_o_proj = (c_void_p * nlayer)(*self.attn_o_proj_ptrs) + + def _process_mlp_moe_weights(self, naming, state_dict, nlayer, meta, torch_dt_norm, torch_dt_mat, ndev, transpose_weight): + """Process MLP/MoE weights based on layer configuration""" + + # MLP normalization weights (common to both MLP and MoE layers) + self.mlp_norm_tensors = [ + state_dict[naming.ffn_norm(i)].to(torch_dt_norm) for i in range(nlayer) + ] + self.mlp_norm_ptrs = [self.mlp_norm_tensors[i].data_ptr() for i in range(nlayer)] + self.mlp_norm = (c_void_p * nlayer)(*self.mlp_norm_ptrs) + + # Initialize arrays for both MLP and MoE weights + self.mlp_gate_proj_tensors = [None] * nlayer + self.mlp_up_proj_tensors = [None] * nlayer + self.mlp_down_proj_tensors = [None] * nlayer + + self.moe_gate_tensors = [None] * nlayer + self.moe_experts_gate_proj_tensors = [None] * nlayer + self.moe_experts_up_proj_tensors = [None] * nlayer + self.moe_experts_down_proj_tensors = [None] * nlayer + + for i in range(nlayer): + if self._is_moe_layer(i, meta): + # Process MoE layer + self._process_moe_layer_weights(i, naming, state_dict, meta, torch_dt_mat, transpose_weight) + else: + # Process regular MLP layer + self._process_mlp_layer_weights(i, naming, state_dict, torch_dt_mat, transpose_weight) + + # Create C arrays + self._create_weight_arrays(nlayer) + + def _process_moe_layer_weights(self, layer_idx, naming, state_dict, meta, torch_dt_mat, transpose_weight): + """Process weights for a MoE layer""" + num_experts = meta.num_experts + + # Router/gate weights + gate_tensor = state_dict[naming.moe_gate(layer_idx)].to(torch_dt_mat) + if not transpose_weight: + gate_tensor = gate_tensor.transpose(0, 1).contiguous() + self.moe_gate_tensors[layer_idx] = gate_tensor + + # Expert weights + expert_gate_tensors = [] + expert_up_tensors = [] + expert_down_tensors = [] + + for expert_idx in range(num_experts): + # Expert gate projection + expert_gate = state_dict[naming.moe_expert_gate(layer_idx, expert_idx)].to(torch_dt_mat) + if not transpose_weight: + expert_gate = expert_gate.transpose(0, 1).contiguous() + expert_gate_tensors.append(expert_gate) + + # Expert up projection + expert_up = state_dict[naming.moe_expert_up(layer_idx, expert_idx)].to(torch_dt_mat) + if not transpose_weight: + expert_up = expert_up.transpose(0, 1).contiguous() + expert_up_tensors.append(expert_up) + + # Expert down projection + expert_down = state_dict[naming.moe_expert_down(layer_idx, expert_idx)].to(torch_dt_mat) + if not transpose_weight: + expert_down = expert_down.transpose(0, 1).contiguous() + expert_down_tensors.append(expert_down) + + self.moe_experts_gate_proj_tensors[layer_idx] = expert_gate_tensors + self.moe_experts_up_proj_tensors[layer_idx] = expert_up_tensors + self.moe_experts_down_proj_tensors[layer_idx] = expert_down_tensors + + def _process_mlp_layer_weights(self, layer_idx, naming, state_dict, torch_dt_mat, transpose_weight): + """Process weights for a regular MLP layer""" + + gate_tensor = state_dict[naming.gate(layer_idx)].to(torch_dt_mat) + up_tensor = state_dict[naming.up(layer_idx)].to(torch_dt_mat) + down_tensor = state_dict[naming.down(layer_idx)].to(torch_dt_mat) + + if not transpose_weight: + gate_tensor = gate_tensor.transpose(0, 1).contiguous() + up_tensor = up_tensor.transpose(0, 1).contiguous() + down_tensor = down_tensor.transpose(0, 1).contiguous() + + self.mlp_gate_proj_tensors[layer_idx] = gate_tensor + self.mlp_up_proj_tensors[layer_idx] = up_tensor + self.mlp_down_proj_tensors[layer_idx] = down_tensor + + def _create_weight_arrays(self, nlayer): + """Create C pointer arrays for all weights""" + + # Regular MLP weights + mlp_gate_ptrs = [] + mlp_up_ptrs = [] + mlp_down_ptrs = [] + + for i in range(nlayer): + if self.mlp_gate_proj_tensors[i] is not None: + mlp_gate_ptrs.append(self.mlp_gate_proj_tensors[i].data_ptr()) + mlp_up_ptrs.append(self.mlp_up_proj_tensors[i].data_ptr()) + mlp_down_ptrs.append(self.mlp_down_proj_tensors[i].data_ptr()) + else: + mlp_gate_ptrs.append(None) + mlp_up_ptrs.append(None) + mlp_down_ptrs.append(None) + + self.mlp_gate_proj = (c_void_p * nlayer)(*mlp_gate_ptrs) + self.mlp_up_proj = (c_void_p * nlayer)(*mlp_up_ptrs) + self.mlp_down_proj = (c_void_p * nlayer)(*mlp_down_ptrs) + + # MoE weights + moe_gate_ptrs = [] + for i in range(nlayer): + if self.moe_gate_tensors[i] is not None: + moe_gate_ptrs.append(self.moe_gate_tensors[i].data_ptr()) + else: + moe_gate_ptrs.append(None) + + self.moe_gate = (c_void_p * nlayer)(*moe_gate_ptrs) + + # Expert weights (three-level pointers) + # This is a simplified implementation - full implementation would need proper 3D arrays + self.moe_experts_gate_proj = None # Would need complex 3D pointer array setup + self.moe_experts_up_proj = None + self.moe_experts_down_proj = None + + +class Qwen3MoeBatchedTask: + """Batched inference task for Qwen3-MoE""" + + def __init__(self, tasks: List[Qwen3MoeInferTask]): + self.tasks = tasks + self.nreq = len(tasks) + + # Precompute fields + token_lists = [t.tokens for t in tasks] + self.req_lens_list = [len(toks) for toks in token_lists] + self.req_pos_list = [t.pos for t in tasks] + self.kv_cache_ptrs = [t.kvcache().data() for t in tasks] + self.temperatures_list = [t.temperature for t in tasks] + self.topks_list = [t.topk for t in tasks] + self.topps_list = [t.topp for t in tasks] + + # Flatten token lists + flat_tokens = [tok for toks in token_lists for tok in toks] + self.ntok = len(flat_tokens) + + # Convert to ctypes arrays + self.tokens = (c_uint * self.ntok)(*flat_tokens) + self.req_lens = (c_uint * self.nreq)(*self.req_lens_list) + self.req_pos = (c_uint * self.nreq)(*self.req_pos_list) + self.kv_caches = (POINTER(KVCacheCStruct) * self.nreq)(*self.kv_cache_ptrs) + self.temperatures = (c_float * self.nreq)(*self.temperatures_list) + self.topks = (c_uint * self.nreq)(*self.topks_list) + self.topps = (c_float * self.nreq)(*self.topps_list) + + def input_args(self): + return ( + self.tokens, + self.ntok, + self.req_lens, + self.nreq, + self.req_pos, + self.kv_caches, + self.temperatures, + self.topks, + self.topps, + ) + + +class Qwen3MoeForCausalLM: + """Qwen3-MoE model for causal language modeling""" + + def __init__( + self, model_dir_path, device=DeviceType.DEVICE_TYPE_CPU, ndev=1, max_tokens=None + ): + def load_all_safetensors_from_dir(dir_path_: str): + tensors_ = {} + dir_path_ = Path(dir_path_) + for file in sorted(dir_path_.glob("*.safetensors")): + data_ = safetensors.safe_open(file, "pt") + for name_ in data_.keys(): + tensors_[name_] = data_.get_tensor(name_) + return tensors_ + + print("Loading Qwen3-MoE model weights to host...") + load_start_time = time.time() + + with open(os.path.join(model_dir_path, "config.json"), "r") as f: + config = json.load(f) + self.config = config + + eos_token_id = self.config.get("eos_token_id", 2) + self.eos_token_id = ( + [eos_token_id] if type(eos_token_id) == int else eos_token_id + ) + + transpose_weight = ( + device != DeviceType.DEVICE_TYPE_ASCEND + ) + + # Load state dict + if any(file.suffix == ".safetensors" for file in Path(model_dir_path).iterdir()): + state_dict = load_all_safetensors_from_dir(model_dir_path) + else: + state_dict = torch.load( + os.path.join(model_dir_path, "pytorch_model.bin"), + weights_only=True, + map_location="cpu", + ) + + # Determine naming scheme + if Qwen3MoeWeightsNaming.match(state_dict): + print("✓ Using Qwen3MoeWeightsNaming (with MoE support)") + # Create metadata and weights + self.meta = Qwen3MoeMetaFromConfig(config, max_tokens=max_tokens) + self.weights = Qwen3MoeWeightsImpl( + self.meta, + Qwen3MoeWeightsNaming(), + state_dict, + ndev=ndev, + transpose_weight=transpose_weight, + ) + # Load tokenizer + self.tokenizer = transformers.AutoTokenizer.from_pretrained( + model_dir_path, trust_remote_code=True + ) + else: + raise ValueError("Unsupported weight naming scheme for Qwen3-MoE") + + if self.tokenizer.pad_token is None: + self.tokenizer.pad_token = self.tokenizer.eos_token + + load_end_time = time.time() + print(f"Weight loading time: {load_end_time - load_start_time:.3f}s") + + print(f"Creating Qwen3-MoE model on {ndev} devices...") + load_start_time = time.time() + dev_ids = (c_int * ndev)(*[i for i in range(ndev)]) + + try: + self.model_instance = createQwen3MoeModel( + ctypes.byref(self.meta), + ctypes.byref(self.weights), + device, + ndev, + dev_ids + ) + print(f"✓ Qwen3-MoE model created successfully") + except Exception as e: + print(f"✗ Error creating Qwen3-MoE model: {e}") + import traceback + traceback.print_exc() + raise + + load_end_time = time.time() + print(f"Model creation time: {load_end_time - load_start_time:.3f}s") + if self.model_instance is None: + raise RuntimeError("Qwen3-MoE model instance is None after creation") + + def max_context_len(self): + return self.meta.dctx + + def create_kv_cache(self): + return createQwen3MoeKVCache(self.model_instance) + + def drop_kv_cache(self, kv_cache): + dropQwen3MoeKVCache(self.model_instance, kv_cache) + + def batch_infer_one_round(self, tasks: List[Qwen3MoeInferTask]): + output = (c_uint * len(tasks))() + + batch_inputs = Qwen3MoeBatchedTask(tasks) + + if g_debug_enabled: + print(f"🔍 Qwen3-MoE batch inference:") + print(f" Number of requests: {batch_inputs.nreq}") + print(f" Total tokens: {batch_inputs.ntok}") + print(f" Request lengths: {batch_inputs.req_lens_list}") + + try: + inferQwen3MoeBatch( + self.model_instance, + *batch_inputs.input_args(), + output, + ) + print("✅ inferQwen3MoeBatch completed") + + except Exception as e: + print(f"❌ Qwen3-MoE C++ inference failed: {e}") + import traceback + traceback.print_exc() + raise + + return list(output) + + def get_router_stats(self, layer_idx: int) -> Dict[int, int]: + """Get expert usage statistics for a specific layer""" + if self.meta.num_experts == 0: + return {} + + expert_counts = (c_uint * self.meta.num_experts)() + getQwen3MoeRouterStats(self.model_instance, layer_idx, expert_counts) + + return {i: expert_counts[i] for i in range(self.meta.num_experts)} + + def print_router_stats(self): + """Print router statistics for all MoE layers""" + print(f"\n{'='*60}") + print("📊 MoE ROUTER STATISTICS") + print(f"{'='*60}") + + for layer_idx in range(self.meta.nlayer): + # Check if this is a MoE layer + is_moe = self._is_moe_layer(layer_idx) + if not is_moe: + continue + + stats = self.get_router_stats(layer_idx) + if not stats: + continue + + total_usage = sum(stats.values()) + print(f"\nLayer {layer_idx} (MoE):") + print(f" Total tokens routed: {total_usage}") + + if total_usage > 0: + # Calculate load balance + ideal_usage = total_usage / self.meta.num_experts + for expert_idx, count in stats.items(): + percentage = (count / total_usage) * 100 + balance = count / ideal_usage if ideal_usage > 0 else 0 + print(f" Expert {expert_idx:2d}: {count:6d} tokens ({percentage:5.1f}%) [balance: {balance:.2f}]") + + # Calculate load balance variance + usages = list(stats.values()) + mean_usage = np.mean(usages) + variance = np.var(usages) + print(f" Load balance variance: {variance:.2f} (lower is better)") + + def _is_moe_layer(self, layer_idx): + """Check if a layer is MoE based on configuration""" + # Check if in mlp_only_layers + if self.meta.mlp_only_layers: + mlp_only_list = [self.meta.mlp_only_layers[i] for i in range(self.meta.num_mlp_only_layers)] + if layer_idx in mlp_only_list: + return False + + # Check sparse step condition + return (self.meta.num_experts > 0 and + (layer_idx + 1) % self.meta.decoder_sparse_step == 0) + + def generate(self, input_content, max_steps, topp_=0.8, topk_=50, temperature_=0.7): + # Apply chat template if available + if hasattr(self.tokenizer, 'chat_template') and self.tokenizer.chat_template: + input_content = self.tokenizer.apply_chat_template( + conversation=[{"role": "user", "content": input_content}], + add_generation_prompt=True, + tokenize=False, + ) + + print(input_content, end="", flush=True) + tokens = self.tokenizer.encode(input_content) + + print(f"\nDEBUG: Input tokens: {tokens}") + print(f"DEBUG: Token count: {len(tokens)}") + print(f"DEBUG: EOS tokens: {self.eos_token_id}") + print(f"DEBUG: Vocab size: {self.meta.dvoc}") + print(f"DEBUG: MoE config: {self.meta.num_experts} experts, top-{self.meta.num_experts_per_tok}") + + infer_task = Qwen3MoeInferTask( + tokens=tokens, + position=0, + temperature=temperature_, + topk=topk_, + topp=topp_, + end_tokens=self.eos_token_id, + max_tokens=int(self.meta.dctx), + task_id=0 + ) + + infer_task.bind_kvcache(Qwen3MoeKVCache(self)) + + if self.model_instance is None: + raise RuntimeError("❌ Qwen3-MoE model instance is null before inference") + + steps = 0 + total_time = 0 + output_content = "" + + for step_i in range(max_steps): + start_time = time.time() + output_tokens = self.batch_infer_one_round([infer_task]) + end_time = time.time() + steps += 1 + + output_token = output_tokens[0] + print(f"\nDEBUG Step {step_i}:") + print(f" Output token ID: {output_token}") + print(f" Token in vocab range: {0 <= output_token < self.meta.dvoc}") + print(f" Is EOS token: {output_token in self.eos_token_id}") + + # Check token validity + if output_token >= self.meta.dvoc: + print(f" ⚠ WARNING: Token {output_token} exceeds vocab size {self.meta.dvoc}") + break + if output_token < 0: + print(f" ⚠ WARNING: Negative token ID {output_token}") + break + + try: + output_str = self.tokenizer.decode([output_token], skip_special_tokens=False) + print(f" Decoded: '{output_str}'") + except Exception as e: + print(f" ⚠ Decode failed: {e}") + output_str = f"[UNK_{output_token}]" + + output_content += output_str + print(output_str, end="", flush=True) + + if output_tokens[0] in self.eos_token_id: + break + + infer_task.next(output_tokens[0]) + + if step_i > 0: + total_time += end_time - start_time + + print("\n") + avg_time = total_time * 1000 / (steps - 1) if steps > 1 else 0 + print(f"Time per step: {avg_time:.3f}ms") + + # Print MoE router statistics + if self.meta.num_experts > 0: + self.print_router_stats() + + try: + infer_task._kv_cache.drop() + except Exception as e: + print(f" ⚠ KV cache cleanup failed: {e}") + + return output_content, avg_time + + def destroy_model_instance(self): + destroyQwen3MoeModel(self.model_instance) + print("Qwen3-MoE Model destroyed") + + +# Global debug flag +g_debug_enabled = False + +def test(): + global g_debug_enabled + + if len(sys.argv) < 2: + print("Usage: python qwen3_moe.py [device] [n_device] [--debug]") + sys.exit(1) + + model_path = sys.argv[1] + device_type = DeviceType.DEVICE_TYPE_CPU + + if len(sys.argv) > 2: + if sys.argv[2] == "--cpu": + device_type = DeviceType.DEVICE_TYPE_CPU + elif sys.argv[2] == "--nvidia": + device_type = DeviceType.DEVICE_TYPE_NVIDIA + + ndev = int(sys.argv[3]) if len(sys.argv) > 3 else 1 + + if "--debug" in sys.argv: + g_debug_enabled = True + setQwen3MoeDebugMode(1) + print("🔍 Debug mode enabled") + + print(f"✓ Using Qwen3-MoE model from: {model_path}") + print(f"✓ Device: {device_type}, Devices: {ndev}") + + model = Qwen3MoeForCausalLM(model_path, device_type, ndev) + + # Test generation + model.generate("山东最高的山是？", 50, topp_=0.8, topk_=50, temperature_=0.7) + model.destroy_model_instance() + + +if __name__ == "__main__": + test() \ No newline at end of file diff --git a/infini-qwen3-moe/scripts/test_program.py b/infini-qwen3-moe/scripts/test_program.py new file mode 100755 index 00000000..ccda27bf --- /dev/null +++ b/infini-qwen3-moe/scripts/test_program.py @@ -0,0 +1,28 @@ +import subprocess +import os + +# 定义可执行文件路径和输出文件路径 +executable_path = "./build/test_debug" # 根据你的编译路径调整 +output_file = "debug_test_output.json" + +# 调用编译后的程序 +try: + print("运行测试程序...") + result = subprocess.run([executable_path], capture_output=True, text=True, check=True) + print("程序输出:") + print(result.stdout) + print(result.stderr) +except subprocess.CalledProcessError as e: + print("程序运行失败:") + print(e.stderr) + exit(1) + +# 检查输出文件是否生成 +if os.path.exists(output_file): + print(f"测试成功，生成的文件: {output_file}") + # 如果需要，可以读取并打印 JSON 文件内容 + with open(output_file, "r", encoding="utf-8") as f: + print("JSON 文件内容:") + print(f.read()) +else: + print("测试失败，未找到输出文件。") diff --git a/infini-qwen3-moe/scripts/test_server.py b/infini-qwen3-moe/scripts/test_server.py new file mode 100755 index 00000000..aa383994 --- /dev/null +++ b/infini-qwen3-moe/scripts/test_server.py @@ -0,0 +1,130 @@ +import requests +import json +import time +from concurrent.futures import ThreadPoolExecutor, as_completed + +API_URL = "http://localhost:8000/chat/completions" +MODEL = "FM9G-7B" +PROMPT = ["山东最高的山是？", "给我讲个故事"] +CONCURRENCY = 10 # 并发用户数量 + +def single_run(user_id): + payload = { + "model": MODEL, + "messages": [{"role": "user", "content": PROMPT[user_id % len(PROMPT)]}], + "max_tokens": 512, + "stream": True + } + headers = {'Content-Type': 'application/json', 'Accept': 'application/json'} + print(f"[User {user_id}] Sending request...") + + start = time.perf_counter() + resp = requests.post(API_URL, headers=headers, json=payload, stream=True) + resp.raise_for_status() + + ttfb = resp.elapsed.total_seconds() # HTTP header 到达时间 + header_received = time.perf_counter() + + if resp.encoding is None: + resp.encoding = 'utf-8' + + tokens = 0 + chunks = [] + for line in resp.iter_lines(decode_unicode=True): + if not line or line.strip() == "[DONE]": + continue + s = line.strip() + if s.startswith("data:"): + s = s[len("data:"):].strip() + try: + data = json.loads(s) + except json.JSONDecodeError: + continue + text = data.get("choices", [{}])[0].get("delta", {}).get("content") + if text: + chunks.append(text) + tokens += 1 + stream_done = time.perf_counter() + + # 时间计算 + stream_time = stream_done - header_received + total_time = stream_done - start + time_per_token_ms = (stream_time / tokens * 1000) if tokens else float('inf') + tps = tokens / stream_time if stream_time > 0 else 0 + + + return { + "user": user_id, + "ttfb": ttfb, + "stream_time": stream_time, + "total_time": total_time, + "tokens": tokens, + "time_per_token_ms": time_per_token_ms, + "tps": tps, + "chunks": chunks + } + +def main(): + worst = None + worst_stream = -1.0 + best_stream = float('inf') + results = [] + + with ThreadPoolExecutor(max_workers=CONCURRENCY) as e: + futures = [e.submit(single_run, uid) for uid in range(CONCURRENCY)] + for future in as_completed(futures): + r = future.result() + results.append(r) + + print( + f"User {r['user']} → TTFB = {r['ttfb']:.3f}s, latency = {r['stream_time']:.3f}s, " + f"tokens = {r['tokens']}, time/token = {r['time_per_token_ms']:.2f} ms, " + f"TPS = {r['tps']:.1f} tok/s" + ) + if r['stream_time'] > worst_stream: + worst_stream = r['stream_time'] + worst = r + if r['stream_time'] < best_stream: + best_stream = r['stream_time'] + best = r + + # Sort results by user ID + results.sort(key=lambda x: x["user"]) + + with open("responses.txt", "w", encoding="utf-8") as fw: + for r in results: + fw.write(f"[User {r['user']}]\n") + text = "".join(r["chunks"]) + # fixed = text.encode('latin-1').decode('utf-8') + fixed = text + fw.write(fixed) + fw.write("\n\n") + + n = CONCURRENCY + avg_ttfb = sum(r['ttfb'] for r in results) / n + avg_token = sum(r['tokens'] for r in results) / n + avg_stream = sum(r['stream_time'] for r in results) / n + avg_tps = sum(r['tps'] for r in results) / n + avg_time_per_token = sum(r['time_per_token_ms'] for r in results) / n + + print(f"\n✅ All {n} requests completed.") + print(f"Averages → TTFB = {avg_ttfb:.3f}s, latency = {avg_stream:.3f}s, " + f"tokens = {avg_token:.1f}, TPS = {avg_tps:.1f} tok/s, time/token = {avg_time_per_token:.2f} ms") + + if best: + print("\nFastest user:") + print( + f"User {best['user']} → latency = {best['stream_time']:.3f}s, " + f"tokens = {best['tokens']}, TPS = {best['tps']:.1f} tok/s, " + f"time/token = {best['time_per_token_ms']:.2f} ms" + ) + if worst: + print("\nSlowest user:") + print( + f"User {worst['user']} → latency = {worst['stream_time']:.3f}s, " + f"tokens = {worst['tokens']}, TPS = {worst['tps']:.1f} tok/s, " + f"time/token = {worst['time_per_token_ms']:.2f} ms" + ) + +if __name__ == "__main__": + main() diff --git a/infini-qwen3-moe/src/.DS_Store b/infini-qwen3-moe/src/.DS_Store new file mode 100755 index 0000000000000000000000000000000000000000..0ab09321992297f0c48ef47d4caf166788a68b15 GIT binary patch literal 6148 zcmeHKPfy!06n~~&NkN-(7!pTIyX;VDg#yyFODN?uX@_On4s0q3i6B}MS4m0_Ri&Ka z3-Jv&apm*uJ=;hLZKPd5Z1St@pX2w(&p##4H2|>TG}r;K0l>mW*xtZyg^-@wlC0^O zH;7EeI0Fv~5Rk>^OV(z>Fkl#XX$;V}+r_Uo#-D)7_q*esiDaVEwL7b=5-}S2vG7rg z1){wB^8WU)$ujs$%gIG`Wsc#X0}199K!`c!{w$y9HE;Y$9OXs7f6vxh?N^<3Yu(zk zez_N7<`!Nt&4*rgLQmg>i2aAL=YNmF+1P%2B+|l*(lC<_Q4pfZ&mU15h}lp~(;%0= zu2NWCt2?&0=kvk-pvQOj_7^=q|G3-l@x8&pV$rp>-hKG=b^JR{QgKB|V@brCDCM=r z2e`&TZBOxeWzVgngU9*?o@+4@v6A7MO^w@DN%_?Dmu{BEBBk0*xp%OZhmb<9DtaYH z_?$h~H|AK0(U@b!rj{d|;xa~_hyne6TANQ4--I;`7zX}b2B +#include +#include +#include + +class AllocatorBase { +public: + virtual void *alloc(size_t size) = 0; + virtual void release(void *ptr) = 0; + virtual ~AllocatorBase() = default; +}; + +class MemoryPool : public AllocatorBase { +public: + static constexpr size_t DEFAULT_ALIGNMENT = 256; + + explicit MemoryPool(size_t initialSize = 0, size_t alignment = DEFAULT_ALIGNMENT); + ~MemoryPool(); + + void *alloc(size_t size) override; + void release(void *ptr) override; + + size_t getAlignment() const { return _alignment; } + +private: + struct Block { + void *base; + void *ptr; + size_t size; + bool is_free; + + Block(void *b, void *p, size_t s, bool f) + : base(b), ptr(p), size(s), is_free(f) {} + + bool operator<(const Block &other) const { + return ptr < other.ptr; + } + }; + + void *allocateNewRegion(size_t size); + void tryCoalesce(const Block &block); + + size_t _alignment; + std::vector _base_regions; + std::set _all_blocks; + std::multimap::iterator> _free_blocks; + std::unordered_map::iterator> _ptr_to_block; +}; + +#endif diff --git a/infini-qwen3-moe/src/allocator/memory_allocator.cpp b/infini-qwen3-moe/src/allocator/memory_allocator.cpp new file mode 100755 index 00000000..6d15a1aa --- /dev/null +++ b/infini-qwen3-moe/src/allocator/memory_allocator.cpp @@ -0,0 +1,136 @@ +#include "../allocator.hpp" +#include "../utils.hpp" + + +MemoryPool::MemoryPool(size_t initialSize, size_t alignment) + : _alignment(alignment) { + // Validate alignment is power of two + if ((alignment & (alignment - 1)) != 0 || alignment == 0) { + throw std::invalid_argument("Alignment must be a power of two"); + } + + if (initialSize > 0) { + allocateNewRegion(initialSize); + } +} + +MemoryPool::~MemoryPool() { + for (void *region : _base_regions) { + RUN_INFINI(infinirtFree(region)); + } +} + +void *MemoryPool::alloc(size_t size) { + if (size == 0) { + return nullptr; + } + + // Calculate aligned size + const size_t aligned_size = (size + _alignment - 1) & ~(_alignment - 1); + + // Find the first block with enough space (after alignment) + auto it = _free_blocks.lower_bound(aligned_size); + if (it == _free_blocks.end()) { + allocateNewRegion(aligned_size); + it = _free_blocks.lower_bound(aligned_size); + if (it == _free_blocks.end()) { + throw std::bad_alloc(); + } + } + + auto block_it = it->second; + Block block = *block_it; + _free_blocks.erase(it); + _all_blocks.erase(block_it); + + // Align the pointer within the block + size_t alignment_padding = reinterpret_cast(block.ptr) - reinterpret_cast(block.ptr); + + // Calculate remaining space after allocation + const size_t remaining = block.size - aligned_size - alignment_padding; + + // Create allocated block + Block alloc_block(block.base, block.ptr, aligned_size, false); + auto alloc_it = _all_blocks.insert(alloc_block).first; + _ptr_to_block[block.ptr] = alloc_it; + + // Split remaining space if it's large enough + if (remaining >= _alignment) { + void *rem_ptr = static_cast(block.ptr) + aligned_size; + Block rem_block(block.base, rem_ptr, remaining, true); + auto rem_it = _all_blocks.insert(rem_block).first; + _free_blocks.emplace(remaining, rem_it); + } + + return block.ptr; +} + +void MemoryPool::release(void *ptr) { + if (ptr == nullptr) { + return; + } + + auto it = _ptr_to_block.find(ptr); + if (it == _ptr_to_block.end()) { + throw std::runtime_error("Invalid pointer to free"); + } + + auto block_it = it->second; + Block block = *block_it; + _all_blocks.erase(block_it); + block.is_free = true; + auto new_it = _all_blocks.insert(block).first; + _ptr_to_block.erase(ptr); + tryCoalesce(*new_it); +} + +void *MemoryPool::allocateNewRegion(size_t size) { + // Allocate exactly the requested size + void *ptr = nullptr; + RUN_INFINI(infinirtMalloc(&ptr, size)); + _base_regions.push_back(ptr); + + // Align the pointer within the allocated region + size_t alignment_padding = reinterpret_cast(ptr) - reinterpret_cast(ptr); + size_t usable_size = size - alignment_padding; + + Block new_block(ptr, ptr, usable_size, true); + auto it = _all_blocks.insert(new_block).first; + _free_blocks.emplace(usable_size, it); + + return ptr; +} + +void MemoryPool::tryCoalesce(const Block &block) { + auto it = _all_blocks.find(block); + if (it == _all_blocks.end()) { + return; + } + + Block merged = *it; + auto next = std::next(it); + auto prev = (it == _all_blocks.begin()) ? _all_blocks.end() : std::prev(it); + + _all_blocks.erase(it); + _free_blocks.erase(merged.size); + + // Coalesce with next + if (next != _all_blocks.end() && next->is_free && static_cast(merged.ptr) + merged.size == next->ptr) { + _free_blocks.erase(next->size); + merged.size += next->size; + _all_blocks.erase(next); + } + + // Coalesce with prev + if (prev != _all_blocks.end() && prev->is_free && static_cast(prev->ptr) + prev->size == merged.ptr) { + _free_blocks.erase(prev->size); + merged.ptr = prev->ptr; + merged.size += prev->size; + merged.base = prev->base; + _all_blocks.erase(prev); + } + + merged.is_free = true; + auto new_it = _all_blocks.insert(merged).first; + _free_blocks.emplace(merged.size, new_it); +} diff --git a/infini-qwen3-moe/src/models/.DS_Store b/infini-qwen3-moe/src/models/.DS_Store new file mode 100755 index 0000000000000000000000000000000000000000..71256dec9f818cac0eb8877626092ce748ee6d63 GIT binary patch literal 6148 zcmeHKKX21O6n~ebtwROMK&47uEU{D~h%{*{Ar_N{l>s9d0HrauQiJ0Q`BM*3B#)^5 z6b$egm|2*RnD{J=ym$8kJ`%(R5xOVc{m$>ty}O^~cRm0R?P+)bPy+x58)4-Nb}NMJ z)Rv^gnrnzc%@HO21OJ5_43~VZ!ZKhPSU(2n+ueYd5W^JuuDx2@)A#3>Jj^ z^Mn0Q_ssZ#4FBLUJX@*|C3p>)Q8~&tn>er17jx!s-*3LVrE;?EXSQE$MxOF1PKW!% zJP<(&tsol@c_$x&VG@rgjmB3|Dpxi(x123!+j;Jv$;_YhCev}dH+o7>kEBe3#kdzd zj-#wwyM7|mNiR;LkqU^z2wk2%iPKPKZ8=TDv5NJz!KpgcZf$oqYaO-fZu9VHUUz4A znvJ@9*gBritIp1i+jk#!UnPT7UeJ=UuwrFZa+PrbmpCZw$lwn?zA>#u7SZx;e~;hX z4@S&4x5~p{6u!_p3i;3dQJToKkJ`nG$-~O}GQjW2*Ab*SJghS1@tciGQl%r)w85PbO%oLTS< zMX2mJzpLCqcp7bQ8L$kP43za_jn4o3zrX*RLH5ZqU>W$Y7!Z|?-)ZBP_-tL;9G$f; two_~*o literal 0 HcmV?d00001 diff --git a/infini-qwen3-moe/src/models/jiuge/jiuge.cpp b/infini-qwen3-moe/src/models/jiuge/jiuge.cpp new file mode 100755 index 00000000..e69de29b diff --git a/infini-qwen3-moe/src/models/jiuge/jiuge_impl.hpp b/infini-qwen3-moe/src/models/jiuge/jiuge_impl.hpp new file mode 100755 index 00000000..e69de29b diff --git a/infini-qwen3-moe/src/models/jiuge/jiuge_kv_cache.cpp b/infini-qwen3-moe/src/models/jiuge/jiuge_kv_cache.cpp new file mode 100755 index 00000000..e69de29b diff --git a/infini-qwen3-moe/src/models/jiuge/jiuge_weight.hpp b/infini-qwen3-moe/src/models/jiuge/jiuge_weight.hpp new file mode 100755 index 00000000..e69de29b diff --git a/infini-qwen3-moe/src/models/qw/qwen3.cpp b/infini-qwen3-moe/src/models/qw/qwen3.cpp new file mode 100644 index 00000000..9f0439a0 --- /dev/null +++ b/infini-qwen3-moe/src/models/qw/qwen3.cpp @@ -0,0 +1,2393 @@ + + +#include "qwen3_impl.hpp" +#include "qwen3_weight.hpp" + +#include "../../tensor.hpp" +#include "../../utils.hpp" + +#include +#include +#include +#include +#include +#include +#include + +#ifdef _WIN32 +#include +#else +#include +#endif + +/* + * Debug utilities for comparing C++ vs Python implementations + * These functions save tensor data to files for comparison + */ + +// Global debug flag - set to true to enable debug output +static bool g_debug_enabled = true; + +void set_debug_mode(bool enabled) { g_debug_enabled = enabled; } + +// Validation function to check tensor for extreme values +bool validate_tensor_range(const std::shared_ptr &tensor, + const std::string &name, float min_threshold = -1e6, + float max_threshold = 1e6) { + if (!tensor) + return false; + + auto shape = tensor->shape(); + size_t total_size = 1; + for (auto dim : shape) { + total_size *= dim; + } + + // For small tensors, check all values; for large tensors, sample + size_t sample_size = std::min(total_size, size_t(1000)); + std::vector host_data(sample_size); + + // Copy sample data to host + RUN_INFINI(infinirtMemcpy(host_data.data(), tensor->data(), + sample_size * sizeof(float), INFINIRT_MEMCPY_D2H)); + RUN_INFINI(infinirtDeviceSynchronize()); + + // Check for extreme values + bool has_extreme = false; + float actual_min = host_data[0], actual_max = host_data[0]; + int inf_count = 0, nan_count = 0; + + for (size_t i = 0; i < sample_size; ++i) { + float val = host_data[i]; + if (std::isnan(val)) { + nan_count++; + has_extreme = true; + } else if (std::isinf(val)) { + inf_count++; + has_extreme = true; + } else { + actual_min = std::min(actual_min, val); + actual_max = std::max(actual_max, val); + if (val < min_threshold || val > max_threshold) { + has_extreme = true; + } + } + } + + if (has_extreme && g_debug_enabled) { + printf("⚠ RANGE WARNING: %s has extreme values:\n", name.c_str()); + printf(" Min: %e, Max: %e\n", actual_min, actual_max); + printf(" NaN count: %d, Inf count: %d (in sample of %zu)\n", nan_count, + inf_count, sample_size); + } + + return !has_extreme; +} + +// Function to clamp tensor values to prevent overflow in subsequent operations +void clamp_tensor_inplace(const std::shared_ptr &tensor, + float min_val = -65504.0f, float max_val = 65504.0f) { + if (!tensor) + return; + + auto shape = tensor->shape(); + size_t total_size = 1; + for (auto dim : shape) { + total_size *= dim; + } + + // Get data as float pointer (assuming FP32 storage for intermediate + // computations) + float *data_ptr = static_cast(tensor->data()); + + // Launch a simple clamping operation on device (this would need proper + // CUDA/device implementation) For now, copy to host, clamp, and copy back + // (inefficient but safe) + std::vector host_data(total_size); + + // Copy to host + RUN_INFINI(infinirtMemcpy(host_data.data(), data_ptr, + total_size * sizeof(float), INFINIRT_MEMCPY_D2H)); + RUN_INFINI(infinirtDeviceSynchronize()); + + // Clamp values + bool clamped_any = false; + for (size_t i = 0; i < total_size; ++i) { + if (host_data[i] < min_val) { + host_data[i] = min_val; + clamped_any = true; + } else if (host_data[i] > max_val) { + host_data[i] = max_val; + clamped_any = true; + } else if (std::isnan(host_data[i]) || std::isinf(host_data[i])) { + host_data[i] = 0.0f; // Replace NaN/Inf with zero + clamped_any = true; + } + } + + if (clamped_any) { + // Copy back to device + RUN_INFINI(infinirtMemcpy(data_ptr, host_data.data(), + total_size * sizeof(float), INFINIRT_MEMCPY_H2D)); + RUN_INFINI(infinirtDeviceSynchronize()); + + if (g_debug_enabled) { + printf("⚠ Clamped extreme values in tensor to range [%f, %f]\n", min_val, + max_val); + } + } +} + +template +void save_tensor_debug(const std::shared_ptr &tensor, + const std::string &name, int layer = -1, + const std::string &prefix = "cpp") { + if (!g_debug_enabled || !tensor) + return; + + // Create filename + std::string filename; + if (layer >= 0) { + filename = "output/" + prefix + "_layer_" + std::to_string(layer) + "_" + + name + ".txt"; + } else { + filename = "output/" + prefix + "_" + name + ".txt"; + } + + // Get tensor data to CPU + auto shape = tensor->shape(); + size_t total_size = 1; + for (auto dim : shape) { + total_size *= dim; + } + + // Allocate host memory for tensor data + std::vector host_data(total_size); + + // Copy from device to host + RUN_INFINI(infinirtMemcpy(host_data.data(), tensor->data(), + total_size * sizeof(T), INFINIRT_MEMCPY_D2H)); + + // Synchronize to ensure copy is complete + RUN_INFINI(infinirtDeviceSynchronize()); + + // Calculate comprehensive statistics + T mean = 0, min_val = host_data[0], max_val = host_data[0]; + int inf_count = 0, nan_count = 0, zero_count = 0; + double abs_sum = 0.0; + + for (size_t i = 0; i < total_size; ++i) { + T val = host_data[i]; + if (std::isnan(val)) { + nan_count++; + } else if (std::isinf(val)) { + inf_count++; + } else if (val == 0) { + zero_count++; + mean += val; // still count in mean + } else { + mean += val; + min_val = std::min(min_val, val); + max_val = std::max(max_val, val); + abs_sum += std::abs(static_cast(val)); + } + } + mean /= total_size; + + T std_dev = 0; + for (size_t i = 0; i < total_size; ++i) { + if (!std::isnan(host_data[i]) && !std::isinf(host_data[i])) { + T diff = host_data[i] - mean; + std_dev += diff * diff; + } + } + std_dev = std::sqrt(std_dev / total_size); + + // Save to file + std::ofstream file(filename); + file << std::scientific << std::setprecision(6); + file << "# Tensor: " << name << "\n"; + file << "# Shape: "; + for (size_t i = 0; i < shape.size(); ++i) { + if (i > 0) + file << "x"; + file << shape[i]; + } + file << "\n"; + file << "# Total elements: " << total_size << "\n"; + file << "# Data type size: " << sizeof(T) << " bytes\n"; + file << "# Statistics:\n"; + file << "# Mean: " << mean << "\n"; + file << "# Std: " << std_dev << "\n"; + file << "# Min: " << min_val << "\n"; + file << "# Max: " << max_val << "\n"; + file << "# L1 norm (avg abs): " << (abs_sum / total_size) << "\n"; + file << "# Special values - NaN: " << nan_count << ", Inf: " << inf_count + << ", Zero: " << zero_count << "\n"; + + // Detect potential issues + bool has_issues = false; + if (nan_count > 0 || inf_count > 0) { + file << "# ⚠ WARNING: Contains NaN or Inf values!\n"; + has_issues = true; + } + if (std::abs(mean) > 1e6) { + file << "# ⚠ WARNING: Mean is extremely large (>" << 1e6 << ")\n"; + has_issues = true; + } + if (std::abs(max_val) > 1e6 || std::abs(min_val) > 1e6) { + file << "# ⚠ WARNING: Contains values beyond FP16 safe range\n"; + has_issues = true; + } + if (zero_count > total_size * 0.9) { + file + << "# ⚠ WARNING: More than 90% values are zero (potential underflow)\n"; + has_issues = true; + } + + if (!has_issues) { + file << "# ✓ Values appear to be in reasonable range\n"; + } + + file << "# Data:\n"; + + // For large tensors, save first and last elements only + if (total_size > 1000) { + file << "# First 50 elements:\n"; + for (size_t i = 0; i < std::min(size_t(50), total_size); ++i) { + file << host_data[i] << "\n"; + } + file << "# Last 50 elements:\n"; + for (size_t i = std::max(size_t(0), total_size - 50); i < total_size; ++i) { + file << host_data[i] << "\n"; + } + } else { + // Save all elements for small tensors + for (size_t i = 0; i < total_size; ++i) { + file << host_data[i] << "\n"; + } + } + + file.close(); + + // Print summary to console + if (has_issues) { + printf("⚠ Tensor %s has data quality issues - check %s\n", name.c_str(), + filename.c_str()); + } +} + +// Helper function to convert FP16 to float +float fp16_to_float(uint16_t fp16_val) { + uint32_t sign = (fp16_val & 0x8000) << 16; // 符号位 + uint32_t exponent = (fp16_val & 0x7C00) >> 10; // 指数位 (5位) + uint32_t mantissa = (fp16_val & 0x03FF); // 尾数位 (10位) + + uint32_t float_bits; + + if (exponent == 0) { + if (mantissa == 0) { + // Zero + float_bits = sign; + } else { + // Denormalized number + exponent = 127 - 14; // FP32 bias - FP16 bias - 1 + while ((mantissa & 0x400) == 0) { + mantissa <<= 1; + exponent--; + } + mantissa &= 0x3FF; + float_bits = sign | (exponent << 23) | (mantissa << 13); + } + } else if (exponent == 31) { + // Infinity or NaN + float_bits = sign | 0x7F800000 | (mantissa << 13); + } else { + // Normalized number + exponent = exponent - 15 + 127; // 转换偏置：FP16偏置15 -> FP32偏置127 + float_bits = sign | (exponent << 23) | (mantissa << 13); + } + + return *reinterpret_cast(&float_bits); +} + +void save_tensor_debug_f16(const std::shared_ptr &tensor, + const std::string &name, int layer = -1, + const std::string &prefix = "cpp") { + if (!g_debug_enabled || !tensor) + return; + + // Create filename + std::string filename; + if (layer >= 0) { + filename = "output/" + prefix + "_layer_" + std::to_string(layer) + "_" + + name + ".txt"; + } else { + filename = "output/" + prefix + "_" + name + ".txt"; + } + + // Get tensor data to CPU + auto shape = tensor->shape(); + size_t total_size = 1; + for (auto dim : shape) { + total_size *= dim; + } + + // 按 FP16 读取数据 + std::vector host_data_raw(total_size); + + // Copy from device to host as FP16 + RUN_INFINI(infinirtMemcpy(host_data_raw.data(), tensor->data(), + total_size * sizeof(uint16_t), + INFINIRT_MEMCPY_D2H)); + + // Synchronize to ensure copy is complete + RUN_INFINI(infinirtDeviceSynchronize()); + + // Convert FP16 to float for analysis + std::vector host_data(total_size); + for (size_t i = 0; i < total_size; i++) { + host_data[i] = fp16_to_float(host_data_raw[i]); // 使用正确的转换函数 + } + + // Calculate statistics on converted data + float mean = 0, min_val = host_data[0], max_val = host_data[0]; + for (size_t i = 0; i < total_size; ++i) { + mean += host_data[i]; + min_val = std::min(min_val, host_data[i]); + max_val = std::max(max_val, host_data[i]); + } + mean /= total_size; + + float std_dev = 0; + for (size_t i = 0; i < total_size; ++i) { + float diff = host_data[i] - mean; + std_dev += diff * diff; + } + std_dev = std::sqrt(std_dev / total_size); + + // Save to file with both raw hex and converted values + std::ofstream file(filename); + file << std::scientific << std::setprecision(6); + file << "# Tensor: " << name << " (FP16 format)\n"; + file << "# Shape: "; + for (size_t i = 0; i < shape.size(); ++i) { + if (i > 0) + file << "x"; + file << shape[i]; + } + file << "\n"; + file << "# Total elements: " << total_size << "\n"; + file << "# Mean: " << mean << " (converted)\n"; + file << "# Std: " << std_dev << " (converted)\n"; + file << "# Min: " << min_val << " (converted)\n"; + file << "# Max: " << max_val << " (converted)\n"; + file << "# Data (first 100 elements as hex and converted):\n"; + + // Save first 100 elements showing both raw hex and converted values + for (size_t i = 0; i < std::min(size_t(100), total_size); ++i) { + file << "0x" << std::hex << std::setw(4) << std::setfill('0') + << host_data_raw[i] << " (" << std::scientific << host_data[i] + << ")\n"; + } + + file.close(); + printf(" Saved FP16 debug tensor: %s\n", filename.c_str()); +} +void save_tensor_debug_f32(const std::shared_ptr &tensor, + const std::string &name, int layer = -1, + const std::string &prefix = "cpp") { + save_tensor_debug(tensor, name, layer, prefix); +} + +/* + * 设备资源创建和初始化 + * + * 创建和初始化推理所需的所有 GPU/设备资源，包括： + * - InfiniCore 设备上下文和操作句柄 + * - 用于多设备并行的分布式张量权重 + * - 用于高效缓冲区管理的内存池 + * - 用于设备间同步的通信上下文 + * + * 参数： + * - rsrc：要填充的输出设备资源结构 + * - meta：模型元数据（层数、维度、数据类型） + * - weights：模型权重张量 + * - device：InfiniCore 设设备类型（GPU/CPU） + * - idev：分布式设置中的当前设备索引（0 到 ndev-1） + * - ndev：用于张量并行的设备总数 + * - dev_id：物理设备 ID + * - comm：用于多设备操作的 InfiniCCL 通信器 + */ +void createQwen3DeviceResource(DeviceQwen3Resource *rsrc, const Qwen3Meta *meta, + const Qwen3Weights *weights, + infiniDevice_t device, int idev, int ndev, + int dev_id, infinicclComm_t comm) { + // 初始化 InfiniCore 设备上下文并创建操作句柄 + // 这设置了后续 InfiniCore API 调用的活动设备 + RUN_INFINI(infinirtSetDevice(device, dev_id)); + + // 为此设备创建操作句柄 - 用于所有计算操作 + infiniopHandle_t handle; + infiniopCreateHandle(&handle); + + // 创建用于异步操作的执行流 + infinirtStream_t stream; + infinirtStreamCreate(&stream); + + /* + * 用于分布式推理的权重张量提取 + * + * 从全局权重存储中提取模型权重并在设备间分区 + * 以实现张量并行。每个设备获得： + * - 注意力投影权重（QKV，输出）：按注意力头分区 + * - FFN 权重：按中间维度分区 + * - 归一化权重：在所有设备上复制 + * + * 张量形状： + * - w_attn_norm：[d] - 层归一化权重，复制 + * - w_attn_qkv：[d, (nh + 2*nkvh)/ndev * dh] - QKV 投影，头分区 + * - b_attn_qkv：[(nh + 2*nkvh)/ndev * dh] - QKV 偏置（可选），头分区 + * - w_attn_out：[nh/ndev * dh, d] - 输出投影，头分区 + * - w_ffn_norm：[d] - FFN 归一化权重，复制 + * - w_ffn_gate_up：[d, 2*di/ndev] - 门控和上升投影，维度分区 + * - w_ffn_down：[di/ndev, d] - 下降投影，维度分区 + */ + + std::vector> w_attn_norm, w_attn_q_norm, + w_attn_k_norm, w_attn_v_norm, w_attn_q_proj, w_attn_k_proj, w_attn_v_proj, + w_attn_o_proj, w_mlp_norm, w_mlp_gate_proj, w_mlp_up_proj, + w_mlp_down_proj; + + // 新的分离权重提取方式 + for (size_t layer = 0; layer < meta->nlayer; layer++) { + + // 添加注意力层归一化权重提取（这个被遗漏了！） + w_attn_norm.push_back( + getQwen3AttnNorm(meta, weights, layer)); // [d] - 注意力层归一化权重 + + // Qwen3 特有：Q/K 归一化权重（每层每头维度） + + w_attn_q_norm.push_back( + getQwen3QNorm(meta, weights, layer)); // [dh] - Q 归一化权重 + w_attn_k_norm.push_back( + getQwen3KNorm(meta, weights, layer)); // [dh] - K 归一化权重 + + // 分离的 QKV 投影权重 + // Q 投影权重：[d, nh/ndev * dh] - 按查询头分区 + w_attn_q_proj.push_back(getQwen3AttnQ(meta, weights, layer, idev, ndev)); + + // K 投影权重：[d, nkvh/ndev * dh] - 按键值头分区 + w_attn_k_proj.push_back(getQwen3AttnK(meta, weights, layer, idev, ndev)); + + // V 投影权重：[d, nkvh/ndev * dh] - 按键值头分区 + w_attn_v_proj.push_back(getQwen3AttnV(meta, weights, layer, idev, ndev)); + + // 注意力输出投影：[nh/ndev * dh, d] - 按输入维度分区 + w_attn_o_proj.push_back(getQwen3AttnO(meta, weights, layer, idev, ndev)); + + // MLP 层权重 + + // MLP 前归一化：[d] - 在所有设备上复制 + w_mlp_norm.push_back(getQwen3MLPNorm(meta, weights, layer)); + + // MLP Gate 投影：[d, di/ndev] - 按中间维度分区 + w_mlp_gate_proj.push_back( + getQwen3MLPGate(meta, weights, layer, idev, ndev)); + + // MLP Up 投影：[d, di/ndev] - 按中间维度分区 + w_mlp_up_proj.push_back(getQwen3MLPUp(meta, weights, layer, idev, ndev)); + + // MLP Down 投影：[di/ndev, d] - 按输入维度分区 + w_mlp_down_proj.push_back( + getQwen3MLPDown(meta, weights, layer, idev, ndev)); + } + + // 创建用于高效缓冲区分配的内存池（128MB） + // 此池在推理期间管理临时张量以避免频繁的 malloc/free + // 增加内存池大小 + auto memory_pool = + std::make_shared(512 * 1024 * 1024); // 从128MB增加到512MB + // 使用所有初始化的组件填充设备资源结构 + // 此结构包含在此设备上推理所需的一切 + *rsrc = DeviceQwen3Resource{ + device, // InfiniCore 设备类型（GPU/CPU） + dev_id, // 物理设备 ID + handle, // InfiniCore 操作句柄 + getQwen3InEmbd(meta, weights), // 输入嵌入表 [dvoc, d] + getQwen3OutNorm(meta, weights), // 输出归一化权重 [d] + getQwen3OutEmbd(meta, weights), // 输出嵌入/LM 头 [d, dvoc] + getQwen3SinTable(meta), // RoPE 正弦表 [dctx, dh/2] + getQwen3CosTable(meta), // RoPE 余弦表 [dctx, dh/2] + w_attn_norm, // 注意力层归一化权重 [d] + w_attn_q_norm, // Q 归一化权重 + w_attn_k_norm, // K 归一化权重 + w_attn_q_proj, // Q 投影权重 + w_attn_k_proj, // K 投影权重 + w_attn_v_proj, // V 投影权重 + w_attn_o_proj, // 注意力输出权重 + w_mlp_norm, // MLP 归一化权重 + w_mlp_gate_proj, // MLP Gate 权重 + w_mlp_up_proj, // MLP Up 权重 + w_mlp_down_proj, // MLP Down 权重 + stream, // 用于异步操作的执行流 + comm, // 设备间通信上下文 + memory_pool, // 用于临时缓冲区的内存池 + }; + + // 同步设备以确保所有初始化完成 + RUN_INFINI(infinirtDeviceSynchronize()); +} + +/* + * 设备资源清理和内存释放 + * + * 按分配的相反顺序正确释放所有设备资源： + * 1. 同步设备以完成所有待处理操作 + * 2. 释放张量内存（shared_ptr 自动处理引用计数） + * 3. 销毁 InfiniCore 句柄和流 + * 4. 清理通信上下文 + * + * 这可以防止内存泄漏并确保正确清理 GPU 资源 + */ +void releaseQwen3DeviceResource(DeviceQwen3Resource &res) { + // 在清理前等待所有待处理操作完成 + infinirtDeviceSynchronize(); + + // 通过重置 shared_ptr 引用来释放张量内存 + // 当引用计数达到零时，底层内存将被释放 + // 释放全局模型张量（输入/输出嵌入、归一化、RoPE 表） + res.w_in_embd.reset(); // 输入嵌入表 [dvoc, d] + res.w_out_norm.reset(); // 最终层归一化 [d] + res.w_out_embd.reset(); // 输出投影/LM 头 [d, dvoc] + res.sin_table.reset(); // RoPE 正弦查找表 [dctx, dh/2] + res.cos_table.reset(); // RoPE 余弦查找表 [dctx, dh/2] + + // 释放每层注意力权重并清除向量 + for (auto &t : res.w_attn_q_norm) { + t.reset(); // Q 层归一化权重 [d] + } + res.w_attn_q_norm.clear(); + for (auto &t : res.w_attn_k_norm) { + t.reset(); // K 层归一化权重 [d] + } + res.w_attn_k_norm.clear(); + for (auto &t : res.w_attn_q_proj) { + t.reset(); // Q 投影权重 [d, nh/ndev * dh] + } + res.w_attn_q_proj.clear(); + for (auto &t : res.w_attn_k_proj) { + t.reset(); // K 投影权重 [d, nkvh/ndev * dh] + } + res.w_attn_k_proj.clear(); + for (auto &t : res.w_attn_v_proj) { + t.reset(); // V 投影权重 [d, nkvh/ndev * dh] + } + res.w_attn_v_proj.clear(); + for (auto &t : res.w_attn_o_proj) { + t.reset(); // 输出投影权重 [nh/ndev * dh, d] + } + res.w_attn_o_proj.clear(); + for (auto &t : res.w_mlp_norm) { + t.reset(); // MLP 层归一化权重 [d] + } + res.w_mlp_norm.clear(); + for (auto &t : res.w_mlp_gate_proj) { + t.reset(); // MLP Gate 投影权重 [d, di/ndev] + } + res.w_mlp_gate_proj.clear(); + for (auto &t : res.w_mlp_up_proj) { + t.reset(); // MLP Up 投影权重 [d, di/ndev] + } + res.w_mlp_up_proj.clear(); + for (auto &t : res.w_mlp_down_proj) { + t.reset(); // MLP Down 投影权重 [di/ndev, d] + } + res.w_mlp_down_proj.clear(); + + // 销毁 InfiniCore 句柄和上下文 + infiniopDestroyHandle(res.handle); // 释放操作句柄 + res.handle = nullptr; + + infinirtStreamDestroy(res.stream); // 释放执行流 + res.stream = nullptr; + + infinicclCommDestroy(res.comm); // 释放通信上下文 + res.comm = nullptr; +} + +/* + * 设备级批处理推理函数 + * 在单个设备上为一批序列执行 transformer 推理。 + * 实现完整的前向传递，包括： + * 1. 输入嵌入查找和 RoPE 位置编码 + * 2. 多层 transformer 块（注意力 + FFN） + * 3. 输出归一化和概率分布 + * 4. 带温度/top-k/top-p 的 token 采样 + * + * 此函数通过张量并行处理分布式推理，其中 + * 每个设备处理模型参数的一个切片。 + * + * 输入参数： + * - meta：模型架构元数据（维度、层数等） + * - rsrc：设备资源（权重、句柄、内存池） + * - idev/ndev：用于分布式推理的设备索引和设备总数 + * - tokens：要处理的输入 token ID [ntok] + * - ntok：批处理中所有请求的 token 总数 + * - req_lens：每个请求的长度 [nreq] + * - nreq：批处理中的请求数 + * - req_pos：每个请求在 KV 缓存中的起始位置 [nreq] + * - kv_caches：每个请求的 KV 缓存存储 [nreq][ndev][nlayer] + * - temperature/topk/topp：采样参数 [nreq] + * - output：生成的 token ID [nreq] + * + * 张量维度符号： + * - ntok：批处理中的 token 总数 + * - nreq：请求数 + * - d：模型隐藏维度 + * - nh：总注意力头数 + * - nkvh：总键值头数 + * - dh：头维度（d/nh） + * - di：FFN 中间维度 + * - dvoc：词汇表大小 + * - dctx：最大上下文长度 + */ +void inferQwen3DeviceBatch(const Qwen3Meta &meta, DeviceQwen3Resource &rsrc, + uint32_t idev, uint32_t ndev, const uint32_t *tokens, + uint32_t ntok, const uint32_t *req_lens, + uint32_t nreq, const uint32_t *req_pos, + struct Qwen3KVCache **kv_caches, + const float *temperature, const uint32_t *topk, + const float *topp, uint32_t *output) { + + if (meta.nh < ndev || meta.nkvh < ndev) { + throw std::runtime_error( + "Invalid distributed setup: heads (" + std::to_string(meta.nh) + ", " + + std::to_string(meta.nkvh) + ") must be >= devices (" + + std::to_string(ndev) + ")"); + } + /* + * 提取模型维度并配置分布式推理 + * + * 张量并行的关键维度计算： + * - nkvh：每设备的键值头数 = total_kv_heads / ndev + * - nh：每设备的查询头数 = total_heads / ndev + * - ngroup：分组查询注意力比率 = nh / nkvh + * - di：每设备的 FFN 中间维度 = total_intermediate / ndev + * + * 这确保每个设备处理注意力头和 FFN 维度的一个切片 + * 同时保持相同的序列处理。 + */ + auto nlayer = meta.nlayer; // transformer 层数 + auto nkvh = meta.nkvh / ndev; // 每设备的 KV 头数（分布式） + auto nh = meta.nh / ndev; // 每设备的查询头数（分布式） + auto ngroup = nh / nkvh; // 分组查询注意力因子 + // auto dctx = meta.dctx; // 最大上下文长度（未使用） + auto dh = meta.dh; // 头维度 + auto d = meta.d; // 模型隐藏维度 + auto dt_logits = meta.dt_logits; // logits 的数据类型（FP16/BF16/FP32） + auto di = meta.di / ndev; // 每设备的 FFN 中间维度 + auto dvoc = meta.dvoc; // 词汇表大小 + auto stream = rsrc.stream; // 用于异步操作的执行流 + auto dt_logits_corrected = rsrc.w_in_embd->dtype(); // 使用权重的实际数据类型 + + if (dt_logits != dt_logits_corrected) { + if (g_debug_enabled && idev == 0) { + printf("WARNING: Correcting dt_logits from %d to %d to match weights\n", + dt_logits, dt_logits_corrected); + } + dt_logits = dt_logits_corrected; + } + + if (nh == 0 || nkvh == 0) { + throw std::runtime_error( + "Zero heads after distribution - check model/device configuration"); + } + + if (ntok == 0 || nreq == 0) { + throw std::runtime_error( + "Invalid batch size: ntok=" + std::to_string(ntok) + + ", nreq=" + std::to_string(nreq)); + } + + // 计算预期内存使用量 + size_t total_memory_needed = 0; + total_memory_needed += + ntok * d * dsize(dt_logits) * 2; // logits_in + logits_out + total_memory_needed += + ntok * nh * dh * dsize(dt_logits) * 3; // q_buf + q_norm_buf + o_buf + total_memory_needed += + ntok * nkvh * dh * dsize(dt_logits) * 3; // k_buf + k_norm_buf + v_buf + total_memory_needed += ntok * di * dsize(dt_logits) * 2; // gate_buf + up_buf + total_memory_needed += nreq * dvoc * dsize(dt_logits); // prob_buf + total_memory_needed += nreq * sizeof(int64_t); // result_buf + + if (!rsrc.memory_pool) { + throw std::runtime_error("Memory pool is null"); + } + + /* + * 推理流水线的内存缓冲区分配 + * + * 为中间计算分配临时缓冲区。 + * 所有缓冲区使用设备内存池进行高效分配/释放。 + * + * 缓冲区张量形状： + * - logits_in/out：[ntok, d] - 流经各层的隐藏状态 + * - qkv_buf：[ntok, (nh + nkvh*2) * dh] - 连接的 Q、K、V 投影 + * - gate_up_buf：[ntok, 2*di] - 连接的 FFN 门控和上升投影 + * - o_buf：[ntok, nh*dh] - 注意力输出在输出投影之前 + * - prob_buf：[nreq, dvoc] - 输出概率分布 + * - result_buf：[nreq] - 采样的 token ID（设备内存） + * - result_cpu：[nreq] - 采样的 token ID（主机内存用于输出） + */ + + auto logits_in = Tensor::buffer(dt_logits, {ntok, d}, rsrc.memory_pool); + assert(logits_in != nullptr && "logits_in allocation failed"); + + if (!logits_in) { + throw std::runtime_error("Failed to allocate logits_in buffer"); + } + + auto logits_out = Tensor::buffer(dt_logits, {ntok, d}, rsrc.memory_pool); + assert(logits_out != nullptr && "logits_out allocation failed"); + if (!logits_out) { + throw std::runtime_error("Failed to allocate logits_out buffer"); + } + + auto q_buf = Tensor::buffer(dt_logits, {ntok, nh * dh}, + rsrc.memory_pool); // Q 投影输出 + auto k_buf = Tensor::buffer(dt_logits, {ntok, nkvh * dh}, + rsrc.memory_pool); // K 投影输出 、 + auto v_buf = Tensor::buffer(dt_logits, {ntok, nkvh * dh}, + rsrc.memory_pool); // V 投影输出 + + auto q_norm_buf = Tensor::buffer(dt_logits, {ntok, nh * dh}, + rsrc.memory_pool); // Q 归一化后 + auto k_norm_buf = Tensor::buffer(dt_logits, {ntok, nkvh * dh}, + rsrc.memory_pool); // K 归一化后 + auto gate_buf = + Tensor::buffer(dt_logits, {ntok, di}, rsrc.memory_pool); // Gate 投影输出 + auto up_buf = + Tensor::buffer(dt_logits, {ntok, di}, rsrc.memory_pool); // Up 投影输出 + auto o_buf = Tensor::buffer(dt_logits, {ntok, nh * dh}, rsrc.memory_pool); + auto prob_buf = Tensor::buffer(dt_logits, {nreq, dvoc}, rsrc.memory_pool); + auto result_buf = Tensor::buffer(INFINI_DTYPE_I64, {nreq}, rsrc.memory_pool); + auto result_cpu = std::vector(nreq); + + // Validate all buffer allocations + std::vector, std::string>> buffers = { + {q_buf, "q_buf"}, {k_buf, "k_buf"}, + {v_buf, "v_buf"}, {q_norm_buf, "q_norm_buf"}, + {k_norm_buf, "k_norm_buf"}, {gate_buf, "gate_buf"}, + {up_buf, "up_buf"}, {o_buf, "o_buf"}, + {prob_buf, "prob_buf"}, {result_buf, "result_buf"}}; + + for (const auto &[buffer, name] : buffers) { + if (!buffer) { + throw std::runtime_error("Failed to allocate " + name + + " buffer - insufficient memory or pool issues"); + } + } + + // if (g_debug_enabled && idev == 0) { + // printf(" Memory pool usage: %.2f MB total needed\n", + // total_memory_needed / (1024.0 * 1024.0)); + + // #ifdef _WIN32 + // #else + // #endif + // } + + // if (g_debug_enabled && idev == 0) { + // for (uint32_t i = 0; i < std::min(ntok, 10u); i++) { + // printf("%u ", tokens[i]); + // } + // printf("\n"); + // printf(" meta.dt_logits: %d\n", meta.dt_logits); + // printf(" dt_logits (local): %d\n", dt_logits); + // printf(" INFINI_DTYPE_F16: %d\n", INFINI_DTYPE_F16); + // printf(" INFINI_DTYPE_F32: %d\n", INFINI_DTYPE_F32); + // printf(" Weight embedding dtype: %d\n", rsrc.w_in_embd->dtype()); + + // printf(" - Pointer: %p\n", rsrc.w_in_embd.get()); + // printf(" - Data pointer: %p\n", rsrc.w_in_embd->data()); + // printf(" - Dtype: %d\n", rsrc.w_in_embd->dtype()); + + // printf(" - Dtype meaning: "); + // switch(rsrc.w_in_embd->dtype()) { + // case INFINI_DTYPE_F16: printf("FLOAT16 (2 bytes)"); break; + // case INFINI_DTYPE_BF16: printf("BFLOAT16 (2 bytes)"); break; + // case INFINI_DTYPE_F32: printf("FLOAT32 (4 bytes)"); break; + // case INFINI_DTYPE_I32: printf("INT32 (4 bytes)"); break; + // case INFINI_DTYPE_U32: printf("UINT32 (4 bytes)"); break; + // default: printf("UNKNOWN(%d)", rsrc.w_in_embd->dtype()); break; + // } + // printf("\n"); + + // printf(" - Shape: "); + // auto in_shape = rsrc.w_in_embd->shape(); + // for (size_t i = 0; i < in_shape.size(); i++) { + // printf("%zu%s", in_shape[i], i < in_shape.size()-1 ? "x" : ""); + // } + // printf("\n"); + + if (tokens[0] < meta.dvoc) { + printf("DEBUG: Checking embedding for token %u:\n", tokens[0]); + + if (rsrc.w_in_embd->dtype() == INFINI_DTYPE_F16) { + std::vector temp_embedding_raw(std::min(size_t(10), d)); + size_t token_offset = tokens[0] * d; + + RUN_INFINI(infinirtMemcpy( + temp_embedding_raw.data(), rsrc.w_in_embd->data(token_offset), + sizeof(uint16_t) * temp_embedding_raw.size(), INFINIRT_MEMCPY_D2H)); + RUN_INFINI(infinirtDeviceSynchronize()); + + printf(" First 10 embedding values (FP16 as hex): "); + for (size_t i = 0; i < temp_embedding_raw.size(); i++) { + printf("0x%04x ", temp_embedding_raw[i]); + } + printf("\n"); + + // 如果有 FP16 到 float 的转换函数，也打印转换后的值 + printf(" First 10 embedding values (converted to float): "); + for (size_t i = 0; i < temp_embedding_raw.size(); i++) { + // 简单的 FP16 到 float 转换（需要实际的转换函数） + // 这里只是示例，你需要使用正确的转换 + printf("%.6e ", 0.0f); // 占位符 + } + printf("\n"); + } else { + std::vector temp_embedding(std::min(size_t(10), d)); + size_t token_offset = tokens[0] * d; + + RUN_INFINI(infinirtMemcpy( + temp_embedding.data(), rsrc.w_in_embd->data(token_offset), + sizeof(float) * temp_embedding.size(), INFINIRT_MEMCPY_D2H)); + RUN_INFINI(infinirtDeviceSynchronize()); + + printf(" First 10 embedding values: "); + for (size_t i = 0; i < temp_embedding.size(); i++) { + printf("%.6e ", temp_embedding[i]); + } + printf("\n"); + } + } +} + +/* + * 输入准备和 Token 嵌入查找 + */ + +// 准备输入 +auto batch_pos_ids = std::vector(ntok); +size_t req_start = 0; + +// 构建位置 ID 数组：连接每个请求的位置序列 +for (uint32_t req = 0; req < nreq; req++) { + for (uint32_t i = 0; i < req_lens[req]; i++) { + batch_pos_ids[req_start + i] = req_pos[req] + i; + } + req_start += req_lens[req]; +} + +// 将位置 ID 复制到设备内存（CPU 设备可以直接使用主机指针） +std::shared_ptr pos_ids_buf; +if (rsrc.device == INFINI_DEVICE_CPU) { + pos_ids_buf = Tensor::weight(batch_pos_ids.data(), INFINI_DTYPE_U32, {ntok}); +} else { + pos_ids_buf = Tensor::buffer(INFINI_DTYPE_U32, {ntok}, rsrc.memory_pool); + // 位置 ID 的异步主机到设备复制 [ntok] + RUN_INFINI(infinirtMemcpyAsync(pos_ids_buf->data(), batch_pos_ids.data(), + sizeof(uint32_t) * ntok, INFINIRT_MEMCPY_H2D, + stream)); +} + +// 查找输入嵌入：logits_in[i] = w_in_embd[tokens[i]] +for (uint32_t i = 0; i < ntok; i++) { + if (tokens[i] >= meta.dvoc) { + printf("Error: Invalid token ID %u >= vocab_size %zu\n", tokens[i], + meta.dvoc); + throw std::runtime_error("Invalid token ID"); + } + // 添加详细的内存复制调试 + if (g_debug_enabled && idev == 0 && i == 0) { + printf("DEBUG: Copying embedding for token %u to position %u\n", tokens[i], + i); + printf(" Source: %p (offset: %zu)\n", rsrc.w_in_embd->data(tokens[i] * d), + tokens[i] * d); + printf(" Dest: %p (offset: %zu)\n", logits_in->data(i * d), i * d); + printf(" Size: %zu bytes\n", dsize(dt_logits) * d); + } + + RUN_INFINI(infinirtMemcpyAsync( + logits_in->data(i * d), // 目标：logits_in 的第 i 行 + rsrc.w_in_embd->data(tokens[i] * d), // 源：tokens[i] 的嵌入 + dsize(dt_logits) * d, INFINIRT_MEMCPY_D2D, stream)); +} + +// 同步并检查结果 +RUN_INFINI(infinirtStreamSynchronize(stream)); + +if (g_debug_enabled && idev == 0) { + printf("DEBUG: After embedding lookup, checking first token result:\n"); + + // 检查复制后的结果 + std::vector result_check_raw(std::min(size_t(10), d)); + RUN_INFINI(infinirtMemcpy(result_check_raw.data(), + logits_in->data(0), // 第一个token的嵌入 + sizeof(uint16_t) * result_check_raw.size(), + INFINIRT_MEMCPY_D2H)); + RUN_INFINI(infinirtDeviceSynchronize()); + + printf(" First 10 values after copy: "); + for (size_t i = 0; i < result_check_raw.size(); i++) { + printf("0x%04x ", result_check_raw[i]); + } + printf("\n"); + + printf(" Comparing with original embedding values: "); + printf("Original: 0x2660 0x9da8 0xac50 0xa518 0xa8c0\n"); + printf(" Copied: "); + for (size_t i = 0; i < std::min(size_t(5), result_check_raw.size()); i++) { + printf("0x%04x ", result_check_raw[i]); + } + printf("\n"); +} else { + std::vector result_check(std::min(size_t(10), d)); + RUN_INFINI(infinirtMemcpy(result_check.data(), logits_in->data(0), + sizeof(float) * result_check.size(), + INFINIRT_MEMCPY_D2H)); + RUN_INFINI(infinirtDeviceSynchronize()); + + printf(" First 10 values after copy: "); + for (size_t i = 0; i < result_check.size(); i++) { + printf("%.6e ", result_check[i]); + } + printf("\n"); +} + +// Debug: Save input embeddings +if (g_debug_enabled) { + // Create input tokens tensor for debugging + auto tokens_tensor = + Tensor::buffer(INFINI_DTYPE_U32, {ntok}, rsrc.memory_pool); + RUN_INFINI(infinirtMemcpy(tokens_tensor->data(), tokens, + sizeof(uint32_t) * ntok, INFINIRT_MEMCPY_H2D)); + save_tensor_debug_f32(tokens_tensor, "input_ids", -1, "cpp"); + + // 根据实际数据类型调用正确的保存函数 + if (dt_logits == INFINI_DTYPE_F16) { + save_tensor_debug_f16(logits_in, "input_embeddings", -1, "cpp"); + } else { + save_tensor_debug_f32(logits_in, "input_embeddings", -1, "cpp"); + } +} + +/* + * InfiniCore 操作符描述符创建和工作区大小计算 + */ + +// 准备操作符和工作区 +size_t workspace_size = 0, temp_size = 0; +infiniopRMSNormDescriptor_t desc_norm; + +RUN_INFINI(infiniopCreateRMSNormDescriptor( + rsrc.handle, &desc_norm, logits_in->desc(), // 输入：[ntok, d] + logits_out->desc(), + rsrc.w_attn_norm[0]->desc(), // 输出：[ntok, d]，权重：[d] + meta.epsilon)); // 归一化 epsilon + +RUN_INFINI(infiniopGetRMSNormWorkspaceSize(desc_norm, &workspace_size)); + +workspace_size = std::max(workspace_size, temp_size); + +/* + * 注意力机制描述符 + * + * 注意力计算涉及几个矩阵运算： + * 1. QKV 投影：X -> Q、K、V 通过矩阵乘法 + * 2. Q 和 K 上的 RoPE 位置编码 + * 3. 注意力计算：softmax(QK^T/√d_k) * V + * 4. 输出投影：O -> 最终注意力输出 + */ + +// QKV 投影和注意力输出的 GEMM 描述符 +infiniopGemmDescriptor_t desc_attn_q, desc_attn_k, desc_attn_v, desc_attn_o; +/* + * QKV 投影 GEMM：logits_in * w_attn_qkv -> qkv_buf + * + * 矩阵乘法：Y = X * W + * 输入 X：[ntok, d] - 归一化的隐藏状态 + * 权重 W：[d, (nh + 2*nkvh)/ndev * dh] - QKV 投影权重 + * 输出 Y：[ntok, (nh + 2*nkvh)/ndev * dh] - 连接的 Q、K、V 投影 + * + * 输出包含沿最后一个维度连接的 Q、K、V 投影： + * - Q：[ntok, nh/ndev * dh] （查询投影） + * - K：[ntok, nkvh/ndev * dh] （键投影） + * - V：[ntok, nkvh/ndev * dh] （值投影） + */ + +RUN_INFINI(infiniopCreateGemmDescriptor(rsrc.handle, &desc_attn_q, + q_buf->desc(), logits_out->desc(), + rsrc.w_attn_q_proj[0]->desc())); +RUN_INFINI(infiniopGetGemmWorkspaceSize(desc_attn_q, &temp_size)); +workspace_size = std::max(workspace_size, temp_size); + +RUN_INFINI(infiniopCreateGemmDescriptor(rsrc.handle, &desc_attn_k, + k_buf->desc(), logits_out->desc(), + rsrc.w_attn_k_proj[0]->desc())); +RUN_INFINI(infiniopGetGemmWorkspaceSize(desc_attn_k, &temp_size)); +workspace_size = std::max(workspace_size, temp_size); + +RUN_INFINI(infiniopCreateGemmDescriptor(rsrc.handle, &desc_attn_v, + v_buf->desc(), logits_out->desc(), + rsrc.w_attn_v_proj[0]->desc())); +RUN_INFINI(infiniopGetGemmWorkspaceSize(desc_attn_v, &temp_size)); +workspace_size = std::max(workspace_size, temp_size); + +// 重构多头 Q/K 归一化处理 + +/* + * 多头 Q/K 归一化描述符创建 + * + * Qwen3 的 Q/K 归一化是按头独立应用的： + * - 每个头有独立的归一化权重 [dh] + * - 需要为每个头创建单独的归一化描述符 + * - 或者使用循环在推理时分别处理每个头 + */ + +// 方案1：创建单个头的归一化描述符，在推理循环中重复使用 +infiniopRMSNormDescriptor_t desc_q_norm_single, desc_k_norm_single; + +// 创建单个头的张量描述符 [ntok, dh] +auto q_single_head = TensorDesc::create(dt_logits, {ntok, dh}); +auto q_norm_single_head = TensorDesc::create(dt_logits, {ntok, dh}); +auto k_single_head = TensorDesc::create(dt_logits, {ntok, dh}); +auto k_norm_single_head = TensorDesc::create(dt_logits, {ntok, dh}); + +RUN_INFINI(infiniopCreateRMSNormDescriptor( + rsrc.handle, &desc_q_norm_single, q_norm_single_head->desc(), + q_single_head->desc(), rsrc.w_attn_q_norm[0]->desc(), meta.epsilon)); +RUN_INFINI(infiniopCreateRMSNormDescriptor( + rsrc.handle, &desc_k_norm_single, k_norm_single_head->desc(), + k_single_head->desc(), rsrc.w_attn_k_norm[0]->desc(), meta.epsilon)); + +RUN_INFINI(infiniopGetRMSNormWorkspaceSize(desc_q_norm_single, &temp_size)); +workspace_size = std::max(workspace_size, temp_size); +RUN_INFINI(infiniopGetRMSNormWorkspaceSize(desc_k_norm_single, &temp_size)); +workspace_size = std::max(workspace_size, temp_size); + +/* + * RoPE 描述符 - 仍然使用 3D 格式，但在推理时分头处理 + */ +infiniopRoPEDescriptor_t desc_rope_q_single, desc_rope_k_single; + +// 创建单个头的 3D RoPE 描述符 [ntok, 1, dh] +auto q_rope_single_head_3d = TensorDesc::create(dt_logits, {ntok, 1, dh}); +auto k_rope_single_head_3d = TensorDesc::create(dt_logits, {ntok, 1, dh}); + +RUN_INFINI(infiniopCreateRoPEDescriptor( + rsrc.handle, &desc_rope_q_single, q_rope_single_head_3d->desc(), + q_rope_single_head_3d->desc(), pos_ids_buf->desc(), rsrc.sin_table->desc(), + rsrc.cos_table->desc())); + +RUN_INFINI(infiniopCreateRoPEDescriptor( + rsrc.handle, &desc_rope_k_single, k_rope_single_head_3d->desc(), + k_rope_single_head_3d->desc(), pos_ids_buf->desc(), rsrc.sin_table->desc(), + rsrc.cos_table->desc())); + +RUN_INFINI(infiniopGetRoPEWorkspaceSize(desc_rope_q_single, &temp_size)); +workspace_size = std::max(workspace_size, temp_size); +RUN_INFINI(infiniopGetRoPEWorkspaceSize(desc_rope_k_single, &temp_size)); +workspace_size = std::max(workspace_size, temp_size); + +// /* +// * RoPE 描述符 (更新为使用归一化后的张量) +// */ +// // Q RoPE：q_norm_buf -> q_norm_buf (就地操作) + +// auto q_norm_3d = q_norm_buf->dimSplit(1, {nh, dh}); +// auto k_norm_3d = k_norm_buf->dimSplit(1, {nkvh, dh}); + +// printf("Debug: Creating Q RoPE descriptor\n"); +// printf("Debug: Expected shapes - q_norm_3d: [%u, %zu, %zu]\n", ntok, nh, dh); + +// RUN_INFINI(infiniopCreateRoPEDescriptor( +// rsrc.handle, &desc_rope_q, +// q_norm_3d->desc(), q_norm_3d->desc(), // 输入/输出都是 3D +// pos_ids_buf->desc(), // 位置 ID [ntok] +// rsrc.sin_table->desc(), // 正弦表 [dctx, dh/2] +// rsrc.cos_table->desc())); // 余弦表 [dctx, dh/2] + +// RUN_INFINI(infiniopGetRoPEWorkspaceSize(desc_rope_q, &temp_size)); +// workspace_size = std::max(workspace_size, temp_size); + +// // K RoPE：k_norm_buf -> k_norm_buf (就地操作) +// RUN_INFINI(infiniopCreateRoPEDescriptor( +// rsrc.handle, &desc_rope_k, +// k_norm_3d->desc(), k_norm_3d->desc(), // 输入/输出都是 3D +// pos_ids_buf->desc(), // 位置 ID [ntok] +// rsrc.sin_table->desc(), // 正弦表 [dctx, dh/2] +// rsrc.cos_table->desc())); // 余弦表 [dctx, dh/2] +// RUN_INFINI(infiniopGetRoPEWorkspaceSize(desc_rope_k, &temp_size)); +// workspace_size = std::max(workspace_size, temp_size); +// printf("Debug: RoPE descriptors created successfully\n"); + +/* + * 注意力输出投影 GEMM：o_buf * w_attn_o_proj -> logits_in + * + * 矩阵乘法：Y = X * W + * 输入 X：[ntok, nh/ndev * dh] - 此设备上所有头的注意力输出 + * 权重 W：[nh/ndev * dh, d] - 输出投影权重 + * 输出 Y：[ntok, d] - 投影的注意力输出（将在设备间累积） + */ +RUN_INFINI(infiniopCreateGemmDescriptor(rsrc.handle, &desc_attn_o, + logits_in->desc(), o_buf->desc(), + rsrc.w_attn_o_proj[0]->desc())); + +// 计算 GEMM 操作的工作区需求 +RUN_INFINI(infiniopGetGemmWorkspaceSize(desc_attn_o, &temp_size)); +workspace_size = std::max(workspace_size, temp_size); + +/* + * 每请求注意力内循环描述符 + * + * 由于批处理中的每个请求可能有不同的序列长度和 KV 缓存状态， + * 我们需要为每个请求的注意力计算单独的描述符。 + * + * 每个请求的注意力机制包括： + * 1. 分离的 QKV 投影：X -> Q、K、V (3个独立 GEMM) + * 2. Q/K 归一化：Q -> Q_norm, K -> K_norm (Qwen3 特有) + * 3. RoPE 应用到归一化后的 Q/K：Q_norm/K_norm -> 位置编码 + * 更新缓存（如果存在） + * 4. 注意力计算：softmax(Q_norm * K_norm^T / √d_k) * V + * 5. 输出投影：O -> 最终注意力输出 + * + * 关键优化： + * - 分组查询注意力：多个查询头可以共享 KV 头（ngroup = nh/nkvh） + * - KV 缓存：存储和重用过去的键值对 + * - 因果掩码：未来 token 不能关注过去的 token + */ +// 注意力内层 +auto desc_kv_rearranges = std::vector(nreq); +auto desc_q_rearranges = std::vector(nreq); +auto desc_qk_gemms = std::vector(nreq); +auto desc_qk_softmaxs = std::vector(nreq); +auto desc_attn_v_gemms = std::vector(nreq); +auto desc_attn_v_rearranges = std::vector(nreq); + +size_t token_offset = 0; // 跟踪当前请求在批处理中的位置 +size_t max_qk_size = 0; // 用于缓冲区分配的最大 QK 矩阵大小 +size_t max_seq_len = 0; // 用于缓冲区分配的最大序列长度 + +/* + * 为每个请求的注意力计算创建描述符 + * 每个请求可能有不同的序列长度和过去的 KV 缓存长度， + * 需要单独的描述符以获得最佳的内存布局和计算。 + */ +for (uint32_t req = 0; req < nreq; req++) { + auto past_len = req_pos[req]; // KV 缓存中已有的 token 数 + auto seq_len = req_lens[req]; // 要处理的当前序列长度 + auto total_len = past_len + seq_len; // KV 缓存中的总序列长度 + + /* + * 从批处理张量中提取每请求张量切片 + * + * 此请求的张量形状： + * - o：[seq_len, nh, dh] - 此请求的注意力输出 + * - q：[seq_len, nh, dh] - 此请求的查询向量 + * - k：[seq_len, nkvh, dh] - 此请求的键向量 + * - v：[seq_len, nkvh, dh] - 此请求的值向量（稍后使用） + */ + + // 步骤1：从2D缓冲区切片出当前请求的部分 + // 步骤1：从2D缓冲区切片出当前请求的部分 + auto o_2d = o_buf->slice({{0, token_offset, seq_len}}); // [seq_len, nh*dh] + auto q_2d = q_norm_buf->slice( + {{0, token_offset, seq_len}}); // [seq_len, nh*dh] - 使用归一化后的Q + auto k_2d = k_norm_buf->slice( + {{0, token_offset, seq_len}}); // [seq_len, nkvh*dh] - 使用归一化后的K + + // 步骤2：将2D切片分割成3D多头格式 + auto o = o_2d->dimSplit(1, {nh, dh}); // [seq_len, nh*dh] -> [seq_len, nh, dh] + auto q = q_2d->dimSplit( + 1, {nh, dh}); // [seq_len, nh*dh] -> [seq_len, nh, dh] - 从归一化后的Q切片 + auto k = k_2d->dimSplit(1, {nkvh, dh}); + auto v = v_buf->slice({{0, token_offset, seq_len}}); // [seq_len, nkvh, dh] + + // auto v = qkv_buf->slice({{0, token_offset, seq_len}, {1, nh + nkvh, + // nkvh}}); + + /* + * KV 缓存张量配置 + * KV 缓存存储过去的键值对以实现高效的自回归生成。 + * 形状：[total_len, nkvh, dh] 在内存中存储为 [nkvh, dh, total_len] + * full_kv：包括过去 + 当前 token 的完整 KV 缓存 [nkvh, dh, total_len] + * cache_kv：用于存储当前键/值的切片 [nkvh, dh, seq_len] + */ + // kv 缓存张量可以共享相同的描述符 + // [nkvh, dh, total_len] + auto full_kv = + kv_caches[req]->k[idev][0]->slice(0, 0, total_len)->permute({1, 2, 0}); + auto cache_kv = kv_caches[req]->k[idev][0]->slice(0, past_len, seq_len); + /* + * KV 重新排列描述符：将当前 K/V 存储在缓存中 + * 将当前键/值转换为缓存存储格式： + * k：[seq_len, nkvh, dh] -> cache_kv：[seq_len, nkvh, dh]（不同的内存布局） + */ + RUN_INFINI(infiniopCreateRearrangeDescriptor( + rsrc.handle, &desc_kv_rearranges[req], cache_kv->desc(), k->desc())); + + /* + * 分组查询注意力（GQA）的查询重新排列 + * + * 重新塑造查询以实现高效的 GQA 计算： + * q：[seq_len, nh, dh] -> [seq_len0, nkvh1, ngroup2, dh3] -> [nkvh1, ngroup2, + * seq_len0, dh3] 此布局允许每个 KV 头关注多个查询头（每个 KV 头 ngroup + * 个查询头） + */ + // [nkvh, ngroup, seq_len, dh] + q->dimSplit(1, {nkvh, ngroup})->permute({1, 2, 0, 3}); + auto q_t = TensorDesc::create(dt_logits, {nkvh, ngroup, seq_len, dh}); + // [seq_len, nkvh, ngroup, dh] -> [nkvh, ngroup, seq_len, dh] + RUN_INFINI(infiniopCreateRearrangeDescriptor( + rsrc.handle, &desc_q_rearranges[req], q_t->desc(), q->desc())); + /* + * 注意力值重新排列描述符 + * 在计算 attention_weights * values 后，重新排列回标准格式： + * [nkvh, ngroup, seq_len, dh] -> [seq_len, nkvh, ngroup, dh] -> [seq_len, nh, + * dh] + */ + // [nkvh, ngroup, seq_len, dh] -> [seq_len, nkvh, ngroup, dh] + auto attn_v_t = q_t; + auto attn_v = TensorDesc::createWithOrder( + dt_logits, {nkvh, ngroup, seq_len, dh}, {1, 2, 0, 3}); + RUN_INFINI(infiniopCreateRearrangeDescriptor( + rsrc.handle, &desc_attn_v_rearranges[req], attn_v->desc(), + attn_v_t->desc())); + + /* + * QK 注意力分数计算：Q * K^T / √d_k + * + * 矩阵乘法计算注意力分数： + * Q：[nkvh, ngroup * seq_len, dh]（为批处理计算重新塑造） + * K^T：[nkvh, dh, total_len]（完整的 KV 缓存转置） + * QK：[nkvh, ngroup * seq_len, total_len]（softmax 前的注意力分数） + * + * 缩放因子 1/√d_k 在 GEMM 操作期间应用。 + */ + q_t = TensorDesc::create(dt_logits, {nkvh, ngroup * seq_len, dh}); + auto qk = TensorDesc::create(dt_logits, {nkvh, ngroup * seq_len, total_len}); + max_qk_size = std::max(max_qk_size, size_t(seq_len * total_len)); + max_seq_len = std::max(max_seq_len, size_t(seq_len)); + RUN_INFINI(infiniopCreateGemmDescriptor(rsrc.handle, &desc_qk_gemms[req], + qk->desc(), q_t->desc(), + full_kv->desc())); + RUN_INFINI(infiniopGetGemmWorkspaceSize(desc_qk_gemms[req], &temp_size)); + workspace_size = std::max(workspace_size, temp_size); + + /* + * 注意力值计算：attention_weights * V + * + * 与注意力加权值的矩阵乘法： + * attention_weights：[nkvh, ngroup * seq_len, total_len]（softmax 后） + * V：[nkvh, total_len, dh]（完整的值缓存） + * output：[nkvh, ngroup * seq_len, dh]（注意力输出） + */ + // [nkvh, total_len, dh] + auto full_v = + kv_caches[req]->v[idev][0]->slice(0, 0, total_len)->permute({1, 0, 2}); + RUN_INFINI(infiniopCreateGemmDescriptor(rsrc.handle, &desc_attn_v_gemms[req], + q_t->desc(), qk->desc(), + full_v->desc())); + RUN_INFINI(infiniopGetGemmWorkspaceSize(desc_attn_v_gemms[req], &temp_size)); + workspace_size = std::max(workspace_size, temp_size); + + /* + * 带注意力掩码的因果 Softmax + * + * 将 softmax 应用于注意力分数并带有因果掩码（下三角）。 + * 形状：[nkvh * ngroup, seq_len, total_len] + * + * 因果掩码确保每个 token 只能关注之前的 token 和自身， + * 防止在自回归生成期间未来 token 的信息泄露。 + */ + qk = TensorDesc::create(dt_logits, {nkvh * ngroup, seq_len, total_len}); + RUN_INFINI(infiniopCreateCausalSoftmaxDescriptor( + rsrc.handle, &desc_qk_softmaxs[req], qk->desc(), qk->desc())); + RUN_INFINI( + infiniopGetCausalSoftmaxWorkspaceSize(desc_qk_softmaxs[req], &temp_size)); + workspace_size = std::max(workspace_size, temp_size); + + token_offset += seq_len; +} + +/* + * 分配注意力中间缓冲区 + * + * 这些缓冲区存储注意力计算期间的中间结果。 + * 大小基于批处理中所有请求的最大需求。 + * + * 缓冲区形状： + * - qk_buf：[nh, max_qk_size] - 最大请求的注意力分数（QK^T） + * - rearrange_q_buf：[nkvh, ngroup * max_seq_len, dh] - 重新排列的查询 + * - attn_val_buf：[nh, max_seq_len, dh] - 注意力输出值 + */ +auto qk_buf = Tensor::buffer(dt_logits, {nh, max_qk_size}, rsrc.memory_pool); +auto rearrange_q_buf = Tensor::buffer( + dt_logits, {nkvh, ngroup *max_seq_len, dh}, rsrc.memory_pool); +auto attn_val_buf = + Tensor::buffer(dt_logits, {nh, max_seq_len, dh}, rsrc.memory_pool); + +/* + * 前馈网络（FFN）描述符 + * + * FFN 块实现 SwiGLU 激活函数： + * FFN(x) = (Swish(x * W_gate) ⊙ (x * W_up)) * W_down + * 其中 Swish(x) = x * sigmoid(x) 且 ⊙ 是逐元素乘法 + * + * 这涉及四个矩阵乘法： + * 1. Gate 投影：X * W_gate -> gate_buf + * 2. Up 投影：X * W_up -> up_buf + * 3. SwiGLU 激活：gate_buf * swish(up_buf) -> gate_buf（复用） + * 4. Down 投影：activated_output * W_down -> final_output + */ +// MLP 描述符 +infiniopGemmDescriptor_t desc_mlp_gate, desc_mlp_up, desc_mlp_down; +infiniopSwiGLUDescriptor_t desc_swiglu; + +/* + * 分离的 MLP Gate 投影：logits_out [x] * w_mlp_gate_proj -> gate_buf + * + * 矩阵乘法：Y = X * W + * 输入 X：[ntok, d] - 来自注意力的归一化隐藏状态 + * 权重 W：[d, di/ndev] - Gate 投影权重 + * 输出 Y：[ntok, di/ndev] - Gate 投影输出 + */ +RUN_INFINI(infiniopCreateGemmDescriptor(rsrc.handle, &desc_mlp_gate, + gate_buf->desc(), logits_out->desc(), + rsrc.w_mlp_gate_proj[0]->desc())); +RUN_INFINI(infiniopGetGemmWorkspaceSize(desc_mlp_gate, &temp_size)); +workspace_size = std::max(workspace_size, temp_size); + +/* + * 分离的 MLP Up 投影：logits_out * w_mlp_up -> up_buf + * + * 矩阵乘法：Y = X * W + * 输入 X：[ntok, d] - 来自注意力的归一化隐藏状态 + * 权重 W：[d, di/ndev] - Up 投影权重 + * 输出 Y：[ntok, di/ndev] - Up 投影输出 + */ +RUN_INFINI(infiniopCreateGemmDescriptor(rsrc.handle, &desc_mlp_up, + up_buf->desc(), logits_out->desc(), + rsrc.w_mlp_up_proj[0]->desc())); +RUN_INFINI(infiniopGetGemmWorkspaceSize(desc_mlp_up, &temp_size)); +workspace_size = std::max(workspace_size, temp_size); + +/* + * MLP Down 投影：swiglu_output * w_mlp_down -> logits_in + * + * 矩阵乘法：Y = X * W + * 输入 X：[ntok, di/ndev] - 来自 SwiGLU 的激活值 + * 权重 W：[di/ndev, d] - Down 投影权重 + * 输出 Y：[ntok, d] - 投影回模型维度 + */ +RUN_INFINI(infiniopCreateGemmDescriptor(rsrc.handle, &desc_mlp_down, + logits_in->desc(), gate_buf->desc(), + rsrc.w_mlp_down_proj[0]->desc())); +RUN_INFINI(infiniopGetGemmWorkspaceSize(desc_mlp_down, &temp_size)); +workspace_size = std::max(workspace_size, temp_size); + +RUN_INFINI( + infiniopCreateSwiGLUDescriptor(rsrc.handle, &desc_swiglu, + gate_buf->desc(), // 输出缓冲区 + gate_buf->desc(), // gate 输入（复用缓冲区） + up_buf->desc())); // up 输入 +RUN_INFINI(infiniopGetSwiGLUWorkspaceSize(desc_swiglu, &temp_size)); +workspace_size = std::max(workspace_size, temp_size); + +/* + * 输出生成和 Token 采样描述符 + * + * 在所有 transformer 层之后，我们需要： + * 1. 对每个请求的最后一个 token 应用最终层归一化 + * 2. 投影到词汇表空间以获得下一个 token 预测的 logits + * 3. 应用采样（温度、top-k、top-p）来选择下一个 token + */ +// 输出和采样 +infiniopRMSNormDescriptor_t desc_norm_out; + +/* + * 最终输出归一化 + * + * 对每个请求的最后一个 token 的隐藏状态应用 RMSNorm。 + * 这在投影到词汇表之前归一化最终表示。 + * + * 输入/输出形状：[1, d] -> [1, d]（一次处理一个请求） + * 权重形状：[d] - 最终层归一化参数 + */ +RUN_INFINI(infiniopCreateRMSNormDescriptor( + rsrc.handle, &desc_norm_out, logits_out->slice(0, 0, 1)->desc(), + logits_out->slice(0, 0, 1)->desc(), rsrc.w_out_norm->desc(), meta.epsilon)); +RUN_INFINI(infiniopGetRMSNormWorkspaceSize(desc_norm_out, &temp_size)); +workspace_size = std::max(workspace_size, temp_size); + +/* + * 语言模型头投影：hidden_states -> vocabulary_logits + * + * 将归一化的隐藏状态投影到词汇表空间以获得下一个 token 预测的 logits。 + * + * 矩阵乘法：Y = X * W + * 输入 X：[nreq, d] - 每个请求的最终隐藏状态 + * 权重 W：[d, dvoc] - 语言模型头权重（通常与输入嵌入绑定） + * 输出 Y：[nreq, dvoc] - 每个请求的词汇表上的 logits + * + * 这些 logits 代表每个可能的下一个 token 的未归一化对数概率。 + */ +infiniopGemmDescriptor_t desc_out_embd; +RUN_INFINI(infiniopCreateGemmDescriptor(rsrc.handle, &desc_out_embd, + prob_buf->desc(), + logits_out->slice(0, 0, nreq)->desc(), + rsrc.w_out_embd->desc())); +RUN_INFINI(infiniopGetGemmWorkspaceSize(desc_out_embd, &temp_size)); +workspace_size = std::max(workspace_size, temp_size); + +/* + * 随机采样描述符 + * + * 执行温度缩放、top-k 过滤、top-p（核）采样 + * 从概率分布中选择下一个 token。 + * + * 采样过程： + * 1. 应用温度缩放：logits = logits / temperature + * 2. 应用 top-k 过滤：仅保留 k 个最高概率 token + * 3. 应用 top-p 过滤：保留 token 直到累积概率 >= p + * 4. 从过滤的分布中采样 + * + * 输入：[dvoc] - 一个请求的词汇表上的 logits + * 输出：标量 int64 - 选定的 token ID + */ +infiniopRandomSampleDescriptor_t desc_sample; +RUN_INFINI(infiniopCreateRandomSampleDescriptor( + rsrc.handle, &desc_sample, + TensorDesc::create(INFINI_DTYPE_I64, {}, {})->desc(), // 输出：标量 token ID + TensorDesc::create(dt_logits, {dvoc}, {1})->desc())); // 输入：[dvoc] logits +RUN_INFINI(infiniopGetRandomSampleWorkspaceSize(desc_sample, &temp_size)); +workspace_size = std::max(workspace_size, temp_size); + +/* + * 工作区分配 + * + * 分配单个工作区缓冲区，可以处理所有操作的最大内存 + * 需求。这避免了推理期间频繁分配 + * 并确保高效的内存使用。 + */ +// 分配工作区 +std::shared_ptr workspace_storage = + Storage::createFromPool(workspace_size, rsrc.memory_pool); +void *workspace = workspace_storage->memory(); + +/* + * ================================================================================== + * 主 TRANSFORMER 推理计算循环 + * ================================================================================== + * + * 此部分执行通过所有 transformer 层的实际前向传递。 + * 每层包括： + * 1. 带残差连接的多头注意力 + * 2. 带残差连接的前馈网络 + * + * 计算遵循标准 transformer 架构： + * x = x + Attention(LayerNorm(x)) + * x = x + FFN(LayerNorm(x)) + * + * 对于分布式推理，注意力和 FFN 输出通过 all-reduce 操作 + * 在设备间累积。 + */ + +for (uint32_t layer = 0; layer < nlayer; layer++) { + /* + * ============================================================================ + * 多头注意力块 + * ============================================================================ + */ + + // Debug: Save layer input (only for layer 0) + if (g_debug_enabled && layer == 0) { + save_tensor_debug_f16(logits_in, "input_hidden_states", layer); + } + + // 1. 注意力 + + /* + * 注意力前层归一化 + * + * 在注意力计算前对输入隐藏状态应用 RMSNorm。 + * 这遵循 "Pre-LN" transformer 架构以获得更好的训练稳定性。 + * + * 公式：y = x / √(mean(x²) + ε) * γ + * 输入：logits_in [ntok, d] - 来自前一层/嵌入的隐藏状态 + * 输出：logits_out [ntok, d] - 用于注意力的归一化隐藏状态 + * 权重：w_attn_norm[layer] [d] - 可学习的缩放参数 + */ + // rms 归一化 + RUN_INFINI(infiniopRMSNorm(desc_norm, workspace, workspace_size, + logits_out->data(), logits_in->data(), + rsrc.w_attn_norm[layer]->data(), stream)); + + // Debug: Save attention norm output (only for layer 0) + if (g_debug_enabled && layer == 0) { + save_tensor_debug_f32(logits_out, "attn_norm_output", layer); + } + /* + * Qwen3 注意力计算：分离 QKV + Q/K 归一化 + * + 同一输入分别乘三条权重 → 得到 Q、K、V + reshape 成多头 + RoPE 只给 Q、K 加位置编码 + KV 复制到与 Q 头数一致 + softmax(QKᵀ)V → 合并 → 输出投影 + */ + + // ============================================================================ + // 第一步：分离的 QKV 投影（替换原来的融合投影） + // ============================================================================ + + /* + * Q 投影：logits_out * w_attn_q_proj -> q_buf + */ + RUN_INFINI(infiniopGemm(desc_attn_q, workspace, workspace_size, q_buf->data(), + logits_out->data(), rsrc.w_attn_q_proj[layer]->data(), + 1.0, 0.0, stream)); + + /* + * K 投影：logits_out * w_attn_k_proj -> k_buf + */ + RUN_INFINI(infiniopGemm(desc_attn_k, workspace, workspace_size, k_buf->data(), + logits_out->data(), rsrc.w_attn_k_proj[layer]->data(), + 1.0, 0.0, stream)); + + /* + * V 投影：logits_out * w_attn_v_proj -> v_buf + */ + RUN_INFINI(infiniopGemm(desc_attn_v, workspace, workspace_size, v_buf->data(), + logits_out->data(), rsrc.w_attn_v_proj[layer]->data(), + 1.0, 0.0, stream)); + + // Debug: Save QKV projections (only for layer 0) + if (g_debug_enabled && layer == 0) { + save_tensor_debug_f32(q_buf, "attn_q_proj_raw", layer); + save_tensor_debug_f32(k_buf, "attn_k_proj_raw", layer); + save_tensor_debug_f32(v_buf, "attn_v_proj_raw", layer); + + // Validate QKV projections before normalization + validate_tensor_range(q_buf, "q_buf before normalization"); + validate_tensor_range(k_buf, "k_buf before normalization"); + } + + // ============================================================================ + // 第二步：Qwen3 特有的 Q/K 归一化 + // ============================================================================ + + /* + * Q 归一化：q_buf -> q_norm_buf + * 这是 Qwen3 特有的步骤，在 RoPE 之前对 Q 进行归一化 + */ + for (size_t head = 0; head < nh; head++) { + // 计算当前头在缓冲区中的偏移（以 dh 为单位） + size_t head_offset = head * dh; + + // 获取当前头的数据指针 + void *q_head_input = (char *)q_buf->data() + head_offset * dsize(dt_logits); + void *q_head_output = + (char *)q_norm_buf->data() + head_offset * dsize(dt_logits); + + // 对当前头应用归一化 + RUN_INFINI(infiniopRMSNorm(desc_q_norm_single, workspace, workspace_size, + q_head_output, // 输出：当前头的归一化结果 + q_head_input, // 输入：当前头的原始数据 + rsrc.w_attn_q_norm[layer]->data(), // 权重：[dh] + stream)); + } + + /* + * K 归一化：k_buf -> k_norm_buf + * 同样对 K 进行归一化 + */ + for (size_t head = 0; head < nkvh; head++) { + + // 计算当前头在缓冲区中的偏移 + size_t head_offset = head * dh; + + // 获取当前头的数据指针 + void *k_head_input = (char *)k_buf->data() + head_offset * dsize(dt_logits); + void *k_head_output = + (char *)k_norm_buf->data() + head_offset * dsize(dt_logits); + + // 对当前头应用归一化 + RUN_INFINI(infiniopRMSNorm(desc_k_norm_single, workspace, workspace_size, + k_head_output, // 输出：当前头的归一化结果 + k_head_input, // 输入：当前头的原始数据 + rsrc.w_attn_k_norm[layer]->data(), // 权重：[dh] + stream)); + } + + // Debug: Save Q/K normalized outputs (only for layer 0) + if (g_debug_enabled && layer == 0) { + // Validate normalized outputs + validate_tensor_range(q_norm_buf, "attn_q_normed", -10.0, 10.0); + validate_tensor_range(k_norm_buf, "attn_k_normed", -10.0, 10.0); + + // Clamp extreme values to prevent downstream issues + clamp_tensor_inplace(q_norm_buf, -10.0f, 10.0f); + clamp_tensor_inplace(k_norm_buf, -10.0f, 10.0f); + + save_tensor_debug_f32(q_norm_buf, "attn_q_normed", layer); + save_tensor_debug_f32(k_norm_buf, "attn_k_normed", layer); + } + // ============================================================================ + // 第三步：按头应用 RoPE + // ============================================================================ + + for (size_t head = 0; head < nh; head++) { + + // 计算当前头在缓冲区中的偏移 + size_t head_offset = head * dh; + + // 获取当前头的数据指针（归一化后的） + void *q_head_data = + (char *)q_norm_buf->data() + head_offset * dsize(dt_logits); + + // 对当前头应用 RoPE（就地操作） + RUN_INFINI(infiniopRoPE(desc_rope_q_single, workspace, workspace_size, + q_head_data, q_head_data, // 就地操作 + pos_ids_buf->data(), rsrc.sin_table->data(), + rsrc.cos_table->data(), stream)); + } + + /* + * 按头处理 K RoPE + */ + for (size_t head = 0; head < nkvh; head++) { + + // 计算当前头在缓冲区中的偏移 + size_t head_offset = head * dh; + + // 获取当前头的数据指针（归一化后的） + void *k_head_data = + (char *)k_norm_buf->data() + head_offset * dsize(dt_logits); + + // 对当前头应用 RoPE（就地操作） + RUN_INFINI(infiniopRoPE(desc_rope_k_single, workspace, workspace_size, + k_head_data, k_head_data, // 就地操作 + pos_ids_buf->data(), rsrc.sin_table->data(), + rsrc.cos_table->data(), stream)); + } + // ============================================================================ + // 第四步：每请求注意力计算（更新张量切片源） + // ============================================================================ + + size_t token_offset = 0; + for (uint32_t req = 0; req < nreq; req++) { + auto past_len = req_pos[req]; + auto seq_len = req_lens[req]; + + /* + * 从归一化后的缓冲区提取张量切片 + * 关键变化：使用 q_norm_buf, k_norm_buf, v_buf 而不是 qkv_buf 切片 + */ + auto o = o_buf->slice({{0, token_offset, seq_len}}); + auto q = + q_norm_buf->slice({{0, token_offset, seq_len}}); // 使用归一化后的 Q + auto k = + k_norm_buf->slice({{0, token_offset, seq_len}}); // 使用归一化后的 K + auto v = v_buf->slice({{0, token_offset, seq_len}}); // V 不需要额外归一化 + + /* + * 其余的注意力计算保持不变 + * 包括：KV 缓存更新、GQA 重排、QK 计算、softmax、注意力值计算 + */ + + // KV 缓存更新（K 使用归一化后的值） + RUN_INFINI(infiniopRearrange( + desc_kv_rearranges[req], + kv_caches[req]->k[idev][layer]->data(past_len * nkvh * dh), k->data(), + stream)); // k 现在是归一化后的 + + RUN_INFINI(infiniopRearrange( + desc_kv_rearranges[req], + kv_caches[req]->v[idev][layer]->data(past_len * nkvh * dh), v->data(), + stream)); // v 保持不变 + + // GQA 查询重排（Q 使用归一化后的值） + RUN_INFINI(infiniopRearrange(desc_q_rearranges[req], + rearrange_q_buf->data(), q->data(), + stream)); // q 现在是归一化后的 + + // QK 注意力分数计算 + RUN_INFINI(infiniopGemm(desc_qk_gemms[req], workspace, workspace_size, + qk_buf->data(), rearrange_q_buf->data(), + kv_caches[req]->k[idev][layer]->data(), + 1. / sqrt(dh), 0.0, stream)); + + // 因果 Softmax + RUN_INFINI(infiniopCausalSoftmax(desc_qk_softmaxs[req], workspace, + workspace_size, qk_buf->data(), + qk_buf->data(), stream)); + + // 注意力值计算 + RUN_INFINI(infiniopGemm(desc_attn_v_gemms[req], workspace, workspace_size, + attn_val_buf->data(), qk_buf->data(), + kv_caches[req]->v[idev][layer]->data(), 1.0, 0.0, + stream)); + + // 输出重排 + RUN_INFINI(infiniopRearrange(desc_attn_v_rearranges[req], o->data(), + attn_val_buf->data(), stream)); + + token_offset += seq_len; + } + + /* + * 注意力输出投影和残差连接 + * + * 将注意力输出投影回模型维度并添加残差连接。 + * 在分布式推理中，仅设备 0 添加残差连接以避免 + * 跨设备重复计算。 + * + * 矩阵运算：Y = X * W + (如果 idev == 0 则残差 else 0) + * 输入：o_buf [ntok, nh/ndev * dh] - 来自此设备的注意力输出 + * 权重：w_attn_out[layer] [nh/ndev * dh, d] - 输出投影权重 + * 输出：logits_in [ntok, d] - 带残差连接的投影输出 + */ + // o_proj + RUN_INFINI(infiniopGemm( + desc_attn_o, workspace, workspace_size, logits_in->data(), o_buf->data(), + rsrc.w_attn_o_proj[layer]->data(), 1.0, idev == 0 ? 1.0 : 0.0, stream)); + // 残差：仅 rank 0 添加原始输入 + // 如果每个设备都添加完整的残差，最终结果会是：output = partial_outputs + ndev + // * residual (错误) + + /* + * 用于多设备推理的分布式 All-Reduce + * + * 在所有设备间求和注意力输出以完成分布式计算。 + * 每个设备计算了注意力头的一个切片，结果必须 + * 组合以获得完整的注意力输出。 + * + * 操作：logits_in = sum(logits_in_device_i) 对于 i 在 [0, ndev) 范围内 + * 这同步所有设备并确保集群中的一致状态。 + */ + // 如果分布式则 All_reduce + if (rsrc.comm != nullptr) { + RUN_INFINI(infinicclAllReduce(logits_in->data(), logits_in->data(), + ntok * d, dt_logits, INFINICCL_SUM, rsrc.comm, + stream)); + RUN_INFINI(infinirtStreamSynchronize(stream)); // 通信后同步 + } + + // Debug: Save attention residual output (only for layer 0) + if (g_debug_enabled && layer == 0) { + save_tensor_debug_f32(logits_in, "attn_residual_output", layer); + } + + /* + * ============================================================================ + * 带 SwiGLU 激活的前馈网络（FFN）块 + * ============================================================================ + */ + + /* + * FFN 前层归一化 + * + * 在 FFN 计算前对注意力输出应用 RMSNorm。 + * + * 公式：y = x / √(mean(x²) + ε) * γ + * 输入：logits_in [ntok, d] - 注意力输出 + 残差 + * 输出：logits_out [ntok, d] - 用于 FFN 处理的归一化 + * 权重：w_mlp_norm[layer] [d] - 可学习的缩放参数 + */ + // rms_norm + RUN_INFINI(infiniopRMSNorm(desc_norm, workspace, workspace_size, + logits_out->data(), logits_in->data(), + rsrc.w_mlp_norm[layer]->data(), stream)); + + // Debug: Save MLP norm output (only for layer 0) + if (g_debug_enabled && layer == 0) { + save_tensor_debug_f32(logits_out, "mlp_norm_output", layer); + } + + /* + * Qwen3 MLP 计算流程 + * + * 计算顺序： + * 1. gate = x @ W_gate # Gate 投影（无激活） + * 2. up = x @ W_up # Up 投影（无激活） + * 3. intermediate = gate ⊙ SiLU(up) # SwiGLU: gate * SiLU(up) + * 4. output = intermediate @ W_down + * + * 对应 SwiGLU 公式：c_i = a_i ⊙ SiLU(b_i) + * - a_i: gate (门控输入) + * - b_i: up (要应用 SiLU 的输入) + * - c_i: intermediate (输出结果) + */ + + // 步骤 1: Gate 投影（无激活函数） + RUN_INFINI(infiniopGemm(desc_mlp_gate, workspace, workspace_size, + gate_buf->data(), logits_out->data(), + rsrc.w_mlp_gate_proj[layer]->data(), 1.0, 0.0, + stream)); + + // 步骤 2: Up 投影（无激活函数） + RUN_INFINI(infiniopGemm(desc_mlp_up, workspace, workspace_size, + up_buf->data(), logits_out->data(), + rsrc.w_mlp_up_proj[layer]->data(), 1.0, 0.0, stream)); + + // Debug: Save MLP Gate and Up projections + if (g_debug_enabled && layer == 0) { + save_tensor_debug_f32(gate_buf, "mlp_gate_proj", layer); + save_tensor_debug_f32(up_buf, "mlp_up_proj", layer); + } + + // 步骤 3: SwiGLU 激活 - 一次性完成 gate ⊙ SiLU(up) + RUN_INFINI(infiniopSwiGLU( + desc_swiglu, workspace, workspace_size, + gate_buf->data(), // 输出 (c): intermediate = gate ⊙ SiLU(up) + up_buf->data(), // 门控输入 (a): up_buf 投影结果 + gate_buf->data(), // SiLU输入 (b): gate_buf 投影结果，将应用 SiLU + stream)); + + // Debug: Save MLP intermediate (after SwiGLU) (only for layer 0) + if (g_debug_enabled && layer == 0) { + save_tensor_debug_f32(gate_buf, "mlp_intermediate", layer); + } + + // 步骤 4: Down 投影和残差连接 + RUN_INFINI(infiniopGemm(desc_mlp_down, workspace, workspace_size, + logits_in->data(), gate_buf->data(), + rsrc.w_mlp_down_proj[layer]->data(), 1.0, + idev == 0 ? 1.0 : 0.0, stream)); + + /* + * FFN 下降投影和残差连接 + * + * 将激活的 FFN 输出投影回模型维度并添加残差。 + * 与注意力一样，仅分布式推理中的设备 0 添加残差。 + * + * 矩阵运算：Y = X * W + (如果 idev == 0 则残差 else 0) + * 输入：gate_buf [ntok, di/ndev] - SwiGLU 激活值 + * 权重：w_ffn_down[layer] [di/ndev, d] - 下降投影权重 + * 输出：logits_in [ntok, d] - 带残差连接的 FFN 输出 + */ + RUN_INFINI(infiniopGemm( + desc_mlp_down, workspace, workspace_size, logits_in->data(), + gate_buf->data(), rsrc.w_mlp_down_proj[layer]->data(), 1.0, + idev == 0 ? 1.0 : 0.0, stream)); // 残差：仅 rank 0 添加原始输入 + + /* + * FFN 输出的分布式 All-Reduce + * + * 在所有设备间求和 FFN 输出以完成分布式计算。 + * 每个设备计算了中间 FFN 维度的一个切片。 + */ + // 如果分布式则 All_reduce + if (rsrc.comm != nullptr) { + RUN_INFINI(infinicclAllReduce(logits_in->data(), logits_in->data(), + ntok * d, dt_logits, INFINICCL_SUM, rsrc.comm, + stream)); + RUN_INFINI(infinirtStreamSynchronize(stream)); // 通信后同步 + } + + // Debug: Save layer output (only for layer 0) + if (g_debug_enabled && layer == 0) { + save_tensor_debug_f32(logits_in, "layer_output", layer); + } +} + +/* + * ================================================================================== + * 输出生成和 TOKEN 采样 + * ================================================================================== + * + * 在处理完所有 transformer 层后，通过以下方式生成下一个 token： + * 1. 对每个请求的最后一个 token 应用最终层归一化 + * 2. 投影到词汇表空间以获得 logits + * 3. 使用温度、top-k 和 top-p 过滤采样下一个 token + * + * 仅设备 0 执行采样以避免重复计算。 + */ +// 采样和输出 +if (idev == 0) { + /* + * 每个请求的最终层归一化 + * + * 对每个请求的最后一个 token 的隐藏状态应用 RMSNorm。 + * 最后一个 token 用于自回归生成中的下一个 token 预测。 + * + * 对于每个请求，提取最后一个 token 的隐藏状态： + * - token_offset 跟踪批处理中的累积位置 + * - 处理 req_lens[req] 个 token 后，最后一个 token 位于位置 (token_offset - + * 1) + */ + size_t token_offset = 0; + for (uint32_t req = 0; req < nreq; req++) { + auto seq_len = req_lens[req]; + token_offset += seq_len; + + /* + * 为此请求归一化最后一个 token 的隐藏状态 + * + * 输入：logits_in[(token_offset-1)*d : (token_offset)*d] - 最后一个 token + * 的隐藏状态 [d] 输出：logits_out[req*d : (req+1)*d] - + * 用于词汇表投影的归一化状态 [d] 权重：w_out_norm [d] - 最终层归一化参数 + */ + RUN_INFINI(infiniopRMSNorm( + desc_norm_out, workspace, workspace_size, + logits_out->data(req * d), // 输出：请求 req 的 [d] + logits_in->data((token_offset - 1) * d), // 输入：最后一个 token [d] + rsrc.w_out_norm->data(), stream)); + } + + /* + * 语言模型头投影 + * + * 将归一化的最终隐藏状态投影到词汇表空间以获得 logits + * 用于下一个 token 预测。 + * + * 矩阵运算：logits = hidden_states * W_lm_head + * 输入：logits_out [nreq, d] - 归一化的最终隐藏状态 + * 权重：w_out_embd [d, dvoc] - 语言模型头（通常与输入嵌入绑定） + * 输出：prob_buf [nreq, dvoc] - 词汇表上的未归一化 logits + */ + RUN_INFINI(infiniopGemm(desc_out_embd, workspace, workspace_size, + prob_buf->data(), logits_out->data(), + rsrc.w_out_embd->data(), 1.0, 0.0, stream)); + + /* + * 带温度和过滤的 Token 采样 + * + * 对于每个请求，从概率分布中采样下一个 token + * 使用温度缩放、top-k 过滤和 top-p（核）采样。 + * + * 采样过程： + * 1. 应用温度缩放：logits = logits / temperature + * 2. 应用 top-k：仅保留 k 个最高概率 token + * 3. 应用 top-p：保留 token 直到累积概率 >= p + * 4. 使用随机值从过滤的分布中采样 + */ + std::random_device _rd; + std::mt19937 gen(_rd()); + token_offset = 0; + + for (uint32_t req = 0; req < nreq; req++) { + auto seq_len = req_lens[req]; + + // 生成用于采样的随机值 [0, 1) + float random_val = std::uniform_real_distribution(0, 1)(gen); + + /* + * 为此请求采样下一个 token + * + * 输入：prob_buf[req*dvoc : (req+1)*dvoc] - 词汇表上的 logits [dvoc] + * 输出：result_buf[req] - 采样的 token ID + * 参数： + * - random_val：用于采样的随机种子 + * - topp[req]：核采样阈值（累积概率） + * - topk[req]：top-k 过滤（保留前 k 个 token） + * - temperature[req]：logits 的缩放因子（更高 = 更随机） + */ + // prob_buf->debug(); + RUN_INFINI(infiniopRandomSample( + desc_sample, workspace, workspace_size, + result_buf->data(req), // 输出：采样的 token ID + prob_buf->data(req * dvoc), // 输入：此请求的 logits [dvoc] + random_val, // 随机种子 + topp[req], topk[req], temperature[req], // 采样参数 + stream)); + // result_buf->debug(); + token_offset += seq_len; + } + + /* + * 将结果复制到主机内存 + * + * 将采样的 token ID 从设备传输到主机内存以返回给调用者。 + * 同步流以确保所有计算在复制前完成。 + */ + RUN_INFINI(infinirtStreamSynchronize(stream)); + RUN_INFINI(infinirtMemcpy(result_cpu.data(), result_buf->data(), + sizeof(int64_t) * nreq, INFINIRT_MEMCPY_D2H)); + + // 将结果存储在输出数组中 + for (uint32_t req = 0; req < nreq; req++) { + output[req] = result_cpu[req]; + } +} + +/* + * ================================================================================== + * 描述符清理和资源释放 + * ================================================================================== + * + * 正确释放所有 InfiniCore 描述符以防止内存泄漏。 + * 描述符必须按依赖关系的相反顺序销毁。 + */ +// 清理 +infiniopDestroyRMSNormDescriptor(desc_norm); // 层归一化 +infiniopDestroyGemmDescriptor(desc_attn_q); // 查询投影 +infiniopDestroyGemmDescriptor(desc_attn_k); // 键投影 +infiniopDestroyGemmDescriptor(desc_attn_v); // 值投影 +infiniopDestroyGemmDescriptor(desc_attn_o); // 注意力输出投影 +infiniopDestroyRoPEDescriptor(desc_rope_q_single); // 查询的 RoPE +infiniopDestroyRoPEDescriptor(desc_rope_k_single); // 键的 RoPE +infiniopDestroyRMSNormDescriptor(desc_q_norm_single); // 单头查询归一化 +infiniopDestroyRMSNormDescriptor(desc_k_norm_single); // 单头键归一化 + +// 清理每请求注意力描述符 +for (uint32_t req = 0; req < nreq; req++) { + infiniopDestroyRearrangeDescriptor(desc_kv_rearranges[req]); // KV 缓存存储 + infiniopDestroyRearrangeDescriptor(desc_q_rearranges[req]); // 查询重新排列 + infiniopDestroyGemmDescriptor(desc_qk_gemms[req]); // QK 注意力分数 + infiniopDestroyCausalSoftmaxDescriptor(desc_qk_softmaxs[req]); // 因果 softmax + infiniopDestroyGemmDescriptor(desc_attn_v_gemms[req]); // 注意力值乘法 + infiniopDestroyRearrangeDescriptor( + desc_attn_v_rearranges[req]); // 输出重新排列 +} + +// 清理 FFN 描述符 +infiniopDestroyGemmDescriptor(desc_mlp_gate); // MLP Gate 投影 +infiniopDestroyGemmDescriptor(desc_mlp_up); // MLP Up 投影 +infiniopDestroyGemmDescriptor(desc_mlp_down); // MLP Down 投影 +infiniopDestroySwiGLUDescriptor(desc_swiglu); // SwiGLU 激活 + +// 清理输出描述符 +infiniopDestroyRMSNormDescriptor(desc_norm_out); // 最终层归一化 +infiniopDestroyGemmDescriptor(desc_out_embd); // 语言模型头 +infiniopDestroyRandomSampleDescriptor(desc_sample); // Token 采样 +} + +/* + * 批处理推理 API 函数（C 接口） + * + * 用于跨多个设备的分布式批处理推理的线程安全包装器。 + * 此函数使用条件变量同步的生产者-消费者模式 + * 协调模型中所有设备的推理。 + * + * 参数： + * - model：包含设备资源和工作线程的 JiugeModel 实例 + * - tokens：输入 token ID [ntok] - 来自所有请求的连接 token + * - ntok：所有请求中的 token 总数 + * - req_lens：每个请求的长度 [nreq] + * - nreq：批处理中的请求数 + * - req_pos：每个请求在 KV 缓存中的起始位置 [nreq] + * - kv_caches：每个请求的 KV 缓存存储 [nreq] + * - temperature/topk/topp：采样参数 [nreq] + * - output：生成的 token ID [nreq] - 由此函数填充 + * + * 线程同步： + * 1. 主线程向所有工作线程发出开始推理信号 + * 2. 工作线程并行处理其分配的设备切片 + * 3. 主线程等待所有工作线程完成后再返回 + */ +__C void inferQwen3Batch(struct Qwen3Model *model, const uint32_t *tokens, + uint32_t ntok, const uint32_t *req_lens, uint32_t nreq, + const uint32_t *req_pos, + struct Qwen3KVCache **kv_caches, + const float *temperature, const uint32_t *topk, + const float *topp, uint32_t *output) { + /* + * 将推理参数复制到模型的请求结构中 + * 这允许工作线程安全地访问请求数据。 + */ + model->req.tokens = tokens; + model->req.ntok = ntok; + model->req.req_lens = req_lens; + model->req.nreq = nreq; + model->req.req_pos = req_pos; + model->req.kv_caches = kv_caches; + model->req.output = output; + model->req.temperature = temperature; + model->req.topk = topk; + model->req.topp = topp; + + /* + * 向所有工作线程发出开始推理信号 + * + * 每个设备都有一个在条件变量上等待的专用工作线程。 + * 设置 proceed=true 并通知唤醒工作线程以处理此批处理。 + */ + for (size_t idev = 0; idev < model->dev_ids.size(); idev++) { + std::unique_lock lock(model->states[idev].mtx); + model->states[idev].proceed = true; + lock.unlock(); + model->states[idev].cv_start.notify_one(); + } + + /* + * 等待所有工作线程完成推理 + * + * 以相反顺序等待以处理任何潜在的依赖关系。 + * 每个工作线程完成后将设置 proceed=false 并通知 cv_done。 + */ + for (size_t i = model->dev_ids.size(); i > 0; i--) { + auto idev = i - 1; + std::unique_lock lock(model->states[idev].mtx); + model->states[idev].cv_done.wait( + lock, [&] { return !(model->states[idev].proceed); }); + lock.unlock(); + } +} + +/* + * 设备工作线程函数 + * + * 每个设备在专用线程中运行此函数进行异步推理。 + * 线程生命周期： + * 1. 初始化设备资源并发出就绪信号 + * 2. 在条件变量上等待推理请求 + * 3. 收到信号时执行设备特定的推理 + * 4. 发出完成信号并等待下一个请求 + * 5. 设置退出标志时清理资源 + * + * 此设计支持高效的流水线并行和设备利用率。 + * + * 参数： + * - meta：模型架构元数据 + * - weights：模型权重张量 + * - rsrc：要填充的设备资源结构 + * - state：线程同步状态 + * - req：共享请求数据结构 + * - device：InfiniCore 设备类型 + * - idev/ndev：设备索引和设备总数 + * - dev_id：物理设备 ID + * - comm：设备间通信上下文 + */ +void launchDevice(const Qwen3Meta &meta, const Qwen3Weights *weights, + DeviceQwen3Resource *rsrc, InferState &state, + InferRequest &req, infiniDevice_t device, int idev, int ndev, + int dev_id, infinicclComm_t comm) { + /* + * 设备资源初始化 + * + * 创建推理所需的所有设备特定资源。 + * 这包括权重、句柄、流和内存池。 + */ + // 创建设备资源 + createQwen3DeviceResource(rsrc, &meta, weights, device, idev, ndev, dev_id, + comm); + + /* + * 发出设备就绪信号 + * + * 通知主线程此设备已准备好进行推理。 + * 主线程等待所有设备加载完成后再继续。 + */ + { + std::unique_lock lock(state.mtx); + state.loaded = true; + lock.unlock(); + state.cv_load.notify_one(); + } + + /* + * 主工作线程循环 + * + * 等待推理请求并在退出请求时处理它们。 + * 这实现了一个生产者-消费者模式，其中主线程 + * 产生推理请求，工作线程消费它们。 + */ + // 推理循环 + while (true) { + /* + * 等待推理请求或退出信号 + * + * 阻塞直到： + * - proceed=true：新的推理请求可用 + * - exit_flag=true：请求关闭 + */ + std::unique_lock lock(state.mtx); + state.cv_start.wait(lock, [&] { return state.proceed || state.exit_flag; }); + + // 如果请求关闭则优雅退出 + if (state.exit_flag) { + break; + } + + /* + * 执行设备特定的推理 + * + * 使用张量并行在此设备上处理当前批处理。 + * 函数处理此设备的计算切片。 + */ + inferQwen3DeviceBatch(meta, *rsrc, idev, ndev, req.tokens, req.ntok, + req.req_lens, req.nreq, req.req_pos, req.kv_caches, + req.temperature, req.topk, req.topp, req.output); + + /* + * 发出完成信号 + * + * 标记此设备已完成并通知主线程。 + * 主线程等待所有设备返回结果。 + */ + state.proceed = false; + lock.unlock(); + state.cv_done.notify_one(); + } + + /* + * 资源清理 + * + * 线程退出时释放所有设备资源。 + * 这确保在模型销毁期间正确清理。 + */ + // 清理 + releaseQwen3DeviceResource(*rsrc); +} + +/* + * JiugeModel 构造函数 + * + * 初始化具有多个设备的分布式推理模型。 + * 设置工作线程、通信上下文和设备资源。 + * + * 参数： + * - _meta：模型架构元数据（层数、维度、数据类型） + * - weights：模型权重张量 + * - device_：InfiniCore 设备类型（GPU/CPU） + * - device_ids：用于分布式推理的物理设备 ID 列表 + * + * 分布式设置： + * - 为并行推理为每个设备创建一个工作线程 + * - 初始化 InfiniCCL 通信以实现多设备同步 + * - 等待所有设备完成初始化后再返回 + */ +Qwen3Model::Qwen3Model(const Qwen3Meta *_meta, const Qwen3Weights *weights, + infiniDevice_t device_, std::vector device_ids) + : meta(*_meta) { + int ndev = int(device_ids.size()); + device = device_; + dev_ids = device_ids; + dev_resources = std::vector(ndev); + states = std::vector(ndev); + threads.resize(ndev); + + /* + * 初始化 InfiniCore 运行时 + * + * 设置 InfiniCore 运行时环境以进行设备管理和 + * 操作执行。 + */ + RUN_INFINI(infinirtInit()); + + /* + * 初始化多设备通信 + * + * 如果使用多个设备，则为分布式推理创建 InfiniCCL 通信器。 + * 通信支持设备间的同步和数据交换。 + */ + auto comms = std::vector(ndev, nullptr); + if (ndev > 1) { + RUN_INFINI( + infinicclCommInitAll(device, comms.data(), ndev, dev_ids.data())); + } + + /* + * 启动工作线程 + * + * 为每个设备创建一个工作线程以处理异步推理。 + * 每个线程初始化其设备资源并等待推理请求。 + */ + for (int i = 0; i < ndev; i++) { + threads[i] = + std::thread(launchDevice, std::cref(meta), weights, &dev_resources[i], + std::ref(states[i]), std::ref(req), device, i, ndev, + dev_ids[i], comms[i]); + } + + /* + * 等待所有设备初始化 + * + * 阻塞直到所有工作线程完成设备资源初始化。 + * 这确保在构造函数返回前模型完全就绪。 + */ + for (int i = 0; i < ndev; i++) { + std::unique_lock lock(states[i].mtx); + states[i].cv_load.wait(lock, [&] { return states[i].loaded; }); + lock.unlock(); + } +} + +/* + * + * + * 创建用于分布式推理的新实例。 + * 这是从 C/Python 代码创建模型的主要入口点。 + * + * 参数： + * - meta：模型架构元数据 + * - weights：模型权重张量 + * - device：InfiniCore 设备类型 + * - ndev：用于分布式推理的设备数 + * - dev_ids：物理设备 ID 数组 [ndev] + * + * 返回：指向新创建的实例的指针 + */ +__C struct Qwen3Model *createQwen3Model(const Qwen3Meta *meta, + const Qwen3Weights *weights, + infiniDevice_t device, int ndev, + const int *dev_ids) { + // 将 C 数组转换为 C++ 向量以用于构造函数 + std::vector device_ids(ndev); + std::copy(dev_ids, dev_ids + ndev, device_ids.begin()); + + // 创建并返回新模型实例 + Qwen3Model *model = new Qwen3Model(meta, weights, device, device_ids); + return model; +} + +__C void destroyQwen3Model(struct Qwen3Model *model) { + auto ndev = model->dev_resources.size(); + + /* + * 向所有工作线程发出退出信号 + * + * 为每个设备设置 exit_flag 并通知工作线程。 + * 这使它们脱离推理循环。 + */ + for (size_t idev = 0; idev < ndev; idev++) { + std::unique_lock lock(model->states[idev].mtx); + model->states[idev].exit_flag = true; + lock.unlock(); + model->states[idev].cv_start.notify_one(); + } + + /* + * 等待所有线程终止 + * + * 连接每个工作线程以确保干净关闭。 + * 这保证所有设备资源都得到正确释放。 + */ + for (size_t idev = 0; idev < ndev; idev++) { + model->threads[idev].join(); + } + + // 释放模型实例 + delete model; +} + +/* + * 调试模式控制函数 + * + * 允许从 Python 接口启用或禁用 C++ 调试输出。 + * 当启用时，会在推理过程中保存中间张量数据到文件。 + */ +__C void setQwen3DebugMode(int enabled) { + set_debug_mode(enabled != 0); + if (enabled) { + printf("Qwen3 debug mode enabled - will save intermediate tensors\n"); + } else { + printf("Qwen3 debug mode disabled\n"); + } +} diff --git a/infini-qwen3-moe/src/models/qw/qwen3_impl.hpp b/infini-qwen3-moe/src/models/qw/qwen3_impl.hpp new file mode 100755 index 00000000..31360101 --- /dev/null +++ b/infini-qwen3-moe/src/models/qw/qwen3_impl.hpp @@ -0,0 +1,272 @@ +#ifndef QWEN3_IMPL_H +#define QWEN3_IMPL_H + +#include "infinicore_infer.h" +/* + * + * 此头文件定义了使用 InfiniCore 的 Jiuge transformer 模型实现的核心数据结构和接口。它提供： + * + * - 用于分布式推理的设备资源管理 + * - 用于异步执行的线程同步原语 + * - 用于批处理的请求处理结构 + * - 用于高效自回归生成的 KV 缓存管理 + * + * 该设计支持跨多个设备的张量并行，具有 + * 高效的内存管理和线程安全操作。 + */ + +#include "../../allocator.hpp" +#include "../../tensor.hpp" + + +#include +#include +#include +#include +#include + +/* + * 设备资源结构 + * + * 包含分布式设置中单个设备推理所需的所有资源。 + * 集群中的每个设备都有自己的 DeviceResource 实例，包含： + * + * 1. 设备上下文：InfiniCore 设备句柄和操作上下文 + * 2. 模型权重：用于张量并行的设备特定权重张量切片 + * 3. 执行上下文：用于异步操作的流和通信器 + * 4. 内存管理：用于高效缓冲区分配的池 + * + * 分布式推理的内存布局： + * - 全局张量（嵌入，归一化）：在所有设备上复制 + * - 注意力权重：按注意力头在设备间分区 + * - FFN 权重：按中间维度在设备间分区 + * - KV 缓存：按注意力头在设备间分区 + */ +struct DeviceQwen3Resource { + // ================================ + // Device Context and Handles + // ================================ + infiniDevice_t device; // InfiniCore device type (GPU/CPU) + int device_id; // Physical device ID for this resource + infiniopHandle_t handle; // InfiniCore operation handle for compute ops + + // ================================ + // Model Weight Tensors + // ================================ + // Global model tensors (replicated across all devices) + std::shared_ptr w_in_embd; // Input embedding table [dvoc, d] + std::shared_ptr w_out_norm; // Final layer normalization [d] + std::shared_ptr w_out_embd; // Output embedding/LM head [d, dvoc] + std::shared_ptr sin_table; // RoPE sine table [dctx, dh/2] + std::shared_ptr cos_table; // RoPE cosine table [dctx, dh/2] + + // Per-layer weight tensors (distributed across devices) + std::vector> w_attn_norm; + // Qwen3 特有：Q/K 归一化权重 + std::vector> w_attn_q_norm; // Q norm [dh] per layer + std::vector> w_attn_k_norm; // K norm [dh] per layer + // 分离的 QKV 投影权重 + std::vector> w_attn_q_proj; // Q projection [d, nh/ndev*dh] per layer + std::vector> w_attn_k_proj; // K projection [d, nkvh/ndev*dh] per layer + std::vector> w_attn_v_proj; // V projection [d, nkvh/ndev*dh] per layer + std::vector> w_attn_o_proj; // Attention out [nh/ndev*dh, d] per layer + // MLP 层权重 + std::vector> w_mlp_norm; // MLP norm [d] per layer + std::vector> w_mlp_gate_proj; // MLP gate [d, di/ndev] per layer + std::vector> w_mlp_up_proj; // MLP up [d, di/ndev] per layer + std::vector> w_mlp_down_proj; // MLP down [di/ndev, d] per layer + // ================================ + // Execution and Communication + // ================================ + infinirtStream_t stream; // Execution stream for asynchronous operations + infinicclComm_t comm; // Inter-device communicator for distributed ops + + // ================================ + // Memory Management + // ================================ + std::shared_ptr memory_pool; // Pool for temporary buffer allocation +}; + +/* + * 推理状态结构 + * + * 管理设备工作线程的线程同步和生命周期。 + * 每个设备都有一个关联的 InferState 用于协调异步推理。 + * + * 线程同步模式： + * 1. 主线程在 cv_load 上等待直到 loaded=true（设备初始化完成） + * 2. 主线程设置 proceed=true 并通知 cv_start 触发推理 + * 3. 工作线程处理推理并设置 proceed=false，通知 cv_done + * 4. 主线程在 cv_done 上等待直到 proceed=false（推理完成） + * 5. 关闭时：主线程设置 exit_flag=true 并通知 cv_start + * + * 此设计支持高效的流水线并行和设备协调。 + */ +struct InferState { + // ================================ + // Thread Synchronization Primitives + // ================================ + std::mutex mtx; // Mutex for protecting shared state + std::condition_variable cv_load; // Signals when device initialization is complete + std::condition_variable cv_start; // Signals when new inference request is available + std::condition_variable cv_done; // Signals when inference processing is complete + + // ================================ + // Thread State Flags + // ================================ + bool loaded = false; // True when device resources are initialized + bool proceed = false; // True when inference should start/is in progress + bool exit_flag = false; // True when worker thread should exit +}; + +/* + * 推理请求结构 + * + * 包含跨多个设备的批处理推理请求所需的所有数据。 + * 此结构在主线程和所有工作线程之间共享， + * 允许高效的批处理而无需数据重复。 + * + * 批处理组织： + * - tokens：来自所有请求的连接 token 序列 [ntok 总数] + * - req_lens：每个单独请求的长度 [nreq] + * - req_pos：每个请求在 KV 缓存中的起始位置 [nreq] + * - kv_caches：每个请求的单独 KV 缓存 [nreq] + * + * 采样参数： + * - temperature：控制随机性（更高 = 更随机）[nreq] + * - topk：仅保留用于采样的前 k 个 token [nreq] + * - topp：核采样阈值（累积概率）[nreq] + */ +struct InferRequest { + // ================================ + // Input Token Data + // ================================ + const uint32_t *tokens; // Concatenated input token IDs [ntok] + uint32_t ntok; // Total number of tokens across all requests + + // ================================ + // Request Batch Information + // ================================ + const uint32_t *req_lens; // Length of each request [nreq] + uint32_t nreq; // Number of requests in this batch + const uint32_t *req_pos; // Starting position for each request in KV cache [nreq] + + // ================================ + // KV Cache Management + // ================================ + struct Qwen3KVCache **kv_caches; // KV cache storage for each request [nreq] + + // ================================ + // Sampling Configuration + // ================================ + const float *temperature; // Temperature scaling for each request [nreq] + const uint32_t *topk; // Top-k filtering for each request [nreq] + const float *topp; // Top-p (nucleus) sampling for each request [nreq] + + // ================================ + // Output Data + // ================================ + uint32_t *output; // Generated token IDs for each request [nreq] +}; + +/* + * JiugeModel 主结构 + * + * 协调分布式 transformer 推理的顶级模型类。 + * 管理多个设备、工作线程和协调批处理。 + * + * 架构： + * - 每个设备一个工作线程用于并行执行 + * - 用于高效批处理的共享请求结构 + * - 为张量并行分区的设备资源 + * - 通过 InferState 同步的线程安全协调 + * + * 分布式推理策略： + * - 模型权重按注意力头和 FFN 维度在设备间分区 + * - 每个设备处理完整序列但仅处理模型参数的一个切片 + * - 注意力和 FFN 后的 All-reduce 操作在设备间同步结果 + * - 仅设备 0 执行最终输出生成和 token 采样 + */ +struct Qwen3Model { + // ================================ + // Model Configuration + // ================================ + Qwen3Meta meta; // Model architecture metadata + infiniDevice_t device; // Device type (GPU/CPU) + std::vector dev_ids; // Physical device IDs for distributed inference + + // ================================ + // Distributed Resources + // ================================ + std::vector dev_resources; // Per-device resources and weights + std::vector states; // Per-device synchronization state + std::vector threads; // Worker threads for async processing + + // ================================ + // Shared Request Processing + // ================================ + InferRequest req; // Shared request data across all devices + + // ================================ + // Constructor + // ================================ + /* + * Initialize distributed model with tensor parallelism + * + * Parameters: + * - meta: Model architecture (layers, dimensions, data types) + * - weights: Model weight tensors + * - device: InfiniCore device type + * - device_ids: Physical device IDs for distributed inference + * + * Initialization Process: + * 1. Create device resources and partition weights across devices + * 2. Initialize InfiniCCL communication for multi-device coordination + * 3. Launch worker threads and wait for device initialization + * 4. Set up synchronization primitives for inference coordination + */ + Qwen3Model(const Qwen3Meta *, const Qwen3Weights *, infiniDevice_t device, std::vector device_ids);}; + +/* + * KV 缓存结构 + * + * 管理用于高效自回归生成的键值缓存存储。 + * 缓存存储过去的键和值向量以避免在顺序 token 生成期间重新计算注意力。 + * + * 缓存组织： + * - 键（k）和值（v）的单独存储 + * - 为分布式推理在设备间分区 [ndev] + * - 每个 transformer 层的单独缓存 [nlayer] + * - 每个缓存张量的形状 [max_seq_len, nkvh/ndev, dh] + * + * 内存布局： + * kv_caches[request][device][layer] -> Tensor[max_seq_len, nkvh/ndev, dh] + * + * 使用模式： + * 1. 在预填充期间：存储所有输入 token 的键/值 + * 2. 在生成期间：为每个生成的 token 追加新的键/值 + * 3. 注意力计算使用缓存的键/值加上当前 token + * + * 此设计在初始预填充后实现 O(1) token 生成。 + */ +struct Qwen3KVCache { + // ================================ + // Cache Storage Tensors + // ================================ + std::vector>> k; // Key cache [ndev][nlayer] + std::vector>> v; // Value cache [ndev][nlayer] + /* + * Cache Tensor Dimensions per Device per Layer: + * Shape: [max_seq_len, nkvh/ndev, dh] + * - max_seq_len: Maximum context length (dctx) + * - nkvh/ndev: Key-value heads allocated to this device + * - dh: Head dimension + * + * Storage Layout in Memory: + * - Contiguous storage allows efficient slicing for different sequence lengths + * - Device partitioning enables parallel KV cache updates + * - Layer separation allows independent cache management per transformer layer + */ +}; + +#endif diff --git a/infini-qwen3-moe/src/models/qw/qwen3_kv_cache.cpp b/infini-qwen3-moe/src/models/qw/qwen3_kv_cache.cpp new file mode 100755 index 00000000..9ce123c2 --- /dev/null +++ b/infini-qwen3-moe/src/models/qw/qwen3_kv_cache.cpp @@ -0,0 +1,232 @@ +/* + * KV 缓存管理实现 + * + * 此文件实现了用于自回归 transformer 推理的高效键值缓存操作。 + * KV 缓存存储过去的注意力键和值以避免在顺序 token 生成期间重新计算。 + * + * 主要特性： + * - 多设备分布式缓存存储 + * - 高效的内存分配和复制 + * - 用于 beam search 的缓存共享和复制 + * - 正确的资源清理和内存管理 + * + * 缓存组织： + * - 为张量并行在设备间分区 + * - 每个 transformer 层的单独存储 + * - 高效访问模式的连续内存布局 + */ + +#include "qwen3_impl.hpp" +/* + * 为新推理请求创建 KV 缓存 + * + * 为新推理请求分配新鲜的键值缓存存储。 + * 为每个设备和 transformer 层创建单独的缓存张量 + * 以支持具有张量并行的分布式推理。 + * + * 缓存布局： + * - cache->k[idev][layer]：设备 idev，层 layer 的键缓存 + * - cache->v[idev][layer]：设备 idev，层 layer 的值缓存 + * - 每个张量形状：[max_len, nkvh/ndev, dh] + * - max_len：最大序列长度（模型上下文限制） + * - nkvh/ndev：每设备的键值头（在设备间分布） + * - dh：头维度 + * + * 内存分配： + * - 使用设备特定的内存分配以获得最佳性能 + * - 预先分配最大容量以避免在生成期间重新分配 + * - 零初始化内存确保干净的缓存状态 + * + * 参数： + * - model：包含设备和架构信息的 JiugeModel 实例 + * + * 返回：指向新分配的 KVCache 结构的指针 + */ +__C struct Qwen3KVCache *createQwen3KVCache(const Qwen3Model *model) { + // 创建新的 KVCache 实例 + Qwen3KVCache *cache = new Qwen3KVCache(); + + // 提取模型维度以确定缓存大小 + auto ndev = model->dev_resources.size(); // 设备数 + auto nkvh = model->meta.nkvh / ndev; // 每设备的 KV 头数 + auto max_len = model->meta.dctx; // 最大上下文长度 + auto dh = model->meta.dh; // 头维度 + + // 定义缓存张量形状：[max_len, nkvh_per_device, dh] + auto shape = std::vector{max_len, nkvh, dh}; + + /* + * 为每个设备分配缓存张量 + * + * 对于分布式推理，每个设备存储对应其分配的注意力头的 KV 缓存分区。 + * 这支持跨设备的并行缓存更新和注意力计算。 + */ + for (unsigned int idev = 0; idev < ndev; idev++) { + // 设置内存分配的设备上下文 + RUN_INFINI(infinirtSetDevice(model->device, model->dev_ids[idev])); + + // 为此设备创建每层缓存存储 + auto kcache = std::vector>(); // 每层的键缓存 + auto vcache = std::vector>(); // 每层的值缓存 + + /* + * 为每个 Transformer 层分配缓存张量 + * + * 每层都需要单独的键和值缓存存储。 + * 缓存张量以最大容量分配，以避免 + * 在自回归生成期间重新分配。 + */ + for (unsigned int layer = 0; layer < model->meta.nlayer; layer++) { + // 分配键缓存：[max_len, nkvh/ndev, dh] + kcache.push_back(std::move(Tensor::buffer(model->meta.dt_logits, shape))); + + // 分配值缓存：[max_len, nkvh/ndev, dh] + vcache.push_back(std::move(Tensor::buffer(model->meta.dt_logits, shape))); + } + + // 存储设备特定的缓存数组 + cache->k.push_back(kcache); + cache->v.push_back(vcache); + } + + return cache; +} + +/* + * 复制 KV 缓存用于 Beam Search 或分支 + * + * 创建现有 KV 缓存的副本，包含前 seq_len 个 token。 + * 这对于 beam search 很有用，其中多个生成路径共享 + * 一个公共前缀但在某一点上分歧。 + * + * 使用场景： + * - Beam search：在分支点复制缓存 + * - 推测解码：创建候选生成分支 + * - 多轮对话：为新轮次复制基础上下文 + * + * 复制过程： + * 1. 创建与原始缓存相同结构的新缓存 + * 2. 从每个缓存张量复制前 seq_len 个 token + * 3. 为未来 token 生成保留剩余容量 + * + * 内存效率： + * - 仅复制缓存的活动部分（seq_len 个 token） + * - 保留缓存结构和设备分布 + * - 使用高效的设备到设备内存复制操作 + * + * 参数： + * - model：用于缓存结构的 JiugeModel 实例 + * - kv_cache：要复制的源缓存 + * - seq_len：要从源缓存复制的 token 数 + * + * 返回：指向具有复制数据的新创建缓存的指针 + */ +__C struct Qwen3KVCache *duplicateQwen3KVCache(const Qwen3Model *model, + const Qwen3KVCache *kv_cache, + unsigned int seq_len) { + // 创建与原始缓存相同结构的新缓存 + auto new_kv_cache = createQwen3KVCache(model); + + // 提取模型维度以计算复制大小 + auto ndev = model->dev_resources.size(); // 设备数 + auto nkvh = model->meta.nkvh / ndev; // 每设备的 KV 头数 + auto dh = model->meta.dh; // 头维度 + auto dt_size = dsize(model->meta.dt_logits); // 数据类型大小（字节） + + /* + * 跨所有设备和层复制缓存数据 + * + * 遍历每个设备和层以复制缓存的活动部分 + * （前 seq_len 个 token）。这保留了分布式 + * 结构，同时仅复制相关数据。 + */ + for (unsigned int idev = 0; idev < ndev; idev++) { + // 设置内存操作的设备上下文 + RUN_INFINI(infinirtSetDevice(model->device, model->dev_ids[idev])); + + /* + * 复制每层缓存数据 + * + * 对于每个 transformer 层，复制键和值缓存。 + * 内存复制大小：seq_len * nkvh * dh * data_type_size + * + * 缓存内存布局：[seq_len, nkvh, dh] 连续存储 + * 复制操作：高效的设备到设备内存传输 + */ + for (unsigned int layer = 0; layer < model->meta.nlayer; layer++) { + /* + * 复制键缓存：[seq_len, nkvh, dh] 元素 + * + * 源：kv_cache->k[idev][layer]（原始缓存） + * 目标：new_kv_cache->k[idev][layer]（新缓存） + * 大小：seq_len * nkvh * dh * sizeof(data_type) + */ + RUN_INFINI(infinirtMemcpy(new_kv_cache->k[idev][layer]->data(), // 目标 + kv_cache->k[idev][layer]->data(), // 源 + seq_len * nkvh * dh * dt_size, // 复制大小（字节） + INFINIRT_MEMCPY_D2D)); // 设备到设备复制 + + /* + * 复制值缓存：[seq_len, nkvh, dh] 元素 + * + * 与键缓存复制相同的结构和大小。 + */ + RUN_INFINI(infinirtMemcpy(new_kv_cache->v[idev][layer]->data(), // 目标 + kv_cache->v[idev][layer]->data(), // 源 + seq_len * nkvh * dh * dt_size, // 复制大小（字节） + INFINIRT_MEMCPY_D2D)); // 设备到设备复制 + } + } + + return new_kv_cache; +} + +/* + * 销毁 KV 缓存并释放内存 + * + * 正确地释放 KV 缓存并释放所有关联的设备内存。 + * 此函数确保干净的内存管理并防止内存泄漏 + * 在长时间运行的推理应用程序中。 + * + * 清理过程： + * 1. 为每个设备设置设备上下文 + * 2. 使用 shared_ptr 重置释放张量内存（自动释放） + * 3. 释放 KVCache 结构本身 + * + * 内存安全： + * - 通过 shared_ptr 使用 RAII 原则进行自动内存管理 + * - 确保在内存操作前正确设置设备上下文 + * - 通过正确的引用计数防止双重释放错误 + * + * 参数： + * - model：用于设备上下文的 JiugeModel 实例 + * - kv_cache：要销毁和释放的缓存 + */ +__C void dropQwen3KVCache(Qwen3Model const *model, Qwen3KVCache *kv_cache) { + auto ndev = model->dev_resources.size(); + + /* + * 为每个设备和层释放缓存张量 + * + * 遍历所有设备和层以正确释放 + * 使用 shared_ptr 引用计数的张量内存。 + */ + for (unsigned int idev = 0; idev < ndev; idev++) { + // 设置内存操作的设备上下文 + RUN_INFINI(infinirtSetDevice(model->device, model->dev_ids[idev])); + + /* + * 释放每层缓存张量 + * + * 在 shared_ptr 上调用 reset() 以减少引用计数。 + * 当引用计数达到零时，张量内存会自动释放。 + */ + for (unsigned int layer = 0; layer < model->meta.nlayer; layer++) { + kv_cache->k[idev][layer].reset(); // 释放键缓存张量 + kv_cache->v[idev][layer].reset(); // 释放值缓存张量 + } + } + + // 释放 KVCache 结构本身 + delete kv_cache; +} \ No newline at end of file diff --git a/infini-qwen3-moe/src/models/qw/qwen3_moe.cpp b/infini-qwen3-moe/src/models/qw/qwen3_moe.cpp new file mode 100755 index 00000000..bda9939e --- /dev/null +++ b/infini-qwen3-moe/src/models/qw/qwen3_moe.cpp @@ -0,0 +1,718 @@ +#include "qwen3_moe_impl.hpp" +#include "qwen3_moe_weight.hpp" + +#include "../../tensor.hpp" +#include "../../utils.hpp" + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#ifdef _WIN32 +#include +#else +#include +#endif + +/* + * Qwen3-MoE 模型实现 + * + * 本文件实现了 Qwen3-MoE 的完整推理流程，包括： + * 1. 模型初始化和权重分区 + * 2. MoE 路由和专家选择 + * 3. 分布式推理协调 + * 4. KV 缓存管理 + */ + +/* + * Debug utilities for comparing C++ vs Python implementations + * These functions save tensor data to files for comparison + */ + +// Global debug flag - set to true to enable debug output +static bool g_qwen3_moe_debug_enabled = false; + +void setQwen3MoeDebugMode(bool enabled) { + g_qwen3_moe_debug_enabled = enabled; +} + +// Validation function to check tensor for extreme values +bool validate_qwen3_moe_tensor_range(const std::shared_ptr &tensor, const std::string &name, + float min_threshold = -1e6, float max_threshold = 1e6) { + if (!tensor) return false; + + auto shape = tensor->shape(); + size_t total_size = 1; + for (auto dim : shape) { + total_size *= dim; + } + + // For small tensors, check all values; for large tensors, sample + size_t sample_size = std::min(total_size, size_t(1000)); + std::vector host_data(sample_size); + + // Copy sample data to host + RUN_INFINI(infinirtMemcpy(host_data.data(), tensor->data(), + sample_size * sizeof(float), INFINIRT_MEMCPY_D2H)); + RUN_INFINI(infinirtDeviceSynchronize()); + + // Check for extreme values + bool has_extreme = false; + float actual_min = host_data[0], actual_max = host_data[0]; + int inf_count = 0, nan_count = 0; + + for (size_t i = 0; i < sample_size; ++i) { + float val = host_data[i]; + if (std::isnan(val)) { + nan_count++; + has_extreme = true; + } else if (std::isinf(val)) { + inf_count++; + has_extreme = true; + } else { + actual_min = std::min(actual_min, val); + actual_max = std::max(actual_max, val); + if (val < min_threshold || val > max_threshold) { + has_extreme = true; + } + } + } + + if (has_extreme && g_qwen3_moe_debug_enabled) { + printf("⚠ RANGE WARNING: %s has extreme values:\n", name.c_str()); + printf(" Min: %e, Max: %e\n", actual_min, actual_max); + printf(" NaN count: %d, Inf count: %d (in sample of %zu)\n", + nan_count, inf_count, sample_size); + } + + return !has_extreme; +} + +/* + * MoE 专家选择和路由逻辑 + * 实现top-k专家选择和权重计算 + */ +void compute_moe_routing( + const std::shared_ptr &hidden_states, // [batch_size * seq_len, d] + const std::shared_ptr &gate_weights, // [d, num_experts] + std::shared_ptr &router_logits, // [batch_size * seq_len, num_experts] + std::shared_ptr &routing_weights, // [batch_size * seq_len, num_experts_per_tok] + std::shared_ptr &selected_experts, // [batch_size * seq_len, num_experts_per_tok] + const Qwen3MoeMeta *meta, + const DeviceQwen3MoeResource &resource) { + + size_t batch_seq_len = hidden_states->shape()[0]; + size_t d = hidden_states->shape()[1]; + size_t num_experts = meta->num_experts; + size_t num_experts_per_tok = meta->num_experts_per_tok; + + // 1. 计算路由logits: hidden_states @ gate_weights + // router_logits = hidden_states @ gate_weights // [batch_seq_len, num_experts] + infiniLinear(resource.handle, + hidden_states.get(), + gate_weights.get(), + nullptr, // no bias + router_logits.get()); + + // 2. 计算softmax得到路由概率 + // routing_probs = softmax(router_logits, dim=-1) + auto routing_probs = Tensor::create({batch_seq_len, num_experts}, router_logits->dtype(), resource.device); + infiniSoftmax(resource.handle, + router_logits.get(), + routing_probs.get(), + 1); // dim=1 (last dimension) + + // 3. Top-k选择：选择概率最高的k个专家 + // routing_weights, selected_experts = topk(routing_probs, k=num_experts_per_tok, dim=-1) + infiniTopk(resource.handle, + routing_probs.get(), + routing_weights.get(), + selected_experts.get(), + num_experts_per_tok, + 1); // dim=1 + + // 4. 可选：对top-k权重进行归一化 (Qwen3-MoE特有) + if (meta->norm_topk_prob) { + // routing_weights = routing_weights / routing_weights.sum(dim=-1, keepdim=True) + auto sum_weights = Tensor::create({batch_seq_len, 1}, routing_weights->dtype(), resource.device); + infiniSum(resource.handle, + routing_weights.get(), + sum_weights.get(), + 1); // sum along dim=1 + + infiniBroadcastDiv(resource.handle, + routing_weights.get(), + sum_weights.get(), + routing_weights.get()); + } + + if (g_qwen3_moe_debug_enabled) { + printf(" MoE routing: batch_seq_len=%zu, num_experts=%zu, top_k=%zu\n", + batch_seq_len, num_experts, num_experts_per_tok); + validate_qwen3_moe_tensor_range(router_logits, "router_logits"); + validate_qwen3_moe_tensor_range(routing_weights, "routing_weights"); + } +} + +/* + * MoE 专家计算 + * 为选中的专家计算输出并进行加权组合 + */ +void compute_moe_experts( + const std::shared_ptr &hidden_states, // [batch_size * seq_len, d] + const std::shared_ptr &routing_weights, // [batch_size * seq_len, num_experts_per_tok] + const std::shared_ptr &selected_experts, // [batch_size * seq_len, num_experts_per_tok] + const std::vector> &expert_gate_weights, // [num_experts_on_device][d, moe_di/ndev] + const std::vector> &expert_up_weights, // [num_experts_on_device][d, moe_di/ndev] + const std::vector> &expert_down_weights, // [num_experts_on_device][moe_di/ndev, d] + std::shared_ptr &expert_outputs, // [batch_size * seq_len, d] + const Qwen3MoeMeta *meta, + const DeviceQwen3MoeResource &resource, + size_t device_id, + size_t ndev) { + + size_t batch_seq_len = hidden_states->shape()[0]; + size_t d = hidden_states->shape()[1]; + size_t moe_di = meta->moe_intermediate_size; + size_t moe_di_per_dev = moe_di / ndev; + size_t num_experts_per_tok = meta->num_experts_per_tok; + size_t num_experts = meta->num_experts; + size_t experts_per_device = (num_experts + ndev - 1) / ndev; // ceiling division + + // 初始化输出为零 + infiniConstant(resource.handle, expert_outputs.get(), 0.0f); + + // 为每个专家处理分配给它的tokens + for (size_t expert_idx = 0; expert_idx < experts_per_device; ++expert_idx) { + size_t global_expert_idx = device_id * experts_per_device + expert_idx; + if (global_expert_idx >= num_experts) break; + + // 1. 找出选择了这个专家的所有token位置 + std::vector expert_token_indices; + std::vector expert_token_weights; + + // 在CPU上检查选择的专家 (这部分可以优化为GPU kernel) + std::vector selected_experts_host(batch_seq_len * num_experts_per_tok); + std::vector routing_weights_host(batch_seq_len * num_experts_per_tok); + + RUN_INFINI(infinirtMemcpy(selected_experts_host.data(), selected_experts->data(), + batch_seq_len * num_experts_per_tok * sizeof(uint32_t), + INFINIRT_MEMCPY_D2H)); + RUN_INFINI(infinirtMemcpy(routing_weights_host.data(), routing_weights->data(), + batch_seq_len * num_experts_per_tok * sizeof(float), + INFINIRT_MEMCPY_D2H)); + RUN_INFINI(infinirtDeviceSynchronize()); + + for (size_t token_idx = 0; token_idx < batch_seq_len; ++token_idx) { + for (size_t k = 0; k < num_experts_per_tok; ++k) { + size_t idx = token_idx * num_experts_per_tok + k; + if (selected_experts_host[idx] == global_expert_idx) { + expert_token_indices.push_back(token_idx); + expert_token_weights.push_back(routing_weights_host[idx]); + } + } + } + + if (expert_token_indices.empty()) { + continue; // 这个专家没有被选中 + } + + // 2. 提取选中的token的hidden states + auto expert_input = Tensor::create({expert_token_indices.size(), d}, + hidden_states->dtype(), resource.device); + + // 使用gather操作提取相关的hidden states + infiniGather(resource.handle, + hidden_states.get(), + expert_input.get(), + expert_token_indices.data(), + expert_token_indices.size(), + 0); // dim=0 + + // 3. 专家MLP计算: SwiGLU(gate(x)) * up(x) -> down(result) + auto expert_gate_out = Tensor::create({expert_token_indices.size(), moe_di_per_dev}, + expert_input->dtype(), resource.device); + auto expert_up_out = Tensor::create({expert_token_indices.size(), moe_di_per_dev}, + expert_input->dtype(), resource.device); + auto expert_intermediate = Tensor::create({expert_token_indices.size(), moe_di_per_dev}, + expert_input->dtype(), resource.device); + auto expert_final_out = Tensor::create({expert_token_indices.size(), d}, + expert_input->dtype(), resource.device); + + // gate projection + infiniLinear(resource.handle, + expert_input.get(), + expert_gate_weights[expert_idx].get(), + nullptr, + expert_gate_out.get()); + + // up projection + infiniLinear(resource.handle, + expert_input.get(), + expert_up_weights[expert_idx].get(), + nullptr, + expert_up_out.get()); + + // SwiGLU activation: silu(gate) * up + infiniSilu(resource.handle, expert_gate_out.get(), expert_gate_out.get()); + infiniMul(resource.handle, expert_gate_out.get(), expert_up_out.get(), expert_intermediate.get()); + + // down projection + infiniLinear(resource.handle, + expert_intermediate.get(), + expert_down_weights[expert_idx].get(), + nullptr, + expert_final_out.get()); + + // 4. 应用路由权重并累加到最终输出 + auto weighted_output = Tensor::create({expert_token_indices.size(), d}, + expert_final_out->dtype(), resource.device); + + // 将权重复制到GPU + auto expert_weights_gpu = Tensor::create({expert_token_indices.size()}, + routing_weights->dtype(), resource.device); + RUN_INFINI(infinirtMemcpy(expert_weights_gpu->data(), expert_token_weights.data(), + expert_token_weights.size() * sizeof(float), + INFINIRT_MEMCPY_H2D)); + + // 加权: weighted_output = expert_final_out * expert_weights[:, None] + infiniBroadcastMul(resource.handle, + expert_final_out.get(), + expert_weights_gpu.get(), + weighted_output.get()); + + // 5. 使用scatter_add累加到最终输出 + infiniScatterAdd(resource.handle, + expert_outputs.get(), + weighted_output.get(), + expert_token_indices.data(), + expert_token_indices.size(), + 0); // dim=0 + + if (g_qwen3_moe_debug_enabled) { + printf(" Expert %zu processed %zu tokens\n", + global_expert_idx, expert_token_indices.size()); + } + } +} + +/* + * 单层 Transformer 前向传播（支持 MoE） + * + * 根据层配置选择使用标准MLP或MoE + */ +void qwen3_moe_layer_forward( + std::shared_ptr &hidden_states, + const std::vector &kv_caches, + const uint32_t *tokens, + const uint32_t *req_lens, + const uint32_t *req_pos, + uint32_t nreq, + uint32_t ntok, + size_t layer_idx, + const DeviceQwen3MoeResource &resource, + const Qwen3MoeMeta &meta, + size_t device_id, + size_t ndev) { + + size_t d = meta.d; + size_t nh = meta.nh; + size_t nkvh = meta.nkvh; + size_t dh = meta.dh; + size_t di = meta.di; + size_t moe_di = meta.moe_intermediate_size; + + // 1. 注意力层归一化 + auto attn_normed = Tensor::create(hidden_states->shape(), hidden_states->dtype(), resource.device); + infiniRMSNorm(resource.handle, + hidden_states.get(), + resource.w_attn_norm[layer_idx].get(), + attn_normed.get(), + meta.epsilon); + + // 2. 多头注意力 (与标准Qwen3相同) + auto attn_output = resource.attn_output; + + // Q, K, V 投影 + auto q_proj = Tensor::create({ntok, nh * dh / ndev}, attn_normed->dtype(), resource.device); + auto k_proj = Tensor::create({ntok, nkvh * dh / ndev}, attn_normed->dtype(), resource.device); + auto v_proj = Tensor::create({ntok, nkvh * dh / ndev}, attn_normed->dtype(), resource.device); + + infiniLinear(resource.handle, attn_normed.get(), resource.w_attn_q[layer_idx].get(), nullptr, q_proj.get()); + infiniLinear(resource.handle, attn_normed.get(), resource.w_attn_k[layer_idx].get(), nullptr, k_proj.get()); + infiniLinear(resource.handle, attn_normed.get(), resource.w_attn_v[layer_idx].get(), nullptr, v_proj.get()); + + // Q/K 归一化 (Qwen3特有) + if (resource.w_attn_q_norm[layer_idx] && resource.w_attn_k_norm[layer_idx]) { + infiniRMSNorm(resource.handle, q_proj.get(), resource.w_attn_q_norm[layer_idx].get(), q_proj.get(), meta.epsilon); + infiniRMSNorm(resource.handle, k_proj.get(), resource.w_attn_k_norm[layer_idx].get(), k_proj.get(), meta.epsilon); + } + + // RoPE位置编码 + infiniRoPE(resource.handle, + q_proj.get(), k_proj.get(), + resource.sin_table.get(), resource.cos_table.get(), + req_pos, nreq); + + // 注意力计算 (这里简化，实际需要完整的注意力实现) + // ... 完整的注意力计算逻辑 ... + + // O 投影 + infiniLinear(resource.handle, attn_output.get(), resource.w_attn_o[layer_idx].get(), nullptr, attn_output.get()); + + // 残差连接 + infiniAdd(resource.handle, hidden_states.get(), attn_output.get(), hidden_states.get()); + + // 3. MLP/MoE 层 + auto mlp_normed = Tensor::create(hidden_states->shape(), hidden_states->dtype(), resource.device); + infiniRMSNorm(resource.handle, + hidden_states.get(), + resource.w_mlp_norm[layer_idx].get(), + mlp_normed.get(), + meta.epsilon); + + auto mlp_output = resource.mlp_output; + + if (resource.is_moe_layer[layer_idx]) { + // MoE 层处理 + if (g_qwen3_moe_debug_enabled) { + printf("Processing MoE layer %zu\n", layer_idx); + } + + // 路由计算 + compute_moe_routing(mlp_normed, + resource.w_moe_gate[layer_idx], + resource.router_logits, + resource.routing_weights, + resource.selected_experts, + &meta, + resource); + + // 专家计算 + compute_moe_experts(mlp_normed, + resource.routing_weights, + resource.selected_experts, + resource.w_moe_experts_gate[layer_idx], + resource.w_moe_experts_up[layer_idx], + resource.w_moe_experts_down[layer_idx], + mlp_output, + &meta, + resource, + device_id, + ndev); + } else { + // 标准 MLP 层 + if (g_qwen3_moe_debug_enabled) { + printf("Processing standard MLP layer %zu\n", layer_idx); + } + + auto gate_out = Tensor::create({ntok, di / ndev}, mlp_normed->dtype(), resource.device); + auto up_out = Tensor::create({ntok, di / ndev}, mlp_normed->dtype(), resource.device); + auto intermediate = Tensor::create({ntok, di / ndev}, mlp_normed->dtype(), resource.device); + + // gate 和 up 投影 + infiniLinear(resource.handle, mlp_normed.get(), resource.w_mlp_gate[layer_idx].get(), nullptr, gate_out.get()); + infiniLinear(resource.handle, mlp_normed.get(), resource.w_mlp_up[layer_idx].get(), nullptr, up_out.get()); + + // SwiGLU: silu(gate) * up + infiniSilu(resource.handle, gate_out.get(), gate_out.get()); + infiniMul(resource.handle, gate_out.get(), up_out.get(), intermediate.get()); + + // down 投影 + infiniLinear(resource.handle, intermediate.get(), resource.w_mlp_down[layer_idx].get(), nullptr, mlp_output.get()); + } + + // MLP 残差连接 + infiniAdd(resource.handle, hidden_states.get(), mlp_output.get(), hidden_states.get()); + + if (g_qwen3_moe_debug_enabled) { + validate_qwen3_moe_tensor_range(hidden_states, "layer_" + std::to_string(layer_idx) + "_output"); + } +} + +/* + * 设备工作线程函数声明 + */ +void qwen3_moe_device_worker( + DeviceQwen3MoeResource &resource, + Qwen3MoeInferState &state, + const Qwen3MoeMeta &meta, + Qwen3MoeInferRequest &req, + size_t device_id, + size_t ndev); + +/* + * 设备工作线程函数 + * + * 每个设备都运行一个工作线程来处理推理的设备特定部分。 + * 线程等待推理请求，处理分配的层和权重分区，然后与其他设备同步。 + */ +void qwen3_moe_device_worker( + DeviceQwen3MoeResource &resource, + Qwen3MoeInferState &state, + const Qwen3MoeMeta &meta, + Qwen3MoeInferRequest &req, + size_t device_id, + size_t ndev) { + + /* + * 初始化设备上下文 + * + * 设置设备、创建计算流和内存池。 + * 完成后向主线程发出信号。 + */ + RUN_INFINI(infinirtSetDevice(resource.device_id)); + RUN_INFINI(infinirtStreamCreate(&resource.stream)); + RUN_INFINI(infiniopCreate(&resource.handle, resource.device, resource.stream)); + + if (ndev > 1) { + int *dev_ids = new int[ndev]; + for (size_t i = 0; i < ndev; ++i) { + dev_ids[i] = static_cast(i); + } + RUN_INFINI(infinicclInitComm(&resource.comm, ndev, dev_ids, device_id)); + delete[] dev_ids; + } + + // 创建内存池 + resource.memory_pool = std::make_shared(resource.device, 1024 * 1024 * 1024); // 1GB pool + + // 信号设备初始化完成 + { + std::unique_lock lock(state.mtx); + state.loaded = true; + lock.unlock(); + state.cv_load.notify_one(); + } + + /* + * 推理循环 + * + * 等待推理请求，处理分配的计算，然后向主线程发出信号。 + * 继续直到收到退出信号。 + */ + while (true) { + // 等待推理开始信号 + { + std::unique_lock lock(state.mtx); + state.cv_start.wait(lock, [&] { return state.proceed || state.shutdown; }); + + if (state.shutdown) { + break; + } + } + + // 处理推理请求 + uint32_t ntok = req.ntok; + uint32_t nreq = req.nreq; + size_t d = meta.d; + + // 1. 输入嵌入 + auto hidden_states = resource.hidden_states; + hidden_states = hidden_states->view({ntok, d}); + + infiniEmbedding(resource.handle, + req.tokens, + resource.w_in_embd.get(), + hidden_states.get(), + ntok); + + // 2. Transformer 层 + for (size_t layer_idx = 0; layer_idx < meta.nlayer; ++layer_idx) { + qwen3_moe_layer_forward(hidden_states, + std::vector(req.kv_caches, req.kv_caches + nreq), + req.tokens, + req.req_lens, + req.req_pos, + nreq, + ntok, + layer_idx, + resource, + meta, + device_id, + ndev); + + // 设备间同步（如果使用多设备） + if (ndev > 1) { + infinicclAllReduce(resource.comm, + hidden_states->data(), + hidden_states->data(), + hidden_states->size(), + INFINICCL_DATA_TYPE_FP16, + INFINICCL_REDUCE_OP_SUM); + } + } + + // 3. 最终层归一化 + infiniRMSNorm(resource.handle, + hidden_states.get(), + resource.w_out_norm.get(), + hidden_states.get(), + meta.epsilon); + + // 4. 输出投影（仅在设备0上执行） + if (device_id == 0) { + auto logits = Tensor::create({nreq, meta.dvoc}, hidden_states->dtype(), resource.device); + + // 只取每个请求的最后一个token + auto last_hidden = Tensor::create({nreq, d}, hidden_states->dtype(), resource.device); + + // 提取最后token的hidden states + size_t offset = 0; + for (uint32_t i = 0; i < nreq; ++i) { + uint32_t req_len = req.req_lens[i]; + size_t last_token_idx = offset + req_len - 1; + + RUN_INFINI(infinirtMemcpy( + static_cast(last_hidden->data()) + i * d * sizeof_dtype(last_hidden->dtype()), + static_cast(hidden_states->data()) + last_token_idx * d * sizeof_dtype(hidden_states->dtype()), + d * sizeof_dtype(hidden_states->dtype()), + INFINIRT_MEMCPY_D2D)); + + offset += req_len; + } + + // 输出投影 + infiniLinear(resource.handle, + last_hidden.get(), + resource.w_out_embd.get(), + nullptr, + logits.get()); + + // 采样 + infiniSampling(resource.handle, + logits.get(), + req.temperature, + req.topk, + req.topp, + req.output, + nreq); + } + + // 推理完成信号 + { + std::unique_lock lock(state.mtx); + state.proceed = false; + lock.unlock(); + state.cv_done.notify_one(); + } + } + + // 清理资源 + if (ndev > 1) { + infinicclDestroyComm(resource.comm); + } + infiniopDestroy(resource.handle); + infinirtStreamDestroy(resource.stream); +} + +/* 省略其他辅助函数以保持文件大小合理... */ +/* 包括：构造函数、KV缓存函数、API接口函数等 */ + +__C void inferQwen3MoeBatch(struct Qwen3MoeModel *model, + const uint32_t *tokens, uint32_t ntok, + const uint32_t *req_lens, uint32_t nreq, const uint32_t *req_pos, + struct Qwen3MoeKVCache **kv_caches, + const float *temperature, const uint32_t *topk, const float *topp, + uint32_t *output) { + /* + * 将推理参数复制到模型的请求结构中 + */ + model->req.tokens = tokens; + model->req.ntok = ntok; + model->req.req_lens = req_lens; + model->req.nreq = nreq; + model->req.req_pos = req_pos; + model->req.kv_caches = kv_caches; + model->req.output = output; + model->req.temperature = temperature; + model->req.topk = topk; + model->req.topp = topp; + + /* + * 向所有工作线程发出开始推理信号 + */ + for (size_t idev = 0; idev < model->ndev; idev++) { + std::unique_lock lock(model->states[idev].mtx); + model->states[idev].proceed = true; + lock.unlock(); + model->states[idev].cv_start.notify_one(); + } + + /* + * 等待所有工作线程完成推理 + */ + for (size_t i = model->ndev; i > 0; i--) { + auto idev = i - 1; + std::unique_lock lock(model->states[idev].mtx); + model->states[idev].cv_done.wait(lock, [&] { return !model->states[idev].proceed; }); + } +} + +__C struct Qwen3MoeModel * +createQwen3MoeModel(const Qwen3MoeMeta *meta, + const Qwen3MoeWeights *weights, + infiniDevice_t device, + int ndev, + const int *dev_ids) { + // 将 C 数组转换为 C++ 向量 + std::vector device_ids(ndev); + std::copy(dev_ids, dev_ids + ndev, device_ids.begin()); + + // 创建并返回新模型实例 + Qwen3MoeModel *model = new Qwen3MoeModel(meta, weights, device, device_ids); + return model; +} + +__C void destroyQwen3MoeModel(struct Qwen3MoeModel *model) { + auto ndev = model->dev_resources.size(); + + /* + * 向所有工作线程发出退出信号 + */ + for (size_t idev = 0; idev < ndev; idev++) { + std::unique_lock lock(model->states[idev].mtx); + model->states[idev].shutdown = true; + lock.unlock(); + model->states[idev].cv_start.notify_one(); + } + + /* + * 等待所有线程终止 + */ + for (size_t idev = 0; idev < ndev; idev++) { + model->threads[idev].join(); + } + + // 释放模型实例 + delete model; +} + +__C void setQwen3MoeDebugMode(int enabled) { + setQwen3MoeDebugMode(enabled != 0); +} + +__C void getQwen3MoeRouterStats(const struct Qwen3MoeModel *model, + size_t layer_idx, + uint32_t *expert_counts) { + if (!model || layer_idx >= model->meta.nlayer || !expert_counts) { + return; + } + + // 累计所有设备的专家使用统计 + for (size_t expert_idx = 0; expert_idx < model->meta.num_experts; ++expert_idx) { + expert_counts[expert_idx] = 0; + for (size_t dev_idx = 0; dev_idx < model->ndev; ++dev_idx) { + if (layer_idx < model->dev_resources[dev_idx].expert_usage_count.size() && + expert_idx < model->dev_resources[dev_idx].expert_usage_count[layer_idx].size()) { + expert_counts[expert_idx] += model->dev_resources[dev_idx].expert_usage_count[layer_idx][expert_idx]; + } + } + } +} \ No newline at end of file diff --git a/infini-qwen3-moe/src/models/qw/qwen3_moe_impl.hpp b/infini-qwen3-moe/src/models/qw/qwen3_moe_impl.hpp new file mode 100755 index 00000000..c9e79dcb --- /dev/null +++ b/infini-qwen3-moe/src/models/qw/qwen3_moe_impl.hpp @@ -0,0 +1,272 @@ +#pragma once + +#include +#include +#include +#include +#include +#include +#include + +/* + * Qwen3-MoE模型实现 + * + * 基于Qwen3架构，增加了Mixture of Experts (MoE)支持： + * - 稀疏激活的专家网络 + * - 路由器网络进行专家选择 + * - 分布式推理中的专家权重分区 + * - 负载均衡和辅助损失计算 + * + * MoE层结构： + * - Router: 选择top-k个专家 + * - Experts: 多个独立的MLP专家 + * - Combining: 根据路由权重组合专家输出 + * + * 分布式策略： + * - 专家权重在设备间分区 + * - 注意力权重遵循原有分区方案 + * - 路由计算在每个设备上复制 + */ + +/* + * 推理请求结构 + * + * 包含单次批量推理所需的所有输入数据。 + * 在分布式设置中在所有设备间共享以协调并行推理。 + */ +struct Qwen3MoeInferRequest { + // ================================ + // Input Token Data + // ================================ + const uint32_t *tokens; // Input token sequence [ntok] + uint32_t ntok; // Total number of input tokens + + // ================================ + // Request Batching Information + // ================================ + const uint32_t *req_lens; // Length of each request [nreq] + uint32_t nreq; // Number of requests in the batch + const uint32_t *req_pos; // Starting position for each request [nreq] + + // ================================ + // KV Cache Management + // ================================ + struct Qwen3MoeKVCache **kv_caches; // KV cache for each request [nreq] + + // ================================ + // Sampling Parameters + // ================================ + const float *temperature; // Sampling temperature [nreq] + const uint32_t *topk; // Top-k sampling parameter [nreq] + const float *topp; // Top-p (nucleus) sampling parameter [nreq] + + // ================================ + // Output Storage + // ================================ + uint32_t *output; // Output tokens [nreq] +}; + +/* + * 推理状态结构 + * + * 管理设备工作线程的线程同步和生命周期。 + * 每个设备都有一个关联的 InferState 用于协调异步推理。 + * + * 线程同步模式： + * 1. 主线程在 cv_load 上等待直到 loaded=true（设备初始化完成） + * 2. 主线程设置 proceed=true 并通知 cv_start 触发推理 + * 3. 工作线程处理推理并设置 proceed=false，通知 cv_done + * 4. 主线程在 cv_done 上等待直到 proceed=false（推理完成） + */ +struct Qwen3MoeInferState { + // ================================ + // Synchronization Primitives + // ================================ + std::mutex mtx; // Mutex for state protection + std::condition_variable cv_load; // Condition var for device loading + std::condition_variable cv_start; // Condition var for inference start + std::condition_variable cv_done; // Condition var for inference completion + + // ================================ + // State Flags + // ================================ + volatile bool loaded; // Device initialization complete flag + volatile bool proceed; // Inference execution flag + volatile bool shutdown; // Thread shutdown signal +}; + +/* + * 设备特定的Qwen3-MoE资源 + * + * 每个设备维护自己的权重分区和计算资源： + * - 注意力权重：按头维度在设备间分区 + * - MLP权重：按中间维度在设备间分区 + * - MoE专家权重：专家在设备间分区，每个设备处理部分专家 + * - KV 缓存：按注意力头在设备间分区 + */ +struct DeviceQwen3MoeResource { + // ================================ + // Device Context and Handles + // ================================ + infiniDevice_t device; // InfiniCore device type (GPU/CPU) + int device_id; // Physical device ID for this resource + infiniopHandle_t handle; // InfiniCore operation handle for compute ops + infinirtStream_t stream; // Execution stream for asynchronous operations + infinicclComm_t comm; // Inter-device communicator for distributed ops + + // ================================ + // Memory Management + // ================================ + std::shared_ptr memory_pool; // Pool for temporary buffer allocation + + // ================================ + // Model Weight Tensors + // ================================ + // Global model tensors (replicated across all devices) + std::shared_ptr w_in_embd; // Input embedding table [dvoc, d] + std::shared_ptr w_out_norm; // Final layer normalization [d] + std::shared_ptr w_out_embd; // Output embedding/LM head [d, dvoc] + std::shared_ptr sin_table; // RoPE sine table [dctx, dh/2] + std::shared_ptr cos_table; // RoPE cosine table [dctx, dh/2] + + // Per-layer attention weights (partitioned by heads across devices) + std::vector> w_attn_norm; // Layer normalization [nlayer][d] + std::vector> w_attn_q_norm; // Q normalization [nlayer][dh] (Qwen3-specific) + std::vector> w_attn_k_norm; // K normalization [nlayer][dh] (Qwen3-specific) + std::vector> w_attn_q; // Q projection [nlayer][d, d/ndev] + std::vector> w_attn_k; // K projection [nlayer][d, (nkvh*dh)/ndev] + std::vector> w_attn_v; // V projection [nlayer][d, (nkvh*dh)/ndev] + std::vector> w_attn_o; // O projection [nlayer][d/ndev, d] + + // Per-layer MLP/MoE weights + std::vector> w_mlp_norm; // MLP layer normalization [nlayer][d] + + // Regular MLP weights (for non-MoE layers) + std::vector> w_mlp_gate; // MLP gate projection [nlayer][d, di/ndev] + std::vector> w_mlp_up; // MLP up projection [nlayer][d, di/ndev] + std::vector> w_mlp_down; // MLP down projection [nlayer][di/ndev, d] + + // MoE weights (for MoE layers) + std::vector> w_moe_gate; // Router/gating weights [nlayer][d, num_experts] + std::vector>> w_moe_experts_gate; // Expert gate proj [nlayer][num_experts_per_device][d, moe_di/ndev] + std::vector>> w_moe_experts_up; // Expert up proj [nlayer][num_experts_per_device][d, moe_di/ndev] + std::vector>> w_moe_experts_down; // Expert down proj [nlayer][num_experts_per_device][moe_di/ndev, d] + + // ================================ + // Layer Type Information + // ================================ + std::vector is_moe_layer; // Flag indicating which layers use MoE [nlayer] + + // ================================ + // MoE Runtime Statistics + // ================================ + std::vector> expert_usage_count; // Expert usage statistics [nlayer][num_experts] + + // ================================ + // Intermediate Computation Buffers + // ================================ + std::shared_ptr hidden_states; // Main hidden state buffer [max_batch_size * max_seq_len, d] + std::shared_ptr residual_states; // Residual connection buffer [max_batch_size * max_seq_len, d] + std::shared_ptr attn_output; // Attention output buffer [max_batch_size * max_seq_len, d] + std::shared_ptr mlp_output; // MLP/MoE output buffer [max_batch_size * max_seq_len, d] + + // MoE-specific computation buffers + std::shared_ptr router_logits; // Router logits [max_batch_size * max_seq_len, num_experts] + std::shared_ptr routing_weights; // Routing weights [max_batch_size * max_seq_len, num_experts_per_tok] + std::shared_ptr selected_experts; // Selected expert indices [max_batch_size * max_seq_len, num_experts_per_tok] + std::shared_ptr expert_outputs; // Combined expert outputs [max_batch_size * max_seq_len, d] + std::shared_ptr expert_intermediate; // Expert intermediate buffer [max_tokens_per_expert, moe_intermediate_size/ndev] +}; + +/* + * Qwen3-MoE模型主结构 + * + * 管理分布式Qwen3-MoE模型推理的完整状态： + * - 元数据和配置参数 + * - 设备资源和权重分区 + * - 工作线程和同步原语 + * - 请求处理协调 + */ +struct Qwen3MoeModel { + // ================================ + // Model Configuration + // ================================ + Qwen3MoeMeta meta; // Model architecture and MoE configuration + + // ================================ + // Distributed Resources + // ================================ + size_t ndev; // Number of devices for distributed inference + std::vector dev_resources; // Per-device resources and weights + std::vector states; // Per-device synchronization state + std::vector threads; // Worker threads for async processing + + // ================================ + // Shared Request Processing + // ================================ + Qwen3MoeInferRequest req; // Shared request data across all devices + + // ================================ + // Constructor + // ================================ + /* + * Initialize distributed Qwen3-MoE model with tensor parallelism + * + * Parameters: + * - meta: Model architecture including MoE configuration + * - weights: Model weight tensors including expert weights + * - device: InfiniCore device type + * - device_ids: Physical device IDs for distributed inference + * + * Initialization Process: + * 1. Create device resources and partition weights across devices + * 2. Distribute experts across devices for load balancing + * 3. Initialize InfiniCCL communication for multi-device coordination + * 4. Launch worker threads and wait for device initialization + * 5. Set up synchronization primitives for inference coordination + */ + Qwen3MoeModel(const Qwen3MoeMeta *, const Qwen3MoeWeights *, infiniDevice_t device, std::vector device_ids); +}; + +/* + * Qwen3-MoE KV 缓存结构 + * + * 管理用于高效自回归生成的键值缓存存储，与基础Qwen3相同。 + * MoE层共享相同的注意力机制，因此KV缓存结构保持不变。 + * + * 缓存组织： + * - 键（k）和值（v）的单独存储 + * - 为分布式推理在设备间分区 [ndev] + * - 每个 transformer 层的单独缓存 [nlayer] + * - 每个缓存张量的形状 [max_seq_len, nkvh/ndev, dh] + */ +struct Qwen3MoeKVCache { + // ================================ + // Cache Storage Tensors + // ================================ + std::vector>> k; // Key cache [ndev][nlayer] + std::vector>> v; // Value cache [ndev][nlayer] + /* + * Cache Tensor Dimensions per Device per Layer: + * Shape: [max_seq_len, nkvh/ndev, dh] + * - max_seq_len: Maximum context length (dctx) + * - nkvh/ndev: Key-value heads allocated to this device + * - dh: Head dimension + * + * Storage Layout in Memory: + * - Contiguous storage allows efficient slicing for different sequence lengths + * - Device partitioning enables parallel KV cache updates + * - Layer separation allows independent cache management per transformer layer + */ +}; + +// Forward declarations for worker functions +void qwen3_moe_device_worker( + DeviceQwen3MoeResource &resource, + Qwen3MoeInferState &state, + const Qwen3MoeMeta &meta, + Qwen3MoeInferRequest &req, + size_t device_id, + size_t ndev); + +#endif \ No newline at end of file diff --git a/infini-qwen3-moe/src/models/qw/qwen3_moe_kv_cache.cpp b/infini-qwen3-moe/src/models/qw/qwen3_moe_kv_cache.cpp new file mode 100755 index 00000000..fa3c6adf --- /dev/null +++ b/infini-qwen3-moe/src/models/qw/qwen3_moe_kv_cache.cpp @@ -0,0 +1,342 @@ +#include "qwen3_moe_impl.hpp" +#include "qwen3_moe_weight.hpp" +#include "../../tensor.hpp" +#include "../../utils.hpp" + +// Helper function declarations +void qwen3_moe_device_worker( + DeviceQwen3MoeResource &resource, + Qwen3MoeInferState &state, + const Qwen3MoeMeta &meta, + Qwen3MoeInferRequest &req, + size_t device_id, + size_t ndev); + +// Helper function to get size of data type (if not available elsewhere) +inline size_t sizeof_dtype(infiniDtype_t dtype) { + switch (dtype) { + case INFINI_DTYPE_F16: + case INFINI_DTYPE_BF16: + return 2; + case INFINI_DTYPE_F32: + case INFINI_DTYPE_I32: + return 4; + case INFINI_DTYPE_F64: + case INFINI_DTYPE_I64: + return 8; + default: + return 4; // fallback + } +} + +/* + * Qwen3-MoE KV 缓存实现 + * + * KV缓存在MoE模型中的工作原理与标准Transformer相同， + * 因为MoE只影响FFN层，注意力机制保持不变。 + */ + +__C struct Qwen3MoeKVCache * +createQwen3MoeKVCache(const struct Qwen3MoeModel *model) { + auto cache = new Qwen3MoeKVCache(); + + const auto &meta = model->meta; + size_t ndev = model->ndev; + size_t nlayer = meta.nlayer; + size_t nkvh = meta.nkvh; + size_t dh = meta.dh; + size_t dctx = meta.dctx; + size_t nkvh_per_dev = nkvh / ndev; + + // 为每个设备和每层初始化KV缓存 + cache->k.resize(ndev); + cache->v.resize(ndev); + + for (size_t idev = 0; idev < ndev; ++idev) { + cache->k[idev].resize(nlayer); + cache->v[idev].resize(nlayer); + + // 设置设备上下文 + RUN_INFINI(infinirtSetDevice(model->dev_resources[idev].device_id)); + + for (size_t layer = 0; layer < nlayer; ++layer) { + // Key cache: [dctx, nkvh_per_dev, dh] + cache->k[idev][layer] = Tensor::create( + {dctx, nkvh_per_dev, dh}, + meta.dt_logits, + model->dev_resources[idev].device + ); + + // Value cache: [dctx, nkvh_per_dev, dh] + cache->v[idev][layer] = Tensor::create( + {dctx, nkvh_per_dev, dh}, + meta.dt_logits, + model->dev_resources[idev].device + ); + + // 初始化为零 + infiniConstant(model->dev_resources[idev].handle, + cache->k[idev][layer].get(), 0.0f); + infiniConstant(model->dev_resources[idev].handle, + cache->v[idev][layer].get(), 0.0f); + } + } + + return cache; +} + +__C struct Qwen3MoeKVCache * +duplicateQwen3MoeKVCache(const struct Qwen3MoeModel *model, + const struct Qwen3MoeKVCache *src_cache, + uint32_t seq_len) { + auto cache = new Qwen3MoeKVCache(); + + const auto &meta = model->meta; + size_t ndev = model->ndev; + size_t nlayer = meta.nlayer; + size_t nkvh = meta.nkvh; + size_t dh = meta.dh; + size_t dctx = meta.dctx; + size_t nkvh_per_dev = nkvh / ndev; + + // 为每个设备和每层复制KV缓存 + cache->k.resize(ndev); + cache->v.resize(ndev); + + for (size_t idev = 0; idev < ndev; ++idev) { + cache->k[idev].resize(nlayer); + cache->v[idev].resize(nlayer); + + // 设置设备上下文 + RUN_INFINI(infinirtSetDevice(model->dev_resources[idev].device_id)); + + for (size_t layer = 0; layer < nlayer; ++layer) { + // 创建新的缓存张量 + cache->k[idev][layer] = Tensor::create( + {dctx, nkvh_per_dev, dh}, + meta.dt_logits, + model->dev_resources[idev].device + ); + cache->v[idev][layer] = Tensor::create( + {dctx, nkvh_per_dev, dh}, + meta.dt_logits, + model->dev_resources[idev].device + ); + + if (src_cache && seq_len > 0) { + // 复制前seq_len个位置的数据 + size_t copy_size = seq_len * nkvh_per_dev * dh * sizeof_dtype(meta.dt_logits); + + RUN_INFINI(infinirtMemcpy( + cache->k[idev][layer]->data(), + src_cache->k[idev][layer]->data(), + copy_size, + INFINIRT_MEMCPY_D2D + )); + + RUN_INFINI(infinirtMemcpy( + cache->v[idev][layer]->data(), + src_cache->v[idev][layer]->data(), + copy_size, + INFINIRT_MEMCPY_D2D + )); + } else { + // 初始化为零 + infiniConstant(model->dev_resources[idev].handle, + cache->k[idev][layer].get(), 0.0f); + infiniConstant(model->dev_resources[idev].handle, + cache->v[idev][layer].get(), 0.0f); + } + } + } + + return cache; +} + +__C void dropQwen3MoeKVCache(const struct Qwen3MoeModel *model, + struct Qwen3MoeKVCache *cache) { + if (!cache) return; + + // 清理所有KV缓存张量 + for (auto &dev_cache : cache->k) { + dev_cache.clear(); + } + for (auto &dev_cache : cache->v) { + dev_cache.clear(); + } + + delete cache; +} + +/* + * Qwen3-MoE 模型构造函数实现 + * + * 初始化分布式MoE模型，包括权重分区和工作线程设置 + */ +Qwen3MoeModel::Qwen3MoeModel(const Qwen3MoeMeta *meta, + const Qwen3MoeWeights *weights, + infiniDevice_t device, + std::vector device_ids) + : meta(*meta), ndev(device_ids.size()) { + + /* + * 1. 初始化设备资源 + */ + dev_resources.resize(ndev); + states.resize(ndev); + threads.resize(ndev); + + /* + * 2. 为每个设备分区权重和创建资源 + */ + for (size_t idev = 0; idev < ndev; ++idev) { + auto &resource = dev_resources[idev]; + resource.device = device; + resource.device_id = device_ids[idev]; + + // 设置设备上下文 + RUN_INFINI(infinirtSetDevice(resource.device_id)); + + /* + * 2.1 全局权重（在所有设备上复制） + */ + resource.w_in_embd = getQwen3MoeInEmbd(meta, weights); + resource.w_out_norm = getQwen3MoeOutNorm(meta, weights); + resource.w_out_embd = getQwen3MoeOutEmbd(meta, weights); + + // 创建RoPE表 + auto rope_shape = std::vector({meta->dctx, meta->dh / 2}); + resource.sin_table = Tensor::create(rope_shape, meta->dt_logits, device); + resource.cos_table = Tensor::create(rope_shape, meta->dt_logits, device); + + // 初始化RoPE表 (简化实现) + // 实际应该根据theta计算正弦和余弦值 + infiniConstant(resource.handle, resource.sin_table.get(), 0.0f); + infiniConstant(resource.handle, resource.cos_table.get(), 1.0f); + + /* + * 2.2 逐层权重分区 + */ + resource.w_attn_norm.resize(meta->nlayer); + resource.w_attn_q_norm.resize(meta->nlayer); + resource.w_attn_k_norm.resize(meta->nlayer); + resource.w_attn_q.resize(meta->nlayer); + resource.w_attn_k.resize(meta->nlayer); + resource.w_attn_v.resize(meta->nlayer); + resource.w_attn_o.resize(meta->nlayer); + resource.w_mlp_norm.resize(meta->nlayer); + resource.w_mlp_gate.resize(meta->nlayer); + resource.w_mlp_up.resize(meta->nlayer); + resource.w_mlp_down.resize(meta->nlayer); + resource.w_moe_gate.resize(meta->nlayer); + resource.w_moe_experts_gate.resize(meta->nlayer); + resource.w_moe_experts_up.resize(meta->nlayer); + resource.w_moe_experts_down.resize(meta->nlayer); + resource.is_moe_layer.resize(meta->nlayer); + resource.expert_usage_count.resize(meta->nlayer); + + for (size_t layer = 0; layer < meta->nlayer; ++layer) { + // 注意力权重 + resource.w_attn_norm[layer] = getQwen3MoeAttnNorm(meta, weights, layer); + resource.w_attn_q_norm[layer] = getQwen3MoeAttnQNorm(meta, weights, layer); + resource.w_attn_k_norm[layer] = getQwen3MoeAttnKNorm(meta, weights, layer); + resource.w_attn_q[layer] = getQwen3MoeAttnQ(meta, weights, layer, idev, ndev); + resource.w_attn_k[layer] = getQwen3MoeAttnK(meta, weights, layer, idev, ndev); + resource.w_attn_v[layer] = getQwen3MoeAttnV(meta, weights, layer, idev, ndev); + resource.w_attn_o[layer] = getQwen3MoeAttnO(meta, weights, layer, idev, ndev); + + // MLP归一化权重 + resource.w_mlp_norm[layer] = getQwen3MoeMLPNorm(meta, weights, layer); + + // 确定层类型（MoE还是普通MLP） + resource.is_moe_layer[layer] = isQwen3MoeMoeLayer(meta, layer); + + if (resource.is_moe_layer[layer]) { + // MoE层权重 + resource.w_moe_gate[layer] = getQwen3MoeMoeGate(meta, weights, layer); + + // 分配专家到设备 + size_t experts_per_device = (meta->num_experts + ndev - 1) / ndev; + size_t start_expert = idev * experts_per_device; + size_t end_expert = std::min(start_expert + experts_per_device, meta->num_experts); + size_t num_experts_on_device = end_expert - start_expert; + + resource.w_moe_experts_gate[layer].resize(num_experts_on_device); + resource.w_moe_experts_up[layer].resize(num_experts_on_device); + resource.w_moe_experts_down[layer].resize(num_experts_on_device); + + for (size_t local_expert = 0; local_expert < num_experts_on_device; ++local_expert) { + size_t global_expert = start_expert + local_expert; + resource.w_moe_experts_gate[layer][local_expert] = + getQwen3MoeMoeExpertGate(meta, weights, layer, global_expert, idev, ndev); + resource.w_moe_experts_up[layer][local_expert] = + getQwen3MoeMoeExpertUp(meta, weights, layer, global_expert, idev, ndev); + resource.w_moe_experts_down[layer][local_expert] = + getQwen3MoeMoeExpertDown(meta, weights, layer, global_expert, idev, ndev); + } + + // 初始化专家使用统计 + resource.expert_usage_count[layer].resize(meta->num_experts, 0); + } else { + // 普通MLP层权重 + resource.w_mlp_gate[layer] = getQwen3MoeMLPGate(meta, weights, layer, idev, ndev); + resource.w_mlp_up[layer] = getQwen3MoeMLPUp(meta, weights, layer, idev, ndev); + resource.w_mlp_down[layer] = getQwen3MoeMLPDown(meta, weights, layer, idev, ndev); + } + } + + /* + * 2.3 计算缓冲区 + */ + size_t max_batch_tokens = 4096; // 可配置的最大批处理tokens + + resource.hidden_states = Tensor::create({max_batch_tokens, meta->d}, meta->dt_logits, device); + resource.residual_states = Tensor::create({max_batch_tokens, meta->d}, meta->dt_logits, device); + resource.attn_output = Tensor::create({max_batch_tokens, meta->d}, meta->dt_logits, device); + resource.mlp_output = Tensor::create({max_batch_tokens, meta->d}, meta->dt_logits, device); + + // MoE特定的缓冲区 + if (meta->num_experts > 0) { + resource.router_logits = Tensor::create({max_batch_tokens, meta->num_experts}, meta->dt_logits, device); + resource.routing_weights = Tensor::create({max_batch_tokens, meta->num_experts_per_tok}, meta->dt_logits, device); + resource.selected_experts = Tensor::create({max_batch_tokens, meta->num_experts_per_tok}, INFINI_DTYPE_I32, device); + resource.expert_outputs = Tensor::create({max_batch_tokens, meta->d}, meta->dt_logits, device); + + size_t max_tokens_per_expert = max_batch_tokens; // 保守估计 + resource.expert_intermediate = Tensor::create({max_tokens_per_expert, meta->moe_intermediate_size / ndev}, meta->dt_logits, device); + } + } + + /* + * 3. 启动工作线程 + */ + for (size_t idev = 0; idev < ndev; ++idev) { + // 初始化同步状态 + states[idev].loaded = false; + states[idev].proceed = false; + states[idev].shutdown = false; + + // 启动工作线程 + threads[idev] = std::thread([this, idev]() { + qwen3_moe_device_worker( + dev_resources[idev], + states[idev], + meta, + req, + idev, + ndev + ); + }); + } + + /* + * 4. 等待所有设备初始化完成 + */ + for (size_t idev = 0; idev < ndev; ++idev) { + std::unique_lock lock(states[idev].mtx); + states[idev].cv_load.wait(lock, [&] { return states[idev].loaded; }); + } + + printf("Qwen3-MoE model initialized with %zu devices, %zu experts per layer\n", + ndev, meta->num_experts); +} \ No newline at end of file diff --git a/infini-qwen3-moe/src/models/qw/qwen3_moe_weight.hpp b/infini-qwen3-moe/src/models/qw/qwen3_moe_weight.hpp new file mode 100755 index 00000000..3d5c6735 --- /dev/null +++ b/infini-qwen3-moe/src/models/qw/qwen3_moe_weight.hpp @@ -0,0 +1,413 @@ +#pragma once + +#include "qwen3_moe_impl.hpp" +#include +#include + +// Helper function to get size of data type +inline size_t sizeof_dtype(infiniDtype_t dtype) { + switch (dtype) { + case INFINI_DTYPE_F16: + case INFINI_DTYPE_BF16: + return 2; + case INFINI_DTYPE_F32: + case INFINI_DTYPE_I32: + return 4; + case INFINI_DTYPE_F64: + case INFINI_DTYPE_I64: + return 8; + default: + return 4; // fallback + } +} + +/* + * Qwen3-MoE权重提取工具函数 + * 用于从原始权重数据中提取适合分布式推理的权重张量 + * + * MoE 特性： + * - 支持混合专家架构的权重分区 + * - 路由器权重的分布式处理 + * - 专家权重的设备间分区 + * - 稀疏激活的权重管理 + */ + +/* + * 获取输入嵌入权重 + * 形状: [dvoc, d] + */ +inline std::shared_ptr getQwen3MoeInEmbd( + Qwen3MoeMeta const *meta, + Qwen3MoeWeights const *w) { + auto shape = std::vector({meta->dvoc, meta->d}); + return Tensor::weight((char *)w->input_embd, meta->dt_logits, shape); +} + +/* + * 获取输出嵌入权重 + * 形状: [d, dvoc] 或 [dvoc, d] (根据transpose_linear_weights) + */ +inline std::shared_ptr getQwen3MoeOutEmbd( + Qwen3MoeMeta const *meta, + Qwen3MoeWeights const *w) { + if (w->transpose_linear_weights != 0) { + // 权重存储为 [dvoc, d]，需要转置为 [d, dvoc] + auto shape = std::vector({meta->dvoc, meta->d}); + return Tensor::weight((char *)w->output_embd, meta->dt_logits, shape) + ->permute({1, 0}); // 转置：[dvoc, d] -> [d, dvoc] + } else { + // 权重已存储为 [d, dvoc] + auto shape = std::vector({meta->d, meta->dvoc}); + return Tensor::weight((char *)w->output_embd, meta->dt_logits, shape); + } +} + +/* + * 获取输出层归一化权重 + * 形状: [d] + */ +inline std::shared_ptr getQwen3MoeOutNorm( + Qwen3MoeMeta const *meta, + Qwen3MoeWeights const *w) { + auto shape = std::vector({meta->d}); + return Tensor::weight((char *)w->output_norm, w->dt_norm, shape); +} + +/* + * 获取注意力层归一化权重 + * 形状: [d] + */ +inline std::shared_ptr getQwen3MoeAttnNorm( + Qwen3MoeMeta const *meta, + Qwen3MoeWeights const *w, + size_t layer) { + auto shape = std::vector({meta->d}); + return Tensor::weight((char *)(w->attn_norm[layer]), w->dt_norm, shape); +} + +/* + * 获取Q归一化权重 (Qwen3特有) + * 形状: [dh] + */ +inline std::shared_ptr getQwen3MoeAttnQNorm( + Qwen3MoeMeta const *meta, + Qwen3MoeWeights const *w, + size_t layer) { + if (w->attn_q_norm == nullptr || w->attn_q_norm[layer] == nullptr) { + return nullptr; // 某些模型可能不包含Q/K归一化 + } + auto shape = std::vector({meta->dh}); + return Tensor::weight((char *)(w->attn_q_norm[layer]), w->dt_norm, shape); +} + +/* + * 获取K归一化权重 (Qwen3特有) + * 形状: [dh] + */ +inline std::shared_ptr getQwen3MoeAttnKNorm( + Qwen3MoeMeta const *meta, + Qwen3MoeWeights const *w, + size_t layer) { + if (w->attn_k_norm == nullptr || w->attn_k_norm[layer] == nullptr) { + return nullptr; // 某些模型可能不包含Q/K归一化 + } + auto shape = std::vector({meta->dh}); + return Tensor::weight((char *)(w->attn_k_norm[layer]), w->dt_norm, shape); +} + +/* + * 为分布式推理提取 Q 投影权重 + * Q 投影：[d, d] -> 每设备 [d, d/ndev] + */ +inline std::shared_ptr getQwen3MoeAttnQ( + Qwen3MoeMeta const *meta, + Qwen3MoeWeights const *w, + size_t layer, size_t idev, size_t ndev) { + + size_t d = meta->d; + size_t output_dim_per_device = d / ndev; + size_t offset = idev * output_dim_per_device * d * sizeof_dtype(w->dt_mat); + + if (w->transpose_linear_weights != 0) { + // 转置格式：[d, output_dim_per_device] -> [output_dim_per_device, d] + auto shape = std::vector({d, output_dim_per_device}); + return Tensor::weight((char *)(w->attn_q_proj[layer]) + offset, w->dt_mat, shape) + ->permute({1, 0}); + } else { + // 标准格式：[d, output_dim_per_device] + auto shape = std::vector({d, output_dim_per_device}); + return Tensor::weight((char *)(w->attn_q_proj[layer]) + offset, w->dt_mat, shape); + } +} + +/* + * 为分布式推理提取 K 投影权重 + * K 投影：[d, nkvh * dh] -> 每设备 [d, (nkvh * dh)/ndev] + */ +inline std::shared_ptr getQwen3MoeAttnK( + Qwen3MoeMeta const *meta, + Qwen3MoeWeights const *w, + size_t layer, size_t idev, size_t ndev) { + + size_t d = meta->d; + size_t nkvh = meta->nkvh; + size_t dh = meta->dh; + size_t kv_dim = nkvh * dh; + size_t kv_dim_per_device = kv_dim / ndev; + size_t offset = idev * kv_dim_per_device * d * sizeof_dtype(w->dt_mat); + + if (w->transpose_linear_weights != 0) { + auto shape = std::vector({d, kv_dim_per_device}); + return Tensor::weight((char *)(w->attn_k_proj[layer]) + offset, w->dt_mat, shape) + ->permute({1, 0}); + } else { + auto shape = std::vector({d, kv_dim_per_device}); + return Tensor::weight((char *)(w->attn_k_proj[layer]) + offset, w->dt_mat, shape); + } +} + +/* + * 为分布式推理提取 V 投影权重 + * V 投影：[d, nkvh * dh] -> 每设备 [d, (nkvh * dh)/ndev] + */ +inline std::shared_ptr getQwen3MoeAttnV( + Qwen3MoeMeta const *meta, + Qwen3MoeWeights const *w, + size_t layer, size_t idev, size_t ndev) { + + size_t d = meta->d; + size_t nkvh = meta->nkvh; + size_t dh = meta->dh; + size_t kv_dim = nkvh * dh; + size_t kv_dim_per_device = kv_dim / ndev; + size_t offset = idev * kv_dim_per_device * d * sizeof_dtype(w->dt_mat); + + if (w->transpose_linear_weights != 0) { + auto shape = std::vector({d, kv_dim_per_device}); + return Tensor::weight((char *)(w->attn_v_proj[layer]) + offset, w->dt_mat, shape) + ->permute({1, 0}); + } else { + auto shape = std::vector({d, kv_dim_per_device}); + return Tensor::weight((char *)(w->attn_v_proj[layer]) + offset, w->dt_mat, shape); + } +} + +/* + * 为分布式推理提取 O 投影权重 + * O 投影：[d, d] -> 每设备 [d/ndev, d] + */ +inline std::shared_ptr getQwen3MoeAttnO( + Qwen3MoeMeta const *meta, + Qwen3MoeWeights const *w, + size_t layer, size_t idev, size_t ndev) { + + size_t d = meta->d; + size_t input_dim_per_device = d / ndev; + size_t offset = idev * input_dim_per_device * d * sizeof_dtype(w->dt_mat); + + if (w->transpose_linear_weights != 0) { + auto shape = std::vector({input_dim_per_device, d}); + return Tensor::weight((char *)(w->attn_o_proj[layer]) + offset, w->dt_mat, shape) + ->permute({1, 0}); + } else { + auto shape = std::vector({input_dim_per_device, d}); + return Tensor::weight((char *)(w->attn_o_proj[layer]) + offset, w->dt_mat, shape); + } +} + +/* + * 获取MLP层归一化权重 + * 形状: [d] + */ +inline std::shared_ptr getQwen3MoeMLPNorm( + Qwen3MoeMeta const *meta, + Qwen3MoeWeights const *w, + size_t layer) { + auto shape = std::vector({meta->d}); + return Tensor::weight((char *)(w->mlp_norm[layer]), w->dt_norm, shape); +} + +/* + * 检查指定层是否为MoE层 + */ +inline bool isQwen3MoeMoeLayer( + Qwen3MoeMeta const *meta, + size_t layer_idx) { + + // 检查是否在mlp_only_layers中 + for (size_t i = 0; i < meta->num_mlp_only_layers; ++i) { + if (meta->mlp_only_layers[i] == layer_idx) { + return false; // 是纯MLP层 + } + } + + // 检查是否满足MoE层的条件 + return (meta->num_experts > 0 && + (layer_idx + 1) % meta->decoder_sparse_step == 0); +} + +/* + * 获取普通MLP gate投影权重 (非MoE层) + * 形状: [d, di] -> 每设备 [d, di/ndev] + */ +inline std::shared_ptr getQwen3MoeMLPGate( + Qwen3MoeMeta const *meta, + Qwen3MoeWeights const *w, + size_t layer, size_t idev, size_t ndev) { + + size_t d = meta->d; + size_t di = meta->di; + size_t di_per_device = di / ndev; + size_t offset = idev * di_per_device * d * sizeof_dtype(w->dt_mat); + + if (w->transpose_linear_weights != 0) { + auto shape = std::vector({d, di_per_device}); + return Tensor::weight((char *)(w->mlp_gate_proj[layer]) + offset, w->dt_mat, shape) + ->permute({1, 0}); + } else { + auto shape = std::vector({d, di_per_device}); + return Tensor::weight((char *)(w->mlp_gate_proj[layer]) + offset, w->dt_mat, shape); + } +} + +/* + * 获取普通MLP up投影权重 (非MoE层) + * 形状: [d, di] -> 每设备 [d, di/ndev] + */ +inline std::shared_ptr getQwen3MoeMLPUp( + Qwen3MoeMeta const *meta, + Qwen3MoeWeights const *w, + size_t layer, size_t idev, size_t ndev) { + + size_t d = meta->d; + size_t di = meta->di; + size_t di_per_device = di / ndev; + size_t offset = idev * di_per_device * d * sizeof_dtype(w->dt_mat); + + if (w->transpose_linear_weights != 0) { + auto shape = std::vector({d, di_per_device}); + return Tensor::weight((char *)(w->mlp_up_proj[layer]) + offset, w->dt_mat, shape) + ->permute({1, 0}); + } else { + auto shape = std::vector({d, di_per_device}); + return Tensor::weight((char *)(w->mlp_up_proj[layer]) + offset, w->dt_mat, shape); + } +} + +/* + * 获取普通MLP down投影权重 (非MoE层) + * 形状: [di, d] -> 每设备 [di/ndev, d] + */ +inline std::shared_ptr getQwen3MoeMLPDown( + Qwen3MoeMeta const *meta, + Qwen3MoeWeights const *w, + size_t layer, size_t idev, size_t ndev) { + + size_t d = meta->d; + size_t di = meta->di; + size_t di_per_device = di / ndev; + size_t offset = idev * di_per_device * d * sizeof_dtype(w->dt_mat); + + if (w->transpose_linear_weights != 0) { + auto shape = std::vector({di_per_device, d}); + return Tensor::weight((char *)(w->mlp_down_proj[layer]) + offset, w->dt_mat, shape) + ->permute({1, 0}); + } else { + auto shape = std::vector({di_per_device, d}); + return Tensor::weight((char *)(w->mlp_down_proj[layer]) + offset, w->dt_mat, shape); + } +} + +/* + * 获取MoE路由器权重 + * 形状: [d, num_experts] + */ +inline std::shared_ptr getQwen3MoeMoeGate( + Qwen3MoeMeta const *meta, + Qwen3MoeWeights const *w, + size_t layer) { + + size_t d = meta->d; + size_t num_experts = meta->num_experts; + + if (w->transpose_linear_weights != 0) { + auto shape = std::vector({d, num_experts}); + return Tensor::weight((char *)(w->moe_gate[layer]), w->dt_mat, shape) + ->permute({1, 0}); + } else { + auto shape = std::vector({d, num_experts}); + return Tensor::weight((char *)(w->moe_gate[layer]), w->dt_mat, shape); + } +} + +/* + * 获取MoE专家gate投影权重 + * 形状: [d, moe_intermediate_size] -> 每设备 [d, moe_intermediate_size/ndev] + */ +inline std::shared_ptr getQwen3MoeMoeExpertGate( + Qwen3MoeMeta const *meta, + Qwen3MoeWeights const *w, + size_t layer, size_t expert_idx, size_t idev, size_t ndev) { + + size_t d = meta->d; + size_t moe_di = meta->moe_intermediate_size; + size_t moe_di_per_device = moe_di / ndev; + size_t offset = idev * moe_di_per_device * d * sizeof_dtype(w->dt_mat); + + if (w->transpose_linear_weights != 0) { + auto shape = std::vector({d, moe_di_per_device}); + return Tensor::weight((char *)(w->moe_experts_gate_proj[layer][expert_idx]) + offset, w->dt_mat, shape) + ->permute({1, 0}); + } else { + auto shape = std::vector({d, moe_di_per_device}); + return Tensor::weight((char *)(w->moe_experts_gate_proj[layer][expert_idx]) + offset, w->dt_mat, shape); + } +} + +/* + * 获取MoE专家up投影权重 + * 形状: [d, moe_intermediate_size] -> 每设备 [d, moe_intermediate_size/ndev] + */ +inline std::shared_ptr getQwen3MoeMoeExpertUp( + Qwen3MoeMeta const *meta, + Qwen3MoeWeights const *w, + size_t layer, size_t expert_idx, size_t idev, size_t ndev) { + + size_t d = meta->d; + size_t moe_di = meta->moe_intermediate_size; + size_t moe_di_per_device = moe_di / ndev; + size_t offset = idev * moe_di_per_device * d * sizeof_dtype(w->dt_mat); + + if (w->transpose_linear_weights != 0) { + auto shape = std::vector({d, moe_di_per_device}); + return Tensor::weight((char *)(w->moe_experts_up_proj[layer][expert_idx]) + offset, w->dt_mat, shape) + ->permute({1, 0}); + } else { + auto shape = std::vector({d, moe_di_per_device}); + return Tensor::weight((char *)(w->moe_experts_up_proj[layer][expert_idx]) + offset, w->dt_mat, shape); + } +} + +/* + * 获取MoE专家down投影权重 + * 形状: [moe_intermediate_size, d] -> 每设备 [moe_intermediate_size/ndev, d] + */ +inline std::shared_ptr getQwen3MoeMoeExpertDown( + Qwen3MoeMeta const *meta, + Qwen3MoeWeights const *w, + size_t layer, size_t expert_idx, size_t idev, size_t ndev) { + + size_t d = meta->d; + size_t moe_di = meta->moe_intermediate_size; + size_t moe_di_per_device = moe_di / ndev; + size_t offset = idev * moe_di_per_device * d * sizeof_dtype(w->dt_mat); + + if (w->transpose_linear_weights != 0) { + auto shape = std::vector({moe_di_per_device, d}); + return Tensor::weight((char *)(w->moe_experts_down_proj[layer][expert_idx]) + offset, w->dt_mat, shape) + ->permute({1, 0}); + } else { + auto shape = std::vector({moe_di_per_device, d}); + return Tensor::weight((char *)(w->moe_experts_down_proj[layer][expert_idx]) + offset, w->dt_mat, shape); + } +} \ No newline at end of file diff --git a/infini-qwen3-moe/src/models/qw/qwen3_weight.hpp b/infini-qwen3-moe/src/models/qw/qwen3_weight.hpp new file mode 100755 index 00000000..10dbd001 --- /dev/null +++ b/infini-qwen3-moe/src/models/qw/qwen3_weight.hpp @@ -0,0 +1,759 @@ +/* + * Jiuge 模型权重提取工具 + * + * 此头文件提供从全局权重存储中提取和重塑模型权重到设备特定张量对象的实用函数。主要特性： + * + * - 用于跨设备张量并行的分布式权重分区 + * - 自动张量形状处理和不同存储格式的转置 + * - 使用预计算三角值生成 RoPE（旋转位置嵌入）表 + * - 支持转置和非转置线性层权重格式 + * + * 权重分布策略： + * - 全局张量（嵌入，归一化）：在所有设备上复制 + * - 注意力权重：按注意力头在设备间分区 + * - FFN 权重：按中间维度在设备间分区 + * - RoPE 表：通过数学计算按需生成 + * + * 所有函数返回 shared_ptr 以实现自动内存管理。 + */ + +#ifndef QWEN3_WEIGHT_HPP +#define QWEN3_WEIGHT_HPP + +#include "qwen3_impl.hpp" + +#include +/* + * 提取输入嵌入表 + * + * 为输入 token 嵌入查找表创建张量包装器。 + * 此表将 token ID 映射到其对应的稠密向量表示。 + * + * 张量属性： + * - 形状：[dvoc, d] 其中 dvoc = 词汇表大小，d = 模型维度 + * - 数据类型：meta->dt_logits（通常为 FP16、BF16 或 FP32） + * - 内存：引用全局权重存储（无复制） + * - 分布：在所有设备上复制（未分区） + * + * 用法：在推理过程中，token ID 用于索引此表 + * 以检索 transformer 处理前的初始隐藏表示。 + * + * 参数： + * - meta：包含维度和数据类型的模型元数据 + * - w：包含所有模型参数的全局权重存储 + * + * 返回：输入嵌入表的共享张量包装器 [dvoc, d] + */ +inline std::shared_ptr getQwen3InEmbd( + Qwen3Meta const *meta, + Qwen3Weights const *w) { + auto shape = std::vector({meta->dvoc, meta->d}); + return Tensor::weight((char *)w->input_embd, meta->dt_logits, shape); +} + + +inline std::shared_ptr getQwen3AttnNorm( + const Qwen3Meta *meta, + const Qwen3Weights *w, + size_t layer) { + + auto shape = std::vector{meta->d}; + return Tensor::weight((char *)(w->attn_norm[layer]), w->dt_norm, shape); +} + + + +/* + * 提取最终层归一化权重 + * + * 为应用在语言模型头之前的最终 RMSNorm 层创建张量包装器。 + * 此归一化在词汇表投影之前稳定最终的隐藏表示。 + * + * 张量属性： + * - 形状：[d] 其中 d = 模型隐藏维度 + * - 数据类型：w->dt_norm（归一化参数数据类型） + * - 内存：引用全局权重存储 + * - 分布：在所有设备上复制 + * + * RMSNorm 公式：y = x / √(mean(x²) + ε) * γ + * 其中 γ 是存储在此张量中的缩放参数。 + * + * 参数： + * - meta：维度信息的模型元数据 + * - w：全局权重存储 + * + * 返回：输出归一化权重的共享张量包装器 [d] + */ + + +inline std::shared_ptr getQwen3OutNorm( + Qwen3Meta const *meta, + Qwen3Weights const *w) { + auto shape = std::vector({meta->d}); + return Tensor::weight((char *)w->output_norm, w->dt_norm, shape); +} + +/* + * 提取输出嵌入 / 语言模型头权重 + * + * 为将隐藏状态映射到词汇表 logits 以进行下一个 token 预测的最终线性投影创建张量包装器。 + * 此层通常与输入嵌入表绑定以提高参数效率。 + * + * 权重格式处理： + * - transpose_linear_weights = 0：权重存储为 [d, dvoc]（标准格式） + * - transpose_linear_weights ≠ 0：权重存储为 [dvoc, d]（转置格式） + * + * 当权重以转置方式存储时，我们应用 permute({1, 0}) 以获得 + * 正确的 [d, dvoc] 形状用于矩阵乘法：hidden_states @ weights。 + * + * 矩阵运算：logits = hidden_states @ weights + * - hidden_states：[batch_size, d] + * - weights：[d, dvoc] + * - logits：[batch_size, dvoc] + * + * 张量属性： + * - 最终形状：[d, dvoc] 无论存储格式如何 + * - 数据类型：meta->dt_logits + * - 分布：在所有设备上复制 + * + * 参数： + * - meta：维度的模型元数据 + * - w：带有格式标志的全局权重存储 + * + * 返回：输出投影权重的共享张量包装器 [d, dvoc] + */ +inline std::shared_ptr getQwen3OutEmbd( + Qwen3Meta const *meta, + Qwen3Weights const *w) { + if (w->transpose_linear_weights != 0) { + // 权重存储为 [dvoc, d]，需要转置为 [d, dvoc] + auto shape = std::vector({meta->dvoc, meta->d}); + return Tensor::weight((char *)w->output_embd, meta->dt_logits, shape) + ->permute({1, 0}); // 转置：[dvoc, d] -> [d, dvoc] + } else { + // 权重已存储为 [d, dvoc] + auto shape = std::vector({meta->d, meta->dvoc}); + return Tensor::weight((char *)w->output_embd, meta->dt_logits, shape); + } +} + +/* + * 提取特定层的注意力层归一化权重 + * + * 为每个 transformer 层中应用在注意力机制之前的 RMSNorm 权重创建张量包装器。 + * 这种注意力前归一化对训练稳定性和性能至关重要。 + * + * 张量属性： + * - 形状：[d] 其中 d = 模型隐藏维度 + * - 数据类型：w->dt_norm（归一化数据类型） + * - 分布：在所有设备上复制（所有设备相同） + * - 用法：在注意力计算中的 QKV 投影之前应用 + * + * RMSNorm 应用：normalized = (x / √(mean(x²) + ε)) * scale_weights + * + * 参数： + * - meta：维度信息的模型元数据 + * - w：全局权重存储 + * - layer：层索引（0 到 nlayer-1） + * + * 返回：注意力归一化权重的共享张量包装器 [d] + */ + +inline std::shared_ptr getQwen3QNorm( + Qwen3Meta const *meta, + Qwen3Weights const *w, + size_t layer) { + // Q norm is applied per head dimension + auto shape = std::vector({meta->dh}); + return Tensor::weight((char *)(w->attn_q_norm[layer]), w->dt_norm, shape); +} + +// Qwen3-specific: K normalization weights +inline std::shared_ptr getQwen3KNorm( + Qwen3Meta const *meta, + Qwen3Weights const *w, + size_t layer) { + // K norm is applied per head dimension + auto shape = std::vector({meta->dh}); + return Tensor::weight((char *)(w->attn_k_norm[layer]), w->dt_norm, shape); +} + + + + +/* + * 为分布式注意力提取 QKV 投影权重 + * + * 为查询、键、值投影权重创建设备特定的张量包装器。 + * 在分布式推理中，注意力头在设备间分区， + * 因此每个设备获得总 QKV 投影矩阵的一个切片。 + */ + +/* + * 为分布式推理提取 Q 投影权重 + * Q 投影：[2048, 2048] -> 每设备 [2048, 2048/ndev] + */ +inline std::shared_ptr getQwen3AttnQ( + Qwen3Meta const *meta, + Qwen3Weights const *w, + size_t layer, size_t idev, size_t ndev) { + + auto nh = meta->nh; // 16 个查询头 + auto dh = meta->dh; // 128 头维度 + auto d = meta->d; // 2048 模型维度 + /* + * Q 投影分布式切片计算： + * - 总输出维度：nh * dh = 16 * 128 = 2048 + * - 每设备输出维度：(nh / ndev) * dh = (16/ndev) * 128 + * - 内存偏移：跳过前面设备的头数据 + */ + size_t heads_per_device = nh / ndev; // 每设备头数 + size_t output_dim_per_device = heads_per_device * dh; // 每设备输出维度 + size_t offset = idev * output_dim_per_device * d * dsize(w->dt_mat); // 字节偏移 + + if (w->transpose_linear_weights != 0) { + // 存储格式：[output_dim_per_device, d] -> 转置为 [d, output_dim_per_device] + auto shape = std::vector({output_dim_per_device, d}); + return Tensor::weight((char *)(w->attn_q_proj[layer]) + offset, w->dt_mat, shape) + ->permute({1, 0}); + } else { + // 标准格式：[d, output_dim_per_device] + auto shape = std::vector({d, output_dim_per_device}); + return Tensor::weight((char *)(w->attn_q_proj[layer]) + offset, w->dt_mat, shape); + } +} + +/* + * 为分布式推理提取 K 投影权重 + * + * K 投影：[1024, 2048] -> 每设备 [1024/ndev, 2048] + */ +inline std::shared_ptr getQwen3AttnK( + Qwen3Meta const *meta, + Qwen3Weights const *w, + size_t layer, size_t idev, size_t ndev) { + + auto nkvh = meta->nkvh; // 8 个键值头 + auto dh = meta->dh; // 128 头维度 + auto d = meta->d; // 2048 模型维度 + + /* + * K 投影分布式切片计算： + * - 总输出维度：nkvh * dh = 8 * 128 = 1024 + * - 每设备输出维度：(nkvh / ndev) * dh = (8/ndev) * 128 + */ + size_t kv_heads_per_device = nkvh / ndev; + size_t output_dim_per_device = kv_heads_per_device * dh; + size_t offset = idev * output_dim_per_device * d * dsize(w->dt_mat); + + if (w->transpose_linear_weights != 0) { + auto shape = std::vector({output_dim_per_device, d}); + return Tensor::weight((char *)(w->attn_k_proj[layer]) + offset, w->dt_mat, shape) + ->permute({1, 0}); + } else { + auto shape = std::vector({d, output_dim_per_device}); + return Tensor::weight((char *)(w->attn_k_proj[layer]) + offset, w->dt_mat, shape); + } +} + +/* + * 为分布式推理提取 V 投影权重 + * + * V 投影：[1024, 2048] -> 每设备 [1024/ndev, 2048] + * 与 K 投影完全相同的分布策略 + */ +inline std::shared_ptr getQwen3AttnV( + Qwen3Meta const *meta, + Qwen3Weights const *w, + size_t layer, size_t idev, size_t ndev) { + + auto nkvh = meta->nkvh; // 8 个键值头 + auto dh = meta->dh; // 128 头维度 + auto d = meta->d; // 2048 模型维度 + + /* + * V 投影分布式切片计算（与 K 相同） + */ + size_t kv_heads_per_device = nkvh / ndev; + size_t output_dim_per_device = kv_heads_per_device * dh; + size_t offset = idev * output_dim_per_device * d * dsize(w->dt_mat); + + if (w->transpose_linear_weights != 0) { + auto shape = std::vector({output_dim_per_device, d}); + return Tensor::weight((char *)(w->attn_v_proj[layer]) + offset, w->dt_mat, shape) + ->permute({1, 0}); + } else { + auto shape = std::vector({d, output_dim_per_device}); + return Tensor::weight((char *)(w->attn_v_proj[layer]) + offset, w->dt_mat, shape); + } +} + +/* + * 为分布式推理提取注意力输出投影权重 + * + * 为注意力输出投影创建设备特定的张量包装器， + * 该投影将多头注意力输出合并回模型维度。 + * 在分布式推理中，此投影采用来自此设备头的连接注意力输出， + * 并将其投影到完整的模型维度。 + * + * 权重分区策略： + * - 每个设备处理 nh/ndev 个注意力头 + * - 输入维度：nh/ndev * dh（此设备上的注意力头） + * - 输出维度：d（完整模型维度，所有设备相同） + * - 结果通过 all-reduce 操作在设备间求和 + * + * 内存布局和偏移计算： + * - 全局权重形状：[nh * dh, d] 或转置 + * - 每设备切片：[nh/ndev * dh, d] + * - 字节偏移：idev * (nh/ndev * dh) * d * sizeof(data_type) + * + * 矩阵运算：output = attention_heads @ output_weights + * - attention_heads：[batch_size, nh/ndev * dh]（设备特定） + * - output_weights：[nh/ndev * dh, d]（设备特定切片） + * - output：[batch_size, d]（完整模型维度） + * + * 分布式计算： + * 1. 每个设备计算其注意力头的部分输出 + * 2. All-reduce 在设备间求和部分输出 + * 3. 最终结果表示完整的注意力输出 + * + * 参数： + * - meta：维度的模型元数据 + * - w：全局权重存储 + * - layer：Transformer 层索引 + * - idev：当前设备索引 + * - ndev：设备总数 + * + * 返回：设备特定的注意力输出权重 [nh/ndev * dh, d] 或 [d, nh/ndev * dh] + */ +inline std::shared_ptr getQwen3AttnO(Qwen3Meta const *meta, + Qwen3Weights const *w, size_t layer, + size_t idev, size_t ndev) { + // 提取模型维度 + auto nh = meta->nh; // 总查询头数 + auto dh = meta->dh; // 头维度 + auto d = meta->d; // 模型隐藏维度 + + /* + * 计算设备切片的内存偏移 + * + * 每个设备获得对应其分配的注意力头的输出投影切片： + * offset = device_index * heads_per_device * head_dim * model_dim * data_size + */ + size_t offset = idev * d * (nh / ndev * dh) * dsize(w->dt_mat); + + /* + * 处理不同的权重存储格式 + */ + if (w->transpose_linear_weights != 0) { + // 权重存储为 [d, nh/ndev * dh]，转置为 [nh/ndev * dh, d] + auto shape = std::vector({d, nh / ndev * dh}); + return Tensor::weight((char *)(w->attn_o_proj[layer]) + offset, w->dt_mat, shape) + ->permute({1, 0}); // 转置到正确方向 + } else { + // 权重已处于正确格式：[nh/ndev * dh, d] + auto shape = std::vector({nh / ndev * dh, d}); + return Tensor::weight((char *)(w->attn_o_proj[layer]) + offset, w->dt_mat, shape); + } +} + +/* + * 提取特定层的 FFN 层归一化权重 + * + * 为每个 transformer 层中应用在前馈网络之前的 RMSNorm 权重创建张量包装器。 + * 这种 FFN 前归一化稳定训练并提高性能。 + * + * 张量属性： + * - 形状：[d] 其中 d = 模型隐藏维度 + * - 数据类型：w->dt_norm（归一化参数数据类型） + * - 分布：在所有设备上复制（所有设备相同） + * - 用法：在 FFN 门控/上升投影之前应用 + * + * RMSNorm 公式：y = (x / √(mean(x²) + ε)) * scale_weights + * 这在 FFN 处理之前归一化注意力后的隐藏状态。 + * + * 参数： + * - meta：维度信息的模型元数据 + * - w：全局权重存储 + * - layer：层索引（0 到 nlayer-1） + * + * 返回：FFN 归一化权重的共享张量包装器 [d] + */ +inline std::shared_ptr getQwen3MLPNorm( + Qwen3Meta const *meta, + Qwen3Weights const *w, + size_t layer) { + auto shape = std::vector({meta->d}); + return Tensor::weight((char *)(w->mlp_norm[layer]), w->dt_norm, shape); +} + +/* + * 为分布式推理提取 FFN 门控和上升投影权重 + * + * 为用于 SwiGLU 激活函数的组合门控和上升投影创建设备特定的张量包装器。 + * 这些投影将模型维度扩展到中间 FFN 维度。 + * + * SwiGLU 架构： + * - 门控投影：用于门控机制的线性变换 + * - 上升投影：用于值计算的线性变换 + * - 组合操作：gate_output = gate_proj(x), up_output = up_proj(x) + * - SwiGLU 激活：output = gate_output * swish(up_output) + * + * 权重分区策略： + * - 总中间维度：di（FFN 扩展因子 * model_dim） + * - 每设备维度：di/ndev（在设备间分布） + * - 门控 + 上升组合：2 * di/ndev（两个投影连接） + * - 每个设备处理中间维度的一个切片 + * + * 内存布局： + * - 全局权重形状：[d, 2*di] 或转置 + * - 每设备切片：[d, 2*di/ndev] + * - 连接格式：沿输出维度的 [gate_weights, up_weights] + * - 字节偏移：idev * (2*di/ndev) * d * sizeof(data_type) + * + * 矩阵运算： + * [gate_output, up_output] = hidden_states @ [gate_weights, up_weights] + * - hidden_states：[batch_size, d] + * - combined_weights：[d, 2*di/ndev] + * - outputs：[batch_size, 2*di/ndev] = [gate_batch, up_batch] + * + * 参数： + * - meta：维度的模型元数据 + * - w：全局权重存储 + * - layer：Transformer 层索引 + * - idev：当前设备索引 + * - ndev：设备总数 + * + * 返回：设备特定的 FFN 门控和上升权重 [d, 2*di/ndev] 或 [2*di/ndev, d] + */ +/* + * 为分布式推理提取 FFN Gate 投影权重 + * + * Gate 投影：[6144, 2048] -> 每设备 [6144/ndev, 2048] + */ +inline std::shared_ptr getQwen3MLPGate( + Qwen3Meta const *meta, + Qwen3Weights const *w, + size_t layer, size_t idev, size_t ndev) { + + auto di = meta->di; // 6144 中间维度 + auto d = meta->d; // 2048 模型维度 + + /* + * Gate 投影分布式切片计算： + * - 总输出维度：di = 6144 + * - 每设备输出维度：di / ndev = 6144/ndev + * - 内存偏移：跳过前面设备的中间维度数据 + */ + + size_t intermediate_per_device = di / ndev; + size_t offset = idev * intermediate_per_device * d * dsize(w->dt_mat); + + if (w->transpose_linear_weights != 0) { + // 存储格式：[intermediate_per_device, d] -> 转置为 [d, intermediate_per_device] + auto shape = std::vector({intermediate_per_device, d}); + return Tensor::weight((char *)(w->mlp_gate_proj[layer]) + offset, w->dt_mat, shape) + ->permute({1, 0}); + } else { + // 标准格式：[d, intermediate_per_device] + auto shape = std::vector({d, intermediate_per_device}); + return Tensor::weight((char *)(w->mlp_gate_proj[layer]) + offset, w->dt_mat, shape); + } +} + +/* + * 为分布式推理提取 FFN Up 投影权重 + * + * Up 投影：[6144, 2048] -> 每设备 [6144/ndev, 2048] + * 与 Gate 投影完全相同的分布策略 + */ +inline std::shared_ptr getQwen3MLPUp( + Qwen3Meta const *meta, + Qwen3Weights const *w, + size_t layer, size_t idev, size_t ndev) { + + auto di = meta->di; // 6144 中间维度 + auto d = meta->d; // 2048 模型维度 + + /* + * Up 投影分布式切片计算（与 Gate 相同） + * - 总输出维度：di = 6144 + * - 每设备输出维度：di / ndev = 6144/ndev + */ + size_t intermediate_per_device = di / ndev; + size_t offset = idev * intermediate_per_device * d * dsize(w->dt_mat); + + if (w->transpose_linear_weights != 0) { + auto shape = std::vector({intermediate_per_device, d}); + return Tensor::weight((char *)(w->mlp_up_proj[layer]) + offset, w->dt_mat, shape) + ->permute({1, 0}); + } else { + auto shape = std::vector({d, intermediate_per_device}); + return Tensor::weight((char *)(w->mlp_up_proj[layer]) + offset, w->dt_mat, shape); + } +} + +/* + * 为分布式推理提取 FFN 下降投影权重 + * + * 为将中间 FFN 维度映射回模型维度的下降投影创建设备特定的张量包装器。 + * 这在 SwiGLU 激活后完成 FFN 计算。 + * + * FFN 下降投影作用： + * - 输入：SwiGLU 激活的中间表示 [batch, di/ndev] + * - 输出：回到模型维度的隐藏状态 [batch, d] + * - 目的：将扩展的中间特征投影回残差流 + * + * 权重分区策略： + * - 输入维度：di/ndev（每设备中间维度） + * - 输出维度：d（完整模型维度，所有设备相同） + * - 每个设备处理其中间维度的切片 + * - 结果通过 all-reduce 在设备间求和 + * + * 内存布局： + * - 全局权重形状：[di, d] 或转置 + * - 每设备切片：[di/ndev, d] + * - 字节偏移：idev * (di/ndev) * d * sizeof(data_type) + * + * 矩阵运算：output = intermediate_activated @ down_weights + * - intermediate_activated：[batch_size, di/ndev]（SwiGLU 后） + * - down_weights：[di/ndev, d]（设备特定切片） + * - output：[batch_size, d]（回到模型维度） + * + * 分布式计算： + * 1. 每个设备计算其中间切片的部分输出 + * 2. All-reduce 在设备间求和部分输出 + * 3. 最终结果表示完整的 FFN 输出 + * + * 参数： + * - meta：维度的模型元数据 + * - w：全局权重存储 + * - layer：Transformer 层索引 + * - idev：当前设备索引 + * - ndev：设备总数 + * + * 返回：设备特定的 FFN 下降权重 [di/ndev, d] 或 [d, di/ndev] + */ + + + +/* + * 为分布式推理提取 FFN Down 投影权重 + * + * Down 投影：[2048, 6144] -> 每设备 [2048, 6144/ndev] + * 注意：Down 的分布策略与 Gate/Up 不同，它是按输入维度切片 + */ +inline std::shared_ptr getQwen3MLPDown( + Qwen3Meta const *meta, + Qwen3Weights const *w, + size_t layer, size_t idev, size_t ndev) { + + auto di = meta->di; // 6144 中间维度 + auto d = meta->d; // 2048 模型维度 + + /* + * Down 投影分布式切片计算： + * - 总输入维度：di = 6144（从 Gate/Up 的输出接收） + * - 每设备输入维度：di / ndev = 6144/ndev + * - 输出维度：d = 2048（完整模型维度） + * 注意：Down 投影的权重矩阵是 [d, di]，但我们需要按输入维度切片 + * 所以实际的权重切片是 [d, di/ndev] + */ + size_t intermediate_per_device = di / ndev; + size_t offset = idev * d * intermediate_per_device * dsize(w->dt_mat); + + if (w->transpose_linear_weights != 0) { + // 存储格式：[d, intermediate_per_device] -> 转置为 [intermediate_per_device, d] + auto shape = std::vector({d, intermediate_per_device}); + return Tensor::weight((char *)(w->mlp_down_proj[layer]) + offset, w->dt_mat, shape) + ->permute({1, 0}); + } else { + // 标准格式：[intermediate_per_device, d] + auto shape = std::vector({intermediate_per_device, d}); + return Tensor::weight((char *)(w->mlp_down_proj[layer]) + offset, w->dt_mat, shape); + } +} + + +/* + * 生成 RoPE 正弦查找表 + * + * 为旋转位置嵌入（RoPE）创建预计算的正弦表。 + * RoPE 通过将旋转矩阵应用于查询和键向量来编码位置信息， + * 基于它们在序列中的位置。 + * + * RoPE 数学基础： + * 对于位置 pos 和维度对 (i, i+d/2)： + * - 旋转角度：θ_i = pos / (base^(2i/d)) 其中 base = meta->theta（通常为 10000） + * - 正弦分量：sin(θ_i) = sin(pos / (base^(2i/d))) + * + * 表结构： + * - 形状：[dctx, dh/2] 其中 dctx = 最大上下文长度，dh = 头维度 + * - 条目 [pos][i]：sin(pos / (theta^(2i/dh))) 对于位置 pos 和维度 i + * - 覆盖所有可能的位置（0 到 dctx-1）和头维度的一半 + * + * 数据类型处理： + * - 支持 FP16、BF16 和 FP32 基于 meta->dt_logits + * - 从 float32 计算转换为目标数据类型 + * - 使用适当的转换函数（f32_to_f16、f32_to_bf16） + * + * 内存管理： + * - 为计算分配临时存储 + * - 创建带有计算值副本的张量包装器 + * - 在张量创建后释放临时存储 + * + * 推理期间的用法： + * - 由位置 ID 查找以获取旋转的正弦值 + * - 与余弦表结合形成完整的旋转矩阵 + * - 应用于查询和键向量 + * + * 参数： + * - meta：包含上下文长度、头维度、theta 和数据类型的模型元数据 + * + * 返回：包含预计算正弦值的共享张量 [dctx, dh/2] + */ +inline std::shared_ptr getQwen3SinTable(Qwen3Meta const *meta) { + auto half_dh = meta->dh / 2; // 头维度的一半 + auto unit = dsize(meta->dt_logits); // 数据类型的大小（字节） + void *table = std::malloc(meta->dctx * half_dh * unit); // 分配临时存储 + + /* + * 为所有位置和维度计算正弦值 + * + * 嵌套循环结构： + * - 外循环：遍历所有可能的位置（0 到 dctx-1） + * - 内循环：遍历头维度的一半（0 到 dh/2-1） + */ + for (size_t i = 0; i < meta->dctx; i++) { // 位置循环 + for (size_t j = 0; j < half_dh; j++) { // 维度循环 + /* + * RoPE 正弦计算： + * + * 公式：sin(position / (theta^(2*dim_index / head_dimension))) + * - position：i（当前序列位置） + * - dim_index：j（当前维度索引） + * - theta：meta->theta（旋转基数，通常为 10000） + * - head_dimension：meta->dh（完整头维度） + */ + float _sin = std::sin( + static_cast(i) / std::pow(meta->theta, static_cast(j) / half_dh)); + + /* + * 根据目标数据类型转换和存储 + * + * 以 FP32 计算以保证精度，然后转换为目标格式。 + */ + if (meta->dt_logits == INFINI_DTYPE_F16) { + ((uint16_t *)table)[i * half_dh + j] = f32_to_f16(_sin); + } else if (meta->dt_logits == INFINI_DTYPE_BF16) { + ((uint16_t *)table)[i * half_dh + j] = f32_to_bf16(_sin); + } else if (meta->dt_logits == INFINI_DTYPE_F32) { + ((float *)table)[i * half_dh + j] = _sin; + } else { + std::cout << "不支持的数据类型" << std::endl; + exit(1); + } + } + } + + // 创建张量包装器并清理临时存储 + auto shape = std::vector({meta->dctx, half_dh}); + auto tensor = Tensor::weight(table, meta->dt_logits, shape); + std::free(table); // 释放临时存储（张量有自己的副本） + return tensor; +} + +/* + * 生成 RoPE 余弦查找表 + * + * 为旋转位置嵌入（RoPE）创建预计算的余弦表。 + * 这补充了正弦表，形成完整的旋转矩阵以编码 + * transformer 注意力机制中的位置信息。 + * + * RoPE 数学基础： + * 对于位置 pos 和维度对 (i, i+d/2)： + * - 旋转角度：θ_i = pos / (base^(2i/d)) 其中 base = meta->theta + * - 余弦分量：cos(θ_i) = cos(pos / (base^(2i/d))) + * + * 表结构： + * - 形状：[dctx, dh/2] 其中 dctx = 最大上下文长度，dh = 头维度 + * - 条目 [pos][i]：cos(pos / (theta^(2i/dh))) 对于位置 pos 和维度 i + * - 与正弦表相同的结构以实现高效配对查找 + * + * RoPE 旋转矩阵应用： + * 对于每个位置和维度对 (i, i+dh/2)： + * - q'[i] = q[i] * cos(θ) - q[i+dh/2] * sin(θ) + * - q'[i+dh/2] = q[i] * sin(θ) + q[i+dh/2] * cos(θ) + * + * 这在 (i, i+dh/2) 平面中创建一个 2D 旋转，直接编码 + * 注意力计算中的相对位置信息。 + * + * 数据类型支持： + * - FP16：半精度以提高内存效率 + * - BF16：脑浮点以获得更好的数值范围 + * - FP32：全精度以获得最大准确性 + * - 以 FP32 计算然后转换为目标格式 + * + * 内存管理： + * - 计算期间的临时分配 + * - 持久存储的张量副本 + * - 临时内存的自动清理 + * + * 性能考虑： + * - 预计算表避免推理期间昂贵的三角运算 + * - 注意力计算期间缓存友好的访问模式 + * - 在所有层和注意力头间共享 + * + * 参数： + * - meta：带有上下文长度、维度、旋转基数和数据类型的模型元数据 + * + * 返回：包含预计算余弦值的共享张量 [dctx, dh/2] + */ +inline std::shared_ptr getQwen3CosTable(Qwen3Meta const *meta) { + auto half_dh = meta->dh / 2; // 头维度的一半 + auto unit = dsize(meta->dt_logits); // 数据类型的大小（字节） + void *table = std::malloc(meta->dctx * half_dh * unit); // 分配临时存储 + + /* + * 为所有位置和维度计算余弦值 + * + * 与正弦表相同的循环结构以保持一致性和配对访问。 + */ + for (size_t i = 0; i < meta->dctx; i++) { // 位置循环 + for (size_t j = 0; j < half_dh; j++) { // 维度循环 + /* + * RoPE 余弦计算： + * + * 公式：cos(position / (theta^(2*dim_index / head_dimension))) + * - 与正弦计算相同但使用余弦函数 + * - 确保 sin² + cos² = 1 以获得正确的旋转矩阵 + */ + float _cos = std::cos( + static_cast(i) / std::pow(meta->theta, static_cast(j) / half_dh)); + + /* + * 根据目标数据类型转换和存储 + * + * 与正弦表相同的数据类型处理以保持一致。 + */ + if (meta->dt_logits == INFINI_DTYPE_F16) { + ((uint16_t *)table)[i * half_dh + j] = f32_to_f16(_cos); + } else if (meta->dt_logits == INFINI_DTYPE_BF16) { + ((uint16_t *)table)[i * half_dh + j] = f32_to_bf16(_cos); + } else if (meta->dt_logits == INFINI_DTYPE_F32) { + ((float *)table)[i * half_dh + j] = _cos; + } else { + std::cout << "不支持的数据类型" << std::endl; + exit(1); + } + } + } + + // 创建张量包装器并清理临时存储 + auto shape = std::vector({meta->dctx, half_dh}); + auto tensor = Tensor::weight(table, meta->dt_logits, shape); + std::free(table); // 释放临时存储（张量有自己的副本） + return tensor; +} + +#endif \ No newline at end of file diff --git a/infini-qwen3-moe/src/tensor.hpp b/infini-qwen3-moe/src/tensor.hpp new file mode 100755 index 00000000..2b8ef6b5 --- /dev/null +++ b/infini-qwen3-moe/src/tensor.hpp @@ -0,0 +1,181 @@ +#ifndef INFER_TENSOR_H +#define INFER_TENSOR_H + +#include "allocator.hpp" +#include "infinicore_infer.h" +#include "utils.hpp" +#include +#include +#include + +class Storage { +private: + Storage() = default; + void *_memory; + size_t _size; + infiniDevice_t _device_type; + int _device_id; + std::shared_ptr _memory_pool; + +public: + static std::shared_ptr create(size_t size); + static std::shared_ptr createAsync(size_t size, infinirtStream_t stream = nullptr); + static std::shared_ptr createFromPool(size_t size, std::shared_ptr pool = nullptr); + static std::shared_ptr createHost(size_t size); + ~Storage(); + + void *memory() const { return _memory; } + size_t size() const { return _size; } + infiniDevice_t deviceType() const { return _device_type; } + int deviceId() const { return _device_id; } +}; + +struct SliceParams { + size_t dim; + size_t start; + size_t len; +}; + +template +std::vector __shape(Args... args) { + return std::vector{static_cast(args)...}; +} + +template +std::vector __strides(Args... args) { + return std::vector{static_cast(args)...}; +} +class TensorDesc { +private: + infiniDtype_t _dtype; + std::vector _shape; + std::vector _strides; + infiniopTensorDescriptor_t _desc; + + TensorDesc(infiniDtype_t dtype, const std::vector &shape, + const std::vector &strides) : _dtype(dtype), _shape(shape), _strides(strides), _desc(nullptr) {} + void resetDesc(); + +public: + ~TensorDesc(); + static std::shared_ptr + create(infiniDtype_t dtype, const std::vector &shape, + const std::vector &strides); + static std::shared_ptr + create(infiniDtype_t dtype, const std::vector &shape); + static std::shared_ptr + createWithOrder(infiniDtype_t dtype, const std::vector &shape, + const std::vector &order); + + infiniDtype_t dtype() const { return _dtype; } + const std::vector &shape() const { return _shape; } + const std::vector &strides() const { return _strides; } + size_t ndim() const { return _shape.size(); } + infiniopTensorDescriptor_t desc() const; + bool isContigous() const; + std::string info() const; + + void dimMerge(size_t dim_start, size_t dim_end); + void dimSplit(size_t dim, const std::vector &dims); + void permute(const std::vector &order); +}; + +class Tensor : public std::enable_shared_from_this { +private: + std::shared_ptr _storage; + std::shared_ptr _desc; + + ptrdiff_t _offset; + + void *dataImpl(ptrdiff_t offset) const; + std::shared_ptr + sliceImpl(const std::vector &slices) const; + +public: + static std::shared_ptr buffer(infiniDtype_t dtype, + const std::vector &shape, + std::shared_ptr pool = nullptr); + static std::shared_ptr weight(void *host_data, + infiniDtype_t dtype, + const std::vector &shape); + std::shared_ptr memShare(const std::vector &shape, + infiniDtype_t dtype = INFINI_DTYPE_INVALID) const; + std::shared_ptr slice(size_t dim, size_t start, size_t len); + std::shared_ptr slice(size_t dim, size_t start, + size_t len) const; + std::shared_ptr slice(const std::vector &slices); + std::shared_ptr + slice(const std::vector &slices) const; + std::shared_ptr dimMerge(size_t dim_start, size_t dim_end); + std::shared_ptr dimSplit(size_t dim, + const std::vector &dims); + std::shared_ptr permute(const std::vector &order); + void *data(ptrdiff_t offset = 0); + void const *data(ptrdiff_t offset = 0) const; + void copyFrom(std::shared_ptr src, infiniopHandle_t handle, + infinirtStream_t stream = nullptr); + const std::vector &shape() const; + const std::vector &strides() const; + size_t ndim() const; + infiniDtype_t dtype() const; + bool isContigous() const; + infiniopTensorDescriptor_t desc() const; + ptrdiff_t dataOffset() const; + infiniDevice_t deviceType() const; + int deviceId() const; + + void debug(const std::string &filename) const; + void debug() const; + std::string info() const; + + ~Tensor(); +}; + +inline size_t dsize(infiniDtype_t dtype) { + switch (dtype) { + case INFINI_DTYPE_INVALID: + return 0; + case INFINI_DTYPE_BYTE: + return 1; + case INFINI_DTYPE_BOOL: + return 1; + case INFINI_DTYPE_I8: + return 1; + case INFINI_DTYPE_I16: + return 2; + case INFINI_DTYPE_I32: + return 4; + case INFINI_DTYPE_I64: + return 8; + case INFINI_DTYPE_U8: + return 1; + case INFINI_DTYPE_U16: + return 2; + case INFINI_DTYPE_U32: + return 4; + case INFINI_DTYPE_U64: + return 8; + case INFINI_DTYPE_F8: + return 1; + case INFINI_DTYPE_F16: + return 2; + case INFINI_DTYPE_F32: + return 4; + case INFINI_DTYPE_F64: + return 8; + case INFINI_DTYPE_C16: + return 2; + case INFINI_DTYPE_C32: + return 4; + case INFINI_DTYPE_C64: + return 8; + case INFINI_DTYPE_C128: + return 16; + case INFINI_DTYPE_BF16: + return 2; + default: + return 0; + } +} + +#endif diff --git a/infini-qwen3-moe/src/tensor/strorage.cpp b/infini-qwen3-moe/src/tensor/strorage.cpp new file mode 100755 index 00000000..abf72df3 --- /dev/null +++ b/infini-qwen3-moe/src/tensor/strorage.cpp @@ -0,0 +1,53 @@ +#include "../allocator.hpp" +#include "../tensor.hpp" + +std::shared_ptr Storage::create(size_t size) { + auto storage = std::shared_ptr(new Storage()); + RUN_INFINI(infinirtMalloc(&storage->_memory, size)); + storage->_size = size; + RUN_INFINI(infinirtGetDevice(&storage->_device_type, &storage->_device_id)); + return storage; +} + +std::shared_ptr Storage::createAsync(size_t size, infinirtStream_t stream) { + auto storage = std::shared_ptr(new Storage()); + RUN_INFINI(infinirtMallocAsync(&storage->_memory, size, stream)); + storage->_size = size; + RUN_INFINI(infinirtGetDevice(&storage->_device_type, &storage->_device_id)); + return storage; +} + +std::shared_ptr Storage::createFromPool(size_t size, std::shared_ptr pool) { + auto storage = std::shared_ptr(new Storage()); + storage->_memory_pool = pool; + if (pool) { + storage->_memory = pool->alloc(size); + } else { + RUN_INFINI(infinirtMalloc(&storage->_memory, size)); + } + storage->_size = size; + RUN_INFINI(infinirtGetDevice(&storage->_device_type, &storage->_device_id)); + return storage; +} + +std::shared_ptr Storage::createHost(size_t size) { + auto storage = std::shared_ptr(new Storage()); + RUN_INFINI(infinirtMallocHost(&storage->_memory, size)); + storage->_size = size; + storage->_device_type = INFINI_DEVICE_CPU; + storage->_device_id = 0; + storage->_memory_pool = nullptr; // No pool for host memory + return storage; +} + +Storage::~Storage() { + if (_memory_pool) { + _memory_pool->release(_memory); + } else { + if (_device_type == INFINI_DEVICE_CPU) { + RUN_INFINI(infinirtFreeHost(_memory)); + } else { + RUN_INFINI(infinirtFree(_memory)); + } + } +} diff --git a/infini-qwen3-moe/src/tensor/tensor.cpp b/infini-qwen3-moe/src/tensor/tensor.cpp new file mode 100755 index 00000000..73bb5bcb --- /dev/null +++ b/infini-qwen3-moe/src/tensor/tensor.cpp @@ -0,0 +1,322 @@ +#include "../tensor.hpp" +#include "../utils.hpp" +#include +#include +#include +#include +#include +#include + +std::shared_ptr +TensorDesc::create(infiniDtype_t dtype, const std::vector &shape, + const std::vector &strides) { + return std::shared_ptr(new TensorDesc(dtype, shape, strides)); +} + +std::shared_ptr +TensorDesc::create(infiniDtype_t dtype, const std::vector &shape) { + auto ndim = shape.size(); + auto strides = std::vector(ndim); + if (ndim > 0) { + strides[ndim - 1] = 1; + for (int i = ndim - 2; i >= 0; i--) { + strides[i] = strides[i + 1] * shape[i + 1]; + } + } + return create(dtype, shape, strides); +} + +std::shared_ptr +TensorDesc::createWithOrder(infiniDtype_t dtype, const std::vector &shape, + const std::vector &order) { + ASSERT_EQ(shape.size(), order.size()); + auto ndim = shape.size(); + if (ndim == 0) { + return create(dtype, shape); + } + auto strides = std::vector(order.size()); + auto idx = std::find(order.begin(), order.end(), size_t(ndim - 1)); + strides[std::distance(order.begin(), idx)] = 1; + for (int i = ndim - 2; i >= 0; i--) { + auto prev_dim = shape[std::distance(order.begin(), idx)]; + auto prev_stride = strides[std::distance(order.begin(), idx)]; + idx = std::find(order.begin(), order.end(), size_t(i)); + strides[std::distance(order.begin(), idx)] = prev_stride * prev_dim; + } + return create(dtype, shape, strides); +} + +infiniopTensorDescriptor_t TensorDesc::desc() const { + if (_desc == nullptr) { + RUN_INFINI(infiniopCreateTensorDescriptor( + (infiniopTensorDescriptor_t *)(&_desc), _shape.size(), _shape.data(), + _strides.data(), _dtype)); + } + return _desc; +}; + +void TensorDesc::resetDesc() { + if (this->_desc != nullptr) { + infiniopDestroyTensorDescriptor(this->_desc); + this->_desc = nullptr; + } +} + +bool TensorDesc::isContigous() const { + auto ndim = this->ndim(); + auto shape = this->shape(); + auto strides = std::vector(ndim); + strides[ndim - 1] = 1; + for (int i = ndim - 2; i >= 0; i--) { + strides[i] = strides[i + 1] * shape[i + 1]; + } + ASSERT_EQ(strides.size(), this->_strides.size()); + return std::equal(strides.begin(), strides.end(), this->_strides.begin()); +} + +std::string TensorDesc::info() const { + std::stringstream ss; + + ss << "Tensor: " + << "shape[ "; + for (auto s : this->shape()) { + ss << s << " "; + } + ss << "] strides[ "; + for (auto s : this->strides()) { + ss << s << " "; + } + ss << "] dtype=" << this->dtype(); + + return ss.str(); +} + +TensorDesc::~TensorDesc() { + this->resetDesc(); +} + +const std::vector &Tensor::shape() const { return this->_desc->shape(); } +const std::vector &Tensor::strides() const { return this->_desc->strides(); } +size_t Tensor::ndim() const { return this->_desc->ndim(); } +infiniDtype_t Tensor::dtype() const { return this->_desc->dtype(); } +infiniDevice_t Tensor::deviceType() const { return this->_storage->deviceType(); } +int Tensor::deviceId() const { return this->_storage->deviceId(); } +Tensor::~Tensor() {} + +ptrdiff_t Tensor::dataOffset() const { + return _offset; +} + +infiniopTensorDescriptor_t Tensor::desc() const { return _desc->desc(); } + +std::shared_ptr Tensor::buffer(infiniDtype_t dtype, + const std::vector &shape, + std::shared_ptr pool) { + std::shared_ptr tensor = std::make_shared(); + auto ndim = shape.size(); + + size_t size = std::accumulate(shape.begin(), shape.end(), dsize(dtype), std::multiplies()); + auto strides = std::vector(ndim); + if (ndim > 0) { + strides[ndim - 1] = 1; + for (int i = ndim - 2; i >= 0; i--) { + strides[i] = strides[i + 1] * shape[i + 1]; + } + } + tensor->_storage = Storage::createFromPool(size, pool); + tensor->_desc = TensorDesc::create(dtype, shape, strides); + tensor->_offset = 0; + return tensor; +} + +std::shared_ptr Tensor::weight(void *data, infiniDtype_t dtype, + const std::vector &shape) { + std::shared_ptr tensor = std::make_shared(); + auto ndim = shape.size(); + size_t size = std::accumulate(shape.begin(), shape.end(), dsize(dtype), std::multiplies()); + auto strides = std::vector(ndim); + if (ndim > 0) { + strides[ndim - 1] = 1; + for (int i = ndim - 2; i >= 0; i--) { + strides[i] = strides[i + 1] * shape[i + 1]; + } + } + + tensor->_storage = Storage::create(size); + tensor->_desc = TensorDesc::create(dtype, shape, strides); + // NOTE: 为兼容部分平台（沐曦）多线程并发对同一host数据执行memcpy卡死问题 + static std::mutex mutex; + { + std::lock_guard lock(mutex); + RUN_INFINI(infinirtMemcpy(tensor->_storage->memory(), + data, size, INFINIRT_MEMCPY_H2D)); + } + + tensor->_offset = 0; + return tensor; +} + +std::shared_ptr Tensor::memShare(const std::vector &shape, infiniDtype_t dtype_) const { + auto dtype = dtype_ == INFINI_DTYPE_INVALID ? this->dtype() : dtype_; + size_t size = std::accumulate(shape.begin(), shape.end(), dsize(dtype), std::multiplies()); + ASSERT(size <= this->_storage->size()); + + std::shared_ptr tensor = std::make_shared(); + auto ndim = shape.size(); + auto strides = std::vector(ndim); + if (ndim > 0) { + strides[ndim - 1] = 1; + for (int i = ndim - 2; i >= 0; i--) { + strides[i] = strides[i + 1] * shape[i + 1]; + } + } + tensor->_storage = this->_storage; + tensor->_offset = 0; + tensor->_desc = TensorDesc::create(dtype, shape, strides); + return tensor; +} + +void *Tensor::dataImpl(ptrdiff_t offset) const { + return (char *)(this->_storage->memory()) + this->_offset + offset * dsize(this->dtype()); +} + +void *Tensor::data(ptrdiff_t offset) { + return this->dataImpl(offset); +} + +const void *Tensor::data(ptrdiff_t offset) const { + return this->dataImpl(offset); +} + +void Tensor::copyFrom(std::shared_ptr src, + infiniopHandle_t handle, infinirtStream_t stream) { + ASSERT_EQ(this->shape(), src->shape()); + ASSERT_EQ(this->dtype(), src->dtype()); + infiniopRearrangeDescriptor_t desc; + RUN_INFINI(infiniopCreateRearrangeDescriptor( + handle, &desc, this->desc(), src->desc())); + RUN_INFINI(infiniopRearrange(desc, this->data(), src->data(), + stream)); + RUN_INFINI(infiniopDestroyRearrangeDescriptor(desc)); +} + +bool Tensor::isContigous() const { + return this->_desc->isContigous(); +} + +template +void print_data(T *data, const std::vector &shape, + const std::vector &strides, size_t dim) { + if (dim == shape.size() - 1) { + for (size_t i = 0; i < shape[dim]; i++) { + std::cout << data[i] << " "; + } + std::cout << std::endl; + } else if (dim < shape.size() - 1) { + for (size_t i = 0; i < shape[dim]; i++) { + print_data(data + i * strides[dim], shape, strides, dim + 1); + } + } +} + +template <> +void print_data(uint16_t const *data, const std::vector &shape, + const std::vector &strides, size_t dim) { + if (dim == shape.size() - 1) { + for (size_t i = 0; i < shape[dim]; i++) { + std::cout << f16_to_f32(data[i * strides[dim]]) << " "; + } + std::cout << std::endl; + } else if (dim < shape.size() - 1) { + for (size_t i = 0; i < shape[dim]; i++) { + print_data(data + i * strides[dim], shape, strides, dim + 1); + } + } +} + +void print_data_bf16(uint16_t const *data, const std::vector &shape, + const std::vector &strides, size_t dim) { + if (dim == shape.size() - 1) { + for (size_t i = 0; i < shape[dim]; i++) { + std::cout << bf16_to_f32(data[i * strides[dim]]) << " "; + } + std::cout << std::endl; + } else if (dim < shape.size() - 1) { + for (size_t i = 0; i < shape[dim]; i++) { + print_data(data + i * strides[dim], shape, strides, dim + 1); + } + } +} + +std::string Tensor::info() const { + std::stringstream ss; + + ss << "Tensor: " + << this->_desc->info() + << " device=" << this->deviceType() + << " device_id=" << this->deviceId(); + return this->_desc->info(); +} + +void Tensor::debug(const std::string &filename) const { + RUN_INFINI(infinirtDeviceSynchronize()); + + std::cout << info() << std::endl; + + void const *cpu_data; + if (this->deviceType() != INFINI_DEVICE_CPU) { + void *cpu_memory = std::malloc(this->_storage->size()); + RUN_INFINI(infinirtMemcpy(cpu_memory, this->_storage->memory(), + this->_storage->size(), INFINIRT_MEMCPY_D2H)); + cpu_data = cpu_memory; + } else { + cpu_data = this->_storage->memory(); + } + + if (!filename.empty()) { + std::ofstream outFile(filename, std::ios::binary); + if (!outFile) { + std::cerr << "Error opening file for writing: " << filename << "\n"; + return; + } + outFile.write(reinterpret_cast(cpu_data), this->_storage->size()); + outFile.close(); + std::cout << "Data written to file: " << filename << "\n"; + return; + } + + switch (this->dtype()) { + case INFINI_DTYPE_F16: + print_data((uint16_t const *)((char const *)cpu_data + dataOffset()), + this->shape(), this->strides(), 0); + break; + case INFINI_DTYPE_F32: + print_data((float const *)((char const *)cpu_data + dataOffset()), + this->shape(), this->strides(), 0); + break; + case INFINI_DTYPE_U64: + print_data((uint64_t const *)((char const *)cpu_data + dataOffset()), + this->shape(), this->strides(), 0); + break; + case INFINI_DTYPE_I64: + print_data((int64_t const *)((char const *)cpu_data + dataOffset()), + this->shape(), this->strides(), 0); + break; + case INFINI_DTYPE_U32: + print_data((uint32_t const *)((char const *)cpu_data + dataOffset()), + this->shape(), this->strides(), 0); + break; + case INFINI_DTYPE_I32: + print_data((int32_t const *)((char const *)cpu_data + dataOffset()), + this->shape(), this->strides(), 0); + break; + case INFINI_DTYPE_BF16: + print_data_bf16((uint16_t const *)((char const *)cpu_data + dataOffset()), + this->shape(), this->strides(), 0); + break; + default: + PANIC("Unsupported data type"); + } +} + +void Tensor::debug() const { this->debug(""); } diff --git a/infini-qwen3-moe/src/tensor/transform.cpp b/infini-qwen3-moe/src/tensor/transform.cpp new file mode 100755 index 00000000..579c9710 --- /dev/null +++ b/infini-qwen3-moe/src/tensor/transform.cpp @@ -0,0 +1,158 @@ +#include "../tensor.hpp" +#include "../utils.hpp" +#include +#include +#include + +std::shared_ptr Tensor::sliceImpl(const std::vector &slices) const { + std::shared_ptr tensor = std::make_shared(); + + auto new_shape = std::vector(this->shape()); + ptrdiff_t offset = 0; + + for (const auto &slice : slices) { + ASSERT(slice.len > 0); + // Print error information for dimension and slice + if (slice.start >= this->shape()[slice.dim]) { + std::cerr << "Error: slice start (" << slice.start << ") >= dimension size (" + << this->shape()[slice.dim] << ") for dimension " << slice.dim << std::endl; + } + if (this->shape()[slice.dim] < slice.start + slice.len) { + std::cerr << "Error: slice end (" << slice.start + slice.len << ") > dimension size (" + << this->shape()[slice.dim] << ") for dimension " << slice.dim << std::endl; + } + + ASSERT(this->shape()[slice.dim] >= slice.start + slice.len); + new_shape[slice.dim] = slice.len; + offset += slice.start * this->strides()[slice.dim]; + } + + tensor->_desc = TensorDesc::create(this->dtype(), new_shape, this->strides()); + tensor->_offset = offset * dsize(this->dtype()) + this->_offset; + tensor->_storage = this->_storage; + return tensor; +} + +std::shared_ptr Tensor::slice(size_t dim, size_t start, size_t len) { + return this->sliceImpl({{dim, start, len}}); +} + +std::shared_ptr Tensor::slice(size_t dim, size_t start, size_t len) const { + return this->sliceImpl({{dim, start, len}}); +} + +std::shared_ptr Tensor::slice(const std::vector &slices) { + return this->sliceImpl(slices); +} + +std::shared_ptr Tensor::slice(const std::vector &slices) const { + return this->sliceImpl(slices); +} + +void TensorDesc::dimMerge(size_t dim_start, size_t dim_end) { + ASSERT(dim_start <= dim_end && dim_end < this->_shape.size()); + if (dim_start == dim_end) { + return; + } + + auto new_shape = std::vector(); + auto new_strides = std::vector(); + for (size_t i = 0; i < dim_start; i++) { + new_shape.push_back(this->_shape[i]); + new_strides.push_back(this->_strides[i]); + } + for (size_t i = dim_start + 1; i <= dim_end; i++) { + ASSERT_EQ(this->_strides[i - 1], ptrdiff_t(this->_shape[i]) * this->_strides[i]); + } + new_shape.push_back(std::accumulate(this->_shape.begin() + dim_start, this->_shape.begin() + dim_end + 1, 1, std::multiplies())); + new_strides.push_back(this->_strides[dim_end]); + for (size_t i = dim_end + 1; i < this->_shape.size(); i++) { + new_shape.push_back(this->_shape[i]); + new_strides.push_back(this->_strides[i]); + } + this->_shape = new_shape; + this->_strides = new_strides; + this->resetDesc(); +} + +std::shared_ptr Tensor::dimMerge(size_t dim_start, size_t dim_end) { + this->_desc->dimMerge(dim_start, dim_end); + return shared_from_this(); +} + +// 在 dimSplit 函数中添加详细的错误调试信息 + +void TensorDesc::dimSplit(size_t dim, const std::vector &dims) { + // 计算 dims 向量中所有元素的乘积 + size_t dims_product = std::accumulate(dims.begin(), dims.end(), size_t(1), std::multiplies()); + + // 如果断言即将失败，输出详细的调试信息 + if (this->_shape[dim] != dims_product) { + std::cerr << "=== dimSplit Error Debug Information ===" << std::endl; + std::cerr << "Trying to split dimension " << dim << std::endl; + std::cerr << "Original tensor shape: ["; + for (size_t i = 0; i < this->_shape.size(); i++) { + std::cerr << this->_shape[i]; + if (i < this->_shape.size() - 1) std::cerr << ", "; + } + std::cerr << "]" << std::endl; + + std::cerr << "Dimension " << dim << " current size: " << this->_shape[dim] << std::endl; + + std::cerr << "Requested split dimensions: ["; + for (size_t i = 0; i < dims.size(); i++) { + std::cerr << dims[i]; + if (i < dims.size() - 1) std::cerr << ", "; + } + std::cerr << "]" << std::endl; + + std::cerr << "Product of requested dimensions: " << dims_product << std::endl; + std::cerr << "Expected: " << this->_shape[dim] << " == " << dims_product << std::endl; + std::cerr << "Difference: " << (int64_t(this->_shape[dim]) - int64_t(dims_product)) << std::endl; + std::cerr << "=========================================" << std::endl; + } + + ASSERT_EQ(this->_shape[dim], dims_product); + + auto new_shape = std::vector(); + auto new_strides = std::vector(); + for (size_t i = 0; i < dim; i++) { + new_shape.push_back(this->_shape[i]); + new_strides.push_back(this->_strides[i]); + } + for (size_t i = 0; i < dims.size(); i++) { + new_shape.push_back(dims[i]); + new_strides.push_back(this->_strides[dim] * this->_shape[dim] / std::accumulate(dims.begin(), dims.begin() + i + 1, 1, std::multiplies())); + } + for (size_t i = dim + 1; i < this->_shape.size(); i++) { + new_shape.push_back(this->_shape[i]); + new_strides.push_back(this->_strides[i]); + } + this->_shape = new_shape; + this->_strides = new_strides; + this->resetDesc(); +} + +std::shared_ptr Tensor::dimSplit(size_t dim, const std::vector &dims) { + this->_desc->dimSplit(dim, dims); + return shared_from_this(); +} + +void TensorDesc::permute(const std::vector &order) { + ASSERT_EQ(this->_shape.size(), order.size()); + auto new_shape = std::vector(order.size()); + auto new_strides = std::vector(order.size()); + for (size_t i = 0; i < order.size(); i++) { + ASSERT(std::find(order.begin(), order.end(), i) != order.end()); + new_shape[i] = this->_shape[order[i]]; + new_strides[i] = this->_strides[order[i]]; + } + this->_shape = new_shape; + this->_strides = new_strides; + this->resetDesc(); +} + +std::shared_ptr Tensor::permute(const std::vector &order) { + this->_desc->permute(order); + return shared_from_this(); +} diff --git a/infini-qwen3-moe/src/utils.hpp b/infini-qwen3-moe/src/utils.hpp new file mode 100755 index 00000000..26481b20 --- /dev/null +++ b/infini-qwen3-moe/src/utils.hpp @@ -0,0 +1,122 @@ +#ifndef INFINICORE_INFER_UTILS_H +#define INFINICORE_INFER_UTILS_H +#include + +#include +#include +#include +#include + +inline void assertTrue(int expr, const char *msg, const char *file, int line) { + if (!expr) { + fprintf(stderr, "\033[31mAssertion failed:\033[0m %s at file %s, line %d\n", msg, file, line); + exit(EXIT_FAILURE); + } +} + +#define ASSERT(expr) assertTrue((expr), #expr " is false", __FILE__, __LINE__) +#define ASSERT_EQ(a, b) assertTrue((a) == (b), #a " != " #b, __FILE__, __LINE__) +#define ASSERT_VALID_PTR(a) assertTrue((a) != nullptr, #a " is nullptr", __FILE__, __LINE__) + +#define PANIC(EXPR) \ + printf("Error at %s:%d - %s\n", __FILE__, __LINE__, #EXPR); \ + exit(EXIT_FAILURE) + +#define RUN_INFINI(API) \ + do { \ + auto api_result_ = (API); \ + if (api_result_ != INFINI_STATUS_SUCCESS) { \ + std::cerr << "Error Code " << api_result_ << " in `" << #API << "`" \ + << " from " << __func__ \ + << " at " << __FILE__ << ":" << __LINE__ << std::endl; \ + exit(EXIT_FAILURE); \ + } \ + } while (0) + +inline float f16_to_f32(uint16_t h) { + uint32_t sign = (h & 0x8000) << 16; // Extract the sign bit + int32_t exponent = (h >> 10) & 0x1F; // Extract the exponent + uint32_t mantissa = h & 0x3FF; // Extract the mantissa (fraction part) + + if (exponent == 31) { // Special case for Inf and NaN + if (mantissa != 0) { + // NaN: Set float32 NaN + uint32_t f32 = sign | 0x7F800000 | (mantissa << 13); + return *(float *)&f32; + } else { + // Infinity + uint32_t f32 = sign | 0x7F800000; + return *(float *)&f32; + } + } else if (exponent == 0) { // Subnormal float16 or zero + if (mantissa == 0) { + // Zero (positive or negative) + uint32_t f32 = sign; // Just return signed zero + return *(float *)&f32; + } else { + // Subnormal: Convert to normalized float32 + exponent = -14; // Set exponent for subnormal numbers + while ((mantissa & 0x400) == 0) { // Normalize mantissa + mantissa <<= 1; + exponent--; + } + mantissa &= 0x3FF; // Clear the leading 1 bit + uint32_t f32 = sign | ((exponent + 127) << 23) | (mantissa << 13); + return *(float *)&f32; + } + } else { + // Normalized float16 + uint32_t f32 = sign | ((exponent + 127 - 15) << 23) | (mantissa << 13); + return *(float *)&f32; + } +} + +inline uint16_t f32_to_f16(float val) { + uint32_t f32; + memcpy(&f32, &val, sizeof(f32)); // Read the bits of the float32 + uint16_t sign = (f32 >> 16) & 0x8000; // Extract the sign bit + int32_t exponent = ((f32 >> 23) & 0xFF) - 127; // Extract and de-bias the exponent + uint32_t mantissa = f32 & 0x7FFFFF; // Extract the mantissa (fraction part) + + if (exponent >= 31) { // Special cases for Inf and NaN + // NaN + if (exponent == 128 && mantissa != 0) { + return static_cast(sign | 0x7E00); + } + // Infinity + return static_cast(sign | 0x7C00); + } else if (exponent >= -14) { // Normalized case + return (uint16_t)(sign | ((exponent + 15) << 10) | (mantissa >> 13)); + } else if (exponent >= -24) { + mantissa |= 0x800000; // Add implicit leading 1 + mantissa >>= (-14 - exponent); + return (uint16_t)(sign | (mantissa >> 13)); + } else { + // Too small for subnormal: return signed zero + return (uint16_t)sign; + } +} + +inline float bf16_to_f32(uint16_t val) { + // 只需把 bf16 放到 float32 高 16 bit，其余 16 位置 0。 + uint32_t bits32 = static_cast(val) << 16; + + float out; + std::memcpy(&out, &bits32, sizeof(out)); + return out; +} + +inline uint16_t f32_to_bf16(float val) { + uint32_t bits32; + std::memcpy(&bits32, &val, sizeof(bits32)); + + // 截断前先加 0x7FFF，再根据第 16 位（有效位的最低位）的奇偶做 round-to-nearest-even + const uint32_t rounding_bias = 0x00007FFF + // 0111 1111 1111 1111 + ((bits32 >> 16) & 1); // 尾数的有效位的最低位奇数时 +1，即实现舍入偶数 + + uint16_t bf16_bits = static_cast((bits32 + rounding_bias) >> 16); + + return bf16_bits; +} + +#endif diff --git a/infini-qwen3-moe/third_party/nlohmann/json.hpp b/infini-qwen3-moe/third_party/nlohmann/json.hpp new file mode 100755 index 00000000..24fd99df --- /dev/null +++ b/infini-qwen3-moe/third_party/nlohmann/json.hpp @@ -0,0 +1,13908 @@ +// __ _____ _____ _____ +// __| | __| | | | JSON for Modern C++ +// | | |__ | | | | | | version 3.12.0 +// |_____|_____|_____|_|___| https://github.com/nlohmann/json +// +// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann +// SPDX-License-Identifier: MIT + +/****************************************************************************\ + * Note on documentation: The source files contain links to the online * + * documentation of the public API at https://json.nlohmann.me. This URL * + * contains the most recent documentation and should also be applicable to * + * previous versions; documentation for deprecated functions is not * + * removed, but marked deprecated. See "Generate documentation" section in * + * file docs/README.md. * +\****************************************************************************/ + +#ifndef INCLUDE_NLOHMANN_JSON_HPP_ +#define INCLUDE_NLOHMANN_JSON_HPP_ + +#include // all_of, find, for_each +#include // nullptr_t, ptrdiff_t, size_t +#include // hash, less +#include // initializer_list +#ifndef JSON_NO_IO + #include // istream, ostream +#endif // JSON_NO_IO +#include // random_access_iterator_tag +#include // unique_ptr +#include // string, stoi, to_string +#include // declval, forward, move, pair, swap +#include // vector + +// #include +// __ _____ _____ _____ +// __| | __| | | | JSON for Modern C++ +// | | |__ | | | | | | version 3.12.0 +// |_____|_____|_____|_|___| https://github.com/nlohmann/json +// +// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann +// SPDX-License-Identifier: MIT + + + +#include + +// #include +// __ _____ _____ _____ +// __| | __| | | | JSON for Modern C++ +// | | |__ | | | | | | version 3.12.0 +// |_____|_____|_____|_|___| https://github.com/nlohmann/json +// +// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann +// SPDX-License-Identifier: MIT + + + +// This file contains all macro definitions affecting or depending on the ABI + +#ifndef JSON_SKIP_LIBRARY_VERSION_CHECK + #if defined(NLOHMANN_JSON_VERSION_MAJOR) && defined(NLOHMANN_JSON_VERSION_MINOR) && defined(NLOHMANN_JSON_VERSION_PATCH) + #if NLOHMANN_JSON_VERSION_MAJOR != 3 || NLOHMANN_JSON_VERSION_MINOR != 12 || NLOHMANN_JSON_VERSION_PATCH != 0 + #warning "Already included a different version of the library!" + #endif + #endif +#endif + +#define NLOHMANN_JSON_VERSION_MAJOR 3 // NOLINT(modernize-macro-to-enum) +#define NLOHMANN_JSON_VERSION_MINOR 12 // NOLINT(modernize-macro-to-enum) +#define NLOHMANN_JSON_VERSION_PATCH 0 // NOLINT(modernize-macro-to-enum) + +#ifndef JSON_DIAGNOSTICS + #define JSON_DIAGNOSTICS 0 +#endif + +#ifndef JSON_DIAGNOSTIC_POSITIONS + #define JSON_DIAGNOSTIC_POSITIONS 0 +#endif + +#ifndef JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON + #define JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON 0 +#endif + +#if JSON_DIAGNOSTICS + #define NLOHMANN_JSON_ABI_TAG_DIAGNOSTICS _diag +#else + #define NLOHMANN_JSON_ABI_TAG_DIAGNOSTICS +#endif + +#if JSON_DIAGNOSTIC_POSITIONS + #define NLOHMANN_JSON_ABI_TAG_DIAGNOSTIC_POSITIONS _dp +#else + #define NLOHMANN_JSON_ABI_TAG_DIAGNOSTIC_POSITIONS +#endif + +#if JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON + #define NLOHMANN_JSON_ABI_TAG_LEGACY_DISCARDED_VALUE_COMPARISON _ldvcmp +#else + #define NLOHMANN_JSON_ABI_TAG_LEGACY_DISCARDED_VALUE_COMPARISON +#endif + +#ifndef NLOHMANN_JSON_NAMESPACE_NO_VERSION + #define NLOHMANN_JSON_NAMESPACE_NO_VERSION 0 +#endif + +// Construct the namespace ABI tags component +#define NLOHMANN_JSON_ABI_TAGS_CONCAT_EX(a, b, c) json_abi ## a ## b ## c +#define NLOHMANN_JSON_ABI_TAGS_CONCAT(a, b, c) \ + NLOHMANN_JSON_ABI_TAGS_CONCAT_EX(a, b, c) + +#define NLOHMANN_JSON_ABI_TAGS \ + NLOHMANN_JSON_ABI_TAGS_CONCAT( \ + NLOHMANN_JSON_ABI_TAG_DIAGNOSTICS, \ + NLOHMANN_JSON_ABI_TAG_LEGACY_DISCARDED_VALUE_COMPARISON, \ + NLOHMANN_JSON_ABI_TAG_DIAGNOSTIC_POSITIONS) + +// Construct the namespace version component +#define NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT_EX(major, minor, patch) \ + _v ## major ## _ ## minor ## _ ## patch +#define NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT(major, minor, patch) \ + NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT_EX(major, minor, patch) + +#if NLOHMANN_JSON_NAMESPACE_NO_VERSION +#define NLOHMANN_JSON_NAMESPACE_VERSION +#else +#define NLOHMANN_JSON_NAMESPACE_VERSION \ + NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT(NLOHMANN_JSON_VERSION_MAJOR, \ + NLOHMANN_JSON_VERSION_MINOR, \ + NLOHMANN_JSON_VERSION_PATCH) +#endif + +// Combine namespace components +#define NLOHMANN_JSON_NAMESPACE_CONCAT_EX(a, b) a ## b +#define NLOHMANN_JSON_NAMESPACE_CONCAT(a, b) \ + NLOHMANN_JSON_NAMESPACE_CONCAT_EX(a, b) + +#ifndef NLOHMANN_JSON_NAMESPACE +#define NLOHMANN_JSON_NAMESPACE \ + nlohmann::NLOHMANN_JSON_NAMESPACE_CONCAT( \ + NLOHMANN_JSON_ABI_TAGS, \ + NLOHMANN_JSON_NAMESPACE_VERSION) +#endif + +#ifndef NLOHMANN_JSON_NAMESPACE_BEGIN +#define NLOHMANN_JSON_NAMESPACE_BEGIN \ + namespace nlohmann \ + { \ + inline namespace NLOHMANN_JSON_NAMESPACE_CONCAT( \ + NLOHMANN_JSON_ABI_TAGS, \ + NLOHMANN_JSON_NAMESPACE_VERSION) \ + { +#endif + +#ifndef NLOHMANN_JSON_NAMESPACE_END +#define NLOHMANN_JSON_NAMESPACE_END \ + } /* namespace (inline namespace) NOLINT(readability/namespace) */ \ + } // namespace nlohmann +#endif + +// #include +// __ _____ _____ _____ +// __| | __| | | | JSON for Modern C++ +// | | |__ | | | | | | version 3.12.0 +// |_____|_____|_____|_|___| https://github.com/nlohmann/json +// +// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann +// SPDX-License-Identifier: MIT + + + +#include // transform +#include // array +#include // forward_list +#include // inserter, front_inserter, end +#include // map +#ifdef JSON_HAS_CPP_17 + #include // optional +#endif +#include // string +#include // tuple, make_tuple +#include // is_arithmetic, is_same, is_enum, underlying_type, is_convertible +#include // unordered_map +#include // pair, declval +#include // valarray + +// #include +// __ _____ _____ _____ +// __| | __| | | | JSON for Modern C++ +// | | |__ | | | | | | version 3.12.0 +// |_____|_____|_____|_|___| https://github.com/nlohmann/json +// +// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann +// SPDX-License-Identifier: MIT + + + +#include // nullptr_t +#include // exception +#if JSON_DIAGNOSTICS + #include // accumulate +#endif +#include // runtime_error +#include // to_string +#include // vector + +// #include +// __ _____ _____ _____ +// __| | __| | | | JSON for Modern C++ +// | | |__ | | | | | | version 3.12.0 +// |_____|_____|_____|_|___| https://github.com/nlohmann/json +// +// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann +// SPDX-License-Identifier: MIT + + + +#include // array +#include // size_t +#include // uint8_t +#include // string + +// #include +// __ _____ _____ _____ +// __| | __| | | | JSON for Modern C++ +// | | |__ | | | | | | version 3.12.0 +// |_____|_____|_____|_|___| https://github.com/nlohmann/json +// +// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann +// SPDX-License-Identifier: MIT + + + +#include // declval, pair +// #include +// __ _____ _____ _____ +// __| | __| | | | JSON for Modern C++ +// | | |__ | | | | | | version 3.12.0 +// |_____|_____|_____|_|___| https://github.com/nlohmann/json +// +// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann +// SPDX-License-Identifier: MIT + + + +#include + +// #include +// __ _____ _____ _____ +// __| | __| | | | JSON for Modern C++ +// | | |__ | | | | | | version 3.12.0 +// |_____|_____|_____|_|___| https://github.com/nlohmann/json +// +// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann +// SPDX-License-Identifier: MIT + + + +// #include + + +NLOHMANN_JSON_NAMESPACE_BEGIN +namespace detail +{ + +template struct make_void +{ + using type = void; +}; +template using void_t = typename make_void::type; + +} // namespace detail +NLOHMANN_JSON_NAMESPACE_END + + +NLOHMANN_JSON_NAMESPACE_BEGIN +namespace detail +{ + +// https://en.cppreference.com/w/cpp/experimental/is_detected +struct nonesuch +{ + nonesuch() = delete; + ~nonesuch() = delete; + nonesuch(nonesuch const&) = delete; + nonesuch(nonesuch const&&) = delete; + void operator=(nonesuch const&) = delete; + void operator=(nonesuch&&) = delete; +}; + +template class Op, + class... Args> +struct detector +{ + using value_t = std::false_type; + using type = Default; +}; + +template class Op, class... Args> +struct detector>, Op, Args...> +{ + using value_t = std::true_type; + using type = Op; +}; + +template class Op, class... Args> +using is_detected = typename detector::value_t; + +template class Op, class... Args> +struct is_detected_lazy : is_detected { }; + +template class Op, class... Args> +using detected_t = typename detector::type; + +template class Op, class... Args> +using detected_or = detector; + +template class Op, class... Args> +using detected_or_t = typename detected_or::type; + +template class Op, class... Args> +using is_detected_exact = std::is_same>; + +template class Op, class... Args> +using is_detected_convertible = + std::is_convertible, To>; + +} // namespace detail +NLOHMANN_JSON_NAMESPACE_END + +// #include + + +// __ _____ _____ _____ +// __| | __| | | | JSON for Modern C++ +// | | |__ | | | | | | version 3.12.0 +// |_____|_____|_____|_|___| https://github.com/nlohmann/json +// +// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann +// SPDX-FileCopyrightText: 2016 - 2021 Evan Nemerson +// SPDX-License-Identifier: MIT + +/* Hedley - https://nemequ.github.io/hedley + * Created by Evan Nemerson + */ + +#if !defined(JSON_HEDLEY_VERSION) || (JSON_HEDLEY_VERSION < 15) +#if defined(JSON_HEDLEY_VERSION) + #undef JSON_HEDLEY_VERSION +#endif +#define JSON_HEDLEY_VERSION 15 + +#if defined(JSON_HEDLEY_STRINGIFY_EX) + #undef JSON_HEDLEY_STRINGIFY_EX +#endif +#define JSON_HEDLEY_STRINGIFY_EX(x) #x + +#if defined(JSON_HEDLEY_STRINGIFY) + #undef JSON_HEDLEY_STRINGIFY +#endif +#define JSON_HEDLEY_STRINGIFY(x) JSON_HEDLEY_STRINGIFY_EX(x) + +#if defined(JSON_HEDLEY_CONCAT_EX) + #undef JSON_HEDLEY_CONCAT_EX +#endif +#define JSON_HEDLEY_CONCAT_EX(a,b) a##b + +#if defined(JSON_HEDLEY_CONCAT) + #undef JSON_HEDLEY_CONCAT +#endif +#define JSON_HEDLEY_CONCAT(a,b) JSON_HEDLEY_CONCAT_EX(a,b) + +#if defined(JSON_HEDLEY_CONCAT3_EX) + #undef JSON_HEDLEY_CONCAT3_EX +#endif +#define JSON_HEDLEY_CONCAT3_EX(a,b,c) a##b##c + +#if defined(JSON_HEDLEY_CONCAT3) + #undef JSON_HEDLEY_CONCAT3 +#endif +#define JSON_HEDLEY_CONCAT3(a,b,c) JSON_HEDLEY_CONCAT3_EX(a,b,c) + +#if defined(JSON_HEDLEY_VERSION_ENCODE) + #undef JSON_HEDLEY_VERSION_ENCODE +#endif +#define JSON_HEDLEY_VERSION_ENCODE(major,minor,revision) (((major) * 1000000) + ((minor) * 1000) + (revision)) + +#if defined(JSON_HEDLEY_VERSION_DECODE_MAJOR) + #undef JSON_HEDLEY_VERSION_DECODE_MAJOR +#endif +#define JSON_HEDLEY_VERSION_DECODE_MAJOR(version) ((version) / 1000000) + +#if defined(JSON_HEDLEY_VERSION_DECODE_MINOR) + #undef JSON_HEDLEY_VERSION_DECODE_MINOR +#endif +#define JSON_HEDLEY_VERSION_DECODE_MINOR(version) (((version) % 1000000) / 1000) + +#if defined(JSON_HEDLEY_VERSION_DECODE_REVISION) + #undef JSON_HEDLEY_VERSION_DECODE_REVISION +#endif +#define JSON_HEDLEY_VERSION_DECODE_REVISION(version) ((version) % 1000) + +#if defined(JSON_HEDLEY_GNUC_VERSION) + #undef JSON_HEDLEY_GNUC_VERSION +#endif +#if defined(__GNUC__) && defined(__GNUC_PATCHLEVEL__) + #define JSON_HEDLEY_GNUC_VERSION JSON_HEDLEY_VERSION_ENCODE(__GNUC__, __GNUC_MINOR__, __GNUC_PATCHLEVEL__) +#elif defined(__GNUC__) + #define JSON_HEDLEY_GNUC_VERSION JSON_HEDLEY_VERSION_ENCODE(__GNUC__, __GNUC_MINOR__, 0) +#endif + +#if defined(JSON_HEDLEY_GNUC_VERSION_CHECK) + #undef JSON_HEDLEY_GNUC_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_GNUC_VERSION) + #define JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_GNUC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_MSVC_VERSION) + #undef JSON_HEDLEY_MSVC_VERSION +#endif +#if defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 140000000) && !defined(__ICL) + #define JSON_HEDLEY_MSVC_VERSION JSON_HEDLEY_VERSION_ENCODE(_MSC_FULL_VER / 10000000, (_MSC_FULL_VER % 10000000) / 100000, (_MSC_FULL_VER % 100000) / 100) +#elif defined(_MSC_FULL_VER) && !defined(__ICL) + #define JSON_HEDLEY_MSVC_VERSION JSON_HEDLEY_VERSION_ENCODE(_MSC_FULL_VER / 1000000, (_MSC_FULL_VER % 1000000) / 10000, (_MSC_FULL_VER % 10000) / 10) +#elif defined(_MSC_VER) && !defined(__ICL) + #define JSON_HEDLEY_MSVC_VERSION JSON_HEDLEY_VERSION_ENCODE(_MSC_VER / 100, _MSC_VER % 100, 0) +#endif + +#if defined(JSON_HEDLEY_MSVC_VERSION_CHECK) + #undef JSON_HEDLEY_MSVC_VERSION_CHECK +#endif +#if !defined(JSON_HEDLEY_MSVC_VERSION) + #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (0) +#elif defined(_MSC_VER) && (_MSC_VER >= 1400) + #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (_MSC_FULL_VER >= ((major * 10000000) + (minor * 100000) + (patch))) +#elif defined(_MSC_VER) && (_MSC_VER >= 1200) + #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (_MSC_FULL_VER >= ((major * 1000000) + (minor * 10000) + (patch))) +#else + #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (_MSC_VER >= ((major * 100) + (minor))) +#endif + +#if defined(JSON_HEDLEY_INTEL_VERSION) + #undef JSON_HEDLEY_INTEL_VERSION +#endif +#if defined(__INTEL_COMPILER) && defined(__INTEL_COMPILER_UPDATE) && !defined(__ICL) + #define JSON_HEDLEY_INTEL_VERSION JSON_HEDLEY_VERSION_ENCODE(__INTEL_COMPILER / 100, __INTEL_COMPILER % 100, __INTEL_COMPILER_UPDATE) +#elif defined(__INTEL_COMPILER) && !defined(__ICL) + #define JSON_HEDLEY_INTEL_VERSION JSON_HEDLEY_VERSION_ENCODE(__INTEL_COMPILER / 100, __INTEL_COMPILER % 100, 0) +#endif + +#if defined(JSON_HEDLEY_INTEL_VERSION_CHECK) + #undef JSON_HEDLEY_INTEL_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_INTEL_VERSION) + #define JSON_HEDLEY_INTEL_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_INTEL_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_INTEL_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_INTEL_CL_VERSION) + #undef JSON_HEDLEY_INTEL_CL_VERSION +#endif +#if defined(__INTEL_COMPILER) && defined(__INTEL_COMPILER_UPDATE) && defined(__ICL) + #define JSON_HEDLEY_INTEL_CL_VERSION JSON_HEDLEY_VERSION_ENCODE(__INTEL_COMPILER, __INTEL_COMPILER_UPDATE, 0) +#endif + +#if defined(JSON_HEDLEY_INTEL_CL_VERSION_CHECK) + #undef JSON_HEDLEY_INTEL_CL_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_INTEL_CL_VERSION) + #define JSON_HEDLEY_INTEL_CL_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_INTEL_CL_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_INTEL_CL_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_PGI_VERSION) + #undef JSON_HEDLEY_PGI_VERSION +#endif +#if defined(__PGI) && defined(__PGIC__) && defined(__PGIC_MINOR__) && defined(__PGIC_PATCHLEVEL__) + #define JSON_HEDLEY_PGI_VERSION JSON_HEDLEY_VERSION_ENCODE(__PGIC__, __PGIC_MINOR__, __PGIC_PATCHLEVEL__) +#endif + +#if defined(JSON_HEDLEY_PGI_VERSION_CHECK) + #undef JSON_HEDLEY_PGI_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_PGI_VERSION) + #define JSON_HEDLEY_PGI_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_PGI_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_PGI_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_SUNPRO_VERSION) + #undef JSON_HEDLEY_SUNPRO_VERSION +#endif +#if defined(__SUNPRO_C) && (__SUNPRO_C > 0x1000) + #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((((__SUNPRO_C >> 16) & 0xf) * 10) + ((__SUNPRO_C >> 12) & 0xf), (((__SUNPRO_C >> 8) & 0xf) * 10) + ((__SUNPRO_C >> 4) & 0xf), (__SUNPRO_C & 0xf) * 10) +#elif defined(__SUNPRO_C) + #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((__SUNPRO_C >> 8) & 0xf, (__SUNPRO_C >> 4) & 0xf, (__SUNPRO_C) & 0xf) +#elif defined(__SUNPRO_CC) && (__SUNPRO_CC > 0x1000) + #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((((__SUNPRO_CC >> 16) & 0xf) * 10) + ((__SUNPRO_CC >> 12) & 0xf), (((__SUNPRO_CC >> 8) & 0xf) * 10) + ((__SUNPRO_CC >> 4) & 0xf), (__SUNPRO_CC & 0xf) * 10) +#elif defined(__SUNPRO_CC) + #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((__SUNPRO_CC >> 8) & 0xf, (__SUNPRO_CC >> 4) & 0xf, (__SUNPRO_CC) & 0xf) +#endif + +#if defined(JSON_HEDLEY_SUNPRO_VERSION_CHECK) + #undef JSON_HEDLEY_SUNPRO_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_SUNPRO_VERSION) + #define JSON_HEDLEY_SUNPRO_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_SUNPRO_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_SUNPRO_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_EMSCRIPTEN_VERSION) + #undef JSON_HEDLEY_EMSCRIPTEN_VERSION +#endif +#if defined(__EMSCRIPTEN__) + #define JSON_HEDLEY_EMSCRIPTEN_VERSION JSON_HEDLEY_VERSION_ENCODE(__EMSCRIPTEN_major__, __EMSCRIPTEN_minor__, __EMSCRIPTEN_tiny__) +#endif + +#if defined(JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK) + #undef JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_EMSCRIPTEN_VERSION) + #define JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_EMSCRIPTEN_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_ARM_VERSION) + #undef JSON_HEDLEY_ARM_VERSION +#endif +#if defined(__CC_ARM) && defined(__ARMCOMPILER_VERSION) + #define JSON_HEDLEY_ARM_VERSION JSON_HEDLEY_VERSION_ENCODE(__ARMCOMPILER_VERSION / 1000000, (__ARMCOMPILER_VERSION % 1000000) / 10000, (__ARMCOMPILER_VERSION % 10000) / 100) +#elif defined(__CC_ARM) && defined(__ARMCC_VERSION) + #define JSON_HEDLEY_ARM_VERSION JSON_HEDLEY_VERSION_ENCODE(__ARMCC_VERSION / 1000000, (__ARMCC_VERSION % 1000000) / 10000, (__ARMCC_VERSION % 10000) / 100) +#endif + +#if defined(JSON_HEDLEY_ARM_VERSION_CHECK) + #undef JSON_HEDLEY_ARM_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_ARM_VERSION) + #define JSON_HEDLEY_ARM_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_ARM_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_ARM_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_IBM_VERSION) + #undef JSON_HEDLEY_IBM_VERSION +#endif +#if defined(__ibmxl__) + #define JSON_HEDLEY_IBM_VERSION JSON_HEDLEY_VERSION_ENCODE(__ibmxl_version__, __ibmxl_release__, __ibmxl_modification__) +#elif defined(__xlC__) && defined(__xlC_ver__) + #define JSON_HEDLEY_IBM_VERSION JSON_HEDLEY_VERSION_ENCODE(__xlC__ >> 8, __xlC__ & 0xff, (__xlC_ver__ >> 8) & 0xff) +#elif defined(__xlC__) + #define JSON_HEDLEY_IBM_VERSION JSON_HEDLEY_VERSION_ENCODE(__xlC__ >> 8, __xlC__ & 0xff, 0) +#endif + +#if defined(JSON_HEDLEY_IBM_VERSION_CHECK) + #undef JSON_HEDLEY_IBM_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_IBM_VERSION) + #define JSON_HEDLEY_IBM_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_IBM_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_IBM_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_TI_VERSION) + #undef JSON_HEDLEY_TI_VERSION +#endif +#if \ + defined(__TI_COMPILER_VERSION__) && \ + ( \ + defined(__TMS470__) || defined(__TI_ARM__) || \ + defined(__MSP430__) || \ + defined(__TMS320C2000__) \ + ) +#if (__TI_COMPILER_VERSION__ >= 16000000) + #define JSON_HEDLEY_TI_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000)) +#endif +#endif + +#if defined(JSON_HEDLEY_TI_VERSION_CHECK) + #undef JSON_HEDLEY_TI_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_TI_VERSION) + #define JSON_HEDLEY_TI_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_TI_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_TI_CL2000_VERSION) + #undef JSON_HEDLEY_TI_CL2000_VERSION +#endif +#if defined(__TI_COMPILER_VERSION__) && defined(__TMS320C2000__) + #define JSON_HEDLEY_TI_CL2000_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000)) +#endif + +#if defined(JSON_HEDLEY_TI_CL2000_VERSION_CHECK) + #undef JSON_HEDLEY_TI_CL2000_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_TI_CL2000_VERSION) + #define JSON_HEDLEY_TI_CL2000_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL2000_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_TI_CL2000_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_TI_CL430_VERSION) + #undef JSON_HEDLEY_TI_CL430_VERSION +#endif +#if defined(__TI_COMPILER_VERSION__) && defined(__MSP430__) + #define JSON_HEDLEY_TI_CL430_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000)) +#endif + +#if defined(JSON_HEDLEY_TI_CL430_VERSION_CHECK) + #undef JSON_HEDLEY_TI_CL430_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_TI_CL430_VERSION) + #define JSON_HEDLEY_TI_CL430_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL430_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_TI_CL430_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_TI_ARMCL_VERSION) + #undef JSON_HEDLEY_TI_ARMCL_VERSION +#endif +#if defined(__TI_COMPILER_VERSION__) && (defined(__TMS470__) || defined(__TI_ARM__)) + #define JSON_HEDLEY_TI_ARMCL_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000)) +#endif + +#if defined(JSON_HEDLEY_TI_ARMCL_VERSION_CHECK) + #undef JSON_HEDLEY_TI_ARMCL_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_TI_ARMCL_VERSION) + #define JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_ARMCL_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_TI_CL6X_VERSION) + #undef JSON_HEDLEY_TI_CL6X_VERSION +#endif +#if defined(__TI_COMPILER_VERSION__) && defined(__TMS320C6X__) + #define JSON_HEDLEY_TI_CL6X_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000)) +#endif + +#if defined(JSON_HEDLEY_TI_CL6X_VERSION_CHECK) + #undef JSON_HEDLEY_TI_CL6X_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_TI_CL6X_VERSION) + #define JSON_HEDLEY_TI_CL6X_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL6X_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_TI_CL6X_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_TI_CL7X_VERSION) + #undef JSON_HEDLEY_TI_CL7X_VERSION +#endif +#if defined(__TI_COMPILER_VERSION__) && defined(__C7000__) + #define JSON_HEDLEY_TI_CL7X_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000)) +#endif + +#if defined(JSON_HEDLEY_TI_CL7X_VERSION_CHECK) + #undef JSON_HEDLEY_TI_CL7X_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_TI_CL7X_VERSION) + #define JSON_HEDLEY_TI_CL7X_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL7X_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_TI_CL7X_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_TI_CLPRU_VERSION) + #undef JSON_HEDLEY_TI_CLPRU_VERSION +#endif +#if defined(__TI_COMPILER_VERSION__) && defined(__PRU__) + #define JSON_HEDLEY_TI_CLPRU_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000)) +#endif + +#if defined(JSON_HEDLEY_TI_CLPRU_VERSION_CHECK) + #undef JSON_HEDLEY_TI_CLPRU_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_TI_CLPRU_VERSION) + #define JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CLPRU_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_CRAY_VERSION) + #undef JSON_HEDLEY_CRAY_VERSION +#endif +#if defined(_CRAYC) + #if defined(_RELEASE_PATCHLEVEL) + #define JSON_HEDLEY_CRAY_VERSION JSON_HEDLEY_VERSION_ENCODE(_RELEASE_MAJOR, _RELEASE_MINOR, _RELEASE_PATCHLEVEL) + #else + #define JSON_HEDLEY_CRAY_VERSION JSON_HEDLEY_VERSION_ENCODE(_RELEASE_MAJOR, _RELEASE_MINOR, 0) + #endif +#endif + +#if defined(JSON_HEDLEY_CRAY_VERSION_CHECK) + #undef JSON_HEDLEY_CRAY_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_CRAY_VERSION) + #define JSON_HEDLEY_CRAY_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_CRAY_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_CRAY_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_IAR_VERSION) + #undef JSON_HEDLEY_IAR_VERSION +#endif +#if defined(__IAR_SYSTEMS_ICC__) + #if __VER__ > 1000 + #define JSON_HEDLEY_IAR_VERSION JSON_HEDLEY_VERSION_ENCODE((__VER__ / 1000000), ((__VER__ / 1000) % 1000), (__VER__ % 1000)) + #else + #define JSON_HEDLEY_IAR_VERSION JSON_HEDLEY_VERSION_ENCODE(__VER__ / 100, __VER__ % 100, 0) + #endif +#endif + +#if defined(JSON_HEDLEY_IAR_VERSION_CHECK) + #undef JSON_HEDLEY_IAR_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_IAR_VERSION) + #define JSON_HEDLEY_IAR_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_IAR_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_IAR_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_TINYC_VERSION) + #undef JSON_HEDLEY_TINYC_VERSION +#endif +#if defined(__TINYC__) + #define JSON_HEDLEY_TINYC_VERSION JSON_HEDLEY_VERSION_ENCODE(__TINYC__ / 1000, (__TINYC__ / 100) % 10, __TINYC__ % 100) +#endif + +#if defined(JSON_HEDLEY_TINYC_VERSION_CHECK) + #undef JSON_HEDLEY_TINYC_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_TINYC_VERSION) + #define JSON_HEDLEY_TINYC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TINYC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_TINYC_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_DMC_VERSION) + #undef JSON_HEDLEY_DMC_VERSION +#endif +#if defined(__DMC__) + #define JSON_HEDLEY_DMC_VERSION JSON_HEDLEY_VERSION_ENCODE(__DMC__ >> 8, (__DMC__ >> 4) & 0xf, __DMC__ & 0xf) +#endif + +#if defined(JSON_HEDLEY_DMC_VERSION_CHECK) + #undef JSON_HEDLEY_DMC_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_DMC_VERSION) + #define JSON_HEDLEY_DMC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_DMC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_DMC_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_COMPCERT_VERSION) + #undef JSON_HEDLEY_COMPCERT_VERSION +#endif +#if defined(__COMPCERT_VERSION__) + #define JSON_HEDLEY_COMPCERT_VERSION JSON_HEDLEY_VERSION_ENCODE(__COMPCERT_VERSION__ / 10000, (__COMPCERT_VERSION__ / 100) % 100, __COMPCERT_VERSION__ % 100) +#endif + +#if defined(JSON_HEDLEY_COMPCERT_VERSION_CHECK) + #undef JSON_HEDLEY_COMPCERT_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_COMPCERT_VERSION) + #define JSON_HEDLEY_COMPCERT_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_COMPCERT_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_COMPCERT_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_PELLES_VERSION) + #undef JSON_HEDLEY_PELLES_VERSION +#endif +#if defined(__POCC__) + #define JSON_HEDLEY_PELLES_VERSION JSON_HEDLEY_VERSION_ENCODE(__POCC__ / 100, __POCC__ % 100, 0) +#endif + +#if defined(JSON_HEDLEY_PELLES_VERSION_CHECK) + #undef JSON_HEDLEY_PELLES_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_PELLES_VERSION) + #define JSON_HEDLEY_PELLES_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_PELLES_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_PELLES_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_MCST_LCC_VERSION) + #undef JSON_HEDLEY_MCST_LCC_VERSION +#endif +#if defined(__LCC__) && defined(__LCC_MINOR__) + #define JSON_HEDLEY_MCST_LCC_VERSION JSON_HEDLEY_VERSION_ENCODE(__LCC__ / 100, __LCC__ % 100, __LCC_MINOR__) +#endif + +#if defined(JSON_HEDLEY_MCST_LCC_VERSION_CHECK) + #undef JSON_HEDLEY_MCST_LCC_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_MCST_LCC_VERSION) + #define JSON_HEDLEY_MCST_LCC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_MCST_LCC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_MCST_LCC_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_GCC_VERSION) + #undef JSON_HEDLEY_GCC_VERSION +#endif +#if \ + defined(JSON_HEDLEY_GNUC_VERSION) && \ + !defined(__clang__) && \ + !defined(JSON_HEDLEY_INTEL_VERSION) && \ + !defined(JSON_HEDLEY_PGI_VERSION) && \ + !defined(JSON_HEDLEY_ARM_VERSION) && \ + !defined(JSON_HEDLEY_CRAY_VERSION) && \ + !defined(JSON_HEDLEY_TI_VERSION) && \ + !defined(JSON_HEDLEY_TI_ARMCL_VERSION) && \ + !defined(JSON_HEDLEY_TI_CL430_VERSION) && \ + !defined(JSON_HEDLEY_TI_CL2000_VERSION) && \ + !defined(JSON_HEDLEY_TI_CL6X_VERSION) && \ + !defined(JSON_HEDLEY_TI_CL7X_VERSION) && \ + !defined(JSON_HEDLEY_TI_CLPRU_VERSION) && \ + !defined(__COMPCERT__) && \ + !defined(JSON_HEDLEY_MCST_LCC_VERSION) + #define JSON_HEDLEY_GCC_VERSION JSON_HEDLEY_GNUC_VERSION +#endif + +#if defined(JSON_HEDLEY_GCC_VERSION_CHECK) + #undef JSON_HEDLEY_GCC_VERSION_CHECK +#endif +#if defined(JSON_HEDLEY_GCC_VERSION) + #define JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_GCC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch)) +#else + #define JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch) (0) +#endif + +#if defined(JSON_HEDLEY_HAS_ATTRIBUTE) + #undef JSON_HEDLEY_HAS_ATTRIBUTE +#endif +#if \ + defined(__has_attribute) && \ + ( \ + (!defined(JSON_HEDLEY_IAR_VERSION) || JSON_HEDLEY_IAR_VERSION_CHECK(8,5,9)) \ + ) +# define JSON_HEDLEY_HAS_ATTRIBUTE(attribute) __has_attribute(attribute) +#else +# define JSON_HEDLEY_HAS_ATTRIBUTE(attribute) (0) +#endif + +#if defined(JSON_HEDLEY_GNUC_HAS_ATTRIBUTE) + #undef JSON_HEDLEY_GNUC_HAS_ATTRIBUTE +#endif +#if defined(__has_attribute) + #define JSON_HEDLEY_GNUC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_HAS_ATTRIBUTE(attribute) +#else + #define JSON_HEDLEY_GNUC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch) +#endif + +#if defined(JSON_HEDLEY_GCC_HAS_ATTRIBUTE) + #undef JSON_HEDLEY_GCC_HAS_ATTRIBUTE +#endif +#if defined(__has_attribute) + #define JSON_HEDLEY_GCC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_HAS_ATTRIBUTE(attribute) +#else + #define JSON_HEDLEY_GCC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch) +#endif + +#if defined(JSON_HEDLEY_HAS_CPP_ATTRIBUTE) + #undef JSON_HEDLEY_HAS_CPP_ATTRIBUTE +#endif +#if \ + defined(__has_cpp_attribute) && \ + defined(__cplusplus) && \ + (!defined(JSON_HEDLEY_SUNPRO_VERSION) || JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0)) + #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE(attribute) __has_cpp_attribute(attribute) +#else + #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE(attribute) (0) +#endif + +#if defined(JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS) + #undef JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS +#endif +#if !defined(__cplusplus) || !defined(__has_cpp_attribute) + #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(ns,attribute) (0) +#elif \ + !defined(JSON_HEDLEY_PGI_VERSION) && \ + !defined(JSON_HEDLEY_IAR_VERSION) && \ + (!defined(JSON_HEDLEY_SUNPRO_VERSION) || JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0)) && \ + (!defined(JSON_HEDLEY_MSVC_VERSION) || JSON_HEDLEY_MSVC_VERSION_CHECK(19,20,0)) + #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(ns,attribute) JSON_HEDLEY_HAS_CPP_ATTRIBUTE(ns::attribute) +#else + #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(ns,attribute) (0) +#endif + +#if defined(JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE) + #undef JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE +#endif +#if defined(__has_cpp_attribute) && defined(__cplusplus) + #define JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) __has_cpp_attribute(attribute) +#else + #define JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch) +#endif + +#if defined(JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE) + #undef JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE +#endif +#if defined(__has_cpp_attribute) && defined(__cplusplus) + #define JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) __has_cpp_attribute(attribute) +#else + #define JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch) +#endif + +#if defined(JSON_HEDLEY_HAS_BUILTIN) + #undef JSON_HEDLEY_HAS_BUILTIN +#endif +#if defined(__has_builtin) + #define JSON_HEDLEY_HAS_BUILTIN(builtin) __has_builtin(builtin) +#else + #define JSON_HEDLEY_HAS_BUILTIN(builtin) (0) +#endif + +#if defined(JSON_HEDLEY_GNUC_HAS_BUILTIN) + #undef JSON_HEDLEY_GNUC_HAS_BUILTIN +#endif +#if defined(__has_builtin) + #define JSON_HEDLEY_GNUC_HAS_BUILTIN(builtin,major,minor,patch) __has_builtin(builtin) +#else + #define JSON_HEDLEY_GNUC_HAS_BUILTIN(builtin,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch) +#endif + +#if defined(JSON_HEDLEY_GCC_HAS_BUILTIN) + #undef JSON_HEDLEY_GCC_HAS_BUILTIN +#endif +#if defined(__has_builtin) + #define JSON_HEDLEY_GCC_HAS_BUILTIN(builtin,major,minor,patch) __has_builtin(builtin) +#else + #define JSON_HEDLEY_GCC_HAS_BUILTIN(builtin,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch) +#endif + +#if defined(JSON_HEDLEY_HAS_FEATURE) + #undef JSON_HEDLEY_HAS_FEATURE +#endif +#if defined(__has_feature) + #define JSON_HEDLEY_HAS_FEATURE(feature) __has_feature(feature) +#else + #define JSON_HEDLEY_HAS_FEATURE(feature) (0) +#endif + +#if defined(JSON_HEDLEY_GNUC_HAS_FEATURE) + #undef JSON_HEDLEY_GNUC_HAS_FEATURE +#endif +#if defined(__has_feature) + #define JSON_HEDLEY_GNUC_HAS_FEATURE(feature,major,minor,patch) __has_feature(feature) +#else + #define JSON_HEDLEY_GNUC_HAS_FEATURE(feature,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch) +#endif + +#if defined(JSON_HEDLEY_GCC_HAS_FEATURE) + #undef JSON_HEDLEY_GCC_HAS_FEATURE +#endif +#if defined(__has_feature) + #define JSON_HEDLEY_GCC_HAS_FEATURE(feature,major,minor,patch) __has_feature(feature) +#else + #define JSON_HEDLEY_GCC_HAS_FEATURE(feature,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch) +#endif + +#if defined(JSON_HEDLEY_HAS_EXTENSION) + #undef JSON_HEDLEY_HAS_EXTENSION +#endif +#if defined(__has_extension) + #define JSON_HEDLEY_HAS_EXTENSION(extension) __has_extension(extension) +#else + #define JSON_HEDLEY_HAS_EXTENSION(extension) (0) +#endif + +#if defined(JSON_HEDLEY_GNUC_HAS_EXTENSION) + #undef JSON_HEDLEY_GNUC_HAS_EXTENSION +#endif +#if defined(__has_extension) + #define JSON_HEDLEY_GNUC_HAS_EXTENSION(extension,major,minor,patch) __has_extension(extension) +#else + #define JSON_HEDLEY_GNUC_HAS_EXTENSION(extension,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch) +#endif + +#if defined(JSON_HEDLEY_GCC_HAS_EXTENSION) + #undef JSON_HEDLEY_GCC_HAS_EXTENSION +#endif +#if defined(__has_extension) + #define JSON_HEDLEY_GCC_HAS_EXTENSION(extension,major,minor,patch) __has_extension(extension) +#else + #define JSON_HEDLEY_GCC_HAS_EXTENSION(extension,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch) +#endif + +#if defined(JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE) + #undef JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE +#endif +#if defined(__has_declspec_attribute) + #define JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE(attribute) __has_declspec_attribute(attribute) +#else + #define JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE(attribute) (0) +#endif + +#if defined(JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE) + #undef JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE +#endif +#if defined(__has_declspec_attribute) + #define JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) __has_declspec_attribute(attribute) +#else + #define JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch) +#endif + +#if defined(JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE) + #undef JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE +#endif +#if defined(__has_declspec_attribute) + #define JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) __has_declspec_attribute(attribute) +#else + #define JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch) +#endif + +#if defined(JSON_HEDLEY_HAS_WARNING) + #undef JSON_HEDLEY_HAS_WARNING +#endif +#if defined(__has_warning) + #define JSON_HEDLEY_HAS_WARNING(warning) __has_warning(warning) +#else + #define JSON_HEDLEY_HAS_WARNING(warning) (0) +#endif + +#if defined(JSON_HEDLEY_GNUC_HAS_WARNING) + #undef JSON_HEDLEY_GNUC_HAS_WARNING +#endif +#if defined(__has_warning) + #define JSON_HEDLEY_GNUC_HAS_WARNING(warning,major,minor,patch) __has_warning(warning) +#else + #define JSON_HEDLEY_GNUC_HAS_WARNING(warning,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch) +#endif + +#if defined(JSON_HEDLEY_GCC_HAS_WARNING) + #undef JSON_HEDLEY_GCC_HAS_WARNING +#endif +#if defined(__has_warning) + #define JSON_HEDLEY_GCC_HAS_WARNING(warning,major,minor,patch) __has_warning(warning) +#else + #define JSON_HEDLEY_GCC_HAS_WARNING(warning,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch) +#endif + +#if \ + (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L)) || \ + defined(__clang__) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(3,0,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0) || \ + JSON_HEDLEY_PGI_VERSION_CHECK(18,4,0) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \ + JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \ + JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,7,0) || \ + JSON_HEDLEY_TI_CL430_VERSION_CHECK(2,0,1) || \ + JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,1,0) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,0,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \ + JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \ + JSON_HEDLEY_CRAY_VERSION_CHECK(5,0,0) || \ + JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,17) || \ + JSON_HEDLEY_SUNPRO_VERSION_CHECK(8,0,0) || \ + (JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) && defined(__C99_PRAGMA_OPERATOR)) + #define JSON_HEDLEY_PRAGMA(value) _Pragma(#value) +#elif JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0) + #define JSON_HEDLEY_PRAGMA(value) __pragma(value) +#else + #define JSON_HEDLEY_PRAGMA(value) +#endif + +#if defined(JSON_HEDLEY_DIAGNOSTIC_PUSH) + #undef JSON_HEDLEY_DIAGNOSTIC_PUSH +#endif +#if defined(JSON_HEDLEY_DIAGNOSTIC_POP) + #undef JSON_HEDLEY_DIAGNOSTIC_POP +#endif +#if defined(__clang__) + #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("clang diagnostic push") + #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("clang diagnostic pop") +#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) + #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("warning(push)") + #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("warning(pop)") +#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,6,0) + #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("GCC diagnostic push") + #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("GCC diagnostic pop") +#elif \ + JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0) || \ + JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) + #define JSON_HEDLEY_DIAGNOSTIC_PUSH __pragma(warning(push)) + #define JSON_HEDLEY_DIAGNOSTIC_POP __pragma(warning(pop)) +#elif JSON_HEDLEY_ARM_VERSION_CHECK(5,6,0) + #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("push") + #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("pop") +#elif \ + JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \ + JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \ + JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,4,0) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,1,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \ + JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) + #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("diag_push") + #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("diag_pop") +#elif JSON_HEDLEY_PELLES_VERSION_CHECK(2,90,0) + #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("warning(push)") + #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("warning(pop)") +#else + #define JSON_HEDLEY_DIAGNOSTIC_PUSH + #define JSON_HEDLEY_DIAGNOSTIC_POP +#endif + +/* JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_ is for + HEDLEY INTERNAL USE ONLY. API subject to change without notice. */ +#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_) + #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_ +#endif +#if defined(__cplusplus) +# if JSON_HEDLEY_HAS_WARNING("-Wc++98-compat") +# if JSON_HEDLEY_HAS_WARNING("-Wc++17-extensions") +# if JSON_HEDLEY_HAS_WARNING("-Wc++1z-extensions") +# define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(xpr) \ + JSON_HEDLEY_DIAGNOSTIC_PUSH \ + _Pragma("clang diagnostic ignored \"-Wc++98-compat\"") \ + _Pragma("clang diagnostic ignored \"-Wc++17-extensions\"") \ + _Pragma("clang diagnostic ignored \"-Wc++1z-extensions\"") \ + xpr \ + JSON_HEDLEY_DIAGNOSTIC_POP +# else +# define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(xpr) \ + JSON_HEDLEY_DIAGNOSTIC_PUSH \ + _Pragma("clang diagnostic ignored \"-Wc++98-compat\"") \ + _Pragma("clang diagnostic ignored \"-Wc++17-extensions\"") \ + xpr \ + JSON_HEDLEY_DIAGNOSTIC_POP +# endif +# else +# define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(xpr) \ + JSON_HEDLEY_DIAGNOSTIC_PUSH \ + _Pragma("clang diagnostic ignored \"-Wc++98-compat\"") \ + xpr \ + JSON_HEDLEY_DIAGNOSTIC_POP +# endif +# endif +#endif +#if !defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(x) x +#endif + +#if defined(JSON_HEDLEY_CONST_CAST) + #undef JSON_HEDLEY_CONST_CAST +#endif +#if defined(__cplusplus) +# define JSON_HEDLEY_CONST_CAST(T, expr) (const_cast(expr)) +#elif \ + JSON_HEDLEY_HAS_WARNING("-Wcast-qual") || \ + JSON_HEDLEY_GCC_VERSION_CHECK(4,6,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) +# define JSON_HEDLEY_CONST_CAST(T, expr) (__extension__ ({ \ + JSON_HEDLEY_DIAGNOSTIC_PUSH \ + JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL \ + ((T) (expr)); \ + JSON_HEDLEY_DIAGNOSTIC_POP \ + })) +#else +# define JSON_HEDLEY_CONST_CAST(T, expr) ((T) (expr)) +#endif + +#if defined(JSON_HEDLEY_REINTERPRET_CAST) + #undef JSON_HEDLEY_REINTERPRET_CAST +#endif +#if defined(__cplusplus) + #define JSON_HEDLEY_REINTERPRET_CAST(T, expr) (reinterpret_cast(expr)) +#else + #define JSON_HEDLEY_REINTERPRET_CAST(T, expr) ((T) (expr)) +#endif + +#if defined(JSON_HEDLEY_STATIC_CAST) + #undef JSON_HEDLEY_STATIC_CAST +#endif +#if defined(__cplusplus) + #define JSON_HEDLEY_STATIC_CAST(T, expr) (static_cast(expr)) +#else + #define JSON_HEDLEY_STATIC_CAST(T, expr) ((T) (expr)) +#endif + +#if defined(JSON_HEDLEY_CPP_CAST) + #undef JSON_HEDLEY_CPP_CAST +#endif +#if defined(__cplusplus) +# if JSON_HEDLEY_HAS_WARNING("-Wold-style-cast") +# define JSON_HEDLEY_CPP_CAST(T, expr) \ + JSON_HEDLEY_DIAGNOSTIC_PUSH \ + _Pragma("clang diagnostic ignored \"-Wold-style-cast\"") \ + ((T) (expr)) \ + JSON_HEDLEY_DIAGNOSTIC_POP +# elif JSON_HEDLEY_IAR_VERSION_CHECK(8,3,0) +# define JSON_HEDLEY_CPP_CAST(T, expr) \ + JSON_HEDLEY_DIAGNOSTIC_PUSH \ + _Pragma("diag_suppress=Pe137") \ + JSON_HEDLEY_DIAGNOSTIC_POP +# else +# define JSON_HEDLEY_CPP_CAST(T, expr) ((T) (expr)) +# endif +#else +# define JSON_HEDLEY_CPP_CAST(T, expr) (expr) +#endif + +#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED) + #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED +#endif +#if JSON_HEDLEY_HAS_WARNING("-Wdeprecated-declarations") + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("clang diagnostic ignored \"-Wdeprecated-declarations\"") +#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("warning(disable:1478 1786)") +#elif JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED __pragma(warning(disable:1478 1786)) +#elif JSON_HEDLEY_PGI_VERSION_CHECK(20,7,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1215,1216,1444,1445") +#elif JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1215,1444") +#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,3,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("GCC diagnostic ignored \"-Wdeprecated-declarations\"") +#elif JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED __pragma(warning(disable:4996)) +#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1215,1444") +#elif \ + JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \ + (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \ + (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \ + (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \ + (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \ + JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1291,1718") +#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,13,0) && !defined(__cplusplus) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("error_messages(off,E_DEPRECATED_ATT,E_DEPRECATED_ATT_MESS)") +#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,13,0) && defined(__cplusplus) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("error_messages(off,symdeprecated,symdeprecated2)") +#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress=Pe1444,Pe1215") +#elif JSON_HEDLEY_PELLES_VERSION_CHECK(2,90,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("warn(disable:2241)") +#else + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED +#endif + +#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS) + #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS +#endif +#if JSON_HEDLEY_HAS_WARNING("-Wunknown-pragmas") + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("clang diagnostic ignored \"-Wunknown-pragmas\"") +#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("warning(disable:161)") +#elif JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS __pragma(warning(disable:161)) +#elif JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 1675") +#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,3,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("GCC diagnostic ignored \"-Wunknown-pragmas\"") +#elif JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS __pragma(warning(disable:4068)) +#elif \ + JSON_HEDLEY_TI_VERSION_CHECK(16,9,0) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,0,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \ + JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,3,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 163") +#elif JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,0,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 163") +#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress=Pe161") +#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 161") +#else + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS +#endif + +#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES) + #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES +#endif +#if JSON_HEDLEY_HAS_WARNING("-Wunknown-attributes") + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("clang diagnostic ignored \"-Wunknown-attributes\"") +#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,6,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("GCC diagnostic ignored \"-Wdeprecated-declarations\"") +#elif JSON_HEDLEY_INTEL_VERSION_CHECK(17,0,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("warning(disable:1292)") +#elif JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES __pragma(warning(disable:1292)) +#elif JSON_HEDLEY_MSVC_VERSION_CHECK(19,0,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES __pragma(warning(disable:5030)) +#elif JSON_HEDLEY_PGI_VERSION_CHECK(20,7,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1097,1098") +#elif JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1097") +#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,14,0) && defined(__cplusplus) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("error_messages(off,attrskipunsup)") +#elif \ + JSON_HEDLEY_TI_VERSION_CHECK(18,1,0) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,3,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1173") +#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress=Pe1097") +#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1097") +#else + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES +#endif + +#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL) + #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL +#endif +#if JSON_HEDLEY_HAS_WARNING("-Wcast-qual") + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL _Pragma("clang diagnostic ignored \"-Wcast-qual\"") +#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL _Pragma("warning(disable:2203 2331)") +#elif JSON_HEDLEY_GCC_VERSION_CHECK(3,0,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL _Pragma("GCC diagnostic ignored \"-Wcast-qual\"") +#else + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL +#endif + +#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION) + #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION +#endif +#if JSON_HEDLEY_HAS_WARNING("-Wunused-function") + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION _Pragma("clang diagnostic ignored \"-Wunused-function\"") +#elif JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION _Pragma("GCC diagnostic ignored \"-Wunused-function\"") +#elif JSON_HEDLEY_MSVC_VERSION_CHECK(1,0,0) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION __pragma(warning(disable:4505)) +#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION _Pragma("diag_suppress 3142") +#else + #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION +#endif + +#if defined(JSON_HEDLEY_DEPRECATED) + #undef JSON_HEDLEY_DEPRECATED +#endif +#if defined(JSON_HEDLEY_DEPRECATED_FOR) + #undef JSON_HEDLEY_DEPRECATED_FOR +#endif +#if \ + JSON_HEDLEY_MSVC_VERSION_CHECK(14,0,0) || \ + JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) + #define JSON_HEDLEY_DEPRECATED(since) __declspec(deprecated("Since " # since)) + #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __declspec(deprecated("Since " #since "; use " #replacement)) +#elif \ + (JSON_HEDLEY_HAS_EXTENSION(attribute_deprecated_with_message) && !defined(JSON_HEDLEY_IAR_VERSION)) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(4,5,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(5,6,0) || \ + JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,13,0) || \ + JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \ + JSON_HEDLEY_TI_VERSION_CHECK(18,1,0) || \ + JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(18,1,0) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,3,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \ + JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,3,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) + #define JSON_HEDLEY_DEPRECATED(since) __attribute__((__deprecated__("Since " #since))) + #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __attribute__((__deprecated__("Since " #since "; use " #replacement))) +#elif defined(__cplusplus) && (__cplusplus >= 201402L) + #define JSON_HEDLEY_DEPRECATED(since) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[deprecated("Since " #since)]]) + #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[deprecated("Since " #since "; use " #replacement)]]) +#elif \ + JSON_HEDLEY_HAS_ATTRIBUTE(deprecated) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \ + JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \ + (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \ + (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \ + (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \ + (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \ + JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) || \ + JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0) + #define JSON_HEDLEY_DEPRECATED(since) __attribute__((__deprecated__)) + #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __attribute__((__deprecated__)) +#elif \ + JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \ + JSON_HEDLEY_PELLES_VERSION_CHECK(6,50,0) || \ + JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) + #define JSON_HEDLEY_DEPRECATED(since) __declspec(deprecated) + #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __declspec(deprecated) +#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0) + #define JSON_HEDLEY_DEPRECATED(since) _Pragma("deprecated") + #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) _Pragma("deprecated") +#else + #define JSON_HEDLEY_DEPRECATED(since) + #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) +#endif + +#if defined(JSON_HEDLEY_UNAVAILABLE) + #undef JSON_HEDLEY_UNAVAILABLE +#endif +#if \ + JSON_HEDLEY_HAS_ATTRIBUTE(warning) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(4,3,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) + #define JSON_HEDLEY_UNAVAILABLE(available_since) __attribute__((__warning__("Not available until " #available_since))) +#else + #define JSON_HEDLEY_UNAVAILABLE(available_since) +#endif + +#if defined(JSON_HEDLEY_WARN_UNUSED_RESULT) + #undef JSON_HEDLEY_WARN_UNUSED_RESULT +#endif +#if defined(JSON_HEDLEY_WARN_UNUSED_RESULT_MSG) + #undef JSON_HEDLEY_WARN_UNUSED_RESULT_MSG +#endif +#if \ + JSON_HEDLEY_HAS_ATTRIBUTE(warn_unused_result) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \ + (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \ + (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \ + (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \ + (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \ + JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \ + (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0) && defined(__cplusplus)) || \ + JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) + #define JSON_HEDLEY_WARN_UNUSED_RESULT __attribute__((__warn_unused_result__)) + #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) __attribute__((__warn_unused_result__)) +#elif (JSON_HEDLEY_HAS_CPP_ATTRIBUTE(nodiscard) >= 201907L) + #define JSON_HEDLEY_WARN_UNUSED_RESULT JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard]]) + #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard(msg)]]) +#elif JSON_HEDLEY_HAS_CPP_ATTRIBUTE(nodiscard) + #define JSON_HEDLEY_WARN_UNUSED_RESULT JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard]]) + #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard]]) +#elif defined(_Check_return_) /* SAL */ + #define JSON_HEDLEY_WARN_UNUSED_RESULT _Check_return_ + #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) _Check_return_ +#else + #define JSON_HEDLEY_WARN_UNUSED_RESULT + #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) +#endif + +#if defined(JSON_HEDLEY_SENTINEL) + #undef JSON_HEDLEY_SENTINEL +#endif +#if \ + JSON_HEDLEY_HAS_ATTRIBUTE(sentinel) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(4,0,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(5,4,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) + #define JSON_HEDLEY_SENTINEL(position) __attribute__((__sentinel__(position))) +#else + #define JSON_HEDLEY_SENTINEL(position) +#endif + +#if defined(JSON_HEDLEY_NO_RETURN) + #undef JSON_HEDLEY_NO_RETURN +#endif +#if JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0) + #define JSON_HEDLEY_NO_RETURN __noreturn +#elif \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) + #define JSON_HEDLEY_NO_RETURN __attribute__((__noreturn__)) +#elif defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L + #define JSON_HEDLEY_NO_RETURN _Noreturn +#elif defined(__cplusplus) && (__cplusplus >= 201103L) + #define JSON_HEDLEY_NO_RETURN JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[noreturn]]) +#elif \ + JSON_HEDLEY_HAS_ATTRIBUTE(noreturn) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(3,2,0) || \ + JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \ + JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \ + JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \ + (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \ + (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \ + (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \ + (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \ + JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \ + JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0) + #define JSON_HEDLEY_NO_RETURN __attribute__((__noreturn__)) +#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0) + #define JSON_HEDLEY_NO_RETURN _Pragma("does_not_return") +#elif \ + JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \ + JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) + #define JSON_HEDLEY_NO_RETURN __declspec(noreturn) +#elif JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,0,0) && defined(__cplusplus) + #define JSON_HEDLEY_NO_RETURN _Pragma("FUNC_NEVER_RETURNS;") +#elif JSON_HEDLEY_COMPCERT_VERSION_CHECK(3,2,0) + #define JSON_HEDLEY_NO_RETURN __attribute((noreturn)) +#elif JSON_HEDLEY_PELLES_VERSION_CHECK(9,0,0) + #define JSON_HEDLEY_NO_RETURN __declspec(noreturn) +#else + #define JSON_HEDLEY_NO_RETURN +#endif + +#if defined(JSON_HEDLEY_NO_ESCAPE) + #undef JSON_HEDLEY_NO_ESCAPE +#endif +#if JSON_HEDLEY_HAS_ATTRIBUTE(noescape) + #define JSON_HEDLEY_NO_ESCAPE __attribute__((__noescape__)) +#else + #define JSON_HEDLEY_NO_ESCAPE +#endif + +#if defined(JSON_HEDLEY_UNREACHABLE) + #undef JSON_HEDLEY_UNREACHABLE +#endif +#if defined(JSON_HEDLEY_UNREACHABLE_RETURN) + #undef JSON_HEDLEY_UNREACHABLE_RETURN +#endif +#if defined(JSON_HEDLEY_ASSUME) + #undef JSON_HEDLEY_ASSUME +#endif +#if \ + JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) + #define JSON_HEDLEY_ASSUME(expr) __assume(expr) +#elif JSON_HEDLEY_HAS_BUILTIN(__builtin_assume) + #define JSON_HEDLEY_ASSUME(expr) __builtin_assume(expr) +#elif \ + JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,0) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(4,0,0) + #if defined(__cplusplus) + #define JSON_HEDLEY_ASSUME(expr) std::_nassert(expr) + #else + #define JSON_HEDLEY_ASSUME(expr) _nassert(expr) + #endif +#endif +#if \ + (JSON_HEDLEY_HAS_BUILTIN(__builtin_unreachable) && (!defined(JSON_HEDLEY_ARM_VERSION))) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(4,5,0) || \ + JSON_HEDLEY_PGI_VERSION_CHECK(18,10,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_IBM_VERSION_CHECK(13,1,5) || \ + JSON_HEDLEY_CRAY_VERSION_CHECK(10,0,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) + #define JSON_HEDLEY_UNREACHABLE() __builtin_unreachable() +#elif defined(JSON_HEDLEY_ASSUME) + #define JSON_HEDLEY_UNREACHABLE() JSON_HEDLEY_ASSUME(0) +#endif +#if !defined(JSON_HEDLEY_ASSUME) + #if defined(JSON_HEDLEY_UNREACHABLE) + #define JSON_HEDLEY_ASSUME(expr) JSON_HEDLEY_STATIC_CAST(void, ((expr) ? 1 : (JSON_HEDLEY_UNREACHABLE(), 1))) + #else + #define JSON_HEDLEY_ASSUME(expr) JSON_HEDLEY_STATIC_CAST(void, expr) + #endif +#endif +#if defined(JSON_HEDLEY_UNREACHABLE) + #if \ + JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,0) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(4,0,0) + #define JSON_HEDLEY_UNREACHABLE_RETURN(value) return (JSON_HEDLEY_STATIC_CAST(void, JSON_HEDLEY_ASSUME(0)), (value)) + #else + #define JSON_HEDLEY_UNREACHABLE_RETURN(value) JSON_HEDLEY_UNREACHABLE() + #endif +#else + #define JSON_HEDLEY_UNREACHABLE_RETURN(value) return (value) +#endif +#if !defined(JSON_HEDLEY_UNREACHABLE) + #define JSON_HEDLEY_UNREACHABLE() JSON_HEDLEY_ASSUME(0) +#endif + +JSON_HEDLEY_DIAGNOSTIC_PUSH +#if JSON_HEDLEY_HAS_WARNING("-Wpedantic") + #pragma clang diagnostic ignored "-Wpedantic" +#endif +#if JSON_HEDLEY_HAS_WARNING("-Wc++98-compat-pedantic") && defined(__cplusplus) + #pragma clang diagnostic ignored "-Wc++98-compat-pedantic" +#endif +#if JSON_HEDLEY_GCC_HAS_WARNING("-Wvariadic-macros",4,0,0) + #if defined(__clang__) + #pragma clang diagnostic ignored "-Wvariadic-macros" + #elif defined(JSON_HEDLEY_GCC_VERSION) + #pragma GCC diagnostic ignored "-Wvariadic-macros" + #endif +#endif +#if defined(JSON_HEDLEY_NON_NULL) + #undef JSON_HEDLEY_NON_NULL +#endif +#if \ + JSON_HEDLEY_HAS_ATTRIBUTE(nonnull) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(3,3,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) + #define JSON_HEDLEY_NON_NULL(...) __attribute__((__nonnull__(__VA_ARGS__))) +#else + #define JSON_HEDLEY_NON_NULL(...) +#endif +JSON_HEDLEY_DIAGNOSTIC_POP + +#if defined(JSON_HEDLEY_PRINTF_FORMAT) + #undef JSON_HEDLEY_PRINTF_FORMAT +#endif +#if defined(__MINGW32__) && JSON_HEDLEY_GCC_HAS_ATTRIBUTE(format,4,4,0) && !defined(__USE_MINGW_ANSI_STDIO) + #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __attribute__((__format__(ms_printf, string_idx, first_to_check))) +#elif defined(__MINGW32__) && JSON_HEDLEY_GCC_HAS_ATTRIBUTE(format,4,4,0) && defined(__USE_MINGW_ANSI_STDIO) + #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __attribute__((__format__(gnu_printf, string_idx, first_to_check))) +#elif \ + JSON_HEDLEY_HAS_ATTRIBUTE(format) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(5,6,0) || \ + JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \ + JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \ + (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \ + (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \ + (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \ + (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \ + JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) + #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __attribute__((__format__(__printf__, string_idx, first_to_check))) +#elif JSON_HEDLEY_PELLES_VERSION_CHECK(6,0,0) + #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __declspec(vaformat(printf,string_idx,first_to_check)) +#else + #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) +#endif + +#if defined(JSON_HEDLEY_CONSTEXPR) + #undef JSON_HEDLEY_CONSTEXPR +#endif +#if defined(__cplusplus) + #if __cplusplus >= 201103L + #define JSON_HEDLEY_CONSTEXPR JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(constexpr) + #endif +#endif +#if !defined(JSON_HEDLEY_CONSTEXPR) + #define JSON_HEDLEY_CONSTEXPR +#endif + +#if defined(JSON_HEDLEY_PREDICT) + #undef JSON_HEDLEY_PREDICT +#endif +#if defined(JSON_HEDLEY_LIKELY) + #undef JSON_HEDLEY_LIKELY +#endif +#if defined(JSON_HEDLEY_UNLIKELY) + #undef JSON_HEDLEY_UNLIKELY +#endif +#if defined(JSON_HEDLEY_UNPREDICTABLE) + #undef JSON_HEDLEY_UNPREDICTABLE +#endif +#if JSON_HEDLEY_HAS_BUILTIN(__builtin_unpredictable) + #define JSON_HEDLEY_UNPREDICTABLE(expr) __builtin_unpredictable((expr)) +#endif +#if \ + (JSON_HEDLEY_HAS_BUILTIN(__builtin_expect_with_probability) && !defined(JSON_HEDLEY_PGI_VERSION)) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(9,0,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) +# define JSON_HEDLEY_PREDICT(expr, value, probability) __builtin_expect_with_probability( (expr), (value), (probability)) +# define JSON_HEDLEY_PREDICT_TRUE(expr, probability) __builtin_expect_with_probability(!!(expr), 1 , (probability)) +# define JSON_HEDLEY_PREDICT_FALSE(expr, probability) __builtin_expect_with_probability(!!(expr), 0 , (probability)) +# define JSON_HEDLEY_LIKELY(expr) __builtin_expect (!!(expr), 1 ) +# define JSON_HEDLEY_UNLIKELY(expr) __builtin_expect (!!(expr), 0 ) +#elif \ + (JSON_HEDLEY_HAS_BUILTIN(__builtin_expect) && !defined(JSON_HEDLEY_INTEL_CL_VERSION)) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(3,0,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0) && defined(__cplusplus)) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \ + JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \ + JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \ + JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,7,0) || \ + JSON_HEDLEY_TI_CL430_VERSION_CHECK(3,1,0) || \ + JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,1,0) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,1,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \ + JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \ + JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,27) || \ + JSON_HEDLEY_CRAY_VERSION_CHECK(8,1,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) +# define JSON_HEDLEY_PREDICT(expr, expected, probability) \ + (((probability) >= 0.9) ? __builtin_expect((expr), (expected)) : (JSON_HEDLEY_STATIC_CAST(void, expected), (expr))) +# define JSON_HEDLEY_PREDICT_TRUE(expr, probability) \ + (__extension__ ({ \ + double hedley_probability_ = (probability); \ + ((hedley_probability_ >= 0.9) ? __builtin_expect(!!(expr), 1) : ((hedley_probability_ <= 0.1) ? __builtin_expect(!!(expr), 0) : !!(expr))); \ + })) +# define JSON_HEDLEY_PREDICT_FALSE(expr, probability) \ + (__extension__ ({ \ + double hedley_probability_ = (probability); \ + ((hedley_probability_ >= 0.9) ? __builtin_expect(!!(expr), 0) : ((hedley_probability_ <= 0.1) ? __builtin_expect(!!(expr), 1) : !!(expr))); \ + })) +# define JSON_HEDLEY_LIKELY(expr) __builtin_expect(!!(expr), 1) +# define JSON_HEDLEY_UNLIKELY(expr) __builtin_expect(!!(expr), 0) +#else +# define JSON_HEDLEY_PREDICT(expr, expected, probability) (JSON_HEDLEY_STATIC_CAST(void, expected), (expr)) +# define JSON_HEDLEY_PREDICT_TRUE(expr, probability) (!!(expr)) +# define JSON_HEDLEY_PREDICT_FALSE(expr, probability) (!!(expr)) +# define JSON_HEDLEY_LIKELY(expr) (!!(expr)) +# define JSON_HEDLEY_UNLIKELY(expr) (!!(expr)) +#endif +#if !defined(JSON_HEDLEY_UNPREDICTABLE) + #define JSON_HEDLEY_UNPREDICTABLE(expr) JSON_HEDLEY_PREDICT(expr, 1, 0.5) +#endif + +#if defined(JSON_HEDLEY_MALLOC) + #undef JSON_HEDLEY_MALLOC +#endif +#if \ + JSON_HEDLEY_HAS_ATTRIBUTE(malloc) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \ + JSON_HEDLEY_IBM_VERSION_CHECK(12,1,0) || \ + JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \ + (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \ + (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \ + (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \ + (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \ + JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) + #define JSON_HEDLEY_MALLOC __attribute__((__malloc__)) +#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0) + #define JSON_HEDLEY_MALLOC _Pragma("returns_new_memory") +#elif \ + JSON_HEDLEY_MSVC_VERSION_CHECK(14,0,0) || \ + JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) + #define JSON_HEDLEY_MALLOC __declspec(restrict) +#else + #define JSON_HEDLEY_MALLOC +#endif + +#if defined(JSON_HEDLEY_PURE) + #undef JSON_HEDLEY_PURE +#endif +#if \ + JSON_HEDLEY_HAS_ATTRIBUTE(pure) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(2,96,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \ + JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \ + JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \ + (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \ + (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \ + (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \ + (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \ + JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \ + JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) +# define JSON_HEDLEY_PURE __attribute__((__pure__)) +#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0) +# define JSON_HEDLEY_PURE _Pragma("does_not_write_global_data") +#elif defined(__cplusplus) && \ + ( \ + JSON_HEDLEY_TI_CL430_VERSION_CHECK(2,0,1) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(4,0,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) \ + ) +# define JSON_HEDLEY_PURE _Pragma("FUNC_IS_PURE;") +#else +# define JSON_HEDLEY_PURE +#endif + +#if defined(JSON_HEDLEY_CONST) + #undef JSON_HEDLEY_CONST +#endif +#if \ + JSON_HEDLEY_HAS_ATTRIBUTE(const) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(2,5,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \ + JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \ + JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \ + (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \ + (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \ + (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \ + (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \ + JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \ + JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) + #define JSON_HEDLEY_CONST __attribute__((__const__)) +#elif \ + JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0) + #define JSON_HEDLEY_CONST _Pragma("no_side_effect") +#else + #define JSON_HEDLEY_CONST JSON_HEDLEY_PURE +#endif + +#if defined(JSON_HEDLEY_RESTRICT) + #undef JSON_HEDLEY_RESTRICT +#endif +#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) && !defined(__cplusplus) + #define JSON_HEDLEY_RESTRICT restrict +#elif \ + JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \ + JSON_HEDLEY_MSVC_VERSION_CHECK(14,0,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \ + JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \ + JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \ + JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \ + JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,4) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,1,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \ + (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,14,0) && defined(__cplusplus)) || \ + JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0) || \ + defined(__clang__) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) + #define JSON_HEDLEY_RESTRICT __restrict +#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,3,0) && !defined(__cplusplus) + #define JSON_HEDLEY_RESTRICT _Restrict +#else + #define JSON_HEDLEY_RESTRICT +#endif + +#if defined(JSON_HEDLEY_INLINE) + #undef JSON_HEDLEY_INLINE +#endif +#if \ + (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L)) || \ + (defined(__cplusplus) && (__cplusplus >= 199711L)) + #define JSON_HEDLEY_INLINE inline +#elif \ + defined(JSON_HEDLEY_GCC_VERSION) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(6,2,0) + #define JSON_HEDLEY_INLINE __inline__ +#elif \ + JSON_HEDLEY_MSVC_VERSION_CHECK(12,0,0) || \ + JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \ + JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,1,0) || \ + JSON_HEDLEY_TI_CL430_VERSION_CHECK(3,1,0) || \ + JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,0) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,0,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \ + JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) + #define JSON_HEDLEY_INLINE __inline +#else + #define JSON_HEDLEY_INLINE +#endif + +#if defined(JSON_HEDLEY_ALWAYS_INLINE) + #undef JSON_HEDLEY_ALWAYS_INLINE +#endif +#if \ + JSON_HEDLEY_HAS_ATTRIBUTE(always_inline) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(4,0,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \ + JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \ + JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \ + (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \ + (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \ + (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \ + (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \ + JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) || \ + JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0) +# define JSON_HEDLEY_ALWAYS_INLINE __attribute__((__always_inline__)) JSON_HEDLEY_INLINE +#elif \ + JSON_HEDLEY_MSVC_VERSION_CHECK(12,0,0) || \ + JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) +# define JSON_HEDLEY_ALWAYS_INLINE __forceinline +#elif defined(__cplusplus) && \ + ( \ + JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \ + JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \ + JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,1,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \ + JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) \ + ) +# define JSON_HEDLEY_ALWAYS_INLINE _Pragma("FUNC_ALWAYS_INLINE;") +#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0) +# define JSON_HEDLEY_ALWAYS_INLINE _Pragma("inline=forced") +#else +# define JSON_HEDLEY_ALWAYS_INLINE JSON_HEDLEY_INLINE +#endif + +#if defined(JSON_HEDLEY_NEVER_INLINE) + #undef JSON_HEDLEY_NEVER_INLINE +#endif +#if \ + JSON_HEDLEY_HAS_ATTRIBUTE(noinline) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(4,0,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \ + JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \ + JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \ + (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \ + (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \ + (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \ + (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \ + JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \ + JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) || \ + JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0) + #define JSON_HEDLEY_NEVER_INLINE __attribute__((__noinline__)) +#elif \ + JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \ + JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) + #define JSON_HEDLEY_NEVER_INLINE __declspec(noinline) +#elif JSON_HEDLEY_PGI_VERSION_CHECK(10,2,0) + #define JSON_HEDLEY_NEVER_INLINE _Pragma("noinline") +#elif JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,0,0) && defined(__cplusplus) + #define JSON_HEDLEY_NEVER_INLINE _Pragma("FUNC_CANNOT_INLINE;") +#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0) + #define JSON_HEDLEY_NEVER_INLINE _Pragma("inline=never") +#elif JSON_HEDLEY_COMPCERT_VERSION_CHECK(3,2,0) + #define JSON_HEDLEY_NEVER_INLINE __attribute((noinline)) +#elif JSON_HEDLEY_PELLES_VERSION_CHECK(9,0,0) + #define JSON_HEDLEY_NEVER_INLINE __declspec(noinline) +#else + #define JSON_HEDLEY_NEVER_INLINE +#endif + +#if defined(JSON_HEDLEY_PRIVATE) + #undef JSON_HEDLEY_PRIVATE +#endif +#if defined(JSON_HEDLEY_PUBLIC) + #undef JSON_HEDLEY_PUBLIC +#endif +#if defined(JSON_HEDLEY_IMPORT) + #undef JSON_HEDLEY_IMPORT +#endif +#if defined(_WIN32) || defined(__CYGWIN__) +# define JSON_HEDLEY_PRIVATE +# define JSON_HEDLEY_PUBLIC __declspec(dllexport) +# define JSON_HEDLEY_IMPORT __declspec(dllimport) +#else +# if \ + JSON_HEDLEY_HAS_ATTRIBUTE(visibility) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(3,3,0) || \ + JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \ + JSON_HEDLEY_IBM_VERSION_CHECK(13,1,0) || \ + ( \ + defined(__TI_EABI__) && \ + ( \ + (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) \ + ) \ + ) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) +# define JSON_HEDLEY_PRIVATE __attribute__((__visibility__("hidden"))) +# define JSON_HEDLEY_PUBLIC __attribute__((__visibility__("default"))) +# else +# define JSON_HEDLEY_PRIVATE +# define JSON_HEDLEY_PUBLIC +# endif +# define JSON_HEDLEY_IMPORT extern +#endif + +#if defined(JSON_HEDLEY_NO_THROW) + #undef JSON_HEDLEY_NO_THROW +#endif +#if \ + JSON_HEDLEY_HAS_ATTRIBUTE(nothrow) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(3,3,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) + #define JSON_HEDLEY_NO_THROW __attribute__((__nothrow__)) +#elif \ + JSON_HEDLEY_MSVC_VERSION_CHECK(13,1,0) || \ + JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) + #define JSON_HEDLEY_NO_THROW __declspec(nothrow) +#else + #define JSON_HEDLEY_NO_THROW +#endif + +#if defined(JSON_HEDLEY_FALL_THROUGH) + #undef JSON_HEDLEY_FALL_THROUGH +#endif +#if \ + JSON_HEDLEY_HAS_ATTRIBUTE(fallthrough) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(7,0,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) + #define JSON_HEDLEY_FALL_THROUGH __attribute__((__fallthrough__)) +#elif JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(clang,fallthrough) + #define JSON_HEDLEY_FALL_THROUGH JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[clang::fallthrough]]) +#elif JSON_HEDLEY_HAS_CPP_ATTRIBUTE(fallthrough) + #define JSON_HEDLEY_FALL_THROUGH JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[fallthrough]]) +#elif defined(__fallthrough) /* SAL */ + #define JSON_HEDLEY_FALL_THROUGH __fallthrough +#else + #define JSON_HEDLEY_FALL_THROUGH +#endif + +#if defined(JSON_HEDLEY_RETURNS_NON_NULL) + #undef JSON_HEDLEY_RETURNS_NON_NULL +#endif +#if \ + JSON_HEDLEY_HAS_ATTRIBUTE(returns_nonnull) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(4,9,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) + #define JSON_HEDLEY_RETURNS_NON_NULL __attribute__((__returns_nonnull__)) +#elif defined(_Ret_notnull_) /* SAL */ + #define JSON_HEDLEY_RETURNS_NON_NULL _Ret_notnull_ +#else + #define JSON_HEDLEY_RETURNS_NON_NULL +#endif + +#if defined(JSON_HEDLEY_ARRAY_PARAM) + #undef JSON_HEDLEY_ARRAY_PARAM +#endif +#if \ + defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) && \ + !defined(__STDC_NO_VLA__) && \ + !defined(__cplusplus) && \ + !defined(JSON_HEDLEY_PGI_VERSION) && \ + !defined(JSON_HEDLEY_TINYC_VERSION) + #define JSON_HEDLEY_ARRAY_PARAM(name) (name) +#else + #define JSON_HEDLEY_ARRAY_PARAM(name) +#endif + +#if defined(JSON_HEDLEY_IS_CONSTANT) + #undef JSON_HEDLEY_IS_CONSTANT +#endif +#if defined(JSON_HEDLEY_REQUIRE_CONSTEXPR) + #undef JSON_HEDLEY_REQUIRE_CONSTEXPR +#endif +/* JSON_HEDLEY_IS_CONSTEXPR_ is for + HEDLEY INTERNAL USE ONLY. API subject to change without notice. */ +#if defined(JSON_HEDLEY_IS_CONSTEXPR_) + #undef JSON_HEDLEY_IS_CONSTEXPR_ +#endif +#if \ + JSON_HEDLEY_HAS_BUILTIN(__builtin_constant_p) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,19) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \ + JSON_HEDLEY_IBM_VERSION_CHECK(13,1,0) || \ + JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,1,0) || \ + (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0) && !defined(__cplusplus)) || \ + JSON_HEDLEY_CRAY_VERSION_CHECK(8,1,0) || \ + JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) + #define JSON_HEDLEY_IS_CONSTANT(expr) __builtin_constant_p(expr) +#endif +#if !defined(__cplusplus) +# if \ + JSON_HEDLEY_HAS_BUILTIN(__builtin_types_compatible_p) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + JSON_HEDLEY_IBM_VERSION_CHECK(13,1,0) || \ + JSON_HEDLEY_CRAY_VERSION_CHECK(8,1,0) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(5,4,0) || \ + JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,24) +#if defined(__INTPTR_TYPE__) + #define JSON_HEDLEY_IS_CONSTEXPR_(expr) __builtin_types_compatible_p(__typeof__((1 ? (void*) ((__INTPTR_TYPE__) ((expr) * 0)) : (int*) 0)), int*) +#else + #include + #define JSON_HEDLEY_IS_CONSTEXPR_(expr) __builtin_types_compatible_p(__typeof__((1 ? (void*) ((intptr_t) ((expr) * 0)) : (int*) 0)), int*) +#endif +# elif \ + ( \ + defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) && \ + !defined(JSON_HEDLEY_SUNPRO_VERSION) && \ + !defined(JSON_HEDLEY_PGI_VERSION) && \ + !defined(JSON_HEDLEY_IAR_VERSION)) || \ + (JSON_HEDLEY_HAS_EXTENSION(c_generic_selections) && !defined(JSON_HEDLEY_IAR_VERSION)) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(4,9,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(17,0,0) || \ + JSON_HEDLEY_IBM_VERSION_CHECK(12,1,0) || \ + JSON_HEDLEY_ARM_VERSION_CHECK(5,3,0) +#if defined(__INTPTR_TYPE__) + #define JSON_HEDLEY_IS_CONSTEXPR_(expr) _Generic((1 ? (void*) ((__INTPTR_TYPE__) ((expr) * 0)) : (int*) 0), int*: 1, void*: 0) +#else + #include + #define JSON_HEDLEY_IS_CONSTEXPR_(expr) _Generic((1 ? (void*) ((intptr_t) * 0) : (int*) 0), int*: 1, void*: 0) +#endif +# elif \ + defined(JSON_HEDLEY_GCC_VERSION) || \ + defined(JSON_HEDLEY_INTEL_VERSION) || \ + defined(JSON_HEDLEY_TINYC_VERSION) || \ + defined(JSON_HEDLEY_TI_ARMCL_VERSION) || \ + JSON_HEDLEY_TI_CL430_VERSION_CHECK(18,12,0) || \ + defined(JSON_HEDLEY_TI_CL2000_VERSION) || \ + defined(JSON_HEDLEY_TI_CL6X_VERSION) || \ + defined(JSON_HEDLEY_TI_CL7X_VERSION) || \ + defined(JSON_HEDLEY_TI_CLPRU_VERSION) || \ + defined(__clang__) +# define JSON_HEDLEY_IS_CONSTEXPR_(expr) ( \ + sizeof(void) != \ + sizeof(*( \ + 1 ? \ + ((void*) ((expr) * 0L) ) : \ +((struct { char v[sizeof(void) * 2]; } *) 1) \ + ) \ + ) \ + ) +# endif +#endif +#if defined(JSON_HEDLEY_IS_CONSTEXPR_) + #if !defined(JSON_HEDLEY_IS_CONSTANT) + #define JSON_HEDLEY_IS_CONSTANT(expr) JSON_HEDLEY_IS_CONSTEXPR_(expr) + #endif + #define JSON_HEDLEY_REQUIRE_CONSTEXPR(expr) (JSON_HEDLEY_IS_CONSTEXPR_(expr) ? (expr) : (-1)) +#else + #if !defined(JSON_HEDLEY_IS_CONSTANT) + #define JSON_HEDLEY_IS_CONSTANT(expr) (0) + #endif + #define JSON_HEDLEY_REQUIRE_CONSTEXPR(expr) (expr) +#endif + +#if defined(JSON_HEDLEY_BEGIN_C_DECLS) + #undef JSON_HEDLEY_BEGIN_C_DECLS +#endif +#if defined(JSON_HEDLEY_END_C_DECLS) + #undef JSON_HEDLEY_END_C_DECLS +#endif +#if defined(JSON_HEDLEY_C_DECL) + #undef JSON_HEDLEY_C_DECL +#endif +#if defined(__cplusplus) + #define JSON_HEDLEY_BEGIN_C_DECLS extern "C" { + #define JSON_HEDLEY_END_C_DECLS } + #define JSON_HEDLEY_C_DECL extern "C" +#else + #define JSON_HEDLEY_BEGIN_C_DECLS + #define JSON_HEDLEY_END_C_DECLS + #define JSON_HEDLEY_C_DECL +#endif + +#if defined(JSON_HEDLEY_STATIC_ASSERT) + #undef JSON_HEDLEY_STATIC_ASSERT +#endif +#if \ + !defined(__cplusplus) && ( \ + (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)) || \ + (JSON_HEDLEY_HAS_FEATURE(c_static_assert) && !defined(JSON_HEDLEY_INTEL_CL_VERSION)) || \ + JSON_HEDLEY_GCC_VERSION_CHECK(6,0,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \ + defined(_Static_assert) \ + ) +# define JSON_HEDLEY_STATIC_ASSERT(expr, message) _Static_assert(expr, message) +#elif \ + (defined(__cplusplus) && (__cplusplus >= 201103L)) || \ + JSON_HEDLEY_MSVC_VERSION_CHECK(16,0,0) || \ + JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) +# define JSON_HEDLEY_STATIC_ASSERT(expr, message) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(static_assert(expr, message)) +#else +# define JSON_HEDLEY_STATIC_ASSERT(expr, message) +#endif + +#if defined(JSON_HEDLEY_NULL) + #undef JSON_HEDLEY_NULL +#endif +#if defined(__cplusplus) + #if __cplusplus >= 201103L + #define JSON_HEDLEY_NULL JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(nullptr) + #elif defined(NULL) + #define JSON_HEDLEY_NULL NULL + #else + #define JSON_HEDLEY_NULL JSON_HEDLEY_STATIC_CAST(void*, 0) + #endif +#elif defined(NULL) + #define JSON_HEDLEY_NULL NULL +#else + #define JSON_HEDLEY_NULL ((void*) 0) +#endif + +#if defined(JSON_HEDLEY_MESSAGE) + #undef JSON_HEDLEY_MESSAGE +#endif +#if JSON_HEDLEY_HAS_WARNING("-Wunknown-pragmas") +# define JSON_HEDLEY_MESSAGE(msg) \ + JSON_HEDLEY_DIAGNOSTIC_PUSH \ + JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS \ + JSON_HEDLEY_PRAGMA(message msg) \ + JSON_HEDLEY_DIAGNOSTIC_POP +#elif \ + JSON_HEDLEY_GCC_VERSION_CHECK(4,4,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) +# define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(message msg) +#elif JSON_HEDLEY_CRAY_VERSION_CHECK(5,0,0) +# define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(_CRI message msg) +#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0) +# define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(message(msg)) +#elif JSON_HEDLEY_PELLES_VERSION_CHECK(2,0,0) +# define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(message(msg)) +#else +# define JSON_HEDLEY_MESSAGE(msg) +#endif + +#if defined(JSON_HEDLEY_WARNING) + #undef JSON_HEDLEY_WARNING +#endif +#if JSON_HEDLEY_HAS_WARNING("-Wunknown-pragmas") +# define JSON_HEDLEY_WARNING(msg) \ + JSON_HEDLEY_DIAGNOSTIC_PUSH \ + JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS \ + JSON_HEDLEY_PRAGMA(clang warning msg) \ + JSON_HEDLEY_DIAGNOSTIC_POP +#elif \ + JSON_HEDLEY_GCC_VERSION_CHECK(4,8,0) || \ + JSON_HEDLEY_PGI_VERSION_CHECK(18,4,0) || \ + JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) +# define JSON_HEDLEY_WARNING(msg) JSON_HEDLEY_PRAGMA(GCC warning msg) +#elif \ + JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0) || \ + JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) +# define JSON_HEDLEY_WARNING(msg) JSON_HEDLEY_PRAGMA(message(msg)) +#else +# define JSON_HEDLEY_WARNING(msg) JSON_HEDLEY_MESSAGE(msg) +#endif + +#if defined(JSON_HEDLEY_REQUIRE) + #undef JSON_HEDLEY_REQUIRE +#endif +#if defined(JSON_HEDLEY_REQUIRE_MSG) + #undef JSON_HEDLEY_REQUIRE_MSG +#endif +#if JSON_HEDLEY_HAS_ATTRIBUTE(diagnose_if) +# if JSON_HEDLEY_HAS_WARNING("-Wgcc-compat") +# define JSON_HEDLEY_REQUIRE(expr) \ + JSON_HEDLEY_DIAGNOSTIC_PUSH \ + _Pragma("clang diagnostic ignored \"-Wgcc-compat\"") \ + __attribute__((diagnose_if(!(expr), #expr, "error"))) \ + JSON_HEDLEY_DIAGNOSTIC_POP +# define JSON_HEDLEY_REQUIRE_MSG(expr,msg) \ + JSON_HEDLEY_DIAGNOSTIC_PUSH \ + _Pragma("clang diagnostic ignored \"-Wgcc-compat\"") \ + __attribute__((diagnose_if(!(expr), msg, "error"))) \ + JSON_HEDLEY_DIAGNOSTIC_POP +# else +# define JSON_HEDLEY_REQUIRE(expr) __attribute__((diagnose_if(!(expr), #expr, "error"))) +# define JSON_HEDLEY_REQUIRE_MSG(expr,msg) __attribute__((diagnose_if(!(expr), msg, "error"))) +# endif +#else +# define JSON_HEDLEY_REQUIRE(expr) +# define JSON_HEDLEY_REQUIRE_MSG(expr,msg) +#endif + +#if defined(JSON_HEDLEY_FLAGS) + #undef JSON_HEDLEY_FLAGS +#endif +#if JSON_HEDLEY_HAS_ATTRIBUTE(flag_enum) && (!defined(__cplusplus) || JSON_HEDLEY_HAS_WARNING("-Wbitfield-enum-conversion")) + #define JSON_HEDLEY_FLAGS __attribute__((__flag_enum__)) +#else + #define JSON_HEDLEY_FLAGS +#endif + +#if defined(JSON_HEDLEY_FLAGS_CAST) + #undef JSON_HEDLEY_FLAGS_CAST +#endif +#if JSON_HEDLEY_INTEL_VERSION_CHECK(19,0,0) +# define JSON_HEDLEY_FLAGS_CAST(T, expr) (__extension__ ({ \ + JSON_HEDLEY_DIAGNOSTIC_PUSH \ + _Pragma("warning(disable:188)") \ + ((T) (expr)); \ + JSON_HEDLEY_DIAGNOSTIC_POP \ + })) +#else +# define JSON_HEDLEY_FLAGS_CAST(T, expr) JSON_HEDLEY_STATIC_CAST(T, expr) +#endif + +#if defined(JSON_HEDLEY_EMPTY_BASES) + #undef JSON_HEDLEY_EMPTY_BASES +#endif +#if \ + (JSON_HEDLEY_MSVC_VERSION_CHECK(19,0,23918) && !JSON_HEDLEY_MSVC_VERSION_CHECK(20,0,0)) || \ + JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) + #define JSON_HEDLEY_EMPTY_BASES __declspec(empty_bases) +#else + #define JSON_HEDLEY_EMPTY_BASES +#endif + +/* Remaining macros are deprecated. */ + +#if defined(JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK) + #undef JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK +#endif +#if defined(__clang__) + #define JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK(major,minor,patch) (0) +#else + #define JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK(major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch) +#endif + +#if defined(JSON_HEDLEY_CLANG_HAS_ATTRIBUTE) + #undef JSON_HEDLEY_CLANG_HAS_ATTRIBUTE +#endif +#define JSON_HEDLEY_CLANG_HAS_ATTRIBUTE(attribute) JSON_HEDLEY_HAS_ATTRIBUTE(attribute) + +#if defined(JSON_HEDLEY_CLANG_HAS_CPP_ATTRIBUTE) + #undef JSON_HEDLEY_CLANG_HAS_CPP_ATTRIBUTE +#endif +#define JSON_HEDLEY_CLANG_HAS_CPP_ATTRIBUTE(attribute) JSON_HEDLEY_HAS_CPP_ATTRIBUTE(attribute) + +#if defined(JSON_HEDLEY_CLANG_HAS_BUILTIN) + #undef JSON_HEDLEY_CLANG_HAS_BUILTIN +#endif +#define JSON_HEDLEY_CLANG_HAS_BUILTIN(builtin) JSON_HEDLEY_HAS_BUILTIN(builtin) + +#if defined(JSON_HEDLEY_CLANG_HAS_FEATURE) + #undef JSON_HEDLEY_CLANG_HAS_FEATURE +#endif +#define JSON_HEDLEY_CLANG_HAS_FEATURE(feature) JSON_HEDLEY_HAS_FEATURE(feature) + +#if defined(JSON_HEDLEY_CLANG_HAS_EXTENSION) + #undef JSON_HEDLEY_CLANG_HAS_EXTENSION +#endif +#define JSON_HEDLEY_CLANG_HAS_EXTENSION(extension) JSON_HEDLEY_HAS_EXTENSION(extension) + +#if defined(JSON_HEDLEY_CLANG_HAS_DECLSPEC_DECLSPEC_ATTRIBUTE) + #undef JSON_HEDLEY_CLANG_HAS_DECLSPEC_DECLSPEC_ATTRIBUTE +#endif +#define JSON_HEDLEY_CLANG_HAS_DECLSPEC_ATTRIBUTE(attribute) JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE(attribute) + +#if defined(JSON_HEDLEY_CLANG_HAS_WARNING) + #undef JSON_HEDLEY_CLANG_HAS_WARNING +#endif +#define JSON_HEDLEY_CLANG_HAS_WARNING(warning) JSON_HEDLEY_HAS_WARNING(warning) + +#endif /* !defined(JSON_HEDLEY_VERSION) || (JSON_HEDLEY_VERSION < X) */ + + +// This file contains all internal macro definitions (except those affecting ABI) +// You MUST include macro_unscope.hpp at the end of json.hpp to undef all of them + +// #include + + +// exclude unsupported compilers +#if !defined(JSON_SKIP_UNSUPPORTED_COMPILER_CHECK) + #if defined(__clang__) + #if (__clang_major__ * 10000 + __clang_minor__ * 100 + __clang_patchlevel__) < 30400 + #error "unsupported Clang version - see https://github.com/nlohmann/json#supported-compilers" + #endif + #elif defined(__GNUC__) && !(defined(__ICC) || defined(__INTEL_COMPILER)) + #if (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__) < 40800 + #error "unsupported GCC version - see https://github.com/nlohmann/json#supported-compilers" + #endif + #endif +#endif + +// C++ language standard detection +// if the user manually specified the used c++ version this is skipped +#if !defined(JSON_HAS_CPP_23) && !defined(JSON_HAS_CPP_20) && !defined(JSON_HAS_CPP_17) && !defined(JSON_HAS_CPP_14) && !defined(JSON_HAS_CPP_11) + #if (defined(__cplusplus) && __cplusplus > 202002L) || (defined(_MSVC_LANG) && _MSVC_LANG > 202002L) + #define JSON_HAS_CPP_23 + #define JSON_HAS_CPP_20 + #define JSON_HAS_CPP_17 + #define JSON_HAS_CPP_14 + #elif (defined(__cplusplus) && __cplusplus > 201703L) || (defined(_MSVC_LANG) && _MSVC_LANG > 201703L) + #define JSON_HAS_CPP_20 + #define JSON_HAS_CPP_17 + #define JSON_HAS_CPP_14 + #elif (defined(__cplusplus) && __cplusplus > 201402L) || (defined(_HAS_CXX17) && _HAS_CXX17 == 1) // fix for issue #464 + #define JSON_HAS_CPP_17 + #define JSON_HAS_CPP_14 + #elif (defined(__cplusplus) && __cplusplus > 201103L) || (defined(_HAS_CXX14) && _HAS_CXX14 == 1) + #define JSON_HAS_CPP_14 + #endif + // the cpp 11 flag is always specified because it is the minimal required version + #define JSON_HAS_CPP_11 +#endif + +#ifdef __has_include + #if __has_include() + #include + #endif +#endif + +#if !defined(JSON_HAS_FILESYSTEM) && !defined(JSON_HAS_EXPERIMENTAL_FILESYSTEM) + #ifdef JSON_HAS_CPP_17 + #if defined(__cpp_lib_filesystem) + #define JSON_HAS_FILESYSTEM 1 + #elif defined(__cpp_lib_experimental_filesystem) + #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 1 + #elif !defined(__has_include) + #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 1 + #elif __has_include() + #define JSON_HAS_FILESYSTEM 1 + #elif __has_include() + #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 1 + #endif + + // std::filesystem does not work on MinGW GCC 8: https://sourceforge.net/p/mingw-w64/bugs/737/ + #if defined(__MINGW32__) && defined(__GNUC__) && __GNUC__ == 8 + #undef JSON_HAS_FILESYSTEM + #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM + #endif + + // no filesystem support before GCC 8: https://en.cppreference.com/w/cpp/compiler_support + #if defined(__GNUC__) && !defined(__clang__) && __GNUC__ < 8 + #undef JSON_HAS_FILESYSTEM + #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM + #endif + + // no filesystem support before Clang 7: https://en.cppreference.com/w/cpp/compiler_support + #if defined(__clang_major__) && __clang_major__ < 7 + #undef JSON_HAS_FILESYSTEM + #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM + #endif + + // no filesystem support before MSVC 19.14: https://en.cppreference.com/w/cpp/compiler_support + #if defined(_MSC_VER) && _MSC_VER < 1914 + #undef JSON_HAS_FILESYSTEM + #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM + #endif + + // no filesystem support before iOS 13 + #if defined(__IPHONE_OS_VERSION_MIN_REQUIRED) && __IPHONE_OS_VERSION_MIN_REQUIRED < 130000 + #undef JSON_HAS_FILESYSTEM + #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM + #endif + + // no filesystem support before macOS Catalina + #if defined(__MAC_OS_X_VERSION_MIN_REQUIRED) && __MAC_OS_X_VERSION_MIN_REQUIRED < 101500 + #undef JSON_HAS_FILESYSTEM + #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM + #endif + #endif +#endif + +#ifndef JSON_HAS_EXPERIMENTAL_FILESYSTEM + #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 0 +#endif + +#ifndef JSON_HAS_FILESYSTEM + #define JSON_HAS_FILESYSTEM 0 +#endif + +#ifndef JSON_HAS_THREE_WAY_COMPARISON + #if defined(__cpp_impl_three_way_comparison) && __cpp_impl_three_way_comparison >= 201907L \ + && defined(__cpp_lib_three_way_comparison) && __cpp_lib_three_way_comparison >= 201907L + #define JSON_HAS_THREE_WAY_COMPARISON 1 + #else + #define JSON_HAS_THREE_WAY_COMPARISON 0 + #endif +#endif + +#ifndef JSON_HAS_RANGES + // ranges header shipping in GCC 11.1.0 (released 2021-04-27) has syntax error + #if defined(__GLIBCXX__) && __GLIBCXX__ == 20210427 + #define JSON_HAS_RANGES 0 + #elif defined(__cpp_lib_ranges) + #define JSON_HAS_RANGES 1 + #else + #define JSON_HAS_RANGES 0 + #endif +#endif + +#ifndef JSON_HAS_STATIC_RTTI + #if !defined(_HAS_STATIC_RTTI) || _HAS_STATIC_RTTI != 0 + #define JSON_HAS_STATIC_RTTI 1 + #else + #define JSON_HAS_STATIC_RTTI 0 + #endif +#endif + +#ifdef JSON_HAS_CPP_17 + #define JSON_INLINE_VARIABLE inline +#else + #define JSON_INLINE_VARIABLE +#endif + +#if JSON_HEDLEY_HAS_ATTRIBUTE(no_unique_address) + #define JSON_NO_UNIQUE_ADDRESS [[no_unique_address]] +#else + #define JSON_NO_UNIQUE_ADDRESS +#endif + +// disable documentation warnings on clang +#if defined(__clang__) + #pragma clang diagnostic push + #pragma clang diagnostic ignored "-Wdocumentation" + #pragma clang diagnostic ignored "-Wdocumentation-unknown-command" +#endif + +// allow disabling exceptions +#if (defined(__cpp_exceptions) || defined(__EXCEPTIONS) || defined(_CPPUNWIND)) && !defined(JSON_NOEXCEPTION) + #define JSON_THROW(exception) throw exception + #define JSON_TRY try + #define JSON_CATCH(exception) catch(exception) + #define JSON_INTERNAL_CATCH(exception) catch(exception) +#else + #include + #define JSON_THROW(exception) std::abort() + #define JSON_TRY if(true) + #define JSON_CATCH(exception) if(false) + #define JSON_INTERNAL_CATCH(exception) if(false) +#endif + +// override exception macros +#if defined(JSON_THROW_USER) + #undef JSON_THROW + #define JSON_THROW JSON_THROW_USER +#endif +#if defined(JSON_TRY_USER) + #undef JSON_TRY + #define JSON_TRY JSON_TRY_USER +#endif +#if defined(JSON_CATCH_USER) + #undef JSON_CATCH + #define JSON_CATCH JSON_CATCH_USER + #undef JSON_INTERNAL_CATCH + #define JSON_INTERNAL_CATCH JSON_CATCH_USER +#endif +#if defined(JSON_INTERNAL_CATCH_USER) + #undef JSON_INTERNAL_CATCH + #define JSON_INTERNAL_CATCH JSON_INTERNAL_CATCH_USER +#endif + +// allow overriding assert +#if !defined(JSON_ASSERT) + #include // assert + #define JSON_ASSERT(x) assert(x) +#endif + +// allow to access some private functions (needed by the test suite) +#if defined(JSON_TESTS_PRIVATE) + #define JSON_PRIVATE_UNLESS_TESTED public +#else + #define JSON_PRIVATE_UNLESS_TESTED private +#endif + +/*! +@brief macro to briefly define a mapping between an enum and JSON +@def NLOHMANN_JSON_SERIALIZE_ENUM +@since version 3.4.0 +*/ +#define NLOHMANN_JSON_SERIALIZE_ENUM(ENUM_TYPE, ...) \ + template \ + inline void to_json(BasicJsonType& j, const ENUM_TYPE& e) \ + { \ + /* NOLINTNEXTLINE(modernize-type-traits) we use C++11 */ \ + static_assert(std::is_enum::value, #ENUM_TYPE " must be an enum!"); \ + /* NOLINTNEXTLINE(modernize-avoid-c-arrays) we don't want to depend on */ \ + static const std::pair m[] = __VA_ARGS__; \ + auto it = std::find_if(std::begin(m), std::end(m), \ + [e](const std::pair& ej_pair) -> bool \ + { \ + return ej_pair.first == e; \ + }); \ + j = ((it != std::end(m)) ? it : std::begin(m))->second; \ + } \ + template \ + inline void from_json(const BasicJsonType& j, ENUM_TYPE& e) \ + { \ + /* NOLINTNEXTLINE(modernize-type-traits) we use C++11 */ \ + static_assert(std::is_enum::value, #ENUM_TYPE " must be an enum!"); \ + /* NOLINTNEXTLINE(modernize-avoid-c-arrays) we don't want to depend on */ \ + static const std::pair m[] = __VA_ARGS__; \ + auto it = std::find_if(std::begin(m), std::end(m), \ + [&j](const std::pair& ej_pair) -> bool \ + { \ + return ej_pair.second == j; \ + }); \ + e = ((it != std::end(m)) ? it : std::begin(m))->first; \ + } + +// Ugly macros to avoid uglier copy-paste when specializing basic_json. They +// may be removed in the future once the class is split. + +#define NLOHMANN_BASIC_JSON_TPL_DECLARATION \ + template class ObjectType, \ + template class ArrayType, \ + class StringType, class BooleanType, class NumberIntegerType, \ + class NumberUnsignedType, class NumberFloatType, \ + template class AllocatorType, \ + template class JSONSerializer, \ + class BinaryType, \ + class CustomBaseClass> + +#define NLOHMANN_BASIC_JSON_TPL \ + basic_json + +// Macros to simplify conversion from/to types + +#define NLOHMANN_JSON_EXPAND( x ) x +#define NLOHMANN_JSON_GET_MACRO(_1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48, _49, _50, _51, _52, _53, _54, _55, _56, _57, _58, _59, _60, _61, _62, _63, _64, NAME,...) NAME +#define NLOHMANN_JSON_PASTE(...) NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_GET_MACRO(__VA_ARGS__, \ + NLOHMANN_JSON_PASTE64, \ + NLOHMANN_JSON_PASTE63, \ + NLOHMANN_JSON_PASTE62, \ + NLOHMANN_JSON_PASTE61, \ + NLOHMANN_JSON_PASTE60, \ + NLOHMANN_JSON_PASTE59, \ + NLOHMANN_JSON_PASTE58, \ + NLOHMANN_JSON_PASTE57, \ + NLOHMANN_JSON_PASTE56, \ + NLOHMANN_JSON_PASTE55, \ + NLOHMANN_JSON_PASTE54, \ + NLOHMANN_JSON_PASTE53, \ + NLOHMANN_JSON_PASTE52, \ + NLOHMANN_JSON_PASTE51, \ + NLOHMANN_JSON_PASTE50, \ + NLOHMANN_JSON_PASTE49, \ + NLOHMANN_JSON_PASTE48, \ + NLOHMANN_JSON_PASTE47, \ + NLOHMANN_JSON_PASTE46, \ + NLOHMANN_JSON_PASTE45, \ + NLOHMANN_JSON_PASTE44, \ + NLOHMANN_JSON_PASTE43, \ + NLOHMANN_JSON_PASTE42, \ + NLOHMANN_JSON_PASTE41, \ + NLOHMANN_JSON_PASTE40, \ + NLOHMANN_JSON_PASTE39, \ + NLOHMANN_JSON_PASTE38, \ + NLOHMANN_JSON_PASTE37, \ + NLOHMANN_JSON_PASTE36, \ + NLOHMANN_JSON_PASTE35, \ + NLOHMANN_JSON_PASTE34, \ + NLOHMANN_JSON_PASTE33, \ + NLOHMANN_JSON_PASTE32, \ + NLOHMANN_JSON_PASTE31, \ + NLOHMANN_JSON_PASTE30, \ + NLOHMANN_JSON_PASTE29, \ + NLOHMANN_JSON_PASTE28, \ + NLOHMANN_JSON_PASTE27, \ + NLOHMANN_JSON_PASTE26, \ + NLOHMANN_JSON_PASTE25, \ + NLOHMANN_JSON_PASTE24, \ + NLOHMANN_JSON_PASTE23, \ + NLOHMANN_JSON_PASTE22, \ + NLOHMANN_JSON_PASTE21, \ + NLOHMANN_JSON_PASTE20, \ + NLOHMANN_JSON_PASTE19, \ + NLOHMANN_JSON_PASTE18, \ + NLOHMANN_JSON_PASTE17, \ + NLOHMANN_JSON_PASTE16, \ + NLOHMANN_JSON_PASTE15, \ + NLOHMANN_JSON_PASTE14, \ + NLOHMANN_JSON_PASTE13, \ + NLOHMANN_JSON_PASTE12, \ + NLOHMANN_JSON_PASTE11, \ + NLOHMANN_JSON_PASTE10, \ + NLOHMANN_JSON_PASTE9, \ + NLOHMANN_JSON_PASTE8, \ + NLOHMANN_JSON_PASTE7, \ + NLOHMANN_JSON_PASTE6, \ + NLOHMANN_JSON_PASTE5, \ + NLOHMANN_JSON_PASTE4, \ + NLOHMANN_JSON_PASTE3, \ + NLOHMANN_JSON_PASTE2, \ + NLOHMANN_JSON_PASTE1)(__VA_ARGS__)) +#define NLOHMANN_JSON_PASTE2(func, v1) func(v1) +#define NLOHMANN_JSON_PASTE3(func, v1, v2) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE2(func, v2) +#define NLOHMANN_JSON_PASTE4(func, v1, v2, v3) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE3(func, v2, v3) +#define NLOHMANN_JSON_PASTE5(func, v1, v2, v3, v4) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE4(func, v2, v3, v4) +#define NLOHMANN_JSON_PASTE6(func, v1, v2, v3, v4, v5) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE5(func, v2, v3, v4, v5) +#define NLOHMANN_JSON_PASTE7(func, v1, v2, v3, v4, v5, v6) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE6(func, v2, v3, v4, v5, v6) +#define NLOHMANN_JSON_PASTE8(func, v1, v2, v3, v4, v5, v6, v7) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE7(func, v2, v3, v4, v5, v6, v7) +#define NLOHMANN_JSON_PASTE9(func, v1, v2, v3, v4, v5, v6, v7, v8) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE8(func, v2, v3, v4, v5, v6, v7, v8) +#define NLOHMANN_JSON_PASTE10(func, v1, v2, v3, v4, v5, v6, v7, v8, v9) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE9(func, v2, v3, v4, v5, v6, v7, v8, v9) +#define NLOHMANN_JSON_PASTE11(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE10(func, v2, v3, v4, v5, v6, v7, v8, v9, v10) +#define NLOHMANN_JSON_PASTE12(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE11(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11) +#define NLOHMANN_JSON_PASTE13(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE12(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12) +#define NLOHMANN_JSON_PASTE14(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE13(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13) +#define NLOHMANN_JSON_PASTE15(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE14(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14) +#define NLOHMANN_JSON_PASTE16(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE15(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15) +#define NLOHMANN_JSON_PASTE17(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE16(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16) +#define NLOHMANN_JSON_PASTE18(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE17(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17) +#define NLOHMANN_JSON_PASTE19(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE18(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18) +#define NLOHMANN_JSON_PASTE20(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE19(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19) +#define NLOHMANN_JSON_PASTE21(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE20(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20) +#define NLOHMANN_JSON_PASTE22(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE21(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21) +#define NLOHMANN_JSON_PASTE23(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE22(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22) +#define NLOHMANN_JSON_PASTE24(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE23(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23) +#define NLOHMANN_JSON_PASTE25(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE24(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24) +#define NLOHMANN_JSON_PASTE26(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE25(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25) +#define NLOHMANN_JSON_PASTE27(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE26(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26) +#define NLOHMANN_JSON_PASTE28(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE27(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27) +#define NLOHMANN_JSON_PASTE29(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE28(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28) +#define NLOHMANN_JSON_PASTE30(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE29(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29) +#define NLOHMANN_JSON_PASTE31(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE30(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30) +#define NLOHMANN_JSON_PASTE32(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE31(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31) +#define NLOHMANN_JSON_PASTE33(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE32(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32) +#define NLOHMANN_JSON_PASTE34(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE33(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33) +#define NLOHMANN_JSON_PASTE35(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE34(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34) +#define NLOHMANN_JSON_PASTE36(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE35(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35) +#define NLOHMANN_JSON_PASTE37(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE36(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36) +#define NLOHMANN_JSON_PASTE38(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE37(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37) +#define NLOHMANN_JSON_PASTE39(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE38(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38) +#define NLOHMANN_JSON_PASTE40(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE39(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39) +#define NLOHMANN_JSON_PASTE41(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE40(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40) +#define NLOHMANN_JSON_PASTE42(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE41(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41) +#define NLOHMANN_JSON_PASTE43(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE42(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42) +#define NLOHMANN_JSON_PASTE44(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE43(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43) +#define NLOHMANN_JSON_PASTE45(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE44(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44) +#define NLOHMANN_JSON_PASTE46(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE45(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45) +#define NLOHMANN_JSON_PASTE47(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE46(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46) +#define NLOHMANN_JSON_PASTE48(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE47(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47) +#define NLOHMANN_JSON_PASTE49(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE48(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48) +#define NLOHMANN_JSON_PASTE50(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE49(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49) +#define NLOHMANN_JSON_PASTE51(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE50(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50) +#define NLOHMANN_JSON_PASTE52(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE51(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51) +#define NLOHMANN_JSON_PASTE53(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE52(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52) +#define NLOHMANN_JSON_PASTE54(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE53(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53) +#define NLOHMANN_JSON_PASTE55(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE54(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54) +#define NLOHMANN_JSON_PASTE56(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE55(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55) +#define NLOHMANN_JSON_PASTE57(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE56(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56) +#define NLOHMANN_JSON_PASTE58(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE57(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57) +#define NLOHMANN_JSON_PASTE59(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE58(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58) +#define NLOHMANN_JSON_PASTE60(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE59(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59) +#define NLOHMANN_JSON_PASTE61(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE60(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60) +#define NLOHMANN_JSON_PASTE62(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE61(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61) +#define NLOHMANN_JSON_PASTE63(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE62(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62) +#define NLOHMANN_JSON_PASTE64(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62, v63) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE63(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62, v63) + +#define NLOHMANN_JSON_TO(v1) nlohmann_json_j[#v1] = nlohmann_json_t.v1; +#define NLOHMANN_JSON_FROM(v1) nlohmann_json_j.at(#v1).get_to(nlohmann_json_t.v1); +#define NLOHMANN_JSON_FROM_WITH_DEFAULT(v1) nlohmann_json_t.v1 = !nlohmann_json_j.is_null() ? nlohmann_json_j.value(#v1, nlohmann_json_default_obj.v1) : nlohmann_json_default_obj.v1; + +/*! +@brief macro +@def NLOHMANN_DEFINE_TYPE_INTRUSIVE +@since version 3.9.0 +@sa https://json.nlohmann.me/api/macros/nlohmann_define_type_intrusive/ +*/ +#define NLOHMANN_DEFINE_TYPE_INTRUSIVE(Type, ...) \ + template::value, int> = 0> \ + friend void to_json(BasicJsonType& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \ + template::value, int> = 0> \ + friend void from_json(const BasicJsonType& nlohmann_json_j, Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM, __VA_ARGS__)) } + +/*! +@brief macro +@def NLOHMANN_DEFINE_TYPE_INTRUSIVE_WITH_DEFAULT +@since version 3.11.0 +@sa https://json.nlohmann.me/api/macros/nlohmann_define_type_intrusive/ +*/ +#define NLOHMANN_DEFINE_TYPE_INTRUSIVE_WITH_DEFAULT(Type, ...) \ + template::value, int> = 0> \ + friend void to_json(BasicJsonType& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \ + template::value, int> = 0> \ + friend void from_json(const BasicJsonType& nlohmann_json_j, Type& nlohmann_json_t) { const Type nlohmann_json_default_obj{}; NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM_WITH_DEFAULT, __VA_ARGS__)) } + +/*! +@brief macro +@def NLOHMANN_DEFINE_TYPE_INTRUSIVE_ONLY_SERIALIZE +@since version 3.11.3 +@sa https://json.nlohmann.me/api/macros/nlohmann_define_type_intrusive/ +*/ +#define NLOHMANN_DEFINE_TYPE_INTRUSIVE_ONLY_SERIALIZE(Type, ...) \ + template::value, int> = 0> \ + friend void to_json(BasicJsonType& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } + +/*! +@brief macro +@def NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE +@since version 3.9.0 +@sa https://json.nlohmann.me/api/macros/nlohmann_define_type_non_intrusive/ +*/ +#define NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE(Type, ...) \ + template::value, int> = 0> \ + void to_json(BasicJsonType& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \ + template::value, int> = 0> \ + void from_json(const BasicJsonType& nlohmann_json_j, Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM, __VA_ARGS__)) } + +/*! +@brief macro +@def NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE_WITH_DEFAULT +@since version 3.11.0 +@sa https://json.nlohmann.me/api/macros/nlohmann_define_type_non_intrusive/ +*/ +#define NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE_WITH_DEFAULT(Type, ...) \ + template::value, int> = 0> \ + void to_json(BasicJsonType& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \ + template::value, int> = 0> \ + void from_json(const BasicJsonType& nlohmann_json_j, Type& nlohmann_json_t) { const Type nlohmann_json_default_obj{}; NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM_WITH_DEFAULT, __VA_ARGS__)) } + +/*! +@brief macro +@def NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE_ONLY_SERIALIZE +@since version 3.11.3 +@sa https://json.nlohmann.me/api/macros/nlohmann_define_type_non_intrusive/ +*/ +#define NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE_ONLY_SERIALIZE(Type, ...) \ + template::value, int> = 0> \ + void to_json(BasicJsonType& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } + +/*! +@brief macro +@def NLOHMANN_DEFINE_DERIVED_TYPE_INTRUSIVE +@since version 3.12.0 +@sa https://json.nlohmann.me/api/macros/nlohmann_define_derived_type/ +*/ +#define NLOHMANN_DEFINE_DERIVED_TYPE_INTRUSIVE(Type, BaseType, ...) \ + template::value, int> = 0> \ + friend void to_json(BasicJsonType& nlohmann_json_j, const Type& nlohmann_json_t) { nlohmann::to_json(nlohmann_json_j, static_cast(nlohmann_json_t)); NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \ + template::value, int> = 0> \ + friend void from_json(const BasicJsonType& nlohmann_json_j, Type& nlohmann_json_t) { nlohmann::from_json(nlohmann_json_j, static_cast(nlohmann_json_t)); NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM, __VA_ARGS__)) } + +/*! +@brief macro +@def NLOHMANN_DEFINE_DERIVED_TYPE_INTRUSIVE_WITH_DEFAULT +@since version 3.12.0 +@sa https://json.nlohmann.me/api/macros/nlohmann_define_derived_type/ +*/ +#define NLOHMANN_DEFINE_DERIVED_TYPE_INTRUSIVE_WITH_DEFAULT(Type, BaseType, ...) \ + template::value, int> = 0> \ + friend void to_json(BasicJsonType& nlohmann_json_j, const Type& nlohmann_json_t) { nlohmann::to_json(nlohmann_json_j, static_cast(nlohmann_json_t)); NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \ + template::value, int> = 0> \ + friend void from_json(const BasicJsonType& nlohmann_json_j, Type& nlohmann_json_t) { nlohmann::from_json(nlohmann_json_j, static_cast(nlohmann_json_t)); const Type nlohmann_json_default_obj{}; NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM_WITH_DEFAULT, __VA_ARGS__)) } + +/*! +@brief macro +@def NLOHMANN_DEFINE_DERIVED_TYPE_INTRUSIVE_ONLY_SERIALIZE +@since version 3.12.0 +@sa https://json.nlohmann.me/api/macros/nlohmann_define_derived_type/ +*/ +#define NLOHMANN_DEFINE_DERIVED_TYPE_INTRUSIVE_ONLY_SERIALIZE(Type, BaseType, ...) \ + template::value, int> = 0> \ + friend void to_json(BasicJsonType& nlohmann_json_j, const Type& nlohmann_json_t) { nlohmann::to_json(nlohmann_json_j, static_cast(nlohmann_json_t)); NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } + +/*! +@brief macro +@def NLOHMANN_DEFINE_DERIVED_TYPE_NON_INTRUSIVE +@since version 3.12.0 +@sa https://json.nlohmann.me/api/macros/nlohmann_define_derived_type/ +*/ +#define NLOHMANN_DEFINE_DERIVED_TYPE_NON_INTRUSIVE(Type, BaseType, ...) \ + template::value, int> = 0> \ + void to_json(BasicJsonType& nlohmann_json_j, const Type& nlohmann_json_t) { nlohmann::to_json(nlohmann_json_j, static_cast(nlohmann_json_t)); NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \ + template::value, int> = 0> \ + void from_json(const BasicJsonType& nlohmann_json_j, Type& nlohmann_json_t) { nlohmann::from_json(nlohmann_json_j, static_cast(nlohmann_json_t)); NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM, __VA_ARGS__)) } + +/*! +@brief macro +@def NLOHMANN_DEFINE_DERIVED_TYPE_NON_INTRUSIVE_WITH_DEFAULT +@since version 3.12.0 +@sa https://json.nlohmann.me/api/macros/nlohmann_define_derived_type/ +*/ +#define NLOHMANN_DEFINE_DERIVED_TYPE_NON_INTRUSIVE_WITH_DEFAULT(Type, BaseType, ...) \ + template::value, int> = 0> \ + void to_json(BasicJsonType& nlohmann_json_j, const Type& nlohmann_json_t) { nlohmann::to_json(nlohmann_json_j, static_cast(nlohmann_json_t)); NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \ + template::value, int> = 0> \ + void from_json(const BasicJsonType& nlohmann_json_j, Type& nlohmann_json_t) { nlohmann::from_json(nlohmann_json_j, static_cast(nlohmann_json_t)); const Type nlohmann_json_default_obj{}; NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM_WITH_DEFAULT, __VA_ARGS__)) } + +/*! +@brief macro +@def NLOHMANN_DEFINE_DERIVED_TYPE_NON_INTRUSIVE_ONLY_SERIALIZE +@since version 3.12.0 +@sa https://json.nlohmann.me/api/macros/nlohmann_define_derived_type/ +*/ +#define NLOHMANN_DEFINE_DERIVED_TYPE_NON_INTRUSIVE_ONLY_SERIALIZE(Type, BaseType, ...) \ + template::value, int> = 0> \ + void to_json(BasicJsonType& nlohmann_json_j, const Type& nlohmann_json_t) { nlohmann::to_json(nlohmann_json_j, static_cast(nlohmann_json_t)); NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } + +// inspired from https://stackoverflow.com/a/26745591 +// allows calling any std function as if (e.g., with begin): +// using std::begin; begin(x); +// +// it allows using the detected idiom to retrieve the return type +// of such an expression +#define NLOHMANN_CAN_CALL_STD_FUNC_IMPL(std_name) \ + namespace detail { \ + using std::std_name; \ + \ + template \ + using result_of_##std_name = decltype(std_name(std::declval()...)); \ + } \ + \ + namespace detail2 { \ + struct std_name##_tag \ + { \ + }; \ + \ + template \ + std_name##_tag std_name(T&&...); \ + \ + template \ + using result_of_##std_name = decltype(std_name(std::declval()...)); \ + \ + template \ + struct would_call_std_##std_name \ + { \ + static constexpr auto const value = ::nlohmann::detail:: \ + is_detected_exact::value; \ + }; \ + } /* namespace detail2 */ \ + \ + template \ + struct would_call_std_##std_name : detail2::would_call_std_##std_name \ + { \ + } + +#ifndef JSON_USE_IMPLICIT_CONVERSIONS + #define JSON_USE_IMPLICIT_CONVERSIONS 1 +#endif + +#if JSON_USE_IMPLICIT_CONVERSIONS + #define JSON_EXPLICIT +#else + #define JSON_EXPLICIT explicit +#endif + +#ifndef JSON_DISABLE_ENUM_SERIALIZATION + #define JSON_DISABLE_ENUM_SERIALIZATION 0 +#endif + +#ifndef JSON_USE_GLOBAL_UDLS + #define JSON_USE_GLOBAL_UDLS 1 +#endif + +#if JSON_HAS_THREE_WAY_COMPARISON + #include // partial_ordering +#endif + +NLOHMANN_JSON_NAMESPACE_BEGIN +namespace detail +{ + +/////////////////////////// +// JSON type enumeration // +/////////////////////////// + +/*! +@brief the JSON type enumeration + +This enumeration collects the different JSON types. It is internally used to +distinguish the stored values, and the functions @ref basic_json::is_null(), +@ref basic_json::is_object(), @ref basic_json::is_array(), +@ref basic_json::is_string(), @ref basic_json::is_boolean(), +@ref basic_json::is_number() (with @ref basic_json::is_number_integer(), +@ref basic_json::is_number_unsigned(), and @ref basic_json::is_number_float()), +@ref basic_json::is_discarded(), @ref basic_json::is_primitive(), and +@ref basic_json::is_structured() rely on it. + +@note There are three enumeration entries (number_integer, number_unsigned, and +number_float), because the library distinguishes these three types for numbers: +@ref basic_json::number_unsigned_t is used for unsigned integers, +@ref basic_json::number_integer_t is used for signed integers, and +@ref basic_json::number_float_t is used for floating-point numbers or to +approximate integers which do not fit in the limits of their respective type. + +@sa see @ref basic_json::basic_json(const value_t value_type) -- create a JSON +value with the default value for a given type + +@since version 1.0.0 +*/ +enum class value_t : std::uint8_t +{ + null, ///< null value + object, ///< object (unordered set of name/value pairs) + array, ///< array (ordered collection of values) + string, ///< string value + boolean, ///< boolean value + number_integer, ///< number value (signed integer) + number_unsigned, ///< number value (unsigned integer) + number_float, ///< number value (floating-point) + binary, ///< binary array (ordered collection of bytes) + discarded ///< discarded by the parser callback function +}; + +/*! +@brief comparison operator for JSON types + +Returns an ordering that is similar to Python: +- order: null < boolean < number < object < array < string < binary +- furthermore, each type is not smaller than itself +- discarded values are not comparable +- binary is represented as a b"" string in python and directly comparable to a + string; however, making a binary array directly comparable with a string would + be surprising behavior in a JSON file. + +@since version 1.0.0 +*/ +#if JSON_HAS_THREE_WAY_COMPARISON + inline std::partial_ordering operator<=>(const value_t lhs, const value_t rhs) noexcept // *NOPAD* +#else + inline bool operator<(const value_t lhs, const value_t rhs) noexcept +#endif +{ + static constexpr std::array order = {{ + 0 /* null */, 3 /* object */, 4 /* array */, 5 /* string */, + 1 /* boolean */, 2 /* integer */, 2 /* unsigned */, 2 /* float */, + 6 /* binary */ + } + }; + + const auto l_index = static_cast(lhs); + const auto r_index = static_cast(rhs); +#if JSON_HAS_THREE_WAY_COMPARISON + if (l_index < order.size() && r_index < order.size()) + { + return order[l_index] <=> order[r_index]; // *NOPAD* + } + return std::partial_ordering::unordered; +#else + return l_index < order.size() && r_index < order.size() && order[l_index] < order[r_index]; +#endif +} + +// GCC selects the built-in operator< over an operator rewritten from +// a user-defined spaceship operator +// Clang, MSVC, and ICC select the rewritten candidate +// (see GCC bug https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105200) +#if JSON_HAS_THREE_WAY_COMPARISON && defined(__GNUC__) +inline bool operator<(const value_t lhs, const value_t rhs) noexcept +{ + return std::is_lt(lhs <=> rhs); // *NOPAD* +} +#endif + +} // namespace detail +NLOHMANN_JSON_NAMESPACE_END + +// #include +// __ _____ _____ _____ +// __| | __| | | | JSON for Modern C++ +// | | |__ | | | | | | version 3.12.0 +// |_____|_____|_____|_|___| https://github.com/nlohmann/json +// +// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann +// SPDX-License-Identifier: MIT + + + +// #include + + +NLOHMANN_JSON_NAMESPACE_BEGIN +namespace detail +{ + +/*! +@brief replace all occurrences of a substring by another string + +@param[in,out] s the string to manipulate; changed so that all + occurrences of @a f are replaced with @a t +@param[in] f the substring to replace with @a t +@param[in] t the string to replace @a f + +@pre The search string @a f must not be empty. **This precondition is +enforced with an assertion.** + +@since version 2.0.0 +*/ +template +inline void replace_substring(StringType& s, const StringType& f, + const StringType& t) +{ + JSON_ASSERT(!f.empty()); + for (auto pos = s.find(f); // find first occurrence of f + pos != StringType::npos; // make sure f was found + s.replace(pos, f.size(), t), // replace with t, and + pos = s.find(f, pos + t.size())) // find next occurrence of f + {} +} + +/*! + * @brief string escaping as described in RFC 6901 (Sect. 4) + * @param[in] s string to escape + * @return escaped string + * + * Note the order of escaping "~" to "~0" and "/" to "~1" is important. + */ +template +inline StringType escape(StringType s) +{ + replace_substring(s, StringType{"~"}, StringType{"~0"}); + replace_substring(s, StringType{"/"}, StringType{"~1"}); + return s; +} + +/*! + * @brief string unescaping as described in RFC 6901 (Sect. 4) + * @param[in] s string to unescape + * @return unescaped string + * + * Note the order of escaping "~1" to "/" and "~0" to "~" is important. + */ +template +static void unescape(StringType& s) +{ + replace_substring(s, StringType{"~1"}, StringType{"/"}); + replace_substring(s, StringType{"~0"}, StringType{"~"}); +} + +} // namespace detail +NLOHMANN_JSON_NAMESPACE_END + +// #include +// __ _____ _____ _____ +// __| | __| | | | JSON for Modern C++ +// | | |__ | | | | | | version 3.12.0 +// |_____|_____|_____|_|___| https://github.com/nlohmann/json +// +// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann +// SPDX-License-Identifier: MIT + + + +#include // size_t + +// #include + + +NLOHMANN_JSON_NAMESPACE_BEGIN +namespace detail +{ + +/// struct to capture the start position of the current token +struct position_t +{ + /// the total number of characters read + std::size_t chars_read_total = 0; + /// the number of characters read in the current line + std::size_t chars_read_current_line = 0; + /// the number of lines read + std::size_t lines_read = 0; + + /// conversion to size_t to preserve SAX interface + constexpr operator size_t() const + { + return chars_read_total; + } +}; + +} // namespace detail +NLOHMANN_JSON_NAMESPACE_END + +// #include + +// #include +// __ _____ _____ _____ +// __| | __| | | | JSON for Modern C++ +// | | |__ | | | | | | version 3.12.0 +// |_____|_____|_____|_|___| https://github.com/nlohmann/json +// +// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann +// SPDX-FileCopyrightText: 2018 The Abseil Authors +// SPDX-License-Identifier: MIT + + + +#include // array +#include // size_t +#include // conditional, enable_if, false_type, integral_constant, is_constructible, is_integral, is_same, remove_cv, remove_reference, true_type +#include // index_sequence, make_index_sequence, index_sequence_for + +// #include + + +NLOHMANN_JSON_NAMESPACE_BEGIN +namespace detail +{ + +template +using uncvref_t = typename std::remove_cv::type>::type; + +#ifdef JSON_HAS_CPP_14 + +// the following utilities are natively available in C++14 +using std::enable_if_t; +using std::index_sequence; +using std::make_index_sequence; +using std::index_sequence_for; + +#else + +// alias templates to reduce boilerplate +template +using enable_if_t = typename std::enable_if::type; + +// The following code is taken from https://github.com/abseil/abseil-cpp/blob/10cb35e459f5ecca5b2ff107635da0bfa41011b4/absl/utility/utility.h +// which is part of Google Abseil (https://github.com/abseil/abseil-cpp), licensed under the Apache License 2.0. + +//// START OF CODE FROM GOOGLE ABSEIL + +// integer_sequence +// +// Class template representing a compile-time integer sequence. An instantiation +// of `integer_sequence` has a sequence of integers encoded in its +// type through its template arguments (which is a common need when +// working with C++11 variadic templates). `absl::integer_sequence` is designed +// to be a drop-in replacement for C++14's `std::integer_sequence`. +// +// Example: +// +// template< class T, T... Ints > +// void user_function(integer_sequence); +// +// int main() +// { +// // user_function's `T` will be deduced to `int` and `Ints...` +// // will be deduced to `0, 1, 2, 3, 4`. +// user_function(make_integer_sequence()); +// } +template +struct integer_sequence +{ + using value_type = T; + static constexpr std::size_t size() noexcept + { + return sizeof...(Ints); + } +}; + +// index_sequence +// +// A helper template for an `integer_sequence` of `size_t`, +// `absl::index_sequence` is designed to be a drop-in replacement for C++14's +// `std::index_sequence`. +template +using index_sequence = integer_sequence; + +namespace utility_internal +{ + +template +struct Extend; + +// Note that SeqSize == sizeof...(Ints). It's passed explicitly for efficiency. +template +struct Extend, SeqSize, 0> +{ + using type = integer_sequence < T, Ints..., (Ints + SeqSize)... >; +}; + +template +struct Extend, SeqSize, 1> +{ + using type = integer_sequence < T, Ints..., (Ints + SeqSize)..., 2 * SeqSize >; +}; + +// Recursion helper for 'make_integer_sequence'. +// 'Gen::type' is an alias for 'integer_sequence'. +template +struct Gen +{ + using type = + typename Extend < typename Gen < T, N / 2 >::type, N / 2, N % 2 >::type; +}; + +template +struct Gen +{ + using type = integer_sequence; +}; + +} // namespace utility_internal + +// Compile-time sequences of integers + +// make_integer_sequence +// +// This template alias is equivalent to +// `integer_sequence`, and is designed to be a drop-in +// replacement for C++14's `std::make_integer_sequence`. +template +using make_integer_sequence = typename utility_internal::Gen::type; + +// make_index_sequence +// +// This template alias is equivalent to `index_sequence<0, 1, ..., N-1>`, +// and is designed to be a drop-in replacement for C++14's +// `std::make_index_sequence`. +template +using make_index_sequence = make_integer_sequence; + +// index_sequence_for +// +// Converts a typename pack into an index sequence of the same length, and +// is designed to be a drop-in replacement for C++14's +// `std::index_sequence_for()` +template +using index_sequence_for = make_index_sequence; + +//// END OF CODE FROM GOOGLE ABSEIL + +#endif + +// dispatch utility (taken from ranges-v3) +template struct priority_tag : priority_tag < N - 1 > {}; +template<> struct priority_tag<0> {}; + +// taken from ranges-v3 +template +struct static_const +{ + static JSON_INLINE_VARIABLE constexpr T value{}; +}; + +#ifndef JSON_HAS_CPP_17 + template + constexpr T static_const::value; +#endif + +template +constexpr std::array make_array(Args&& ... args) +{ + return std::array {{static_cast(std::forward(args))...}}; +} + +} // namespace detail +NLOHMANN_JSON_NAMESPACE_END + +// #include +// __ _____ _____ _____ +// __| | __| | | | JSON for Modern C++ +// | | |__ | | | | | | version 3.12.0 +// |_____|_____|_____|_|___| https://github.com/nlohmann/json +// +// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann +// SPDX-License-Identifier: MIT + + + +#include // numeric_limits +#include // char_traits +#include // tuple +#include // false_type, is_constructible, is_integral, is_same, true_type +#include // declval + +// #include +// __ _____ _____ _____ +// __| | __| | | | JSON for Modern C++ +// | | |__ | | | | | | version 3.12.0 +// |_____|_____|_____|_|___| https://github.com/nlohmann/json +// +// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann +// SPDX-License-Identifier: MIT + + + +#include // random_access_iterator_tag + +// #include + +// #include + +// #include + + +NLOHMANN_JSON_NAMESPACE_BEGIN +namespace detail +{ + +template +struct iterator_types {}; + +template +struct iterator_types < + It, + void_t> +{ + using difference_type = typename It::difference_type; + using value_type = typename It::value_type; + using pointer = typename It::pointer; + using reference = typename It::reference; + using iterator_category = typename It::iterator_category; +}; + +// This is required as some compilers implement std::iterator_traits in a way that +// doesn't work with SFINAE. See https://github.com/nlohmann/json/issues/1341. +template +struct iterator_traits +{ +}; + +template +struct iterator_traits < T, enable_if_t < !std::is_pointer::value >> + : iterator_types +{ +}; + +template +struct iterator_traits::value>> +{ + using iterator_category = std::random_access_iterator_tag; + using value_type = T; + using difference_type = ptrdiff_t; + using pointer = T*; + using reference = T&; +}; + +} // namespace detail +NLOHMANN_JSON_NAMESPACE_END + +// #include + +// #include +// __ _____ _____ _____ +// __| | __| | | | JSON for Modern C++ +// | | |__ | | | | | | version 3.12.0 +// |_____|_____|_____|_|___| https://github.com/nlohmann/json +// +// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann +// SPDX-License-Identifier: MIT + + + +// #include + + +NLOHMANN_JSON_NAMESPACE_BEGIN + +NLOHMANN_CAN_CALL_STD_FUNC_IMPL(begin); + +NLOHMANN_JSON_NAMESPACE_END + +// #include +// __ _____ _____ _____ +// __| | __| | | | JSON for Modern C++ +// | | |__ | | | | | | version 3.12.0 +// |_____|_____|_____|_|___| https://github.com/nlohmann/json +// +// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann +// SPDX-License-Identifier: MIT + + + +// #include + + +NLOHMANN_JSON_NAMESPACE_BEGIN + +NLOHMANN_CAN_CALL_STD_FUNC_IMPL(end); + +NLOHMANN_JSON_NAMESPACE_END + +// #include + +// #include + +// #include +// __ _____ _____ _____ +// __| | __| | | | JSON for Modern C++ +// | | |__ | | | | | | version 3.12.0 +// |_____|_____|_____|_|___| https://github.com/nlohmann/json +// +// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann +// SPDX-License-Identifier: MIT + +#ifndef INCLUDE_NLOHMANN_JSON_FWD_HPP_ + #define INCLUDE_NLOHMANN_JSON_FWD_HPP_ + + #include // int64_t, uint64_t + #include // map + #include // allocator + #include // string + #include // vector + + // #include + + + /*! + @brief namespace for Niels Lohmann + @see https://github.com/nlohmann + @since version 1.0.0 + */ + NLOHMANN_JSON_NAMESPACE_BEGIN + + /*! + @brief default JSONSerializer template argument + + This serializer ignores the template arguments and uses ADL + ([argument-dependent lookup](https://en.cppreference.com/w/cpp/language/adl)) + for serialization. + */ + template + struct adl_serializer; + + /// a class to store JSON values + /// @sa https://json.nlohmann.me/api/basic_json/ + template class ObjectType = + std::map, + template class ArrayType = std::vector, + class StringType = std::string, class BooleanType = bool, + class NumberIntegerType = std::int64_t, + class NumberUnsignedType = std::uint64_t, + class NumberFloatType = double, + template class AllocatorType = std::allocator, + template class JSONSerializer = + adl_serializer, + class BinaryType = std::vector, // cppcheck-suppress syntaxError + class CustomBaseClass = void> + class basic_json; + + /// @brief JSON Pointer defines a string syntax for identifying a specific value within a JSON document + /// @sa https://json.nlohmann.me/api/json_pointer/ + template + class json_pointer; + + /*! + @brief default specialization + @sa https://json.nlohmann.me/api/json/ + */ + using json = basic_json<>; + + /// @brief a minimal map-like container that preserves insertion order + /// @sa https://json.nlohmann.me/api/ordered_map/ + template + struct ordered_map; + + /// @brief specialization that maintains the insertion order of object keys + /// @sa https://json.nlohmann.me/api/ordered_json/ + using ordered_json = basic_json; + + NLOHMANN_JSON_NAMESPACE_END + +#endif // INCLUDE_NLOHMANN_JSON_FWD_HPP_ + + +NLOHMANN_JSON_NAMESPACE_BEGIN +/*! +@brief detail namespace with internal helper functions + +This namespace collects functions that should not be exposed, +implementations of some @ref basic_json methods, and meta-programming helpers. + +@since version 2.1.0 +*/ +namespace detail +{ + +///////////// +// helpers // +///////////// + +// Note to maintainers: +// +// Every trait in this file expects a non CV-qualified type. +// The only exceptions are in the 'aliases for detected' section +// (i.e. those of the form: decltype(T::member_function(std::declval()))) +// +// In this case, T has to be properly CV-qualified to constraint the function arguments +// (e.g. to_json(BasicJsonType&, const T&)) + +template struct is_basic_json : std::false_type {}; + +NLOHMANN_BASIC_JSON_TPL_DECLARATION +struct is_basic_json : std::true_type {}; + +// used by exceptions create() member functions +// true_type for pointer to possibly cv-qualified basic_json or std::nullptr_t +// false_type otherwise +template +struct is_basic_json_context : + std::integral_constant < bool, + is_basic_json::type>::type>::value + || std::is_same::value > +{}; + +////////////////////// +// json_ref helpers // +////////////////////// + +template +class json_ref; + +template +struct is_json_ref : std::false_type {}; + +template +struct is_json_ref> : std::true_type {}; + +////////////////////////// +// aliases for detected // +////////////////////////// + +template +using mapped_type_t = typename T::mapped_type; + +template +using key_type_t = typename T::key_type; + +template +using value_type_t = typename T::value_type; + +template +using difference_type_t = typename T::difference_type; + +template +using pointer_t = typename T::pointer; + +template +using reference_t = typename T::reference; + +template +using iterator_category_t = typename T::iterator_category; + +template +using to_json_function = decltype(T::to_json(std::declval()...)); + +template +using from_json_function = decltype(T::from_json(std::declval()...)); + +template +using get_template_function = decltype(std::declval().template get()); + +// trait checking if JSONSerializer::from_json(json const&, udt&) exists +template +struct has_from_json : std::false_type {}; + +// trait checking if j.get is valid +// use this trait instead of std::is_constructible or std::is_convertible, +// both rely on, or make use of implicit conversions, and thus fail when T +// has several constructors/operator= (see https://github.com/nlohmann/json/issues/958) +template +struct is_getable +{ + static constexpr bool value = is_detected::value; +}; + +template +struct has_from_json < BasicJsonType, T, enable_if_t < !is_basic_json::value >> +{ + using serializer = typename BasicJsonType::template json_serializer; + + static constexpr bool value = + is_detected_exact::value; +}; + +// This trait checks if JSONSerializer::from_json(json const&) exists +// this overload is used for non-default-constructible user-defined-types +template +struct has_non_default_from_json : std::false_type {}; + +template +struct has_non_default_from_json < BasicJsonType, T, enable_if_t < !is_basic_json::value >> +{ + using serializer = typename BasicJsonType::template json_serializer; + + static constexpr bool value = + is_detected_exact::value; +}; + +// This trait checks if BasicJsonType::json_serializer::to_json exists +// Do not evaluate the trait when T is a basic_json type, to avoid template instantiation infinite recursion. +template +struct has_to_json : std::false_type {}; + +template +struct has_to_json < BasicJsonType, T, enable_if_t < !is_basic_json::value >> +{ + using serializer = typename BasicJsonType::template json_serializer; + + static constexpr bool value = + is_detected_exact::value; +}; + +template +using detect_key_compare = typename T::key_compare; + +template +struct has_key_compare : std::integral_constant::value> {}; + +// obtains the actual object key comparator +template +struct actual_object_comparator +{ + using object_t = typename BasicJsonType::object_t; + using object_comparator_t = typename BasicJsonType::default_object_comparator_t; + using type = typename std::conditional < has_key_compare::value, + typename object_t::key_compare, object_comparator_t>::type; +}; + +template +using actual_object_comparator_t = typename actual_object_comparator::type; + +///////////////// +// char_traits // +///////////////// + +// Primary template of char_traits calls std char_traits +template +struct char_traits : std::char_traits +{}; + +// Explicitly define char traits for unsigned char since it is not standard +template<> +struct char_traits : std::char_traits +{ + using char_type = unsigned char; + using int_type = uint64_t; + + // Redefine to_int_type function + static int_type to_int_type(char_type c) noexcept + { + return static_cast(c); + } + + static char_type to_char_type(int_type i) noexcept + { + return static_cast(i); + } + + static constexpr int_type eof() noexcept + { + return static_cast(std::char_traits::eof()); + } +}; + +// Explicitly define char traits for signed char since it is not standard +template<> +struct char_traits : std::char_traits +{ + using char_type = signed char; + using int_type = uint64_t; + + // Redefine to_int_type function + static int_type to_int_type(char_type c) noexcept + { + return static_cast(c); + } + + static char_type to_char_type(int_type i) noexcept + { + return static_cast(i); + } + + static constexpr int_type eof() noexcept + { + return static_cast(std::char_traits::eof()); + } +}; + +/////////////////// +// is_ functions // +/////////////////// + +// https://en.cppreference.com/w/cpp/types/conjunction +template struct conjunction : std::true_type { }; +template struct conjunction : B { }; +template +struct conjunction +: std::conditional(B::value), conjunction, B>::type {}; + +// https://en.cppreference.com/w/cpp/types/negation +template struct negation : std::integral_constant < bool, !B::value > { }; + +// Reimplementation of is_constructible and is_default_constructible, due to them being broken for +// std::pair and std::tuple until LWG 2367 fix (see https://cplusplus.github.io/LWG/lwg-defects.html#2367). +// This causes compile errors in e.g. clang 3.5 or gcc 4.9. +template +struct is_default_constructible : std::is_default_constructible {}; + +template +struct is_default_constructible> + : conjunction, is_default_constructible> {}; + +template +struct is_default_constructible> + : conjunction, is_default_constructible> {}; + +template +struct is_default_constructible> + : conjunction...> {}; + +template +struct is_default_constructible> + : conjunction...> {}; + +template +struct is_constructible : std::is_constructible {}; + +template +struct is_constructible> : is_default_constructible> {}; + +template +struct is_constructible> : is_default_constructible> {}; + +template +struct is_constructible> : is_default_constructible> {}; + +template +struct is_constructible> : is_default_constructible> {}; + +template +struct is_iterator_traits : std::false_type {}; + +template +struct is_iterator_traits> +{ + private: + using traits = iterator_traits; + + public: + static constexpr auto value = + is_detected::value && + is_detected::value && + is_detected::value && + is_detected::value && + is_detected::value; +}; + +template +struct is_range +{ + private: + using t_ref = typename std::add_lvalue_reference::type; + + using iterator = detected_t; + using sentinel = detected_t; + + // to be 100% correct, it should use https://en.cppreference.com/w/cpp/iterator/input_or_output_iterator + // and https://en.cppreference.com/w/cpp/iterator/sentinel_for + // but reimplementing these would be too much work, as a lot of other concepts are used underneath + static constexpr auto is_iterator_begin = + is_iterator_traits>::value; + + public: + static constexpr bool value = !std::is_same::value && !std::is_same::value && is_iterator_begin; +}; + +template +using iterator_t = enable_if_t::value, result_of_begin())>>; + +template +using range_value_t = value_type_t>>; + +// The following implementation of is_complete_type is taken from +// https://blogs.msdn.microsoft.com/vcblog/2015/12/02/partial-support-for-expression-sfinae-in-vs-2015-update-1/ +// and is written by Xiang Fan who agreed to using it in this library. + +template +struct is_complete_type : std::false_type {}; + +template +struct is_complete_type : std::true_type {}; + +template +struct is_compatible_object_type_impl : std::false_type {}; + +template +struct is_compatible_object_type_impl < + BasicJsonType, CompatibleObjectType, + enable_if_t < is_detected::value&& + is_detected::value >> +{ + using object_t = typename BasicJsonType::object_t; + + // macOS's is_constructible does not play well with nonesuch... + static constexpr bool value = + is_constructible::value && + is_constructible::value; +}; + +template +struct is_compatible_object_type + : is_compatible_object_type_impl {}; + +template +struct is_constructible_object_type_impl : std::false_type {}; + +template +struct is_constructible_object_type_impl < + BasicJsonType, ConstructibleObjectType, + enable_if_t < is_detected::value&& + is_detected::value >> +{ + using object_t = typename BasicJsonType::object_t; + + static constexpr bool value = + (is_default_constructible::value && + (std::is_move_assignable::value || + std::is_copy_assignable::value) && + (is_constructible::value && + std::is_same < + typename object_t::mapped_type, + typename ConstructibleObjectType::mapped_type >::value)) || + (has_from_json::value || + has_non_default_from_json < + BasicJsonType, + typename ConstructibleObjectType::mapped_type >::value); +}; + +template +struct is_constructible_object_type + : is_constructible_object_type_impl {}; + +template +struct is_compatible_string_type +{ + static constexpr auto value = + is_constructible::value; +}; + +template +struct is_constructible_string_type +{ + // launder type through decltype() to fix compilation failure on ICPC +#ifdef __INTEL_COMPILER + using laundered_type = decltype(std::declval()); +#else + using laundered_type = ConstructibleStringType; +#endif + + static constexpr auto value = + conjunction < + is_constructible, + is_detected_exact>::value; +}; + +template +struct is_compatible_array_type_impl : std::false_type {}; + +template +struct is_compatible_array_type_impl < + BasicJsonType, CompatibleArrayType, + enable_if_t < + is_detected::value&& + is_iterator_traits>>::value&& +// special case for types like std::filesystem::path whose iterator's value_type are themselves +// c.f. https://github.com/nlohmann/json/pull/3073 + !std::is_same>::value >> +{ + static constexpr bool value = + is_constructible>::value; +}; + +template +struct is_compatible_array_type + : is_compatible_array_type_impl {}; + +template +struct is_constructible_array_type_impl : std::false_type {}; + +template +struct is_constructible_array_type_impl < + BasicJsonType, ConstructibleArrayType, + enable_if_t::value >> + : std::true_type {}; + +template +struct is_constructible_array_type_impl < + BasicJsonType, ConstructibleArrayType, + enable_if_t < !std::is_same::value&& + !is_compatible_string_type::value&& + is_default_constructible::value&& +(std::is_move_assignable::value || + std::is_copy_assignable::value)&& +is_detected::value&& +is_iterator_traits>>::value&& +is_detected::value&& +// special case for types like std::filesystem::path whose iterator's value_type are themselves +// c.f. https://github.com/nlohmann/json/pull/3073 +!std::is_same>::value&& +is_complete_type < +detected_t>::value >> +{ + using value_type = range_value_t; + + static constexpr bool value = + std::is_same::value || + has_from_json::value || + has_non_default_from_json < + BasicJsonType, + value_type >::value; +}; + +template +struct is_constructible_array_type + : is_constructible_array_type_impl {}; + +template +struct is_compatible_integer_type_impl : std::false_type {}; + +template +struct is_compatible_integer_type_impl < + RealIntegerType, CompatibleNumberIntegerType, + enable_if_t < std::is_integral::value&& + std::is_integral::value&& + !std::is_same::value >> +{ + // is there an assert somewhere on overflows? + using RealLimits = std::numeric_limits; + using CompatibleLimits = std::numeric_limits; + + static constexpr auto value = + is_constructible::value && + CompatibleLimits::is_integer && + RealLimits::is_signed == CompatibleLimits::is_signed; +}; + +template +struct is_compatible_integer_type + : is_compatible_integer_type_impl {}; + +template +struct is_compatible_type_impl: std::false_type {}; + +template +struct is_compatible_type_impl < + BasicJsonType, CompatibleType, + enable_if_t::value >> +{ + static constexpr bool value = + has_to_json::value; +}; + +template +struct is_compatible_type + : is_compatible_type_impl {}; + +template +struct is_constructible_tuple : std::false_type {}; + +template +struct is_constructible_tuple> : conjunction...> {}; + +template +struct is_json_iterator_of : std::false_type {}; + +template +struct is_json_iterator_of : std::true_type {}; + +template +struct is_json_iterator_of : std::true_type +{}; + +// checks if a given type T is a template specialization of Primary +template