From bba994f2cdec81d6a09e64c97042f2a8fc2e0403 Mon Sep 17 00:00:00 2001 From: Royi Rassin Date: Wed, 25 Mar 2026 08:31:35 +0200 Subject: [PATCH 1/4] Record: 5-expert Hedge Mixer + TTT (3-seed mean val_bpb=1.0745) --- submission-2026-03-24/README.md | 110 ++ submission-2026-03-24/log_seed1337.txt | 173 +++ submission-2026-03-24/log_seed42.txt | 173 +++ submission-2026-03-24/log_seed7.txt | 173 +++ submission-2026-03-24/requirements.txt | 5 + submission-2026-03-24/submission.json | 14 + submission-2026-03-24/train_gpt.py | 1966 ++++++++++++++++++++++++ 7 files changed, 2614 insertions(+) create mode 100644 submission-2026-03-24/README.md create mode 100644 submission-2026-03-24/log_seed1337.txt create mode 100644 submission-2026-03-24/log_seed42.txt create mode 100644 submission-2026-03-24/log_seed7.txt create mode 100644 submission-2026-03-24/requirements.txt create mode 100644 submission-2026-03-24/submission.json create mode 100644 submission-2026-03-24/train_gpt.py diff --git a/submission-2026-03-24/README.md b/submission-2026-03-24/README.md new file mode 100644 index 000000000..2753515cc --- /dev/null +++ b/submission-2026-03-24/README.md @@ -0,0 +1,110 @@ +# 5-expert Hedge Mixer + TTT + +**val_bpb: 1.0745** (3-seed mean) | **<15.5 MB** | 8xH100 SXM + +## Results (8xH100 80GB SXM) + +| Seed | steps | step_avg | Pre-TTT bpb | **Post-TTT bpb** | TTT gain | Eval time | Artifact | +|------|-------|----------|-------------|-----------------|----------|-----------|----------| +| 1337 | 5,997 | 97.1ms | 1.1248 | **1.0560** | -0.0688 | 563s | 15.48 MB | +| 42 | 5,997 | 97.1ms | 1.1257 | **1.0970** | -0.0287 | 563s | 15.41 MB | +| 7 | 5,983 | 97.3ms | 1.1251 | **1.0704** | -0.0547 | 561s | 15.43 MB | +| **Mean** | | | **1.1252** | **1.0745** | **-0.0507** | | | + +## Key Contribution: 5-expert Logistic Context Mixer + +GPU-vectorized online context mixing using the Hedge/multiplicative-weights algorithm. Five experts blend predictions in log-probability space: + +| Expert | Source | Description | +|--------|--------|-------------| +| 0 | Neural | Base model log-softmax | +| 1 | Unigram | Token frequency from scored tokens | +| 2 | Bigram | P(next \| prev) from scored tokens | +| 3 | Trigram | Hashed P(next \| prev2, prev1) with 64K buckets | +| 4 | Entropy | Neural model entropy as confidence regularizer | + +Expert weights are updated online via Hedge: `log_w -= eta * loss`. N-gram tables are built incrementally from already-scored tokens only (legal). + +## Architecture + +PR #606 base with the following additions: + +| Component | Setting | +|-----------|---------| +| Layers | 11 (512d, 8H, 8KV) | +| MLP | 3x with **LeakyReLU(0.5)^2** | +| BigramHash | 6144 (dim=128) | +| XSA | All 11 layers (ws=8) | +| RoPE | Partial (16/64 dims) | +| LN Scale | 1/sqrt(layer+1) | +| VE128 | Layers 9-10 | +| Weight avg | EMA(0.997) | +| Quantization | Full GPTQ int5 + zstd (level 22) | +| Pruning | 3% magnitude | + +## Legal Score-First TTT + +Backward-looking adaptation with GPTQ-calibrated model: + +1. Validation tokens split into 474 chunks of 131K tokens each +2. For each chunk: + - **SCORE**: Sliding window eval (stride=32, seq_len=2048) with 5-expert mixer blending + - **TRAIN**: AdamW(lr=0.0001) on already-scored chunk. 3 epochs, last 2 blocks unfrozen + norms + lm_head, cosine LR decay, Polyak averaging +3. Last chunk scored but never trained on + +### TTT Hyperparameters + +| Parameter | Value | +|-----------|-------| +| Chunk size | 131,072 tokens | +| Optimizer | AdamW (lr=0.0001) | +| Epochs per chunk | 3 | +| Frozen blocks | First 9 (last 2 + norms + head unfrozen) | +| Polyak decay | 0.998 | +| Adaptive LR | max_mult=3.0 | +| Mixer eta | 0.1 | + +### Training Budget + +GPTQ calibration runs within the 600s training budget (18s reserved from training loop for EMA selection + calibration + quantization). + +| Phase | Time | +|-------|------| +| Training loop | 582s | +| EMA + GPTQ calibration + quantization | ~18s | +| **Total training** | **~600s** | +| Sliding window eval | ~165s | +| TTT eval with mixer | ~562s | +| **Total eval** | **~562s** | + +## Reproduction + +```bash +# Install dependencies +pip install -r requirements.txt +# Build FA3 Hopper kernels (required) +cd /tmp && git clone https://github.com/Dao-AILab/flash-attention +cd flash-attention/hopper && python setup.py install + +# Run training + eval (single seed) +DATA_PATH=./data/datasets/fineweb10B_sp1024 \ +TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model \ +SEED=1337 MAX_WALLCLOCK_SECONDS=600 \ +USE_MIXER=1 TTT_LR=0.0001 TTT_CHUNK_TOKENS=131072 \ + torchrun --standalone --nproc_per_node=8 train_gpt.py + +# Run all 3 seeds +for SEED in 1337 42 7; do + DATA_PATH=./data/datasets/fineweb10B_sp1024 \ + TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model \ + SEED=$SEED MAX_WALLCLOCK_SECONDS=600 \ + USE_MIXER=1 TTT_LR=0.0001 TTT_CHUNK_TOKENS=131072 \ + torchrun --standalone --nproc_per_node=8 train_gpt.py +done +``` + +## Credits + +- **Base model**: PR #606 by @gowtham0992 +- **TTT recipe**: PR #461 by @Christopher-Lee-McClendon +- **Mixer inspiration**: PAQ compression (context mixing) + Hedge algorithm diff --git a/submission-2026-03-24/log_seed1337.txt b/submission-2026-03-24/log_seed1337.txt new file mode 100644 index 000000000..656321ad9 --- /dev/null +++ b/submission-2026-03-24/log_seed1337.txt @@ -0,0 +1,173 @@ +W0325 00:33:17.141000 673732 torch/distributed/run.py:852] +W0325 00:33:17.141000 673732 torch/distributed/run.py:852] ***************************************** +W0325 00:33:17.141000 673732 torch/distributed/run.py:852] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0325 00:33:17.141000 673732 torch/distributed/run.py:852] ***************************************** +logs/31484e99-50c6-404c-8bb8-635f618baafa.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=../data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=../data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +mixed_precision: 68 int5 layers, 0 int6 layers (last 0 blocks) +model_params:33317980 +XSA:[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] ws:8 gqa:8/8 +lr:embed=0.035 matrix=0.025 scalar=0.025 batch:786432 wall:600s seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9285 val_bpb:4.1034 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9305 train_time:148ms step_avg:147.68ms +step:2/20000 train_loss:8.6412 train_time:239ms step_avg:119.37ms +step:3/20000 train_loss:7.7278 train_time:333ms step_avg:111.04ms +step:4/20000 train_loss:7.2812 train_time:428ms step_avg:106.95ms +step:5/20000 train_loss:7.0672 train_time:524ms step_avg:104.81ms +step:6/20000 train_loss:6.9647 train_time:619ms step_avg:103.14ms +step:7/20000 train_loss:6.8519 train_time:714ms step_avg:102.01ms +step:8/20000 train_loss:6.7091 train_time:809ms step_avg:101.08ms +step:9/20000 train_loss:6.3640 train_time:903ms step_avg:100.36ms +step:10/20000 train_loss:6.0314 train_time:998ms step_avg:99.77ms +step:500/20000 train_loss:2.3594 train_time:48287ms step_avg:96.57ms +step:1000/20000 train_loss:2.2366 train_time:96650ms step_avg:96.65ms +step:1500/20000 train_loss:2.1871 train_time:145067ms step_avg:96.71ms +step:2000/20000 train_loss:2.0272 train_time:193559ms step_avg:96.78ms +step:2500/20000 train_loss:2.1332 train_time:242098ms step_avg:96.84ms +step:3000/20000 train_loss:2.1140 train_time:290638ms step_avg:96.88ms +step:3500/20000 train_loss:2.1198 train_time:339166ms step_avg:96.90ms +step:4000/20000 train_loss:1.9079 train_time:387684ms step_avg:96.92ms +step:4000/20000 val_loss:2.0000 val_bpb:1.1845 train_time:387689ms step_avg:96.92ms +late_qat:enabled step:4255 scale:0.4998 +step:4500/20000 train_loss:2.0575 train_time:436219ms step_avg:96.94ms +step:5000/20000 train_loss:2.0307 train_time:484743ms step_avg:96.95ms +swa:start step:5350 +step:5500/20000 train_loss:1.9404 train_time:533445ms step_avg:96.99ms +step:5997/20000 val_loss:1.9019 val_bpb:1.1264 train_time:582051ms step_avg:97.06ms +stopping_early: wallclock_cap train_time:582051ms step:5997/20000 +peak memory allocated: 26200 MiB reserved: 26782 MiB +ema:applying EMA weights (skipping diagnostic evals) +gptq:calibrating with training data... +gptq:calibrated 68 layers in 1.9s +Serialized model: 130432585 bytes +Code size: 96428 bytes +pruning:3.0% magnitude pruning applied +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +Serialized model int6+zstd: 15387977 bytes +Total submission size int6+zstd: 15484405 bytes + ttt: pre-compiling forward+backward kernels... + ttt: pre-compile done +final_int6_sliding_window val_loss:1.8992 val_bpb:1.1248 stride:32 eval_time:164968ms +final_int6_sliding_window_exact val_loss:1.89920781 val_bpb:1.12482157 +TTT: epochs=3 lr=0.0001 freeze_first=2 chunk=131072 opt=adamw +TTT temperature: 0.98 +PPM alpha: 0.85, Byte-weighted TTT: True + Logistic context mixer enabled: eta=0.1 + Adaptive LR enabled: max_mult=3.0 +ttt:start chunks=474 chunk_tokens=131072 windows=1938176 stride=32 lr=0.0001 epochs=3 opt=adamw freeze_first=2 +ttt:params unfrozen=5780500 frozen=27537480 + Polyak averaging enabled: decay=0.998 + ttt_train [1] seqs=64 start_train... + ttt_train [1] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.3437 + ttt_train [1] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.3101 + ttt_train [1] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.3086 + ttt_chunk [1/474] bpb=1.202687 time=1.3s + ttt_train [2] seqs=64 start_train... + ttt_train [2] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.1361 + ttt_train [2] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.1334 + ttt_train [2] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.1283 + ttt_chunk [2/474] bpb=1.128760 time=2.5s + ttt_train [3] seqs=64 start_train... + ttt_train [3] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.0525 + ttt_train [3] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.0510 + ttt_train [3] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.0484 + ttt_chunk [3/474] bpb=1.080018 time=3.7s + ttt_chunk [4/474] bpb=1.076079 time=4.8s + ttt_chunk [5/474] bpb=1.066728 time=6.0s + ttt_chunk [11/474] bpb=1.031430 time=13.1s + ttt_chunk [21/474] bpb=1.018884 time=25.0s + ttt_chunk [31/474] bpb=1.015939 time=36.9s + ttt_chunk [41/474] bpb=1.022519 time=48.7s + ttt_chunk [51/474] bpb=1.028238 time=60.6s + ttt_chunk [61/474] bpb=1.025430 time=72.5s + ttt_chunk [71/474] bpb=1.026528 time=84.3s + ttt_chunk [81/474] bpb=1.026915 time=96.2s + ttt_chunk [91/474] bpb=1.028643 time=108.1s + ttt_chunk [101/474] bpb=1.025107 time=119.9s + ttt_chunk [111/474] bpb=1.024927 time=131.8s + ttt_chunk [121/474] bpb=1.027728 time=143.7s + ttt_chunk [131/474] bpb=1.027818 time=155.5s + ttt_chunk [141/474] bpb=1.026649 time=167.4s + ttt_chunk [151/474] bpb=1.024273 time=179.3s + ttt_chunk [161/474] bpb=1.024436 time=191.2s + ttt_chunk [171/474] bpb=1.022907 time=203.0s + ttt_chunk [181/474] bpb=1.023651 time=214.9s + ttt_chunk [191/474] bpb=1.022518 time=226.8s + ttt_chunk [201/474] bpb=1.021847 time=238.6s + ttt_chunk [211/474] bpb=1.020784 time=250.5s + ttt_chunk [221/474] bpb=1.021277 time=262.4s + ttt_chunk [231/474] bpb=1.020969 time=274.2s + ttt_chunk [241/474] bpb=1.020429 time=286.1s + ttt_chunk [251/474] bpb=1.022389 time=298.0s + ttt_chunk [261/474] bpb=1.024492 time=309.8s + ttt_chunk [271/474] bpb=1.024714 time=321.7s + ttt_chunk [281/474] bpb=1.026263 time=333.6s + ttt_chunk [291/474] bpb=1.027062 time=345.4s + ttt_chunk [301/474] bpb=1.029658 time=357.3s + ttt_chunk [311/474] bpb=1.031598 time=369.2s + ttt_chunk [321/474] bpb=1.032553 time=381.0s + ttt_chunk [331/474] bpb=1.034020 time=392.9s + ttt_chunk [341/474] bpb=1.035727 time=404.8s + ttt_chunk [351/474] bpb=1.036601 time=416.6s + ttt_chunk [361/474] bpb=1.039560 time=428.5s + ttt_chunk [371/474] bpb=1.041485 time=440.4s + ttt_chunk [381/474] bpb=1.044560 time=452.3s + ttt_chunk [391/474] bpb=1.047764 time=464.1s + ttt_chunk [401/474] bpb=1.050371 time=476.0s + ttt_chunk [411/474] bpb=1.052690 time=487.9s + ttt_chunk [421/474] bpb=1.055963 time=499.8s + ttt_chunk [431/474] bpb=1.056364 time=511.6s + ttt_chunk [441/474] bpb=1.057641 time=523.5s + ttt_chunk [451/474] bpb=1.058576 time=535.4s + ttt_chunk [461/474] bpb=1.060302 time=547.2s + ttt_chunk [471/474] bpb=1.061844 time=559.1s + ttt_chunk [474/474] bpb=1.062011 time=561.7s +ttt:done val_loss=1.783077 val_bpb=1.056042 elapsed=561.7s +final_int6_ttt val_loss:1.7831 val_bpb:1.0560 stride:32 eval_time:562538ms +final_int6_ttt_exact val_loss:1.78307674 val_bpb:1.05604198 diff --git a/submission-2026-03-24/log_seed42.txt b/submission-2026-03-24/log_seed42.txt new file mode 100644 index 000000000..a185217fe --- /dev/null +++ b/submission-2026-03-24/log_seed42.txt @@ -0,0 +1,173 @@ +W0325 00:58:00.120000 677084 torch/distributed/run.py:852] +W0325 00:58:00.120000 677084 torch/distributed/run.py:852] ***************************************** +W0325 00:58:00.120000 677084 torch/distributed/run.py:852] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0325 00:58:00.120000 677084 torch/distributed/run.py:852] ***************************************** +logs/e0b62b26-4d67-4f38-8e55-7f891da41982.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=../data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=../data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +mixed_precision: 68 int5 layers, 0 int6 layers (last 0 blocks) +model_params:33317980 +XSA:[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] ws:8 gqa:8/8 +lr:embed=0.035 matrix=0.025 scalar=0.025 batch:786432 wall:600s seed:42 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9301 val_bpb:4.1044 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9309 train_time:146ms step_avg:146.46ms +step:2/20000 train_loss:8.7072 train_time:237ms step_avg:118.44ms +step:3/20000 train_loss:7.7698 train_time:331ms step_avg:110.45ms +step:4/20000 train_loss:7.2987 train_time:427ms step_avg:106.64ms +step:5/20000 train_loss:7.0366 train_time:521ms step_avg:104.22ms +step:6/20000 train_loss:6.9492 train_time:616ms step_avg:102.74ms +step:7/20000 train_loss:6.8511 train_time:711ms step_avg:101.54ms +step:8/20000 train_loss:6.7338 train_time:805ms step_avg:100.67ms +step:9/20000 train_loss:6.3596 train_time:900ms step_avg:100.05ms +step:10/20000 train_loss:6.0340 train_time:996ms step_avg:99.58ms +step:500/20000 train_loss:2.3619 train_time:48288ms step_avg:96.58ms +step:1000/20000 train_loss:2.2442 train_time:96649ms step_avg:96.65ms +step:1500/20000 train_loss:2.1847 train_time:145067ms step_avg:96.71ms +step:2000/20000 train_loss:2.0315 train_time:193558ms step_avg:96.78ms +step:2500/20000 train_loss:2.1358 train_time:242084ms step_avg:96.83ms +step:3000/20000 train_loss:2.1167 train_time:290616ms step_avg:96.87ms +step:3500/20000 train_loss:2.1233 train_time:339152ms step_avg:96.90ms +step:4000/20000 train_loss:1.9137 train_time:387678ms step_avg:96.92ms +step:4000/20000 val_loss:2.0017 val_bpb:1.1855 train_time:387683ms step_avg:96.92ms +late_qat:enabled step:4255 scale:0.4998 +step:4500/20000 train_loss:2.0589 train_time:436214ms step_avg:96.94ms +step:5000/20000 train_loss:2.0322 train_time:484739ms step_avg:96.95ms +swa:start step:5350 +step:5500/20000 train_loss:1.9442 train_time:533403ms step_avg:96.98ms +step:5997/20000 val_loss:1.9035 val_bpb:1.1274 train_time:582007ms step_avg:97.05ms +stopping_early: wallclock_cap train_time:582007ms step:5997/20000 +peak memory allocated: 26197 MiB reserved: 26782 MiB +ema:applying EMA weights (skipping diagnostic evals) +gptq:calibrating with training data... +gptq:calibrated 68 layers in 1.9s +Serialized model: 130432585 bytes +Code size: 96428 bytes +pruning:3.0% magnitude pruning applied +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +Serialized model int6+zstd: 15312167 bytes +Total submission size int6+zstd: 15408595 bytes + ttt: pre-compiling forward+backward kernels... + ttt: pre-compile done +final_int6_sliding_window val_loss:1.9006 val_bpb:1.1257 stride:32 eval_time:164957ms +final_int6_sliding_window_exact val_loss:1.90062694 val_bpb:1.12566206 +TTT: epochs=3 lr=0.0001 freeze_first=2 chunk=131072 opt=adamw +TTT temperature: 0.98 +PPM alpha: 0.85, Byte-weighted TTT: True + Logistic context mixer enabled: eta=0.1 + Adaptive LR enabled: max_mult=3.0 +ttt:start chunks=474 chunk_tokens=131072 windows=1938176 stride=32 lr=0.0001 epochs=3 opt=adamw freeze_first=2 +ttt:params unfrozen=5780500 frozen=27537480 + Polyak averaging enabled: decay=0.998 + ttt_train [1] seqs=64 start_train... + ttt_train [1] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.3462 + ttt_train [1] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.3103 + ttt_train [1] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.3076 + ttt_chunk [1/474] bpb=1.202959 time=1.2s + ttt_train [2] seqs=64 start_train... + ttt_train [2] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.1330 + ttt_train [2] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.1310 + ttt_train [2] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.1268 + ttt_chunk [2/474] bpb=1.129301 time=2.4s + ttt_train [3] seqs=64 start_train... + ttt_train [3] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.0619 + ttt_train [3] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.0610 + ttt_train [3] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.0582 + ttt_chunk [3/474] bpb=1.079049 time=3.6s + ttt_chunk [4/474] bpb=1.074949 time=4.8s + ttt_chunk [5/474] bpb=1.064662 time=6.0s + ttt_chunk [11/474] bpb=1.025191 time=13.1s + ttt_chunk [21/474] bpb=1.008678 time=25.0s + ttt_chunk [31/474] bpb=1.002098 time=36.9s + ttt_chunk [41/474] bpb=1.007125 time=48.7s + ttt_chunk [51/474] bpb=1.012485 time=60.6s + ttt_chunk [61/474] bpb=1.010685 time=72.5s + ttt_chunk [71/474] bpb=1.013076 time=84.4s + ttt_chunk [81/474] bpb=1.015652 time=96.2s + ttt_chunk [91/474] bpb=1.019922 time=108.1s + ttt_chunk [101/474] bpb=1.019995 time=120.0s + ttt_chunk [111/474] bpb=1.024232 time=131.9s + ttt_chunk [121/474] bpb=1.031908 time=143.7s + ttt_chunk [131/474] bpb=1.037128 time=155.6s + ttt_chunk [141/474] bpb=1.041120 time=167.5s + ttt_chunk [151/474] bpb=1.044067 time=179.4s + ttt_chunk [161/474] bpb=1.049148 time=191.3s + ttt_chunk [171/474] bpb=1.052038 time=203.1s + ttt_chunk [181/474] bpb=1.057080 time=215.0s + ttt_chunk [191/474] bpb=1.059619 time=226.9s + ttt_chunk [201/474] bpb=1.062378 time=238.8s + ttt_chunk [211/474] bpb=1.064340 time=250.6s + ttt_chunk [221/474] bpb=1.067700 time=262.5s + ttt_chunk [231/474] bpb=1.069761 time=274.4s + ttt_chunk [241/474] bpb=1.071136 time=286.3s + ttt_chunk [251/474] bpb=1.074548 time=298.2s + ttt_chunk [261/474] bpb=1.077509 time=310.0s + ttt_chunk [271/474] bpb=1.078241 time=321.9s + ttt_chunk [281/474] bpb=1.079767 time=333.8s + ttt_chunk [291/474] bpb=1.080218 time=345.7s + ttt_chunk [301/474] bpb=1.082217 time=357.5s + ttt_chunk [311/474] bpb=1.083294 time=369.4s + ttt_chunk [321/474] bpb=1.083302 time=381.3s + ttt_chunk [331/474] bpb=1.083667 time=393.2s + ttt_chunk [341/474] bpb=1.084241 time=405.0s + ttt_chunk [351/474] bpb=1.083985 time=416.9s + ttt_chunk [361/474] bpb=1.085885 time=428.8s + ttt_chunk [371/474] bpb=1.086654 time=440.7s + ttt_chunk [381/474] bpb=1.088639 time=452.5s + ttt_chunk [391/474] bpb=1.090740 time=464.4s + ttt_chunk [401/474] bpb=1.092278 time=476.3s + ttt_chunk [411/474] bpb=1.093605 time=488.2s + ttt_chunk [421/474] bpb=1.095927 time=500.1s + ttt_chunk [431/474] bpb=1.095382 time=511.9s + ttt_chunk [441/474] bpb=1.095777 time=523.8s + ttt_chunk [451/474] bpb=1.095835 time=535.7s + ttt_chunk [461/474] bpb=1.096740 time=547.6s + ttt_chunk [471/474] bpb=1.097514 time=559.5s + ttt_chunk [474/474] bpb=1.097488 time=562.1s +ttt:done val_loss=1.852181 val_bpb=1.096969 elapsed=562.1s +final_int6_ttt val_loss:1.8522 val_bpb:1.0970 stride:32 eval_time:562886ms +final_int6_ttt_exact val_loss:1.85218082 val_bpb:1.09696945 diff --git a/submission-2026-03-24/log_seed7.txt b/submission-2026-03-24/log_seed7.txt new file mode 100644 index 000000000..fcaf81b2e --- /dev/null +++ b/submission-2026-03-24/log_seed7.txt @@ -0,0 +1,173 @@ +W0325 01:22:56.941000 679893 torch/distributed/run.py:852] +W0325 01:22:56.941000 679893 torch/distributed/run.py:852] ***************************************** +W0325 01:22:56.941000 679893 torch/distributed/run.py:852] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0325 01:22:56.941000 679893 torch/distributed/run.py:852] ***************************************** +logs/fb9c44fd-95a9-4cb3-a79d-3bf240580a1a.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=../data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=../data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +mixed_precision: 68 int5 layers, 0 int6 layers (last 0 blocks) +model_params:33317980 +XSA:[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] ws:8 gqa:8/8 +lr:embed=0.035 matrix=0.025 scalar=0.025 batch:786432 wall:600s seed:7 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9294 val_bpb:4.1040 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9310 train_time:151ms step_avg:150.83ms +step:2/20000 train_loss:8.7138 train_time:240ms step_avg:120.17ms +step:3/20000 train_loss:7.7624 train_time:335ms step_avg:111.68ms +step:4/20000 train_loss:7.2817 train_time:430ms step_avg:107.46ms +step:5/20000 train_loss:7.1027 train_time:525ms step_avg:104.94ms +step:6/20000 train_loss:6.9241 train_time:619ms step_avg:103.24ms +step:7/20000 train_loss:6.8669 train_time:714ms step_avg:102.05ms +step:8/20000 train_loss:6.6918 train_time:809ms step_avg:101.14ms +step:9/20000 train_loss:6.3601 train_time:904ms step_avg:100.46ms +step:10/20000 train_loss:6.0035 train_time:999ms step_avg:99.88ms +step:500/20000 train_loss:2.3626 train_time:48426ms step_avg:96.85ms +step:1000/20000 train_loss:2.2381 train_time:96920ms step_avg:96.92ms +step:1500/20000 train_loss:2.1845 train_time:145469ms step_avg:96.98ms +step:2000/20000 train_loss:2.0268 train_time:194077ms step_avg:97.04ms +step:2500/20000 train_loss:2.1315 train_time:242724ms step_avg:97.09ms +step:3000/20000 train_loss:2.1123 train_time:291375ms step_avg:97.13ms +step:3500/20000 train_loss:2.1194 train_time:340018ms step_avg:97.15ms +step:4000/20000 train_loss:1.9084 train_time:388663ms step_avg:97.17ms +step:4000/20000 val_loss:1.9988 val_bpb:1.1838 train_time:388668ms step_avg:97.17ms +late_qat:enabled step:4240 scale:0.4998 +step:4500/20000 train_loss:2.0531 train_time:437292ms step_avg:97.18ms +step:5000/20000 train_loss:2.0297 train_time:485905ms step_avg:97.18ms +swa:start step:5300 +step:5500/20000 train_loss:1.9406 train_time:534723ms step_avg:97.22ms +step:5983/20000 val_loss:1.9008 val_bpb:1.1258 train_time:582062ms step_avg:97.29ms +stopping_early: wallclock_cap train_time:582062ms step:5983/20000 +peak memory allocated: 26197 MiB reserved: 26782 MiB +ema:applying EMA weights (skipping diagnostic evals) +gptq:calibrating with training data... +gptq:calibrated 68 layers in 2.0s +Serialized model: 130432585 bytes +Code size: 96428 bytes +pruning:3.0% magnitude pruning applied +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +Serialized model int6+zstd: 15331691 bytes +Total submission size int6+zstd: 15428119 bytes + ttt: pre-compiling forward+backward kernels... + ttt: pre-compile done +final_int6_sliding_window val_loss:1.8997 val_bpb:1.1251 stride:32 eval_time:165026ms +final_int6_sliding_window_exact val_loss:1.89974539 val_bpb:1.12513995 +TTT: epochs=3 lr=0.0001 freeze_first=2 chunk=131072 opt=adamw +TTT temperature: 0.98 +PPM alpha: 0.85, Byte-weighted TTT: True + Logistic context mixer enabled: eta=0.1 + Adaptive LR enabled: max_mult=3.0 +ttt:start chunks=474 chunk_tokens=131072 windows=1938176 stride=32 lr=0.0001 epochs=3 opt=adamw freeze_first=2 +ttt:params unfrozen=5780500 frozen=27537480 + Polyak averaging enabled: decay=0.998 + ttt_train [1] seqs=64 start_train... + ttt_train [1] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.3476 + ttt_train [1] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.3121 + ttt_train [1] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.3099 + ttt_chunk [1/474] bpb=1.199612 time=1.2s + ttt_train [2] seqs=64 start_train... + ttt_train [2] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.1304 + ttt_train [2] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.1283 + ttt_train [2] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.1239 + ttt_chunk [2/474] bpb=1.128925 time=2.4s + ttt_train [3] seqs=64 start_train... + ttt_train [3] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.0552 + ttt_train [3] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.0540 + ttt_train [3] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.0517 + ttt_chunk [3/474] bpb=1.081615 time=3.6s + ttt_chunk [4/474] bpb=1.078973 time=4.8s + ttt_chunk [5/474] bpb=1.069501 time=5.9s + ttt_chunk [11/474] bpb=1.037589 time=13.0s + ttt_chunk [21/474] bpb=1.028394 time=24.9s + ttt_chunk [31/474] bpb=1.028114 time=36.7s + ttt_chunk [41/474] bpb=1.038811 time=48.6s + ttt_chunk [51/474] bpb=1.048907 time=60.4s + ttt_chunk [61/474] bpb=1.050285 time=72.2s + ttt_chunk [71/474] bpb=1.055243 time=84.1s + ttt_chunk [81/474] bpb=1.059893 time=95.9s + ttt_chunk [91/474] bpb=1.065566 time=107.7s + ttt_chunk [101/474] bpb=1.065416 time=119.6s + ttt_chunk [111/474] bpb=1.068729 time=131.4s + ttt_chunk [121/474] bpb=1.074869 time=143.3s + ttt_chunk [131/474] bpb=1.078175 time=155.1s + ttt_chunk [141/474] bpb=1.080151 time=166.9s + ttt_chunk [151/474] bpb=1.080074 time=178.8s + ttt_chunk [161/474] bpb=1.082475 time=190.6s + ttt_chunk [171/474] bpb=1.082532 time=202.5s + ttt_chunk [181/474] bpb=1.084595 time=214.3s + ttt_chunk [191/474] bpb=1.084726 time=226.1s + ttt_chunk [201/474] bpb=1.084927 time=238.0s + ttt_chunk [211/474] bpb=1.084565 time=249.8s + ttt_chunk [221/474] bpb=1.085631 time=261.6s + ttt_chunk [231/474] bpb=1.085931 time=273.5s + ttt_chunk [241/474] bpb=1.085693 time=285.3s + ttt_chunk [251/474] bpb=1.087461 time=297.1s + ttt_chunk [261/474] bpb=1.088614 time=309.0s + ttt_chunk [271/474] bpb=1.087539 time=320.8s + ttt_chunk [281/474] bpb=1.087260 time=332.7s + ttt_chunk [291/474] bpb=1.085792 time=344.5s + ttt_chunk [301/474] bpb=1.085870 time=356.3s + ttt_chunk [311/474] bpb=1.084865 time=368.2s + ttt_chunk [321/474] bpb=1.082823 time=380.0s + ttt_chunk [331/474] bpb=1.081088 time=391.9s + ttt_chunk [341/474] bpb=1.079459 time=403.7s + ttt_chunk [351/474] bpb=1.076696 time=415.5s + ttt_chunk [361/474] bpb=1.076032 time=427.4s + ttt_chunk [371/474] bpb=1.074447 time=439.2s + ttt_chunk [381/474] bpb=1.073773 time=451.0s + ttt_chunk [391/474] bpb=1.073248 time=462.9s + ttt_chunk [401/474] bpb=1.072241 time=474.7s + ttt_chunk [411/474] bpb=1.071242 time=486.6s + ttt_chunk [421/474] bpb=1.071481 time=498.4s + ttt_chunk [431/474] bpb=1.069344 time=510.2s + ttt_chunk [441/474] bpb=1.068440 time=522.1s + ttt_chunk [451/474] bpb=1.067601 time=533.9s + ttt_chunk [461/474] bpb=1.068029 time=545.7s + ttt_chunk [471/474] bpb=1.068628 time=557.6s + ttt_chunk [474/474] bpb=1.068643 time=560.2s +ttt:done val_loss=1.807296 val_bpb=1.070386 elapsed=560.2s +final_int6_ttt val_loss:1.8073 val_bpb:1.0704 stride:32 eval_time:561081ms +final_int6_ttt_exact val_loss:1.80729571 val_bpb:1.07038587 diff --git a/submission-2026-03-24/requirements.txt b/submission-2026-03-24/requirements.txt new file mode 100644 index 000000000..a124ed4f5 --- /dev/null +++ b/submission-2026-03-24/requirements.txt @@ -0,0 +1,5 @@ +torch>=2.10.0 +numpy +sentencepiece +zstandard +flash-attn-3 diff --git a/submission-2026-03-24/submission.json b/submission-2026-03-24/submission.json new file mode 100644 index 000000000..2835282df --- /dev/null +++ b/submission-2026-03-24/submission.json @@ -0,0 +1,14 @@ +{ + "track": "10min_16mb", + "date": "2026-03-24", + "name": "5-expert Hedge Mixer + TTT", + "author": "notapplica", + "seed_results": { + "1337": {"val_loss": 1.78307674, "val_bpb": 1.05604198, "artifact_bytes": 15484405}, + "42": {"val_loss": 1.85218082, "val_bpb": 1.09696945, "artifact_bytes": 15408595}, + "7": {"val_loss": 1.80729571, "val_bpb": 1.07038587, "artifact_bytes": 15428119} + }, + "mean_val_loss": 1.81418442, + "mean_val_bpb": 1.07446577, + "code_bytes": 96416 +} diff --git a/submission-2026-03-24/train_gpt.py b/submission-2026-03-24/train_gpt.py new file mode 100644 index 000000000..7460ef9be --- /dev/null +++ b/submission-2026-03-24/train_gpt.py @@ -0,0 +1,1966 @@ +"""V25: LeakyReLU^2 + TempCal + Mixed int5/int6 + 33.6M model.""" +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path +try: + import zstandard + _COMPRESSOR = "zstd" +except ImportError: + _COMPRESSOR = "zlib" +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +try: + from flash_attn_interface import flash_attn_func as flash_attn_3_func + _HAS_FA3 = True +except ImportError: + try: + from flash_attn import flash_attn_func as flash_attn_3_func + _HAS_FA3 = True + except ImportError: + _HAS_FA3 = False + flash_attn_3_func = None + +class LogisticContextMixer: + """GPU-vectorized logistic context mixing (inspired by PAQ compression). + + Maintains GPU-resident n-gram count tables and learns online mixing weights + using the Hedge/multiplicative-weights algorithm. + + Experts: + 0: Neural model (logits passed in) + 1: Unigram frequencies from scored tokens + 2: Bigram frequencies (prev_token → next_token) + 3: FastPPM (orders 0-4, CPU-side) + 4: ExactMatchCache (high-order exact matches, CPU-side) + """ + + def __init__(self, vocab_size: int = 1024, device: str = 'cuda', eta: float = 0.1): + self.V = vocab_size + self.device = device + self.eta = eta # Hedge learning rate + self.K = 5 # number of experts + + # Expert weights (log-domain for numerical stability) + self.log_weights = torch.zeros(self.K, device=device) + # Bias toward neural model initially + self.log_weights[0] = 2.0 + + # N-gram count tables (GPU-resident) + self.uni_counts = torch.zeros(vocab_size, device=device) + self.bi_counts = torch.zeros(vocab_size, vocab_size, device=device) + self.total_tokens = 0 + + # GPU Trigram: hashed table [HASH_SIZE, V] to keep memory reasonable + self.TRI_HASH = 65536 # 64K hash buckets for (prev2, prev1) pairs + self.tri_counts = torch.zeros(self.TRI_HASH, vocab_size, device=device) + self.tri_row_totals = torch.zeros(self.TRI_HASH, device=device) + + def update(self, tokens): + """Update all expert statistics with newly scored tokens.""" + if hasattr(tokens, 'cpu'): + t = tokens.to(self.device).long() + else: + t = torch.tensor(tokens, device=self.device, dtype=torch.long) + + n = t.numel() + if n == 0: + return + self.total_tokens += n + + # Unigram: in-place scatter_add + ones = torch.ones(n, device=self.device) + self.uni_counts.scatter_add_(0, t, ones) + + # Bigram: in-place scatter_add on flattened view (no temporary 1M tensor) + if n >= 2: + ctx = t[:-1] + nxt = t[1:] + bi_idx = ctx * self.V + nxt + ones_bi = torch.ones(n - 1, device=self.device) + self.bi_counts.reshape(-1).scatter_add_(0, bi_idx, ones_bi) + + # Trigram: in-place scatter_add on flattened view (no temporary 67M tensor) + if n >= 3: + prev2 = t[:-2] + prev1 = t[1:-1] + nxt3 = t[2:] + tri_ctx = ((prev2 * 36313) ^ (prev1 * 27191)) % self.TRI_HASH + tri_idx = tri_ctx * self.V + nxt3 + ones_tri = torch.ones(n - 2, device=self.device) + self.tri_counts.reshape(-1).scatter_add_(0, tri_idx, ones_tri) + self.tri_row_totals.scatter_add_(0, tri_ctx, ones_tri) + + def get_expert_log_probs(self, neural_logits, x_batch, y_batch, wlens): + """Get log-probability of targets from each expert. All GPU-vectorized. + + Args: + neural_logits: [bsz, seq_len, V] neural model logits + x_batch: [bsz, seq_len] input tokens (context) + y_batch: [bsz, seq_len] target tokens + wlens: list of actual lengths per sequence + + Returns: + expert_nll: [bsz, seq_len, K] NLL from each expert + """ + bsz, slen, V = neural_logits.shape + uniform_nll = math.log(self.V) + has_data = self.total_tokens > 0 # Python int — no GPU-CPU sync + + # Expert 0: Neural model — compute log_softmax once, reuse for entropy + neural_lp = F.log_softmax(neural_logits, dim=-1) + neural_nll = -neural_lp.gather(2, y_batch.unsqueeze(2)).squeeze(2) # [bsz, slen] + + # Expert 1: Unigram + if has_data: + uni_probs = (self.uni_counts + 0.1) / (self.total_tokens + 0.1 * self.V) + uni_nll = -uni_probs.log()[y_batch] # [bsz, slen] + else: + uni_nll = torch.full((bsz, slen), uniform_nll, device=self.device) + + # Expert 2: Bigram P(next | prev) + if has_data: + bi_total = self.bi_counts.sum(dim=1, keepdim=True) # [V, 1] + bi_probs = (self.bi_counts + 0.1) / (bi_total + 0.1 * self.V) # [V, V] + prev_flat = x_batch.reshape(-1) + next_flat = y_batch.reshape(-1) + bi_nll = -bi_probs.log()[prev_flat, next_flat].reshape(bsz, slen) + else: + bi_nll = torch.full((bsz, slen), uniform_nll, device=self.device) + + # Expert 3: GPU Trigram P(next | hash(prev2, prev1)) — vectorized + if has_data and slen >= 2: + prev2 = torch.zeros_like(x_batch) + prev2[:, 1:] = x_batch[:, :-1] + ctx_hash = ((prev2 * 36313) ^ (x_batch * 27191)) % self.TRI_HASH + ctx_flat = ctx_hash.reshape(-1).long() + next_flat = y_batch.reshape(-1).long() + tri_count = self.tri_counts[ctx_flat, next_flat] + tri_total = self.tri_row_totals[ctx_flat].clamp(min=1) + tri_prob = (tri_count + 0.01) / (tri_total + 0.01 * self.V) + tri_nll = -tri_prob.log().reshape(bsz, slen) + else: + tri_nll = torch.full((bsz, slen), uniform_nll, device=self.device) + + # Expert 4: Neural entropy — reuse neural_lp (no redundant softmax) + entropy_nll = -(neural_lp.exp() * neural_lp).sum(-1) # [bsz, slen] + + # Stack: [bsz, slen, K] + return torch.stack([neural_nll, uni_nll, bi_nll, tri_nll, entropy_nll], dim=-1) + + def mix_and_score(self, neural_logits, x_batch, y_batch, wlens): + """Compute mixed NLL using current expert weights. + + Returns (mixed_nll [bsz, slen], expert_nll [bsz, slen, K] or None). + Caller should pass expert_nll to update_weights() to avoid recomputation. + """ + if self.total_tokens < 10000: + # Not enough data for n-grams — just use neural + nll = F.cross_entropy( + neural_logits.reshape(-1, neural_logits.size(-1)), + y_batch.reshape(-1), reduction="none" + ).reshape(neural_logits.shape[0], neural_logits.shape[1]) + return nll, None + + expert_nll = self.get_expert_log_probs(neural_logits, x_batch, y_batch, wlens) # [bsz, slen, K] + + # Log-domain mixing: log(sum_k w_k * p_k) = logsumexp(log_w_k + log_p_k) + log_w = self.log_weights - self.log_weights.logsumexp(0) # normalize + mixed_lp = (-expert_nll + log_w.unsqueeze(0).unsqueeze(0)).logsumexp(dim=-1) # [bsz, slen] + + return -mixed_lp, expert_nll # mixed NLL + cached expert NLL + + def update_weights(self, expert_nll, wlens): + """Update expert weights using Hedge algorithm on pre-computed expert NLLs.""" + if expert_nll is None: + return + + with torch.no_grad(): + # Vectorized mask: compare position index against window lengths + bsz, slen = expert_nll.shape[0], expert_nll.shape[1] + wlens_t = torch.tensor(wlens, device=self.device, dtype=torch.long) + mask = torch.arange(slen, device=self.device).unsqueeze(0) < wlens_t.unsqueeze(1) # [bsz, slen] bool + + # Masked mean NLL per expert + masked_nll = expert_nll * mask.unsqueeze(-1).float() + expert_mean_loss = masked_nll.sum(dim=(0, 1)) / mask.sum().clamp(min=1) # [K] + + # Hedge update: log_w -= eta * loss + self.log_weights -= self.eta * expert_mean_loss + + +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 8)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.5)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 32)) + int6_last_n = int(os.environ.get("INT6_LAST_N", 0)) # all int5 (saves ~300KB vs int6 for last 2 blocks) + ttt_temperature = float(os.environ.get("TTT_TEMPERATURE", 0.98)) # post-TTT temperature calibration + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 6144)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 11)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.5)) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + prune_pct = float(os.environ.get("PRUNE_PCT", 0.03)) + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X + +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] + +def eval_val(args: Hyperparameters, model: nn.Module, rank: int, world_size: int, + device: torch.device, grad_accum_steps: int, val_tokens: Tensor, + base_bytes_lut: Tensor, has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, eval_seq_len: int | None = None) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale", + ).split(",") + if pattern +) +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_Q = 0.9999984 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + _soft_round_alpha: float = 1.0 # temperature for soft-round (annealed during training) + _use_soft_round: bool = False # enable soft-round QAT instead of STE + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self._clip_range = 15 # default int5, set to 31 for int6 layers + + @staticmethod + def soft_round(y: Tensor, alpha: float) -> Tensor: + """Differentiable approximation to round() from Agustsson & Theis (NeurIPS 2020). + s_alpha(y) = floor(y) + 0.5 * tanh(alpha * r) / tanh(alpha/2) + 0.5 + where r = y - floor(y) - 0.5 (centered fractional part) + """ + fl = torch.floor(y) + r = y - fl - 0.5 + return fl + 0.5 * torch.tanh(alpha * r) / (math.tanh(alpha / 2) + 1e-10) + 0.5 + + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + cr = self._clip_range + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + if CastedLinear._use_soft_round: + # Soft-Round QAT: differentiable rounding with temperature annealing + w32 = self.weight.float() + row_clip = torch.quantile(w32.abs(), 0.9995, dim=1) + scale = (row_clip / float(cr)).clamp_min(1.0 / float(cr)) + w_scaled = w32 / scale[:, None] + w_rounded = CastedLinear.soft_round(w_scaled, CastedLinear._soft_round_alpha) + w_q = (torch.clamp(w_rounded, -(cr+1), cr) * scale[:, None]).to(x.dtype) + w = w_q # fully differentiable path + else: + # Original STE QAT + with torch.no_grad(): + w32 = self.weight.float() + row_clip = torch.quantile(w32.abs(), 0.9995, dim=1) + scale = (row_clip / float(cr)).clamp_min(1.0 / float(cr)) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -(cr+1), cr) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, + rope_base: float, qk_gain_init: float): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False + + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + B, T, H, D = y.shape + Hkv = v.size(-2) + y_g = y.reshape(B, T, Hkv, H // Hkv, D) + vn = F.normalize(v, dim=-1).unsqueeze(-2) + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + + def forward(self, x: Tensor, v_embed: Tensor | None = None) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = self.c_v(x) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if _HAS_FA3: + y = flash_attn_3_func(q, k, v, causal=True).contiguous() + else: + y = F.scaled_dot_product_attention( + q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), + attn_mask=None, is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + return self.proj(y) + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class ValueEmbedding(nn.Module): + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + return self.proj(F.leaky_relu(self.fc(x), negative_slope=0.5).square()) + +class Block(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, mlp_mult: int, + rope_base: float, qk_gain_init: float, layer_idx: int = 0, + ln_scale: bool = False, dtg: bool = False): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + + def forward(self, x: Tensor, x0: Tensor, v_embed: Tensor | None = None) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, v_embed=v_embed) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out + +class GPT(nn.Module): + def __init__(self, vocab_size: int, num_layers: int, model_dim: int, num_heads: int, + num_kv_heads: int, mlp_mult: int, tie_embeddings: bool, tied_embed_init_std: float, + logit_softcap: float, rope_base: float, qk_gain_init: float, + bigram_vocab_size: int = 0, bigram_dim: int = 128, xsa_last_n: int = 0, + rope_dims: int = 0, ln_scale: bool = False, dtg: bool = False, + ve_enabled: bool = False, ve_dim: int = 128, ve_layers: str = "9,10"): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList([ + Block(model_dim, num_heads, num_kv_heads, mlp_mult, rope_base, + qk_gain_init, layer_idx=i, ln_scale=ln_scale, dtg=dtg) + for i in range(num_layers) + ]) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + num_layers = len(self.blocks) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + if ".proj." in name or name.endswith(".proj"): + with torch.no_grad(): + module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) + + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x = self.blocks[i](x, x0, v_embed=ve) + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x = self.blocks[bi](x, x0, v_embed=ve) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return F.cross_entropy(logits.float(), targets, reduction="mean") + + def forward_logits(self, input_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x = self.blocks[i](x, x0, v_embed=ve) + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x = self.blocks[bi](x, x0, v_embed=ve) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + +def eval_val_sliding(args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, eval_seq_len: int | None = None) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + + # Pre-compile: dummy forward+backward with TTT shapes to warm the compile cache + if rank == 0: + print(" ttt: pre-compiling forward+backward kernels...", flush=True) + _dummy_x = torch.zeros(1, seq_len, dtype=torch.int64, device=device) + _dummy_y = torch.zeros(1, seq_len, dtype=torch.int64, device=device) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + _dummy_logits = base_model.forward_logits(_dummy_x) + _dummy_loss = F.cross_entropy(_dummy_logits.reshape(-1, _dummy_logits.size(-1)), _dummy_y.reshape(-1)) + _dummy_loss.backward() + base_model.zero_grad(set_to_none=True) + if rank == 0: + print(" ttt: pre-compile done", flush=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + +def eval_val_sliding_ttt( + args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, ttt_epochs: int = 3, ttt_lr: float = 0.001, + ttt_momentum: float = 0.9, ttt_freeze_blocks: int = 2, + batch_seqs: int = 32, eval_seq_len: int | None = None, + ttt_chunk_tokens: int = 32768, ttt_optimizer: str = "adamw", + ttt_temp: float = 1.0, + ppm_alpha: float = 0.85, + byte_weighted_ttt: bool = True, + use_cache: bool = True, + cache_alpha: float = 0.3, + adaptive_lr: bool = True, + adaptive_lr_max_mult: float = 3.0, +) -> tuple[float, float]: + """Legal score-first TTT: score each chunk, then train on it. + Every token scored BEFORE any update that could use it.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + + # Initialize GPU-vectorized logistic context mixer + use_mixer = os.environ.get("USE_MIXER", "1") == "1" + mixer = LogisticContextMixer( + vocab_size=val_tokens.to(torch.int32).max().item() + 1, + device=device, + eta=float(os.environ.get("MIXER_ETA", "0.1")), + ) if use_mixer else None + if use_mixer and rank == 0: + print(f" Logistic context mixer enabled: eta={mixer.eta}") + if adaptive_lr and rank == 0: + print(f" Adaptive LR enabled: max_mult={adaptive_lr_max_mult}") + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on scored token position + num_chunks = (total_tokens + ttt_chunk_tokens - 1) // ttt_chunk_tokens + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk_tokens, num_chunks - 1) + chunk_windows[ci].append(ws) + + if rank == 0: + print(f"ttt:start chunks={num_chunks} chunk_tokens={ttt_chunk_tokens} " + f"windows={len(window_starts)} stride={stride} " + f"lr={ttt_lr} epochs={ttt_epochs} opt={ttt_optimizer} " + f"freeze_first={ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze everything, then selectively unfreeze for TTT + num_blocks = len(base_model.blocks) + for p in base_model.parameters(): + p.requires_grad_(False) + ttt_params = [] + ttt_param_ids = set() + use_qttt = os.environ.get("QTTT", "0") == "1" + if use_qttt: + # qTTT: only unfreeze Q projections in last N blocks + norms + head + for i in range(max(0, num_blocks - ttt_freeze_blocks), num_blocks): + for name, p in base_model.blocks[i].named_parameters(): + if "c_q" in name: + p.requires_grad_(True) + ttt_params.append(p) + ttt_param_ids.add(id(p)) + else: + # Standard: unfreeze all params in last N blocks + for i in range(max(0, num_blocks - ttt_freeze_blocks), num_blocks): + for p in base_model.blocks[i].parameters(): + p.requires_grad_(True) + ttt_params.append(p) + ttt_param_ids.add(id(p)) + # Unfreeze norms, scales, lm_head + for name, p in base_model.named_parameters(): + if "norm" in name or "scale" in name or "lm_head" in name: + p.requires_grad_(True) + if id(p) not in ttt_param_ids: + ttt_params.append(p) + ttt_param_ids.add(id(p)) + + if rank == 0: + n_unfrozen = sum(p.numel() for p in ttt_params) + n_frozen = sum(p.numel() for p in base_model.parameters() if not p.requires_grad) + print(f"ttt:params unfrozen={n_unfrozen} frozen={n_frozen}") + + if ttt_optimizer == "adamw": + optimizer = torch.optim.AdamW(ttt_params, lr=ttt_lr, weight_decay=0.0, betas=(0.9, 0.999)) + else: + optimizer = torch.optim.SGD(ttt_params, lr=ttt_lr, momentum=ttt_momentum) + + # Polyak averaging (TTT weight EMA) for smoother scoring + use_polyak = os.environ.get("USE_POLYAK", "1") == "1" + polyak_decay = float(os.environ.get("POLYAK_DECAY", "0.998")) + if use_polyak: + polyak_state = {id(p): p.data.clone() for p in ttt_params} + if rank == 0: + print(f" Polyak averaging enabled: decay={polyak_decay}") + + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + + # --- Phase 1: SCORE this chunk (inference_mode, no grad) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + # Swap in Polyak-averaged weights for scoring + if use_polyak and ci > 0: + _saved_weights = {} + for p in ttt_params: + _saved_weights[id(p)] = p.data.clone() + p.data.copy_(polyak_state[id(p)]) + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + logits_scaled = logits.float() / ttt_temp + + # Adaptive temperature: sharpen confident predictions more + if ttt_temp != 1.0: + with torch.no_grad(): + probs_for_entropy = F.softmax(logits.float(), dim=-1) + token_entropy = -(probs_for_entropy * (probs_for_entropy + 1e-10).log()).sum(-1) + max_ent = math.log(logits.size(-1)) + # Confident tokens (low entropy) get more sharpening + adaptive_temp = 1.0 - (1.0 - ttt_temp) * (1.0 - token_entropy / max_ent) + adaptive_temp = adaptive_temp.clamp(min=0.9, max=1.05) + logits_scaled = logits.float() / adaptive_temp.unsqueeze(-1) + + # Logistic context mixing (GPU-vectorized) or plain CE + if mixer is not None: + nll, expert_nll = mixer.mix_and_score(logits_scaled, x_batch, y_batch, wlens) + mixer.update_weights(expert_nll, wlens) + else: + nll = F.cross_entropy( + logits_scaled.reshape(-1, logits_scaled.size(-1)), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # --- Update context mixer with scored chunk tokens (GPU-vectorized) --- + chunk_start_tok = ci * ttt_chunk_tokens + chunk_end_tok = min((ci + 1) * ttt_chunk_tokens, total_tokens) + if mixer is not None: + mixer.update(val_tokens[chunk_start_tok:chunk_end_tok + 1]) + + # Swap back training weights after scoring + if use_polyak and ci > 0: + for p in ttt_params: + p.data.copy_(_saved_weights[id(p)]) + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and ttt_epochs > 0: + chunk_start = ci * ttt_chunk_tokens + chunk_end = min((ci + 1) * ttt_chunk_tokens, total_tokens) + chunk_seqs = (chunk_end - chunk_start) // seq_len + if rank == 0 and ci < 3: + print(f" ttt_train [{ci+1}] seqs={chunk_seqs} start_train...", flush=True) + if chunk_seqs > 0: + # Cosine LR across chunks + adaptive scaling + cos_lr = ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + if adaptive_lr: + # Increase LR as we've seen more data (more confident adaptation) + progress = min(ci / max(num_chunks * 0.3, 1), 1.0) # ramp over first 30% of chunks + lr_mult = 1.0 + (adaptive_lr_max_mult - 1.0) * progress + cos_lr = cos_lr * lr_mult + for pg in optimizer.param_groups: + pg["lr"] = cos_lr + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(ttt_epochs): + if rank == 0 and ci < 3: + print(f" ttt_train [{ci+1}] epoch={_ep+1}/{ttt_epochs} batches={my_chunk_seqs} ...", flush=True) + for bs in range(0, my_chunk_seqs, batch_seqs): + be = min(bs + batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + if byte_weighted_ttt: + # Byte-weighted loss: tokens covering more bytes matter more + ttt_logits = base_model.forward_logits(x) + per_token_loss = F.cross_entropy( + ttt_logits.reshape(-1, ttt_logits.size(-1)), + y.reshape(-1), reduction='none' + ).reshape(y.shape) + byte_weights = base_bytes_lut[y].float() + byte_weights = byte_weights + (has_leading_space_lut[y] & ~is_boundary_token_lut[x]).float() + ttt_loss = (per_token_loss * byte_weights).sum() / byte_weights.sum() + else: + ttt_loss = base_model(x, y) + ttt_loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, 1.0) + optimizer.step() + # Update Polyak EMA after each step + if use_polyak: + for p in ttt_params: + polyak_state[id(p)].lerp_(p.data, 1.0 - polyak_decay) + if rank == 0 and ci < 3: + print(f" step done ep={_ep+1} bs={bs} loss={ttt_loss.item():.4f}", flush=True) + + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1 or ci < 5): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + print(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s", flush=True) + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + if rank == 0: + print(f"ttt:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" + +def quantize_int6_per_row(t: Tensor, clip_range: int = 15) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale + + +def _find_best_row_scales(W: Tensor, clip_range: int = 15) -> Tensor: + """Find optimal per-row scales by searching percentile clipping thresholds.""" + t32 = W.float() + best_s = t32.abs().amax(dim=1) / clip_range + best_s = best_s.clamp_min(1.0 / clip_range) + best_err = torch.full((t32.shape[0],), float('inf')) + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range) + q = torch.clamp(torch.round(t32 / s[:, None]), -clip_range, clip_range) + recon = q * s[:, None] + err = (t32 - recon).pow(2).mean(dim=1) + improved = err < best_err + best_s[improved] = s[improved] + best_err[improved] = err[improved] + return best_s + +def gptq_quantize_weight(W: Tensor, H: Tensor, clip_range: int = 15, + block_size: int = 128, percdamp: float = 0.01) -> tuple[Tensor, Tensor]: + """GPTQ: quantize weight matrix W using Hessian H = X^T X for error compensation.""" + W = W.float().clone() + rows, cols = W.shape + row_scale = _find_best_row_scales(W, clip_range) + H = H.float().clone() + damp = percdamp * H.diag().mean() + H.diagonal().add_(damp) + perm = torch.argsort(H.diag()) + invperm = torch.argsort(perm) + W = W[:, perm] + H = H[perm][:, perm] + try: + L = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(L) + except torch._C._LinAlgError: + Hinv = torch.diag(1.0 / H.diag().clamp_min(1e-6)) + Q = torch.zeros(rows, cols, dtype=torch.int8) + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + W_block = W[:, i1:i2].clone() + Hinv_block = Hinv[i1:i2, i1:i2] + Err = torch.zeros_like(W_block) + for j in range(i2 - i1): + w_col = W_block[:, j] + h_inv_jj = Hinv_block[j, j].clamp_min(1e-8) + q_col = torch.clamp(torch.round(w_col / row_scale), -clip_range, clip_range) + deq_col = q_col * row_scale + Q[:, i1 + j] = q_col.to(torch.int8) + err = (w_col - deq_col) / h_inv_jj + Err[:, j] = err + if j + 1 < i2 - i1: + W_block[:, j + 1:] -= err.unsqueeze(1) * Hinv_block[j, j + 1:].unsqueeze(0) + if i2 < cols: + W[:, i2:] -= Err @ Hinv[i1:i2, i2:] + Q = Q[:, invperm] + return Q, row_scale.to(torch.float16) + +def gptq_calibrate(model: nn.Module, train_pattern: str, device: torch.device, + n_samples: int = 256, seq_len: int = 2048) -> dict[str, Tensor]: + """Collect Hessian H = X^T X for each linear layer using training data.""" + hessians: dict[str, Tensor] = {} + n_seen: dict[str, int] = {} + hooks = [] + def make_hook(name: str): + def hook_fn(module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + if name not in hessians: + hessians[name] = torch.zeros(x.shape[1], x.shape[1], device=x.device, dtype=torch.float32) + n_seen[name] = 0 + hessians[name].addmm_(x.t(), x) + n_seen[name] += x.shape[0] + return hook_fn + for name, module in model.named_modules(): + if isinstance(module, (nn.Linear, CastedLinear)): + hooks.append(module.register_forward_hook(make_hook(name))) + stream = TokenStream(train_pattern) + model.eval() + with torch.no_grad(): + for _ in range(n_samples): + tokens = stream.take(seq_len + 1).to(device=device, dtype=torch.int64) + x = tokens[:-1].unsqueeze(0) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + model.forward_logits(x) + for h in hooks: + h.remove() + for name in hessians: + hessians[name] /= max(n_seen[name], 1) + return hessians + +def _get_layer_clip_range(name: str, num_layers: int, int6_last_n: int) -> int: + """Return clip_range based on which layer the param belongs to.""" + import re + m = re.search(r'blocks\.(\d+)\.', name) + if m: + layer_idx = int(m.group(1)) + if layer_idx >= num_layers - int6_last_n: + return 31 # int6 + return 15 # int5 + +def mixed_quantize_int6_gptq(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor], + num_layers: int = 11, int6_last_n: int = 2) -> tuple[dict, dict]: + """GPTQ quantization with mixed int5/int6 precision. int6 for last int6_last_n layers, int5 for rest.""" + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + gptq_count, naive_count = 0, 0 + int5_params, int6_params = 0, 0 + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + cr = _get_layer_clip_range(name, num_layers, int6_last_n) + if cr == 31: + int6_params += t.numel() + else: + int5_params += t.numel() + if cat in int6_cats and t.ndim == 2: + module_name = name.rsplit(".weight", 1)[0] if name.endswith(".weight") else name + H = hessians.get(module_name) + if H is not None and H.shape[0] == t.shape[1]: + q, s = gptq_quantize_weight(t, H.cpu(), clip_range=cr) + gptq_count += 1 + else: + q, s = quantize_int6_per_row(t, clip_range=cr) + naive_count += 1 + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": f"int{'6' if cr == 31 else '5'}"} + elif cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t, clip_range=cr) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": f"int{'6' if cr == 31 else '5'}"} + naive_count += 1 + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + print(f"gptq_quantize: {gptq_count} GPTQ layers, {naive_count} naive layers", flush=True) + print(f"mixed_precision: {int5_params} int5 params, {int6_params} int6 params", flush=True) + return result, meta + + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta + +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +def main() -> None: + global zeropower_via_newtonschulz5 + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0(f"Python {sys.version} PyTorch {torch.__version__}", console=False) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + CastedLinear._qat_enabled = args.qat_enabled + base_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, rope_dims=args.rope_dims, ln_scale=args.ln_scale, + dtg=args.dtg_enabled, ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + matrix_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + # Set int6 clip_range for last N layers (mixed precision) + int6_start = args.num_layers - args.int6_last_n + for i, block in enumerate(base_model.blocks): + if i >= int6_start: + for m in block.modules(): + if isinstance(m, CastedLinear): + m._clip_range = 31 # int6 + if master_process: + int5_count = sum(1 for m in base_model.modules() if isinstance(m, CastedLinear) and m._clip_range == 15) + int6_count = sum(1 for m in base_model.modules() if isinstance(m, CastedLinear) and m._clip_range == 31) + log0(f"mixed_precision: {int5_count} int5 layers, {int6_count} int6 layers (last {args.int6_last_n} blocks)") + log0(f"model_params:{n_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:{xsa_layers} ws:{world_size} gqa:{args.num_heads}/{args.num_kv_heads}") + log0(f"lr:embed={token_lr} matrix={args.matrix_lr} scalar={args.scalar_lr} batch:{args.train_batch_tokens} wall:{args.max_wallclock_seconds:.0f}s seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + train_reserve_ms = 18000 # reserve 18s for EMA + GPTQ calibration + quantization + save + effective_train_ms = (max_wallclock_ms - train_reserve_ms) if max_wallclock_ms is not None else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if effective_train_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(effective_train_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + # TTT_ONLY mode: skip training, load saved model, run TTT eval + if os.environ.get("TTT_ONLY", "0") == "1": + log0("TTT_ONLY mode: skipping training, loading saved model...") + sd_cpu = {k: v.cpu() for k, v in torch.load("final_model.pt", map_location="cpu").items()} + if args.prune_pct > 0: + for k, v in sd_cpu.items(): + if v.ndim == 2 and v.numel() > 65536: + thresh = torch.quantile(v.abs().float(), args.prune_pct) + v[v.abs() < thresh] = 0.0 + log0(f"pruning:{args.prune_pct*100:.1f}% magnitude pruning applied") + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(zstandard.ZstdDecompressor().decompress(quant_blob_disk) if _COMPRESSOR == "zstd" else zlib.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + sw_seq_len = int(os.environ.get("EVAL_SEQ_LEN", str(effective_eval_seq_len))) + log0(f"TTT_ONLY: model loaded, starting TTT eval...") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_epochs = int(os.environ.get("TTT_EPOCHS", "3")) + ttt_lr = float(os.environ.get("TTT_LR", "0.0005")) + ttt_freeze = int(os.environ.get("TTT_FREEZE_BLOCKS", "2")) + ttt_chunk = int(os.environ.get("TTT_CHUNK_TOKENS", "32768")) + ttt_opt = os.environ.get("TTT_OPTIMIZER", "adamw") + log0(f"TTT: epochs={ttt_epochs} lr={ttt_lr} freeze_first={ttt_freeze} chunk={ttt_chunk} opt={ttt_opt}") + ttt_temp = args.ttt_temperature + log0(f"TTT temperature: {ttt_temp}") + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, ttt_epochs=ttt_epochs, ttt_lr=ttt_lr, + ttt_freeze_blocks=ttt_freeze, eval_seq_len=sw_seq_len, + ttt_chunk_tokens=ttt_chunk, ttt_optimizer=ttt_opt, + ttt_temp=ttt_temp, + ppm_alpha=float(os.environ.get("PPM_ALPHA", "0.85")), + byte_weighted_ttt=os.environ.get("BYTE_WEIGHTED_TTT", "1") == "1", + use_cache=os.environ.get("USE_CACHE", "1") == "1", + cache_alpha=float(os.environ.get("CACHE_ALPHA", "0.3")), + adaptive_lr=os.environ.get("ADAPTIVE_LR", "1") == "1", + adaptive_lr_max_mult=float(os.environ.get("ADAPTIVE_LR_MAX", "3.0")), + ) + torch.cuda.synchronize() + log0( + f"final_int6_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" + ) + log0(f"final_int6_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() + return + + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + # Anneal soft-round alpha based on QAT progress + if CastedLinear._use_soft_round and CastedLinear._qat_enabled: + qat_progress = max(0.0, 1.0 - scale / max(args.late_qat_threshold, 0.01)) + CastedLinear._soft_round_alpha = 1.0 + 15.0 * qat_progress + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + CastedLinear._use_soft_round = os.environ.get("SOFT_ROUND_QAT", "0") == "1" + if CastedLinear._use_soft_round and master_process: + log0(f"soft_round_qat:enabled initial_alpha=1.0") + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + reached_cap = effective_train_ms is not None and approx_training_time_ms >= effective_train_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + # Apply EMA weights directly (skip diagnostic evals to save ~5s of reserve) + log0("ema:applying EMA weights (skipping diagnostic evals)") + current_state = base_model.state_dict() + ema_sd = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(ema_sd, strict=True) + # GPTQ calibration on final model (within reserved training budget) + log0("gptq:calibrating with training data...") + t_gptq = time.perf_counter() + gptq_hessians = gptq_calibrate(base_model, args.train_files, device, n_samples=128, seq_len=args.train_seq_len) + log0(f"gptq:calibrated {len(gptq_hessians)} layers in {time.perf_counter()-t_gptq:.1f}s") + export_sd = base_model.state_dict() + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + if args.prune_pct > 0: + for k, v in sd_cpu.items(): + if v.ndim == 2 and v.numel() > 65536: + thresh = torch.quantile(v.abs().float(), args.prune_pct) + v[v.abs() < thresh] = 0.0 + if master_process: + log0(f"pruning:{args.prune_pct*100:.1f}% magnitude pruning applied") + quant_result, quant_meta = mixed_quantize_int6_gptq(sd_cpu, {"mlp", "attn"}, gptq_hessians, num_layers=args.num_layers, int6_last_n=args.int6_last_n) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) if _COMPRESSOR == "zstd" else zlib.compress(quant_raw, 9) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+{_COMPRESSOR}: {quant_file_bytes} bytes") + log0(f"Total submission size int6+{_COMPRESSOR}: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(zstandard.ZstdDecompressor().decompress(quant_blob_disk) if _COMPRESSOR == "zstd" else zlib.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + sw_seq_len = int(os.environ.get("EVAL_SEQ_LEN", str(effective_eval_seq_len))) + if sw_seq_len != effective_eval_seq_len and rank == 0: + log0(f"Eval seq_len override: {effective_eval_seq_len} -> {sw_seq_len}") + if args.eval_stride > 0 and args.eval_stride < sw_seq_len and not os.environ.get("SKIP_SLIDING"): + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_epochs = int(os.environ.get("TTT_EPOCHS", "3")) + ttt_lr = float(os.environ.get("TTT_LR", "0.0005")) + ttt_freeze = int(os.environ.get("TTT_FREEZE_BLOCKS", "2")) + ttt_chunk = int(os.environ.get("TTT_CHUNK_TOKENS", "32768")) + ttt_opt = os.environ.get("TTT_OPTIMIZER", "adamw") + log0(f"TTT: epochs={ttt_epochs} lr={ttt_lr} freeze_first={ttt_freeze} chunk={ttt_chunk} opt={ttt_opt}") + ttt_temp = args.ttt_temperature + log0(f"TTT temperature: {ttt_temp}") + ppm_alpha_val = float(os.environ.get("PPM_ALPHA", "0.85")) + bw_ttt = os.environ.get("BYTE_WEIGHTED_TTT", "1") == "1" + log0(f"PPM alpha: {ppm_alpha_val}, Byte-weighted TTT: {bw_ttt}") + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, ttt_epochs=ttt_epochs, ttt_lr=ttt_lr, + ttt_freeze_blocks=ttt_freeze, eval_seq_len=sw_seq_len, + ttt_chunk_tokens=ttt_chunk, ttt_optimizer=ttt_opt, + ttt_temp=ttt_temp, + ppm_alpha=float(os.environ.get("PPM_ALPHA", "0.85")), + byte_weighted_ttt=os.environ.get("BYTE_WEIGHTED_TTT", "1") == "1", + use_cache=os.environ.get("USE_CACHE", "1") == "1", + cache_alpha=float(os.environ.get("CACHE_ALPHA", "0.3")), + adaptive_lr=os.environ.get("ADAPTIVE_LR", "1") == "1", + adaptive_lr_max_mult=float(os.environ.get("ADAPTIVE_LR_MAX", "3.0")), + ) + torch.cuda.synchronize() + log0( + f"final_int6_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" + ) + log0(f"final_int6_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() From a7afc01e6507b7bc9115191123655829d7195165 Mon Sep 17 00:00:00 2001 From: Royi Rassin Date: Wed, 25 Mar 2026 08:42:06 +0200 Subject: [PATCH 2/4] Move submission to records/track_10min_16mb/ --- .../2026-03-24_HedgeMixer_TTT/README.md | 110 + .../log_seed1337.txt | 173 ++ .../2026-03-24_HedgeMixer_TTT/log_seed42.txt | 173 ++ .../2026-03-24_HedgeMixer_TTT/log_seed7.txt | 173 ++ .../requirements.txt | 5 + .../2026-03-24_HedgeMixer_TTT/submission.json | 14 + .../2026-03-24_HedgeMixer_TTT/train_gpt.py | 1966 +++++++++++++++++ 7 files changed, 2614 insertions(+) create mode 100644 records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/README.md create mode 100644 records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/log_seed1337.txt create mode 100644 records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/log_seed42.txt create mode 100644 records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/log_seed7.txt create mode 100644 records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/requirements.txt create mode 100644 records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/submission.json create mode 100644 records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/train_gpt.py diff --git a/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/README.md b/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/README.md new file mode 100644 index 000000000..2753515cc --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/README.md @@ -0,0 +1,110 @@ +# 5-expert Hedge Mixer + TTT + +**val_bpb: 1.0745** (3-seed mean) | **<15.5 MB** | 8xH100 SXM + +## Results (8xH100 80GB SXM) + +| Seed | steps | step_avg | Pre-TTT bpb | **Post-TTT bpb** | TTT gain | Eval time | Artifact | +|------|-------|----------|-------------|-----------------|----------|-----------|----------| +| 1337 | 5,997 | 97.1ms | 1.1248 | **1.0560** | -0.0688 | 563s | 15.48 MB | +| 42 | 5,997 | 97.1ms | 1.1257 | **1.0970** | -0.0287 | 563s | 15.41 MB | +| 7 | 5,983 | 97.3ms | 1.1251 | **1.0704** | -0.0547 | 561s | 15.43 MB | +| **Mean** | | | **1.1252** | **1.0745** | **-0.0507** | | | + +## Key Contribution: 5-expert Logistic Context Mixer + +GPU-vectorized online context mixing using the Hedge/multiplicative-weights algorithm. Five experts blend predictions in log-probability space: + +| Expert | Source | Description | +|--------|--------|-------------| +| 0 | Neural | Base model log-softmax | +| 1 | Unigram | Token frequency from scored tokens | +| 2 | Bigram | P(next \| prev) from scored tokens | +| 3 | Trigram | Hashed P(next \| prev2, prev1) with 64K buckets | +| 4 | Entropy | Neural model entropy as confidence regularizer | + +Expert weights are updated online via Hedge: `log_w -= eta * loss`. N-gram tables are built incrementally from already-scored tokens only (legal). + +## Architecture + +PR #606 base with the following additions: + +| Component | Setting | +|-----------|---------| +| Layers | 11 (512d, 8H, 8KV) | +| MLP | 3x with **LeakyReLU(0.5)^2** | +| BigramHash | 6144 (dim=128) | +| XSA | All 11 layers (ws=8) | +| RoPE | Partial (16/64 dims) | +| LN Scale | 1/sqrt(layer+1) | +| VE128 | Layers 9-10 | +| Weight avg | EMA(0.997) | +| Quantization | Full GPTQ int5 + zstd (level 22) | +| Pruning | 3% magnitude | + +## Legal Score-First TTT + +Backward-looking adaptation with GPTQ-calibrated model: + +1. Validation tokens split into 474 chunks of 131K tokens each +2. For each chunk: + - **SCORE**: Sliding window eval (stride=32, seq_len=2048) with 5-expert mixer blending + - **TRAIN**: AdamW(lr=0.0001) on already-scored chunk. 3 epochs, last 2 blocks unfrozen + norms + lm_head, cosine LR decay, Polyak averaging +3. Last chunk scored but never trained on + +### TTT Hyperparameters + +| Parameter | Value | +|-----------|-------| +| Chunk size | 131,072 tokens | +| Optimizer | AdamW (lr=0.0001) | +| Epochs per chunk | 3 | +| Frozen blocks | First 9 (last 2 + norms + head unfrozen) | +| Polyak decay | 0.998 | +| Adaptive LR | max_mult=3.0 | +| Mixer eta | 0.1 | + +### Training Budget + +GPTQ calibration runs within the 600s training budget (18s reserved from training loop for EMA selection + calibration + quantization). + +| Phase | Time | +|-------|------| +| Training loop | 582s | +| EMA + GPTQ calibration + quantization | ~18s | +| **Total training** | **~600s** | +| Sliding window eval | ~165s | +| TTT eval with mixer | ~562s | +| **Total eval** | **~562s** | + +## Reproduction + +```bash +# Install dependencies +pip install -r requirements.txt +# Build FA3 Hopper kernels (required) +cd /tmp && git clone https://github.com/Dao-AILab/flash-attention +cd flash-attention/hopper && python setup.py install + +# Run training + eval (single seed) +DATA_PATH=./data/datasets/fineweb10B_sp1024 \ +TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model \ +SEED=1337 MAX_WALLCLOCK_SECONDS=600 \ +USE_MIXER=1 TTT_LR=0.0001 TTT_CHUNK_TOKENS=131072 \ + torchrun --standalone --nproc_per_node=8 train_gpt.py + +# Run all 3 seeds +for SEED in 1337 42 7; do + DATA_PATH=./data/datasets/fineweb10B_sp1024 \ + TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model \ + SEED=$SEED MAX_WALLCLOCK_SECONDS=600 \ + USE_MIXER=1 TTT_LR=0.0001 TTT_CHUNK_TOKENS=131072 \ + torchrun --standalone --nproc_per_node=8 train_gpt.py +done +``` + +## Credits + +- **Base model**: PR #606 by @gowtham0992 +- **TTT recipe**: PR #461 by @Christopher-Lee-McClendon +- **Mixer inspiration**: PAQ compression (context mixing) + Hedge algorithm diff --git a/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/log_seed1337.txt b/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/log_seed1337.txt new file mode 100644 index 000000000..656321ad9 --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/log_seed1337.txt @@ -0,0 +1,173 @@ +W0325 00:33:17.141000 673732 torch/distributed/run.py:852] +W0325 00:33:17.141000 673732 torch/distributed/run.py:852] ***************************************** +W0325 00:33:17.141000 673732 torch/distributed/run.py:852] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0325 00:33:17.141000 673732 torch/distributed/run.py:852] ***************************************** +logs/31484e99-50c6-404c-8bb8-635f618baafa.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=../data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=../data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +mixed_precision: 68 int5 layers, 0 int6 layers (last 0 blocks) +model_params:33317980 +XSA:[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] ws:8 gqa:8/8 +lr:embed=0.035 matrix=0.025 scalar=0.025 batch:786432 wall:600s seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9285 val_bpb:4.1034 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9305 train_time:148ms step_avg:147.68ms +step:2/20000 train_loss:8.6412 train_time:239ms step_avg:119.37ms +step:3/20000 train_loss:7.7278 train_time:333ms step_avg:111.04ms +step:4/20000 train_loss:7.2812 train_time:428ms step_avg:106.95ms +step:5/20000 train_loss:7.0672 train_time:524ms step_avg:104.81ms +step:6/20000 train_loss:6.9647 train_time:619ms step_avg:103.14ms +step:7/20000 train_loss:6.8519 train_time:714ms step_avg:102.01ms +step:8/20000 train_loss:6.7091 train_time:809ms step_avg:101.08ms +step:9/20000 train_loss:6.3640 train_time:903ms step_avg:100.36ms +step:10/20000 train_loss:6.0314 train_time:998ms step_avg:99.77ms +step:500/20000 train_loss:2.3594 train_time:48287ms step_avg:96.57ms +step:1000/20000 train_loss:2.2366 train_time:96650ms step_avg:96.65ms +step:1500/20000 train_loss:2.1871 train_time:145067ms step_avg:96.71ms +step:2000/20000 train_loss:2.0272 train_time:193559ms step_avg:96.78ms +step:2500/20000 train_loss:2.1332 train_time:242098ms step_avg:96.84ms +step:3000/20000 train_loss:2.1140 train_time:290638ms step_avg:96.88ms +step:3500/20000 train_loss:2.1198 train_time:339166ms step_avg:96.90ms +step:4000/20000 train_loss:1.9079 train_time:387684ms step_avg:96.92ms +step:4000/20000 val_loss:2.0000 val_bpb:1.1845 train_time:387689ms step_avg:96.92ms +late_qat:enabled step:4255 scale:0.4998 +step:4500/20000 train_loss:2.0575 train_time:436219ms step_avg:96.94ms +step:5000/20000 train_loss:2.0307 train_time:484743ms step_avg:96.95ms +swa:start step:5350 +step:5500/20000 train_loss:1.9404 train_time:533445ms step_avg:96.99ms +step:5997/20000 val_loss:1.9019 val_bpb:1.1264 train_time:582051ms step_avg:97.06ms +stopping_early: wallclock_cap train_time:582051ms step:5997/20000 +peak memory allocated: 26200 MiB reserved: 26782 MiB +ema:applying EMA weights (skipping diagnostic evals) +gptq:calibrating with training data... +gptq:calibrated 68 layers in 1.9s +Serialized model: 130432585 bytes +Code size: 96428 bytes +pruning:3.0% magnitude pruning applied +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +Serialized model int6+zstd: 15387977 bytes +Total submission size int6+zstd: 15484405 bytes + ttt: pre-compiling forward+backward kernels... + ttt: pre-compile done +final_int6_sliding_window val_loss:1.8992 val_bpb:1.1248 stride:32 eval_time:164968ms +final_int6_sliding_window_exact val_loss:1.89920781 val_bpb:1.12482157 +TTT: epochs=3 lr=0.0001 freeze_first=2 chunk=131072 opt=adamw +TTT temperature: 0.98 +PPM alpha: 0.85, Byte-weighted TTT: True + Logistic context mixer enabled: eta=0.1 + Adaptive LR enabled: max_mult=3.0 +ttt:start chunks=474 chunk_tokens=131072 windows=1938176 stride=32 lr=0.0001 epochs=3 opt=adamw freeze_first=2 +ttt:params unfrozen=5780500 frozen=27537480 + Polyak averaging enabled: decay=0.998 + ttt_train [1] seqs=64 start_train... + ttt_train [1] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.3437 + ttt_train [1] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.3101 + ttt_train [1] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.3086 + ttt_chunk [1/474] bpb=1.202687 time=1.3s + ttt_train [2] seqs=64 start_train... + ttt_train [2] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.1361 + ttt_train [2] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.1334 + ttt_train [2] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.1283 + ttt_chunk [2/474] bpb=1.128760 time=2.5s + ttt_train [3] seqs=64 start_train... + ttt_train [3] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.0525 + ttt_train [3] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.0510 + ttt_train [3] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.0484 + ttt_chunk [3/474] bpb=1.080018 time=3.7s + ttt_chunk [4/474] bpb=1.076079 time=4.8s + ttt_chunk [5/474] bpb=1.066728 time=6.0s + ttt_chunk [11/474] bpb=1.031430 time=13.1s + ttt_chunk [21/474] bpb=1.018884 time=25.0s + ttt_chunk [31/474] bpb=1.015939 time=36.9s + ttt_chunk [41/474] bpb=1.022519 time=48.7s + ttt_chunk [51/474] bpb=1.028238 time=60.6s + ttt_chunk [61/474] bpb=1.025430 time=72.5s + ttt_chunk [71/474] bpb=1.026528 time=84.3s + ttt_chunk [81/474] bpb=1.026915 time=96.2s + ttt_chunk [91/474] bpb=1.028643 time=108.1s + ttt_chunk [101/474] bpb=1.025107 time=119.9s + ttt_chunk [111/474] bpb=1.024927 time=131.8s + ttt_chunk [121/474] bpb=1.027728 time=143.7s + ttt_chunk [131/474] bpb=1.027818 time=155.5s + ttt_chunk [141/474] bpb=1.026649 time=167.4s + ttt_chunk [151/474] bpb=1.024273 time=179.3s + ttt_chunk [161/474] bpb=1.024436 time=191.2s + ttt_chunk [171/474] bpb=1.022907 time=203.0s + ttt_chunk [181/474] bpb=1.023651 time=214.9s + ttt_chunk [191/474] bpb=1.022518 time=226.8s + ttt_chunk [201/474] bpb=1.021847 time=238.6s + ttt_chunk [211/474] bpb=1.020784 time=250.5s + ttt_chunk [221/474] bpb=1.021277 time=262.4s + ttt_chunk [231/474] bpb=1.020969 time=274.2s + ttt_chunk [241/474] bpb=1.020429 time=286.1s + ttt_chunk [251/474] bpb=1.022389 time=298.0s + ttt_chunk [261/474] bpb=1.024492 time=309.8s + ttt_chunk [271/474] bpb=1.024714 time=321.7s + ttt_chunk [281/474] bpb=1.026263 time=333.6s + ttt_chunk [291/474] bpb=1.027062 time=345.4s + ttt_chunk [301/474] bpb=1.029658 time=357.3s + ttt_chunk [311/474] bpb=1.031598 time=369.2s + ttt_chunk [321/474] bpb=1.032553 time=381.0s + ttt_chunk [331/474] bpb=1.034020 time=392.9s + ttt_chunk [341/474] bpb=1.035727 time=404.8s + ttt_chunk [351/474] bpb=1.036601 time=416.6s + ttt_chunk [361/474] bpb=1.039560 time=428.5s + ttt_chunk [371/474] bpb=1.041485 time=440.4s + ttt_chunk [381/474] bpb=1.044560 time=452.3s + ttt_chunk [391/474] bpb=1.047764 time=464.1s + ttt_chunk [401/474] bpb=1.050371 time=476.0s + ttt_chunk [411/474] bpb=1.052690 time=487.9s + ttt_chunk [421/474] bpb=1.055963 time=499.8s + ttt_chunk [431/474] bpb=1.056364 time=511.6s + ttt_chunk [441/474] bpb=1.057641 time=523.5s + ttt_chunk [451/474] bpb=1.058576 time=535.4s + ttt_chunk [461/474] bpb=1.060302 time=547.2s + ttt_chunk [471/474] bpb=1.061844 time=559.1s + ttt_chunk [474/474] bpb=1.062011 time=561.7s +ttt:done val_loss=1.783077 val_bpb=1.056042 elapsed=561.7s +final_int6_ttt val_loss:1.7831 val_bpb:1.0560 stride:32 eval_time:562538ms +final_int6_ttt_exact val_loss:1.78307674 val_bpb:1.05604198 diff --git a/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/log_seed42.txt b/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/log_seed42.txt new file mode 100644 index 000000000..a185217fe --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/log_seed42.txt @@ -0,0 +1,173 @@ +W0325 00:58:00.120000 677084 torch/distributed/run.py:852] +W0325 00:58:00.120000 677084 torch/distributed/run.py:852] ***************************************** +W0325 00:58:00.120000 677084 torch/distributed/run.py:852] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0325 00:58:00.120000 677084 torch/distributed/run.py:852] ***************************************** +logs/e0b62b26-4d67-4f38-8e55-7f891da41982.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=../data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=../data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +mixed_precision: 68 int5 layers, 0 int6 layers (last 0 blocks) +model_params:33317980 +XSA:[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] ws:8 gqa:8/8 +lr:embed=0.035 matrix=0.025 scalar=0.025 batch:786432 wall:600s seed:42 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9301 val_bpb:4.1044 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9309 train_time:146ms step_avg:146.46ms +step:2/20000 train_loss:8.7072 train_time:237ms step_avg:118.44ms +step:3/20000 train_loss:7.7698 train_time:331ms step_avg:110.45ms +step:4/20000 train_loss:7.2987 train_time:427ms step_avg:106.64ms +step:5/20000 train_loss:7.0366 train_time:521ms step_avg:104.22ms +step:6/20000 train_loss:6.9492 train_time:616ms step_avg:102.74ms +step:7/20000 train_loss:6.8511 train_time:711ms step_avg:101.54ms +step:8/20000 train_loss:6.7338 train_time:805ms step_avg:100.67ms +step:9/20000 train_loss:6.3596 train_time:900ms step_avg:100.05ms +step:10/20000 train_loss:6.0340 train_time:996ms step_avg:99.58ms +step:500/20000 train_loss:2.3619 train_time:48288ms step_avg:96.58ms +step:1000/20000 train_loss:2.2442 train_time:96649ms step_avg:96.65ms +step:1500/20000 train_loss:2.1847 train_time:145067ms step_avg:96.71ms +step:2000/20000 train_loss:2.0315 train_time:193558ms step_avg:96.78ms +step:2500/20000 train_loss:2.1358 train_time:242084ms step_avg:96.83ms +step:3000/20000 train_loss:2.1167 train_time:290616ms step_avg:96.87ms +step:3500/20000 train_loss:2.1233 train_time:339152ms step_avg:96.90ms +step:4000/20000 train_loss:1.9137 train_time:387678ms step_avg:96.92ms +step:4000/20000 val_loss:2.0017 val_bpb:1.1855 train_time:387683ms step_avg:96.92ms +late_qat:enabled step:4255 scale:0.4998 +step:4500/20000 train_loss:2.0589 train_time:436214ms step_avg:96.94ms +step:5000/20000 train_loss:2.0322 train_time:484739ms step_avg:96.95ms +swa:start step:5350 +step:5500/20000 train_loss:1.9442 train_time:533403ms step_avg:96.98ms +step:5997/20000 val_loss:1.9035 val_bpb:1.1274 train_time:582007ms step_avg:97.05ms +stopping_early: wallclock_cap train_time:582007ms step:5997/20000 +peak memory allocated: 26197 MiB reserved: 26782 MiB +ema:applying EMA weights (skipping diagnostic evals) +gptq:calibrating with training data... +gptq:calibrated 68 layers in 1.9s +Serialized model: 130432585 bytes +Code size: 96428 bytes +pruning:3.0% magnitude pruning applied +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +Serialized model int6+zstd: 15312167 bytes +Total submission size int6+zstd: 15408595 bytes + ttt: pre-compiling forward+backward kernels... + ttt: pre-compile done +final_int6_sliding_window val_loss:1.9006 val_bpb:1.1257 stride:32 eval_time:164957ms +final_int6_sliding_window_exact val_loss:1.90062694 val_bpb:1.12566206 +TTT: epochs=3 lr=0.0001 freeze_first=2 chunk=131072 opt=adamw +TTT temperature: 0.98 +PPM alpha: 0.85, Byte-weighted TTT: True + Logistic context mixer enabled: eta=0.1 + Adaptive LR enabled: max_mult=3.0 +ttt:start chunks=474 chunk_tokens=131072 windows=1938176 stride=32 lr=0.0001 epochs=3 opt=adamw freeze_first=2 +ttt:params unfrozen=5780500 frozen=27537480 + Polyak averaging enabled: decay=0.998 + ttt_train [1] seqs=64 start_train... + ttt_train [1] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.3462 + ttt_train [1] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.3103 + ttt_train [1] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.3076 + ttt_chunk [1/474] bpb=1.202959 time=1.2s + ttt_train [2] seqs=64 start_train... + ttt_train [2] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.1330 + ttt_train [2] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.1310 + ttt_train [2] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.1268 + ttt_chunk [2/474] bpb=1.129301 time=2.4s + ttt_train [3] seqs=64 start_train... + ttt_train [3] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.0619 + ttt_train [3] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.0610 + ttt_train [3] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.0582 + ttt_chunk [3/474] bpb=1.079049 time=3.6s + ttt_chunk [4/474] bpb=1.074949 time=4.8s + ttt_chunk [5/474] bpb=1.064662 time=6.0s + ttt_chunk [11/474] bpb=1.025191 time=13.1s + ttt_chunk [21/474] bpb=1.008678 time=25.0s + ttt_chunk [31/474] bpb=1.002098 time=36.9s + ttt_chunk [41/474] bpb=1.007125 time=48.7s + ttt_chunk [51/474] bpb=1.012485 time=60.6s + ttt_chunk [61/474] bpb=1.010685 time=72.5s + ttt_chunk [71/474] bpb=1.013076 time=84.4s + ttt_chunk [81/474] bpb=1.015652 time=96.2s + ttt_chunk [91/474] bpb=1.019922 time=108.1s + ttt_chunk [101/474] bpb=1.019995 time=120.0s + ttt_chunk [111/474] bpb=1.024232 time=131.9s + ttt_chunk [121/474] bpb=1.031908 time=143.7s + ttt_chunk [131/474] bpb=1.037128 time=155.6s + ttt_chunk [141/474] bpb=1.041120 time=167.5s + ttt_chunk [151/474] bpb=1.044067 time=179.4s + ttt_chunk [161/474] bpb=1.049148 time=191.3s + ttt_chunk [171/474] bpb=1.052038 time=203.1s + ttt_chunk [181/474] bpb=1.057080 time=215.0s + ttt_chunk [191/474] bpb=1.059619 time=226.9s + ttt_chunk [201/474] bpb=1.062378 time=238.8s + ttt_chunk [211/474] bpb=1.064340 time=250.6s + ttt_chunk [221/474] bpb=1.067700 time=262.5s + ttt_chunk [231/474] bpb=1.069761 time=274.4s + ttt_chunk [241/474] bpb=1.071136 time=286.3s + ttt_chunk [251/474] bpb=1.074548 time=298.2s + ttt_chunk [261/474] bpb=1.077509 time=310.0s + ttt_chunk [271/474] bpb=1.078241 time=321.9s + ttt_chunk [281/474] bpb=1.079767 time=333.8s + ttt_chunk [291/474] bpb=1.080218 time=345.7s + ttt_chunk [301/474] bpb=1.082217 time=357.5s + ttt_chunk [311/474] bpb=1.083294 time=369.4s + ttt_chunk [321/474] bpb=1.083302 time=381.3s + ttt_chunk [331/474] bpb=1.083667 time=393.2s + ttt_chunk [341/474] bpb=1.084241 time=405.0s + ttt_chunk [351/474] bpb=1.083985 time=416.9s + ttt_chunk [361/474] bpb=1.085885 time=428.8s + ttt_chunk [371/474] bpb=1.086654 time=440.7s + ttt_chunk [381/474] bpb=1.088639 time=452.5s + ttt_chunk [391/474] bpb=1.090740 time=464.4s + ttt_chunk [401/474] bpb=1.092278 time=476.3s + ttt_chunk [411/474] bpb=1.093605 time=488.2s + ttt_chunk [421/474] bpb=1.095927 time=500.1s + ttt_chunk [431/474] bpb=1.095382 time=511.9s + ttt_chunk [441/474] bpb=1.095777 time=523.8s + ttt_chunk [451/474] bpb=1.095835 time=535.7s + ttt_chunk [461/474] bpb=1.096740 time=547.6s + ttt_chunk [471/474] bpb=1.097514 time=559.5s + ttt_chunk [474/474] bpb=1.097488 time=562.1s +ttt:done val_loss=1.852181 val_bpb=1.096969 elapsed=562.1s +final_int6_ttt val_loss:1.8522 val_bpb:1.0970 stride:32 eval_time:562886ms +final_int6_ttt_exact val_loss:1.85218082 val_bpb:1.09696945 diff --git a/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/log_seed7.txt b/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/log_seed7.txt new file mode 100644 index 000000000..fcaf81b2e --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/log_seed7.txt @@ -0,0 +1,173 @@ +W0325 01:22:56.941000 679893 torch/distributed/run.py:852] +W0325 01:22:56.941000 679893 torch/distributed/run.py:852] ***************************************** +W0325 01:22:56.941000 679893 torch/distributed/run.py:852] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0325 01:22:56.941000 679893 torch/distributed/run.py:852] ***************************************** +logs/fb9c44fd-95a9-4cb3-a79d-3bf240580a1a.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=../data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=../data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +mixed_precision: 68 int5 layers, 0 int6 layers (last 0 blocks) +model_params:33317980 +XSA:[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] ws:8 gqa:8/8 +lr:embed=0.035 matrix=0.025 scalar=0.025 batch:786432 wall:600s seed:7 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9294 val_bpb:4.1040 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9310 train_time:151ms step_avg:150.83ms +step:2/20000 train_loss:8.7138 train_time:240ms step_avg:120.17ms +step:3/20000 train_loss:7.7624 train_time:335ms step_avg:111.68ms +step:4/20000 train_loss:7.2817 train_time:430ms step_avg:107.46ms +step:5/20000 train_loss:7.1027 train_time:525ms step_avg:104.94ms +step:6/20000 train_loss:6.9241 train_time:619ms step_avg:103.24ms +step:7/20000 train_loss:6.8669 train_time:714ms step_avg:102.05ms +step:8/20000 train_loss:6.6918 train_time:809ms step_avg:101.14ms +step:9/20000 train_loss:6.3601 train_time:904ms step_avg:100.46ms +step:10/20000 train_loss:6.0035 train_time:999ms step_avg:99.88ms +step:500/20000 train_loss:2.3626 train_time:48426ms step_avg:96.85ms +step:1000/20000 train_loss:2.2381 train_time:96920ms step_avg:96.92ms +step:1500/20000 train_loss:2.1845 train_time:145469ms step_avg:96.98ms +step:2000/20000 train_loss:2.0268 train_time:194077ms step_avg:97.04ms +step:2500/20000 train_loss:2.1315 train_time:242724ms step_avg:97.09ms +step:3000/20000 train_loss:2.1123 train_time:291375ms step_avg:97.13ms +step:3500/20000 train_loss:2.1194 train_time:340018ms step_avg:97.15ms +step:4000/20000 train_loss:1.9084 train_time:388663ms step_avg:97.17ms +step:4000/20000 val_loss:1.9988 val_bpb:1.1838 train_time:388668ms step_avg:97.17ms +late_qat:enabled step:4240 scale:0.4998 +step:4500/20000 train_loss:2.0531 train_time:437292ms step_avg:97.18ms +step:5000/20000 train_loss:2.0297 train_time:485905ms step_avg:97.18ms +swa:start step:5300 +step:5500/20000 train_loss:1.9406 train_time:534723ms step_avg:97.22ms +step:5983/20000 val_loss:1.9008 val_bpb:1.1258 train_time:582062ms step_avg:97.29ms +stopping_early: wallclock_cap train_time:582062ms step:5983/20000 +peak memory allocated: 26197 MiB reserved: 26782 MiB +ema:applying EMA weights (skipping diagnostic evals) +gptq:calibrating with training data... +gptq:calibrated 68 layers in 2.0s +Serialized model: 130432585 bytes +Code size: 96428 bytes +pruning:3.0% magnitude pruning applied +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33161216 int5 params, 0 int6 params +Serialized model int6+zstd: 15331691 bytes +Total submission size int6+zstd: 15428119 bytes + ttt: pre-compiling forward+backward kernels... + ttt: pre-compile done +final_int6_sliding_window val_loss:1.8997 val_bpb:1.1251 stride:32 eval_time:165026ms +final_int6_sliding_window_exact val_loss:1.89974539 val_bpb:1.12513995 +TTT: epochs=3 lr=0.0001 freeze_first=2 chunk=131072 opt=adamw +TTT temperature: 0.98 +PPM alpha: 0.85, Byte-weighted TTT: True + Logistic context mixer enabled: eta=0.1 + Adaptive LR enabled: max_mult=3.0 +ttt:start chunks=474 chunk_tokens=131072 windows=1938176 stride=32 lr=0.0001 epochs=3 opt=adamw freeze_first=2 +ttt:params unfrozen=5780500 frozen=27537480 + Polyak averaging enabled: decay=0.998 + ttt_train [1] seqs=64 start_train... + ttt_train [1] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.3476 + ttt_train [1] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.3121 + ttt_train [1] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.3099 + ttt_chunk [1/474] bpb=1.199612 time=1.2s + ttt_train [2] seqs=64 start_train... + ttt_train [2] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.1304 + ttt_train [2] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.1283 + ttt_train [2] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.1239 + ttt_chunk [2/474] bpb=1.128925 time=2.4s + ttt_train [3] seqs=64 start_train... + ttt_train [3] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.0552 + ttt_train [3] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.0540 + ttt_train [3] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.0517 + ttt_chunk [3/474] bpb=1.081615 time=3.6s + ttt_chunk [4/474] bpb=1.078973 time=4.8s + ttt_chunk [5/474] bpb=1.069501 time=5.9s + ttt_chunk [11/474] bpb=1.037589 time=13.0s + ttt_chunk [21/474] bpb=1.028394 time=24.9s + ttt_chunk [31/474] bpb=1.028114 time=36.7s + ttt_chunk [41/474] bpb=1.038811 time=48.6s + ttt_chunk [51/474] bpb=1.048907 time=60.4s + ttt_chunk [61/474] bpb=1.050285 time=72.2s + ttt_chunk [71/474] bpb=1.055243 time=84.1s + ttt_chunk [81/474] bpb=1.059893 time=95.9s + ttt_chunk [91/474] bpb=1.065566 time=107.7s + ttt_chunk [101/474] bpb=1.065416 time=119.6s + ttt_chunk [111/474] bpb=1.068729 time=131.4s + ttt_chunk [121/474] bpb=1.074869 time=143.3s + ttt_chunk [131/474] bpb=1.078175 time=155.1s + ttt_chunk [141/474] bpb=1.080151 time=166.9s + ttt_chunk [151/474] bpb=1.080074 time=178.8s + ttt_chunk [161/474] bpb=1.082475 time=190.6s + ttt_chunk [171/474] bpb=1.082532 time=202.5s + ttt_chunk [181/474] bpb=1.084595 time=214.3s + ttt_chunk [191/474] bpb=1.084726 time=226.1s + ttt_chunk [201/474] bpb=1.084927 time=238.0s + ttt_chunk [211/474] bpb=1.084565 time=249.8s + ttt_chunk [221/474] bpb=1.085631 time=261.6s + ttt_chunk [231/474] bpb=1.085931 time=273.5s + ttt_chunk [241/474] bpb=1.085693 time=285.3s + ttt_chunk [251/474] bpb=1.087461 time=297.1s + ttt_chunk [261/474] bpb=1.088614 time=309.0s + ttt_chunk [271/474] bpb=1.087539 time=320.8s + ttt_chunk [281/474] bpb=1.087260 time=332.7s + ttt_chunk [291/474] bpb=1.085792 time=344.5s + ttt_chunk [301/474] bpb=1.085870 time=356.3s + ttt_chunk [311/474] bpb=1.084865 time=368.2s + ttt_chunk [321/474] bpb=1.082823 time=380.0s + ttt_chunk [331/474] bpb=1.081088 time=391.9s + ttt_chunk [341/474] bpb=1.079459 time=403.7s + ttt_chunk [351/474] bpb=1.076696 time=415.5s + ttt_chunk [361/474] bpb=1.076032 time=427.4s + ttt_chunk [371/474] bpb=1.074447 time=439.2s + ttt_chunk [381/474] bpb=1.073773 time=451.0s + ttt_chunk [391/474] bpb=1.073248 time=462.9s + ttt_chunk [401/474] bpb=1.072241 time=474.7s + ttt_chunk [411/474] bpb=1.071242 time=486.6s + ttt_chunk [421/474] bpb=1.071481 time=498.4s + ttt_chunk [431/474] bpb=1.069344 time=510.2s + ttt_chunk [441/474] bpb=1.068440 time=522.1s + ttt_chunk [451/474] bpb=1.067601 time=533.9s + ttt_chunk [461/474] bpb=1.068029 time=545.7s + ttt_chunk [471/474] bpb=1.068628 time=557.6s + ttt_chunk [474/474] bpb=1.068643 time=560.2s +ttt:done val_loss=1.807296 val_bpb=1.070386 elapsed=560.2s +final_int6_ttt val_loss:1.8073 val_bpb:1.0704 stride:32 eval_time:561081ms +final_int6_ttt_exact val_loss:1.80729571 val_bpb:1.07038587 diff --git a/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/requirements.txt b/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/requirements.txt new file mode 100644 index 000000000..a124ed4f5 --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/requirements.txt @@ -0,0 +1,5 @@ +torch>=2.10.0 +numpy +sentencepiece +zstandard +flash-attn-3 diff --git a/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/submission.json b/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/submission.json new file mode 100644 index 000000000..2835282df --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/submission.json @@ -0,0 +1,14 @@ +{ + "track": "10min_16mb", + "date": "2026-03-24", + "name": "5-expert Hedge Mixer + TTT", + "author": "notapplica", + "seed_results": { + "1337": {"val_loss": 1.78307674, "val_bpb": 1.05604198, "artifact_bytes": 15484405}, + "42": {"val_loss": 1.85218082, "val_bpb": 1.09696945, "artifact_bytes": 15408595}, + "7": {"val_loss": 1.80729571, "val_bpb": 1.07038587, "artifact_bytes": 15428119} + }, + "mean_val_loss": 1.81418442, + "mean_val_bpb": 1.07446577, + "code_bytes": 96416 +} diff --git a/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/train_gpt.py b/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/train_gpt.py new file mode 100644 index 000000000..7460ef9be --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/train_gpt.py @@ -0,0 +1,1966 @@ +"""V25: LeakyReLU^2 + TempCal + Mixed int5/int6 + 33.6M model.""" +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path +try: + import zstandard + _COMPRESSOR = "zstd" +except ImportError: + _COMPRESSOR = "zlib" +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +try: + from flash_attn_interface import flash_attn_func as flash_attn_3_func + _HAS_FA3 = True +except ImportError: + try: + from flash_attn import flash_attn_func as flash_attn_3_func + _HAS_FA3 = True + except ImportError: + _HAS_FA3 = False + flash_attn_3_func = None + +class LogisticContextMixer: + """GPU-vectorized logistic context mixing (inspired by PAQ compression). + + Maintains GPU-resident n-gram count tables and learns online mixing weights + using the Hedge/multiplicative-weights algorithm. + + Experts: + 0: Neural model (logits passed in) + 1: Unigram frequencies from scored tokens + 2: Bigram frequencies (prev_token → next_token) + 3: FastPPM (orders 0-4, CPU-side) + 4: ExactMatchCache (high-order exact matches, CPU-side) + """ + + def __init__(self, vocab_size: int = 1024, device: str = 'cuda', eta: float = 0.1): + self.V = vocab_size + self.device = device + self.eta = eta # Hedge learning rate + self.K = 5 # number of experts + + # Expert weights (log-domain for numerical stability) + self.log_weights = torch.zeros(self.K, device=device) + # Bias toward neural model initially + self.log_weights[0] = 2.0 + + # N-gram count tables (GPU-resident) + self.uni_counts = torch.zeros(vocab_size, device=device) + self.bi_counts = torch.zeros(vocab_size, vocab_size, device=device) + self.total_tokens = 0 + + # GPU Trigram: hashed table [HASH_SIZE, V] to keep memory reasonable + self.TRI_HASH = 65536 # 64K hash buckets for (prev2, prev1) pairs + self.tri_counts = torch.zeros(self.TRI_HASH, vocab_size, device=device) + self.tri_row_totals = torch.zeros(self.TRI_HASH, device=device) + + def update(self, tokens): + """Update all expert statistics with newly scored tokens.""" + if hasattr(tokens, 'cpu'): + t = tokens.to(self.device).long() + else: + t = torch.tensor(tokens, device=self.device, dtype=torch.long) + + n = t.numel() + if n == 0: + return + self.total_tokens += n + + # Unigram: in-place scatter_add + ones = torch.ones(n, device=self.device) + self.uni_counts.scatter_add_(0, t, ones) + + # Bigram: in-place scatter_add on flattened view (no temporary 1M tensor) + if n >= 2: + ctx = t[:-1] + nxt = t[1:] + bi_idx = ctx * self.V + nxt + ones_bi = torch.ones(n - 1, device=self.device) + self.bi_counts.reshape(-1).scatter_add_(0, bi_idx, ones_bi) + + # Trigram: in-place scatter_add on flattened view (no temporary 67M tensor) + if n >= 3: + prev2 = t[:-2] + prev1 = t[1:-1] + nxt3 = t[2:] + tri_ctx = ((prev2 * 36313) ^ (prev1 * 27191)) % self.TRI_HASH + tri_idx = tri_ctx * self.V + nxt3 + ones_tri = torch.ones(n - 2, device=self.device) + self.tri_counts.reshape(-1).scatter_add_(0, tri_idx, ones_tri) + self.tri_row_totals.scatter_add_(0, tri_ctx, ones_tri) + + def get_expert_log_probs(self, neural_logits, x_batch, y_batch, wlens): + """Get log-probability of targets from each expert. All GPU-vectorized. + + Args: + neural_logits: [bsz, seq_len, V] neural model logits + x_batch: [bsz, seq_len] input tokens (context) + y_batch: [bsz, seq_len] target tokens + wlens: list of actual lengths per sequence + + Returns: + expert_nll: [bsz, seq_len, K] NLL from each expert + """ + bsz, slen, V = neural_logits.shape + uniform_nll = math.log(self.V) + has_data = self.total_tokens > 0 # Python int — no GPU-CPU sync + + # Expert 0: Neural model — compute log_softmax once, reuse for entropy + neural_lp = F.log_softmax(neural_logits, dim=-1) + neural_nll = -neural_lp.gather(2, y_batch.unsqueeze(2)).squeeze(2) # [bsz, slen] + + # Expert 1: Unigram + if has_data: + uni_probs = (self.uni_counts + 0.1) / (self.total_tokens + 0.1 * self.V) + uni_nll = -uni_probs.log()[y_batch] # [bsz, slen] + else: + uni_nll = torch.full((bsz, slen), uniform_nll, device=self.device) + + # Expert 2: Bigram P(next | prev) + if has_data: + bi_total = self.bi_counts.sum(dim=1, keepdim=True) # [V, 1] + bi_probs = (self.bi_counts + 0.1) / (bi_total + 0.1 * self.V) # [V, V] + prev_flat = x_batch.reshape(-1) + next_flat = y_batch.reshape(-1) + bi_nll = -bi_probs.log()[prev_flat, next_flat].reshape(bsz, slen) + else: + bi_nll = torch.full((bsz, slen), uniform_nll, device=self.device) + + # Expert 3: GPU Trigram P(next | hash(prev2, prev1)) — vectorized + if has_data and slen >= 2: + prev2 = torch.zeros_like(x_batch) + prev2[:, 1:] = x_batch[:, :-1] + ctx_hash = ((prev2 * 36313) ^ (x_batch * 27191)) % self.TRI_HASH + ctx_flat = ctx_hash.reshape(-1).long() + next_flat = y_batch.reshape(-1).long() + tri_count = self.tri_counts[ctx_flat, next_flat] + tri_total = self.tri_row_totals[ctx_flat].clamp(min=1) + tri_prob = (tri_count + 0.01) / (tri_total + 0.01 * self.V) + tri_nll = -tri_prob.log().reshape(bsz, slen) + else: + tri_nll = torch.full((bsz, slen), uniform_nll, device=self.device) + + # Expert 4: Neural entropy — reuse neural_lp (no redundant softmax) + entropy_nll = -(neural_lp.exp() * neural_lp).sum(-1) # [bsz, slen] + + # Stack: [bsz, slen, K] + return torch.stack([neural_nll, uni_nll, bi_nll, tri_nll, entropy_nll], dim=-1) + + def mix_and_score(self, neural_logits, x_batch, y_batch, wlens): + """Compute mixed NLL using current expert weights. + + Returns (mixed_nll [bsz, slen], expert_nll [bsz, slen, K] or None). + Caller should pass expert_nll to update_weights() to avoid recomputation. + """ + if self.total_tokens < 10000: + # Not enough data for n-grams — just use neural + nll = F.cross_entropy( + neural_logits.reshape(-1, neural_logits.size(-1)), + y_batch.reshape(-1), reduction="none" + ).reshape(neural_logits.shape[0], neural_logits.shape[1]) + return nll, None + + expert_nll = self.get_expert_log_probs(neural_logits, x_batch, y_batch, wlens) # [bsz, slen, K] + + # Log-domain mixing: log(sum_k w_k * p_k) = logsumexp(log_w_k + log_p_k) + log_w = self.log_weights - self.log_weights.logsumexp(0) # normalize + mixed_lp = (-expert_nll + log_w.unsqueeze(0).unsqueeze(0)).logsumexp(dim=-1) # [bsz, slen] + + return -mixed_lp, expert_nll # mixed NLL + cached expert NLL + + def update_weights(self, expert_nll, wlens): + """Update expert weights using Hedge algorithm on pre-computed expert NLLs.""" + if expert_nll is None: + return + + with torch.no_grad(): + # Vectorized mask: compare position index against window lengths + bsz, slen = expert_nll.shape[0], expert_nll.shape[1] + wlens_t = torch.tensor(wlens, device=self.device, dtype=torch.long) + mask = torch.arange(slen, device=self.device).unsqueeze(0) < wlens_t.unsqueeze(1) # [bsz, slen] bool + + # Masked mean NLL per expert + masked_nll = expert_nll * mask.unsqueeze(-1).float() + expert_mean_loss = masked_nll.sum(dim=(0, 1)) / mask.sum().clamp(min=1) # [K] + + # Hedge update: log_w -= eta * loss + self.log_weights -= self.eta * expert_mean_loss + + +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 8)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.5)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 32)) + int6_last_n = int(os.environ.get("INT6_LAST_N", 0)) # all int5 (saves ~300KB vs int6 for last 2 blocks) + ttt_temperature = float(os.environ.get("TTT_TEMPERATURE", 0.98)) # post-TTT temperature calibration + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 6144)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 11)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.5)) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + prune_pct = float(os.environ.get("PRUNE_PCT", 0.03)) + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X + +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] + +def eval_val(args: Hyperparameters, model: nn.Module, rank: int, world_size: int, + device: torch.device, grad_accum_steps: int, val_tokens: Tensor, + base_bytes_lut: Tensor, has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, eval_seq_len: int | None = None) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale", + ).split(",") + if pattern +) +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_Q = 0.9999984 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + _soft_round_alpha: float = 1.0 # temperature for soft-round (annealed during training) + _use_soft_round: bool = False # enable soft-round QAT instead of STE + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self._clip_range = 15 # default int5, set to 31 for int6 layers + + @staticmethod + def soft_round(y: Tensor, alpha: float) -> Tensor: + """Differentiable approximation to round() from Agustsson & Theis (NeurIPS 2020). + s_alpha(y) = floor(y) + 0.5 * tanh(alpha * r) / tanh(alpha/2) + 0.5 + where r = y - floor(y) - 0.5 (centered fractional part) + """ + fl = torch.floor(y) + r = y - fl - 0.5 + return fl + 0.5 * torch.tanh(alpha * r) / (math.tanh(alpha / 2) + 1e-10) + 0.5 + + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + cr = self._clip_range + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + if CastedLinear._use_soft_round: + # Soft-Round QAT: differentiable rounding with temperature annealing + w32 = self.weight.float() + row_clip = torch.quantile(w32.abs(), 0.9995, dim=1) + scale = (row_clip / float(cr)).clamp_min(1.0 / float(cr)) + w_scaled = w32 / scale[:, None] + w_rounded = CastedLinear.soft_round(w_scaled, CastedLinear._soft_round_alpha) + w_q = (torch.clamp(w_rounded, -(cr+1), cr) * scale[:, None]).to(x.dtype) + w = w_q # fully differentiable path + else: + # Original STE QAT + with torch.no_grad(): + w32 = self.weight.float() + row_clip = torch.quantile(w32.abs(), 0.9995, dim=1) + scale = (row_clip / float(cr)).clamp_min(1.0 / float(cr)) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -(cr+1), cr) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, + rope_base: float, qk_gain_init: float): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False + + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + B, T, H, D = y.shape + Hkv = v.size(-2) + y_g = y.reshape(B, T, Hkv, H // Hkv, D) + vn = F.normalize(v, dim=-1).unsqueeze(-2) + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + + def forward(self, x: Tensor, v_embed: Tensor | None = None) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = self.c_v(x) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if _HAS_FA3: + y = flash_attn_3_func(q, k, v, causal=True).contiguous() + else: + y = F.scaled_dot_product_attention( + q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), + attn_mask=None, is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + return self.proj(y) + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class ValueEmbedding(nn.Module): + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + return self.proj(F.leaky_relu(self.fc(x), negative_slope=0.5).square()) + +class Block(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, mlp_mult: int, + rope_base: float, qk_gain_init: float, layer_idx: int = 0, + ln_scale: bool = False, dtg: bool = False): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + + def forward(self, x: Tensor, x0: Tensor, v_embed: Tensor | None = None) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, v_embed=v_embed) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out + +class GPT(nn.Module): + def __init__(self, vocab_size: int, num_layers: int, model_dim: int, num_heads: int, + num_kv_heads: int, mlp_mult: int, tie_embeddings: bool, tied_embed_init_std: float, + logit_softcap: float, rope_base: float, qk_gain_init: float, + bigram_vocab_size: int = 0, bigram_dim: int = 128, xsa_last_n: int = 0, + rope_dims: int = 0, ln_scale: bool = False, dtg: bool = False, + ve_enabled: bool = False, ve_dim: int = 128, ve_layers: str = "9,10"): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList([ + Block(model_dim, num_heads, num_kv_heads, mlp_mult, rope_base, + qk_gain_init, layer_idx=i, ln_scale=ln_scale, dtg=dtg) + for i in range(num_layers) + ]) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + num_layers = len(self.blocks) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + if ".proj." in name or name.endswith(".proj"): + with torch.no_grad(): + module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) + + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x = self.blocks[i](x, x0, v_embed=ve) + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x = self.blocks[bi](x, x0, v_embed=ve) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return F.cross_entropy(logits.float(), targets, reduction="mean") + + def forward_logits(self, input_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x = self.blocks[i](x, x0, v_embed=ve) + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x = self.blocks[bi](x, x0, v_embed=ve) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + +def eval_val_sliding(args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, eval_seq_len: int | None = None) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + + # Pre-compile: dummy forward+backward with TTT shapes to warm the compile cache + if rank == 0: + print(" ttt: pre-compiling forward+backward kernels...", flush=True) + _dummy_x = torch.zeros(1, seq_len, dtype=torch.int64, device=device) + _dummy_y = torch.zeros(1, seq_len, dtype=torch.int64, device=device) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + _dummy_logits = base_model.forward_logits(_dummy_x) + _dummy_loss = F.cross_entropy(_dummy_logits.reshape(-1, _dummy_logits.size(-1)), _dummy_y.reshape(-1)) + _dummy_loss.backward() + base_model.zero_grad(set_to_none=True) + if rank == 0: + print(" ttt: pre-compile done", flush=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + +def eval_val_sliding_ttt( + args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, ttt_epochs: int = 3, ttt_lr: float = 0.001, + ttt_momentum: float = 0.9, ttt_freeze_blocks: int = 2, + batch_seqs: int = 32, eval_seq_len: int | None = None, + ttt_chunk_tokens: int = 32768, ttt_optimizer: str = "adamw", + ttt_temp: float = 1.0, + ppm_alpha: float = 0.85, + byte_weighted_ttt: bool = True, + use_cache: bool = True, + cache_alpha: float = 0.3, + adaptive_lr: bool = True, + adaptive_lr_max_mult: float = 3.0, +) -> tuple[float, float]: + """Legal score-first TTT: score each chunk, then train on it. + Every token scored BEFORE any update that could use it.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + + # Initialize GPU-vectorized logistic context mixer + use_mixer = os.environ.get("USE_MIXER", "1") == "1" + mixer = LogisticContextMixer( + vocab_size=val_tokens.to(torch.int32).max().item() + 1, + device=device, + eta=float(os.environ.get("MIXER_ETA", "0.1")), + ) if use_mixer else None + if use_mixer and rank == 0: + print(f" Logistic context mixer enabled: eta={mixer.eta}") + if adaptive_lr and rank == 0: + print(f" Adaptive LR enabled: max_mult={adaptive_lr_max_mult}") + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on scored token position + num_chunks = (total_tokens + ttt_chunk_tokens - 1) // ttt_chunk_tokens + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk_tokens, num_chunks - 1) + chunk_windows[ci].append(ws) + + if rank == 0: + print(f"ttt:start chunks={num_chunks} chunk_tokens={ttt_chunk_tokens} " + f"windows={len(window_starts)} stride={stride} " + f"lr={ttt_lr} epochs={ttt_epochs} opt={ttt_optimizer} " + f"freeze_first={ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze everything, then selectively unfreeze for TTT + num_blocks = len(base_model.blocks) + for p in base_model.parameters(): + p.requires_grad_(False) + ttt_params = [] + ttt_param_ids = set() + use_qttt = os.environ.get("QTTT", "0") == "1" + if use_qttt: + # qTTT: only unfreeze Q projections in last N blocks + norms + head + for i in range(max(0, num_blocks - ttt_freeze_blocks), num_blocks): + for name, p in base_model.blocks[i].named_parameters(): + if "c_q" in name: + p.requires_grad_(True) + ttt_params.append(p) + ttt_param_ids.add(id(p)) + else: + # Standard: unfreeze all params in last N blocks + for i in range(max(0, num_blocks - ttt_freeze_blocks), num_blocks): + for p in base_model.blocks[i].parameters(): + p.requires_grad_(True) + ttt_params.append(p) + ttt_param_ids.add(id(p)) + # Unfreeze norms, scales, lm_head + for name, p in base_model.named_parameters(): + if "norm" in name or "scale" in name or "lm_head" in name: + p.requires_grad_(True) + if id(p) not in ttt_param_ids: + ttt_params.append(p) + ttt_param_ids.add(id(p)) + + if rank == 0: + n_unfrozen = sum(p.numel() for p in ttt_params) + n_frozen = sum(p.numel() for p in base_model.parameters() if not p.requires_grad) + print(f"ttt:params unfrozen={n_unfrozen} frozen={n_frozen}") + + if ttt_optimizer == "adamw": + optimizer = torch.optim.AdamW(ttt_params, lr=ttt_lr, weight_decay=0.0, betas=(0.9, 0.999)) + else: + optimizer = torch.optim.SGD(ttt_params, lr=ttt_lr, momentum=ttt_momentum) + + # Polyak averaging (TTT weight EMA) for smoother scoring + use_polyak = os.environ.get("USE_POLYAK", "1") == "1" + polyak_decay = float(os.environ.get("POLYAK_DECAY", "0.998")) + if use_polyak: + polyak_state = {id(p): p.data.clone() for p in ttt_params} + if rank == 0: + print(f" Polyak averaging enabled: decay={polyak_decay}") + + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + + # --- Phase 1: SCORE this chunk (inference_mode, no grad) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + # Swap in Polyak-averaged weights for scoring + if use_polyak and ci > 0: + _saved_weights = {} + for p in ttt_params: + _saved_weights[id(p)] = p.data.clone() + p.data.copy_(polyak_state[id(p)]) + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + logits_scaled = logits.float() / ttt_temp + + # Adaptive temperature: sharpen confident predictions more + if ttt_temp != 1.0: + with torch.no_grad(): + probs_for_entropy = F.softmax(logits.float(), dim=-1) + token_entropy = -(probs_for_entropy * (probs_for_entropy + 1e-10).log()).sum(-1) + max_ent = math.log(logits.size(-1)) + # Confident tokens (low entropy) get more sharpening + adaptive_temp = 1.0 - (1.0 - ttt_temp) * (1.0 - token_entropy / max_ent) + adaptive_temp = adaptive_temp.clamp(min=0.9, max=1.05) + logits_scaled = logits.float() / adaptive_temp.unsqueeze(-1) + + # Logistic context mixing (GPU-vectorized) or plain CE + if mixer is not None: + nll, expert_nll = mixer.mix_and_score(logits_scaled, x_batch, y_batch, wlens) + mixer.update_weights(expert_nll, wlens) + else: + nll = F.cross_entropy( + logits_scaled.reshape(-1, logits_scaled.size(-1)), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # --- Update context mixer with scored chunk tokens (GPU-vectorized) --- + chunk_start_tok = ci * ttt_chunk_tokens + chunk_end_tok = min((ci + 1) * ttt_chunk_tokens, total_tokens) + if mixer is not None: + mixer.update(val_tokens[chunk_start_tok:chunk_end_tok + 1]) + + # Swap back training weights after scoring + if use_polyak and ci > 0: + for p in ttt_params: + p.data.copy_(_saved_weights[id(p)]) + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and ttt_epochs > 0: + chunk_start = ci * ttt_chunk_tokens + chunk_end = min((ci + 1) * ttt_chunk_tokens, total_tokens) + chunk_seqs = (chunk_end - chunk_start) // seq_len + if rank == 0 and ci < 3: + print(f" ttt_train [{ci+1}] seqs={chunk_seqs} start_train...", flush=True) + if chunk_seqs > 0: + # Cosine LR across chunks + adaptive scaling + cos_lr = ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + if adaptive_lr: + # Increase LR as we've seen more data (more confident adaptation) + progress = min(ci / max(num_chunks * 0.3, 1), 1.0) # ramp over first 30% of chunks + lr_mult = 1.0 + (adaptive_lr_max_mult - 1.0) * progress + cos_lr = cos_lr * lr_mult + for pg in optimizer.param_groups: + pg["lr"] = cos_lr + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(ttt_epochs): + if rank == 0 and ci < 3: + print(f" ttt_train [{ci+1}] epoch={_ep+1}/{ttt_epochs} batches={my_chunk_seqs} ...", flush=True) + for bs in range(0, my_chunk_seqs, batch_seqs): + be = min(bs + batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + if byte_weighted_ttt: + # Byte-weighted loss: tokens covering more bytes matter more + ttt_logits = base_model.forward_logits(x) + per_token_loss = F.cross_entropy( + ttt_logits.reshape(-1, ttt_logits.size(-1)), + y.reshape(-1), reduction='none' + ).reshape(y.shape) + byte_weights = base_bytes_lut[y].float() + byte_weights = byte_weights + (has_leading_space_lut[y] & ~is_boundary_token_lut[x]).float() + ttt_loss = (per_token_loss * byte_weights).sum() / byte_weights.sum() + else: + ttt_loss = base_model(x, y) + ttt_loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, 1.0) + optimizer.step() + # Update Polyak EMA after each step + if use_polyak: + for p in ttt_params: + polyak_state[id(p)].lerp_(p.data, 1.0 - polyak_decay) + if rank == 0 and ci < 3: + print(f" step done ep={_ep+1} bs={bs} loss={ttt_loss.item():.4f}", flush=True) + + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1 or ci < 5): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + print(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s", flush=True) + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + if rank == 0: + print(f"ttt:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" + +def quantize_int6_per_row(t: Tensor, clip_range: int = 15) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale + + +def _find_best_row_scales(W: Tensor, clip_range: int = 15) -> Tensor: + """Find optimal per-row scales by searching percentile clipping thresholds.""" + t32 = W.float() + best_s = t32.abs().amax(dim=1) / clip_range + best_s = best_s.clamp_min(1.0 / clip_range) + best_err = torch.full((t32.shape[0],), float('inf')) + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range) + q = torch.clamp(torch.round(t32 / s[:, None]), -clip_range, clip_range) + recon = q * s[:, None] + err = (t32 - recon).pow(2).mean(dim=1) + improved = err < best_err + best_s[improved] = s[improved] + best_err[improved] = err[improved] + return best_s + +def gptq_quantize_weight(W: Tensor, H: Tensor, clip_range: int = 15, + block_size: int = 128, percdamp: float = 0.01) -> tuple[Tensor, Tensor]: + """GPTQ: quantize weight matrix W using Hessian H = X^T X for error compensation.""" + W = W.float().clone() + rows, cols = W.shape + row_scale = _find_best_row_scales(W, clip_range) + H = H.float().clone() + damp = percdamp * H.diag().mean() + H.diagonal().add_(damp) + perm = torch.argsort(H.diag()) + invperm = torch.argsort(perm) + W = W[:, perm] + H = H[perm][:, perm] + try: + L = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(L) + except torch._C._LinAlgError: + Hinv = torch.diag(1.0 / H.diag().clamp_min(1e-6)) + Q = torch.zeros(rows, cols, dtype=torch.int8) + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + W_block = W[:, i1:i2].clone() + Hinv_block = Hinv[i1:i2, i1:i2] + Err = torch.zeros_like(W_block) + for j in range(i2 - i1): + w_col = W_block[:, j] + h_inv_jj = Hinv_block[j, j].clamp_min(1e-8) + q_col = torch.clamp(torch.round(w_col / row_scale), -clip_range, clip_range) + deq_col = q_col * row_scale + Q[:, i1 + j] = q_col.to(torch.int8) + err = (w_col - deq_col) / h_inv_jj + Err[:, j] = err + if j + 1 < i2 - i1: + W_block[:, j + 1:] -= err.unsqueeze(1) * Hinv_block[j, j + 1:].unsqueeze(0) + if i2 < cols: + W[:, i2:] -= Err @ Hinv[i1:i2, i2:] + Q = Q[:, invperm] + return Q, row_scale.to(torch.float16) + +def gptq_calibrate(model: nn.Module, train_pattern: str, device: torch.device, + n_samples: int = 256, seq_len: int = 2048) -> dict[str, Tensor]: + """Collect Hessian H = X^T X for each linear layer using training data.""" + hessians: dict[str, Tensor] = {} + n_seen: dict[str, int] = {} + hooks = [] + def make_hook(name: str): + def hook_fn(module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + if name not in hessians: + hessians[name] = torch.zeros(x.shape[1], x.shape[1], device=x.device, dtype=torch.float32) + n_seen[name] = 0 + hessians[name].addmm_(x.t(), x) + n_seen[name] += x.shape[0] + return hook_fn + for name, module in model.named_modules(): + if isinstance(module, (nn.Linear, CastedLinear)): + hooks.append(module.register_forward_hook(make_hook(name))) + stream = TokenStream(train_pattern) + model.eval() + with torch.no_grad(): + for _ in range(n_samples): + tokens = stream.take(seq_len + 1).to(device=device, dtype=torch.int64) + x = tokens[:-1].unsqueeze(0) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + model.forward_logits(x) + for h in hooks: + h.remove() + for name in hessians: + hessians[name] /= max(n_seen[name], 1) + return hessians + +def _get_layer_clip_range(name: str, num_layers: int, int6_last_n: int) -> int: + """Return clip_range based on which layer the param belongs to.""" + import re + m = re.search(r'blocks\.(\d+)\.', name) + if m: + layer_idx = int(m.group(1)) + if layer_idx >= num_layers - int6_last_n: + return 31 # int6 + return 15 # int5 + +def mixed_quantize_int6_gptq(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor], + num_layers: int = 11, int6_last_n: int = 2) -> tuple[dict, dict]: + """GPTQ quantization with mixed int5/int6 precision. int6 for last int6_last_n layers, int5 for rest.""" + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + gptq_count, naive_count = 0, 0 + int5_params, int6_params = 0, 0 + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + cr = _get_layer_clip_range(name, num_layers, int6_last_n) + if cr == 31: + int6_params += t.numel() + else: + int5_params += t.numel() + if cat in int6_cats and t.ndim == 2: + module_name = name.rsplit(".weight", 1)[0] if name.endswith(".weight") else name + H = hessians.get(module_name) + if H is not None and H.shape[0] == t.shape[1]: + q, s = gptq_quantize_weight(t, H.cpu(), clip_range=cr) + gptq_count += 1 + else: + q, s = quantize_int6_per_row(t, clip_range=cr) + naive_count += 1 + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": f"int{'6' if cr == 31 else '5'}"} + elif cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t, clip_range=cr) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": f"int{'6' if cr == 31 else '5'}"} + naive_count += 1 + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + print(f"gptq_quantize: {gptq_count} GPTQ layers, {naive_count} naive layers", flush=True) + print(f"mixed_precision: {int5_params} int5 params, {int6_params} int6 params", flush=True) + return result, meta + + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta + +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +def main() -> None: + global zeropower_via_newtonschulz5 + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0(f"Python {sys.version} PyTorch {torch.__version__}", console=False) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + CastedLinear._qat_enabled = args.qat_enabled + base_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, rope_dims=args.rope_dims, ln_scale=args.ln_scale, + dtg=args.dtg_enabled, ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + matrix_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + # Set int6 clip_range for last N layers (mixed precision) + int6_start = args.num_layers - args.int6_last_n + for i, block in enumerate(base_model.blocks): + if i >= int6_start: + for m in block.modules(): + if isinstance(m, CastedLinear): + m._clip_range = 31 # int6 + if master_process: + int5_count = sum(1 for m in base_model.modules() if isinstance(m, CastedLinear) and m._clip_range == 15) + int6_count = sum(1 for m in base_model.modules() if isinstance(m, CastedLinear) and m._clip_range == 31) + log0(f"mixed_precision: {int5_count} int5 layers, {int6_count} int6 layers (last {args.int6_last_n} blocks)") + log0(f"model_params:{n_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:{xsa_layers} ws:{world_size} gqa:{args.num_heads}/{args.num_kv_heads}") + log0(f"lr:embed={token_lr} matrix={args.matrix_lr} scalar={args.scalar_lr} batch:{args.train_batch_tokens} wall:{args.max_wallclock_seconds:.0f}s seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + train_reserve_ms = 18000 # reserve 18s for EMA + GPTQ calibration + quantization + save + effective_train_ms = (max_wallclock_ms - train_reserve_ms) if max_wallclock_ms is not None else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if effective_train_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(effective_train_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + # TTT_ONLY mode: skip training, load saved model, run TTT eval + if os.environ.get("TTT_ONLY", "0") == "1": + log0("TTT_ONLY mode: skipping training, loading saved model...") + sd_cpu = {k: v.cpu() for k, v in torch.load("final_model.pt", map_location="cpu").items()} + if args.prune_pct > 0: + for k, v in sd_cpu.items(): + if v.ndim == 2 and v.numel() > 65536: + thresh = torch.quantile(v.abs().float(), args.prune_pct) + v[v.abs() < thresh] = 0.0 + log0(f"pruning:{args.prune_pct*100:.1f}% magnitude pruning applied") + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(zstandard.ZstdDecompressor().decompress(quant_blob_disk) if _COMPRESSOR == "zstd" else zlib.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + sw_seq_len = int(os.environ.get("EVAL_SEQ_LEN", str(effective_eval_seq_len))) + log0(f"TTT_ONLY: model loaded, starting TTT eval...") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_epochs = int(os.environ.get("TTT_EPOCHS", "3")) + ttt_lr = float(os.environ.get("TTT_LR", "0.0005")) + ttt_freeze = int(os.environ.get("TTT_FREEZE_BLOCKS", "2")) + ttt_chunk = int(os.environ.get("TTT_CHUNK_TOKENS", "32768")) + ttt_opt = os.environ.get("TTT_OPTIMIZER", "adamw") + log0(f"TTT: epochs={ttt_epochs} lr={ttt_lr} freeze_first={ttt_freeze} chunk={ttt_chunk} opt={ttt_opt}") + ttt_temp = args.ttt_temperature + log0(f"TTT temperature: {ttt_temp}") + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, ttt_epochs=ttt_epochs, ttt_lr=ttt_lr, + ttt_freeze_blocks=ttt_freeze, eval_seq_len=sw_seq_len, + ttt_chunk_tokens=ttt_chunk, ttt_optimizer=ttt_opt, + ttt_temp=ttt_temp, + ppm_alpha=float(os.environ.get("PPM_ALPHA", "0.85")), + byte_weighted_ttt=os.environ.get("BYTE_WEIGHTED_TTT", "1") == "1", + use_cache=os.environ.get("USE_CACHE", "1") == "1", + cache_alpha=float(os.environ.get("CACHE_ALPHA", "0.3")), + adaptive_lr=os.environ.get("ADAPTIVE_LR", "1") == "1", + adaptive_lr_max_mult=float(os.environ.get("ADAPTIVE_LR_MAX", "3.0")), + ) + torch.cuda.synchronize() + log0( + f"final_int6_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" + ) + log0(f"final_int6_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() + return + + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + # Anneal soft-round alpha based on QAT progress + if CastedLinear._use_soft_round and CastedLinear._qat_enabled: + qat_progress = max(0.0, 1.0 - scale / max(args.late_qat_threshold, 0.01)) + CastedLinear._soft_round_alpha = 1.0 + 15.0 * qat_progress + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + CastedLinear._use_soft_round = os.environ.get("SOFT_ROUND_QAT", "0") == "1" + if CastedLinear._use_soft_round and master_process: + log0(f"soft_round_qat:enabled initial_alpha=1.0") + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + reached_cap = effective_train_ms is not None and approx_training_time_ms >= effective_train_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + # Apply EMA weights directly (skip diagnostic evals to save ~5s of reserve) + log0("ema:applying EMA weights (skipping diagnostic evals)") + current_state = base_model.state_dict() + ema_sd = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(ema_sd, strict=True) + # GPTQ calibration on final model (within reserved training budget) + log0("gptq:calibrating with training data...") + t_gptq = time.perf_counter() + gptq_hessians = gptq_calibrate(base_model, args.train_files, device, n_samples=128, seq_len=args.train_seq_len) + log0(f"gptq:calibrated {len(gptq_hessians)} layers in {time.perf_counter()-t_gptq:.1f}s") + export_sd = base_model.state_dict() + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + if args.prune_pct > 0: + for k, v in sd_cpu.items(): + if v.ndim == 2 and v.numel() > 65536: + thresh = torch.quantile(v.abs().float(), args.prune_pct) + v[v.abs() < thresh] = 0.0 + if master_process: + log0(f"pruning:{args.prune_pct*100:.1f}% magnitude pruning applied") + quant_result, quant_meta = mixed_quantize_int6_gptq(sd_cpu, {"mlp", "attn"}, gptq_hessians, num_layers=args.num_layers, int6_last_n=args.int6_last_n) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) if _COMPRESSOR == "zstd" else zlib.compress(quant_raw, 9) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+{_COMPRESSOR}: {quant_file_bytes} bytes") + log0(f"Total submission size int6+{_COMPRESSOR}: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(zstandard.ZstdDecompressor().decompress(quant_blob_disk) if _COMPRESSOR == "zstd" else zlib.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + sw_seq_len = int(os.environ.get("EVAL_SEQ_LEN", str(effective_eval_seq_len))) + if sw_seq_len != effective_eval_seq_len and rank == 0: + log0(f"Eval seq_len override: {effective_eval_seq_len} -> {sw_seq_len}") + if args.eval_stride > 0 and args.eval_stride < sw_seq_len and not os.environ.get("SKIP_SLIDING"): + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_epochs = int(os.environ.get("TTT_EPOCHS", "3")) + ttt_lr = float(os.environ.get("TTT_LR", "0.0005")) + ttt_freeze = int(os.environ.get("TTT_FREEZE_BLOCKS", "2")) + ttt_chunk = int(os.environ.get("TTT_CHUNK_TOKENS", "32768")) + ttt_opt = os.environ.get("TTT_OPTIMIZER", "adamw") + log0(f"TTT: epochs={ttt_epochs} lr={ttt_lr} freeze_first={ttt_freeze} chunk={ttt_chunk} opt={ttt_opt}") + ttt_temp = args.ttt_temperature + log0(f"TTT temperature: {ttt_temp}") + ppm_alpha_val = float(os.environ.get("PPM_ALPHA", "0.85")) + bw_ttt = os.environ.get("BYTE_WEIGHTED_TTT", "1") == "1" + log0(f"PPM alpha: {ppm_alpha_val}, Byte-weighted TTT: {bw_ttt}") + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, ttt_epochs=ttt_epochs, ttt_lr=ttt_lr, + ttt_freeze_blocks=ttt_freeze, eval_seq_len=sw_seq_len, + ttt_chunk_tokens=ttt_chunk, ttt_optimizer=ttt_opt, + ttt_temp=ttt_temp, + ppm_alpha=float(os.environ.get("PPM_ALPHA", "0.85")), + byte_weighted_ttt=os.environ.get("BYTE_WEIGHTED_TTT", "1") == "1", + use_cache=os.environ.get("USE_CACHE", "1") == "1", + cache_alpha=float(os.environ.get("CACHE_ALPHA", "0.3")), + adaptive_lr=os.environ.get("ADAPTIVE_LR", "1") == "1", + adaptive_lr_max_mult=float(os.environ.get("ADAPTIVE_LR_MAX", "3.0")), + ) + torch.cuda.synchronize() + log0( + f"final_int6_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" + ) + log0(f"final_int6_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() From 9a1d2fdd71b32dc3c2fe8654a76c4398f1f0f8c0 Mon Sep 17 00:00:00 2001 From: Royi Rassin Date: Wed, 25 Mar 2026 08:44:20 +0200 Subject: [PATCH 3/4] Remove old submission directory --- submission-2026-03-24/README.md | 110 -- submission-2026-03-24/log_seed1337.txt | 173 --- submission-2026-03-24/log_seed42.txt | 173 --- submission-2026-03-24/log_seed7.txt | 173 --- submission-2026-03-24/requirements.txt | 5 - submission-2026-03-24/submission.json | 14 - submission-2026-03-24/train_gpt.py | 1966 ------------------------ 7 files changed, 2614 deletions(-) delete mode 100644 submission-2026-03-24/README.md delete mode 100644 submission-2026-03-24/log_seed1337.txt delete mode 100644 submission-2026-03-24/log_seed42.txt delete mode 100644 submission-2026-03-24/log_seed7.txt delete mode 100644 submission-2026-03-24/requirements.txt delete mode 100644 submission-2026-03-24/submission.json delete mode 100644 submission-2026-03-24/train_gpt.py diff --git a/submission-2026-03-24/README.md b/submission-2026-03-24/README.md deleted file mode 100644 index 2753515cc..000000000 --- a/submission-2026-03-24/README.md +++ /dev/null @@ -1,110 +0,0 @@ -# 5-expert Hedge Mixer + TTT - -**val_bpb: 1.0745** (3-seed mean) | **<15.5 MB** | 8xH100 SXM - -## Results (8xH100 80GB SXM) - -| Seed | steps | step_avg | Pre-TTT bpb | **Post-TTT bpb** | TTT gain | Eval time | Artifact | -|------|-------|----------|-------------|-----------------|----------|-----------|----------| -| 1337 | 5,997 | 97.1ms | 1.1248 | **1.0560** | -0.0688 | 563s | 15.48 MB | -| 42 | 5,997 | 97.1ms | 1.1257 | **1.0970** | -0.0287 | 563s | 15.41 MB | -| 7 | 5,983 | 97.3ms | 1.1251 | **1.0704** | -0.0547 | 561s | 15.43 MB | -| **Mean** | | | **1.1252** | **1.0745** | **-0.0507** | | | - -## Key Contribution: 5-expert Logistic Context Mixer - -GPU-vectorized online context mixing using the Hedge/multiplicative-weights algorithm. Five experts blend predictions in log-probability space: - -| Expert | Source | Description | -|--------|--------|-------------| -| 0 | Neural | Base model log-softmax | -| 1 | Unigram | Token frequency from scored tokens | -| 2 | Bigram | P(next \| prev) from scored tokens | -| 3 | Trigram | Hashed P(next \| prev2, prev1) with 64K buckets | -| 4 | Entropy | Neural model entropy as confidence regularizer | - -Expert weights are updated online via Hedge: `log_w -= eta * loss`. N-gram tables are built incrementally from already-scored tokens only (legal). - -## Architecture - -PR #606 base with the following additions: - -| Component | Setting | -|-----------|---------| -| Layers | 11 (512d, 8H, 8KV) | -| MLP | 3x with **LeakyReLU(0.5)^2** | -| BigramHash | 6144 (dim=128) | -| XSA | All 11 layers (ws=8) | -| RoPE | Partial (16/64 dims) | -| LN Scale | 1/sqrt(layer+1) | -| VE128 | Layers 9-10 | -| Weight avg | EMA(0.997) | -| Quantization | Full GPTQ int5 + zstd (level 22) | -| Pruning | 3% magnitude | - -## Legal Score-First TTT - -Backward-looking adaptation with GPTQ-calibrated model: - -1. Validation tokens split into 474 chunks of 131K tokens each -2. For each chunk: - - **SCORE**: Sliding window eval (stride=32, seq_len=2048) with 5-expert mixer blending - - **TRAIN**: AdamW(lr=0.0001) on already-scored chunk. 3 epochs, last 2 blocks unfrozen + norms + lm_head, cosine LR decay, Polyak averaging -3. Last chunk scored but never trained on - -### TTT Hyperparameters - -| Parameter | Value | -|-----------|-------| -| Chunk size | 131,072 tokens | -| Optimizer | AdamW (lr=0.0001) | -| Epochs per chunk | 3 | -| Frozen blocks | First 9 (last 2 + norms + head unfrozen) | -| Polyak decay | 0.998 | -| Adaptive LR | max_mult=3.0 | -| Mixer eta | 0.1 | - -### Training Budget - -GPTQ calibration runs within the 600s training budget (18s reserved from training loop for EMA selection + calibration + quantization). - -| Phase | Time | -|-------|------| -| Training loop | 582s | -| EMA + GPTQ calibration + quantization | ~18s | -| **Total training** | **~600s** | -| Sliding window eval | ~165s | -| TTT eval with mixer | ~562s | -| **Total eval** | **~562s** | - -## Reproduction - -```bash -# Install dependencies -pip install -r requirements.txt -# Build FA3 Hopper kernels (required) -cd /tmp && git clone https://github.com/Dao-AILab/flash-attention -cd flash-attention/hopper && python setup.py install - -# Run training + eval (single seed) -DATA_PATH=./data/datasets/fineweb10B_sp1024 \ -TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model \ -SEED=1337 MAX_WALLCLOCK_SECONDS=600 \ -USE_MIXER=1 TTT_LR=0.0001 TTT_CHUNK_TOKENS=131072 \ - torchrun --standalone --nproc_per_node=8 train_gpt.py - -# Run all 3 seeds -for SEED in 1337 42 7; do - DATA_PATH=./data/datasets/fineweb10B_sp1024 \ - TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model \ - SEED=$SEED MAX_WALLCLOCK_SECONDS=600 \ - USE_MIXER=1 TTT_LR=0.0001 TTT_CHUNK_TOKENS=131072 \ - torchrun --standalone --nproc_per_node=8 train_gpt.py -done -``` - -## Credits - -- **Base model**: PR #606 by @gowtham0992 -- **TTT recipe**: PR #461 by @Christopher-Lee-McClendon -- **Mixer inspiration**: PAQ compression (context mixing) + Hedge algorithm diff --git a/submission-2026-03-24/log_seed1337.txt b/submission-2026-03-24/log_seed1337.txt deleted file mode 100644 index 656321ad9..000000000 --- a/submission-2026-03-24/log_seed1337.txt +++ /dev/null @@ -1,173 +0,0 @@ -W0325 00:33:17.141000 673732 torch/distributed/run.py:852] -W0325 00:33:17.141000 673732 torch/distributed/run.py:852] ***************************************** -W0325 00:33:17.141000 673732 torch/distributed/run.py:852] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. -W0325 00:33:17.141000 673732 torch/distributed/run.py:852] ***************************************** -logs/31484e99-50c6-404c-8bb8-635f618baafa.txt -val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=../data/tokenizers/fineweb_1024_bpe.model -train_loader:dataset:fineweb10B_sp1024 train_shards:80 -val_loader:shards pattern=../data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 -mixed_precision: 68 int5 layers, 0 int6 layers (last 0 blocks) -model_params:33317980 -XSA:[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] ws:8 gqa:8/8 -lr:embed=0.035 matrix=0.025 scalar=0.025 batch:786432 wall:600s seed:1337 -warmup_step:1/20 -warmup_step:2/20 -warmup_step:3/20 -warmup_step:4/20 -warmup_step:5/20 -warmup_step:6/20 -warmup_step:7/20 -warmup_step:8/20 -warmup_step:9/20 -warmup_step:10/20 -warmup_step:11/20 -warmup_step:12/20 -warmup_step:13/20 -warmup_step:14/20 -warmup_step:15/20 -warmup_step:16/20 -warmup_step:17/20 -warmup_step:18/20 -warmup_step:19/20 -warmup_step:20/20 -step:0/20000 val_loss:6.9285 val_bpb:4.1034 train_time:0ms step_avg:0.01ms -step:1/20000 train_loss:6.9305 train_time:148ms step_avg:147.68ms -step:2/20000 train_loss:8.6412 train_time:239ms step_avg:119.37ms -step:3/20000 train_loss:7.7278 train_time:333ms step_avg:111.04ms -step:4/20000 train_loss:7.2812 train_time:428ms step_avg:106.95ms -step:5/20000 train_loss:7.0672 train_time:524ms step_avg:104.81ms -step:6/20000 train_loss:6.9647 train_time:619ms step_avg:103.14ms -step:7/20000 train_loss:6.8519 train_time:714ms step_avg:102.01ms -step:8/20000 train_loss:6.7091 train_time:809ms step_avg:101.08ms -step:9/20000 train_loss:6.3640 train_time:903ms step_avg:100.36ms -step:10/20000 train_loss:6.0314 train_time:998ms step_avg:99.77ms -step:500/20000 train_loss:2.3594 train_time:48287ms step_avg:96.57ms -step:1000/20000 train_loss:2.2366 train_time:96650ms step_avg:96.65ms -step:1500/20000 train_loss:2.1871 train_time:145067ms step_avg:96.71ms -step:2000/20000 train_loss:2.0272 train_time:193559ms step_avg:96.78ms -step:2500/20000 train_loss:2.1332 train_time:242098ms step_avg:96.84ms -step:3000/20000 train_loss:2.1140 train_time:290638ms step_avg:96.88ms -step:3500/20000 train_loss:2.1198 train_time:339166ms step_avg:96.90ms -step:4000/20000 train_loss:1.9079 train_time:387684ms step_avg:96.92ms -step:4000/20000 val_loss:2.0000 val_bpb:1.1845 train_time:387689ms step_avg:96.92ms -late_qat:enabled step:4255 scale:0.4998 -step:4500/20000 train_loss:2.0575 train_time:436219ms step_avg:96.94ms -step:5000/20000 train_loss:2.0307 train_time:484743ms step_avg:96.95ms -swa:start step:5350 -step:5500/20000 train_loss:1.9404 train_time:533445ms step_avg:96.99ms -step:5997/20000 val_loss:1.9019 val_bpb:1.1264 train_time:582051ms step_avg:97.06ms -stopping_early: wallclock_cap train_time:582051ms step:5997/20000 -peak memory allocated: 26200 MiB reserved: 26782 MiB -ema:applying EMA weights (skipping diagnostic evals) -gptq:calibrating with training data... -gptq:calibrated 68 layers in 1.9s -Serialized model: 130432585 bytes -Code size: 96428 bytes -pruning:3.0% magnitude pruning applied -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -Serialized model int6+zstd: 15387977 bytes -Total submission size int6+zstd: 15484405 bytes - ttt: pre-compiling forward+backward kernels... - ttt: pre-compile done -final_int6_sliding_window val_loss:1.8992 val_bpb:1.1248 stride:32 eval_time:164968ms -final_int6_sliding_window_exact val_loss:1.89920781 val_bpb:1.12482157 -TTT: epochs=3 lr=0.0001 freeze_first=2 chunk=131072 opt=adamw -TTT temperature: 0.98 -PPM alpha: 0.85, Byte-weighted TTT: True - Logistic context mixer enabled: eta=0.1 - Adaptive LR enabled: max_mult=3.0 -ttt:start chunks=474 chunk_tokens=131072 windows=1938176 stride=32 lr=0.0001 epochs=3 opt=adamw freeze_first=2 -ttt:params unfrozen=5780500 frozen=27537480 - Polyak averaging enabled: decay=0.998 - ttt_train [1] seqs=64 start_train... - ttt_train [1] epoch=1/3 batches=8 ... - step done ep=1 bs=0 loss=2.3437 - ttt_train [1] epoch=2/3 batches=8 ... - step done ep=2 bs=0 loss=2.3101 - ttt_train [1] epoch=3/3 batches=8 ... - step done ep=3 bs=0 loss=2.3086 - ttt_chunk [1/474] bpb=1.202687 time=1.3s - ttt_train [2] seqs=64 start_train... - ttt_train [2] epoch=1/3 batches=8 ... - step done ep=1 bs=0 loss=2.1361 - ttt_train [2] epoch=2/3 batches=8 ... - step done ep=2 bs=0 loss=2.1334 - ttt_train [2] epoch=3/3 batches=8 ... - step done ep=3 bs=0 loss=2.1283 - ttt_chunk [2/474] bpb=1.128760 time=2.5s - ttt_train [3] seqs=64 start_train... - ttt_train [3] epoch=1/3 batches=8 ... - step done ep=1 bs=0 loss=2.0525 - ttt_train [3] epoch=2/3 batches=8 ... - step done ep=2 bs=0 loss=2.0510 - ttt_train [3] epoch=3/3 batches=8 ... - step done ep=3 bs=0 loss=2.0484 - ttt_chunk [3/474] bpb=1.080018 time=3.7s - ttt_chunk [4/474] bpb=1.076079 time=4.8s - ttt_chunk [5/474] bpb=1.066728 time=6.0s - ttt_chunk [11/474] bpb=1.031430 time=13.1s - ttt_chunk [21/474] bpb=1.018884 time=25.0s - ttt_chunk [31/474] bpb=1.015939 time=36.9s - ttt_chunk [41/474] bpb=1.022519 time=48.7s - ttt_chunk [51/474] bpb=1.028238 time=60.6s - ttt_chunk [61/474] bpb=1.025430 time=72.5s - ttt_chunk [71/474] bpb=1.026528 time=84.3s - ttt_chunk [81/474] bpb=1.026915 time=96.2s - ttt_chunk [91/474] bpb=1.028643 time=108.1s - ttt_chunk [101/474] bpb=1.025107 time=119.9s - ttt_chunk [111/474] bpb=1.024927 time=131.8s - ttt_chunk [121/474] bpb=1.027728 time=143.7s - ttt_chunk [131/474] bpb=1.027818 time=155.5s - ttt_chunk [141/474] bpb=1.026649 time=167.4s - ttt_chunk [151/474] bpb=1.024273 time=179.3s - ttt_chunk [161/474] bpb=1.024436 time=191.2s - ttt_chunk [171/474] bpb=1.022907 time=203.0s - ttt_chunk [181/474] bpb=1.023651 time=214.9s - ttt_chunk [191/474] bpb=1.022518 time=226.8s - ttt_chunk [201/474] bpb=1.021847 time=238.6s - ttt_chunk [211/474] bpb=1.020784 time=250.5s - ttt_chunk [221/474] bpb=1.021277 time=262.4s - ttt_chunk [231/474] bpb=1.020969 time=274.2s - ttt_chunk [241/474] bpb=1.020429 time=286.1s - ttt_chunk [251/474] bpb=1.022389 time=298.0s - ttt_chunk [261/474] bpb=1.024492 time=309.8s - ttt_chunk [271/474] bpb=1.024714 time=321.7s - ttt_chunk [281/474] bpb=1.026263 time=333.6s - ttt_chunk [291/474] bpb=1.027062 time=345.4s - ttt_chunk [301/474] bpb=1.029658 time=357.3s - ttt_chunk [311/474] bpb=1.031598 time=369.2s - ttt_chunk [321/474] bpb=1.032553 time=381.0s - ttt_chunk [331/474] bpb=1.034020 time=392.9s - ttt_chunk [341/474] bpb=1.035727 time=404.8s - ttt_chunk [351/474] bpb=1.036601 time=416.6s - ttt_chunk [361/474] bpb=1.039560 time=428.5s - ttt_chunk [371/474] bpb=1.041485 time=440.4s - ttt_chunk [381/474] bpb=1.044560 time=452.3s - ttt_chunk [391/474] bpb=1.047764 time=464.1s - ttt_chunk [401/474] bpb=1.050371 time=476.0s - ttt_chunk [411/474] bpb=1.052690 time=487.9s - ttt_chunk [421/474] bpb=1.055963 time=499.8s - ttt_chunk [431/474] bpb=1.056364 time=511.6s - ttt_chunk [441/474] bpb=1.057641 time=523.5s - ttt_chunk [451/474] bpb=1.058576 time=535.4s - ttt_chunk [461/474] bpb=1.060302 time=547.2s - ttt_chunk [471/474] bpb=1.061844 time=559.1s - ttt_chunk [474/474] bpb=1.062011 time=561.7s -ttt:done val_loss=1.783077 val_bpb=1.056042 elapsed=561.7s -final_int6_ttt val_loss:1.7831 val_bpb:1.0560 stride:32 eval_time:562538ms -final_int6_ttt_exact val_loss:1.78307674 val_bpb:1.05604198 diff --git a/submission-2026-03-24/log_seed42.txt b/submission-2026-03-24/log_seed42.txt deleted file mode 100644 index a185217fe..000000000 --- a/submission-2026-03-24/log_seed42.txt +++ /dev/null @@ -1,173 +0,0 @@ -W0325 00:58:00.120000 677084 torch/distributed/run.py:852] -W0325 00:58:00.120000 677084 torch/distributed/run.py:852] ***************************************** -W0325 00:58:00.120000 677084 torch/distributed/run.py:852] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. -W0325 00:58:00.120000 677084 torch/distributed/run.py:852] ***************************************** -logs/e0b62b26-4d67-4f38-8e55-7f891da41982.txt -val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=../data/tokenizers/fineweb_1024_bpe.model -train_loader:dataset:fineweb10B_sp1024 train_shards:80 -val_loader:shards pattern=../data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 -mixed_precision: 68 int5 layers, 0 int6 layers (last 0 blocks) -model_params:33317980 -XSA:[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] ws:8 gqa:8/8 -lr:embed=0.035 matrix=0.025 scalar=0.025 batch:786432 wall:600s seed:42 -warmup_step:1/20 -warmup_step:2/20 -warmup_step:3/20 -warmup_step:4/20 -warmup_step:5/20 -warmup_step:6/20 -warmup_step:7/20 -warmup_step:8/20 -warmup_step:9/20 -warmup_step:10/20 -warmup_step:11/20 -warmup_step:12/20 -warmup_step:13/20 -warmup_step:14/20 -warmup_step:15/20 -warmup_step:16/20 -warmup_step:17/20 -warmup_step:18/20 -warmup_step:19/20 -warmup_step:20/20 -step:0/20000 val_loss:6.9301 val_bpb:4.1044 train_time:0ms step_avg:0.01ms -step:1/20000 train_loss:6.9309 train_time:146ms step_avg:146.46ms -step:2/20000 train_loss:8.7072 train_time:237ms step_avg:118.44ms -step:3/20000 train_loss:7.7698 train_time:331ms step_avg:110.45ms -step:4/20000 train_loss:7.2987 train_time:427ms step_avg:106.64ms -step:5/20000 train_loss:7.0366 train_time:521ms step_avg:104.22ms -step:6/20000 train_loss:6.9492 train_time:616ms step_avg:102.74ms -step:7/20000 train_loss:6.8511 train_time:711ms step_avg:101.54ms -step:8/20000 train_loss:6.7338 train_time:805ms step_avg:100.67ms -step:9/20000 train_loss:6.3596 train_time:900ms step_avg:100.05ms -step:10/20000 train_loss:6.0340 train_time:996ms step_avg:99.58ms -step:500/20000 train_loss:2.3619 train_time:48288ms step_avg:96.58ms -step:1000/20000 train_loss:2.2442 train_time:96649ms step_avg:96.65ms -step:1500/20000 train_loss:2.1847 train_time:145067ms step_avg:96.71ms -step:2000/20000 train_loss:2.0315 train_time:193558ms step_avg:96.78ms -step:2500/20000 train_loss:2.1358 train_time:242084ms step_avg:96.83ms -step:3000/20000 train_loss:2.1167 train_time:290616ms step_avg:96.87ms -step:3500/20000 train_loss:2.1233 train_time:339152ms step_avg:96.90ms -step:4000/20000 train_loss:1.9137 train_time:387678ms step_avg:96.92ms -step:4000/20000 val_loss:2.0017 val_bpb:1.1855 train_time:387683ms step_avg:96.92ms -late_qat:enabled step:4255 scale:0.4998 -step:4500/20000 train_loss:2.0589 train_time:436214ms step_avg:96.94ms -step:5000/20000 train_loss:2.0322 train_time:484739ms step_avg:96.95ms -swa:start step:5350 -step:5500/20000 train_loss:1.9442 train_time:533403ms step_avg:96.98ms -step:5997/20000 val_loss:1.9035 val_bpb:1.1274 train_time:582007ms step_avg:97.05ms -stopping_early: wallclock_cap train_time:582007ms step:5997/20000 -peak memory allocated: 26197 MiB reserved: 26782 MiB -ema:applying EMA weights (skipping diagnostic evals) -gptq:calibrating with training data... -gptq:calibrated 68 layers in 1.9s -Serialized model: 130432585 bytes -Code size: 96428 bytes -pruning:3.0% magnitude pruning applied -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -Serialized model int6+zstd: 15312167 bytes -Total submission size int6+zstd: 15408595 bytes - ttt: pre-compiling forward+backward kernels... - ttt: pre-compile done -final_int6_sliding_window val_loss:1.9006 val_bpb:1.1257 stride:32 eval_time:164957ms -final_int6_sliding_window_exact val_loss:1.90062694 val_bpb:1.12566206 -TTT: epochs=3 lr=0.0001 freeze_first=2 chunk=131072 opt=adamw -TTT temperature: 0.98 -PPM alpha: 0.85, Byte-weighted TTT: True - Logistic context mixer enabled: eta=0.1 - Adaptive LR enabled: max_mult=3.0 -ttt:start chunks=474 chunk_tokens=131072 windows=1938176 stride=32 lr=0.0001 epochs=3 opt=adamw freeze_first=2 -ttt:params unfrozen=5780500 frozen=27537480 - Polyak averaging enabled: decay=0.998 - ttt_train [1] seqs=64 start_train... - ttt_train [1] epoch=1/3 batches=8 ... - step done ep=1 bs=0 loss=2.3462 - ttt_train [1] epoch=2/3 batches=8 ... - step done ep=2 bs=0 loss=2.3103 - ttt_train [1] epoch=3/3 batches=8 ... - step done ep=3 bs=0 loss=2.3076 - ttt_chunk [1/474] bpb=1.202959 time=1.2s - ttt_train [2] seqs=64 start_train... - ttt_train [2] epoch=1/3 batches=8 ... - step done ep=1 bs=0 loss=2.1330 - ttt_train [2] epoch=2/3 batches=8 ... - step done ep=2 bs=0 loss=2.1310 - ttt_train [2] epoch=3/3 batches=8 ... - step done ep=3 bs=0 loss=2.1268 - ttt_chunk [2/474] bpb=1.129301 time=2.4s - ttt_train [3] seqs=64 start_train... - ttt_train [3] epoch=1/3 batches=8 ... - step done ep=1 bs=0 loss=2.0619 - ttt_train [3] epoch=2/3 batches=8 ... - step done ep=2 bs=0 loss=2.0610 - ttt_train [3] epoch=3/3 batches=8 ... - step done ep=3 bs=0 loss=2.0582 - ttt_chunk [3/474] bpb=1.079049 time=3.6s - ttt_chunk [4/474] bpb=1.074949 time=4.8s - ttt_chunk [5/474] bpb=1.064662 time=6.0s - ttt_chunk [11/474] bpb=1.025191 time=13.1s - ttt_chunk [21/474] bpb=1.008678 time=25.0s - ttt_chunk [31/474] bpb=1.002098 time=36.9s - ttt_chunk [41/474] bpb=1.007125 time=48.7s - ttt_chunk [51/474] bpb=1.012485 time=60.6s - ttt_chunk [61/474] bpb=1.010685 time=72.5s - ttt_chunk [71/474] bpb=1.013076 time=84.4s - ttt_chunk [81/474] bpb=1.015652 time=96.2s - ttt_chunk [91/474] bpb=1.019922 time=108.1s - ttt_chunk [101/474] bpb=1.019995 time=120.0s - ttt_chunk [111/474] bpb=1.024232 time=131.9s - ttt_chunk [121/474] bpb=1.031908 time=143.7s - ttt_chunk [131/474] bpb=1.037128 time=155.6s - ttt_chunk [141/474] bpb=1.041120 time=167.5s - ttt_chunk [151/474] bpb=1.044067 time=179.4s - ttt_chunk [161/474] bpb=1.049148 time=191.3s - ttt_chunk [171/474] bpb=1.052038 time=203.1s - ttt_chunk [181/474] bpb=1.057080 time=215.0s - ttt_chunk [191/474] bpb=1.059619 time=226.9s - ttt_chunk [201/474] bpb=1.062378 time=238.8s - ttt_chunk [211/474] bpb=1.064340 time=250.6s - ttt_chunk [221/474] bpb=1.067700 time=262.5s - ttt_chunk [231/474] bpb=1.069761 time=274.4s - ttt_chunk [241/474] bpb=1.071136 time=286.3s - ttt_chunk [251/474] bpb=1.074548 time=298.2s - ttt_chunk [261/474] bpb=1.077509 time=310.0s - ttt_chunk [271/474] bpb=1.078241 time=321.9s - ttt_chunk [281/474] bpb=1.079767 time=333.8s - ttt_chunk [291/474] bpb=1.080218 time=345.7s - ttt_chunk [301/474] bpb=1.082217 time=357.5s - ttt_chunk [311/474] bpb=1.083294 time=369.4s - ttt_chunk [321/474] bpb=1.083302 time=381.3s - ttt_chunk [331/474] bpb=1.083667 time=393.2s - ttt_chunk [341/474] bpb=1.084241 time=405.0s - ttt_chunk [351/474] bpb=1.083985 time=416.9s - ttt_chunk [361/474] bpb=1.085885 time=428.8s - ttt_chunk [371/474] bpb=1.086654 time=440.7s - ttt_chunk [381/474] bpb=1.088639 time=452.5s - ttt_chunk [391/474] bpb=1.090740 time=464.4s - ttt_chunk [401/474] bpb=1.092278 time=476.3s - ttt_chunk [411/474] bpb=1.093605 time=488.2s - ttt_chunk [421/474] bpb=1.095927 time=500.1s - ttt_chunk [431/474] bpb=1.095382 time=511.9s - ttt_chunk [441/474] bpb=1.095777 time=523.8s - ttt_chunk [451/474] bpb=1.095835 time=535.7s - ttt_chunk [461/474] bpb=1.096740 time=547.6s - ttt_chunk [471/474] bpb=1.097514 time=559.5s - ttt_chunk [474/474] bpb=1.097488 time=562.1s -ttt:done val_loss=1.852181 val_bpb=1.096969 elapsed=562.1s -final_int6_ttt val_loss:1.8522 val_bpb:1.0970 stride:32 eval_time:562886ms -final_int6_ttt_exact val_loss:1.85218082 val_bpb:1.09696945 diff --git a/submission-2026-03-24/log_seed7.txt b/submission-2026-03-24/log_seed7.txt deleted file mode 100644 index fcaf81b2e..000000000 --- a/submission-2026-03-24/log_seed7.txt +++ /dev/null @@ -1,173 +0,0 @@ -W0325 01:22:56.941000 679893 torch/distributed/run.py:852] -W0325 01:22:56.941000 679893 torch/distributed/run.py:852] ***************************************** -W0325 01:22:56.941000 679893 torch/distributed/run.py:852] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. -W0325 01:22:56.941000 679893 torch/distributed/run.py:852] ***************************************** -logs/fb9c44fd-95a9-4cb3-a79d-3bf240580a1a.txt -val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=../data/tokenizers/fineweb_1024_bpe.model -train_loader:dataset:fineweb10B_sp1024 train_shards:80 -val_loader:shards pattern=../data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 -mixed_precision: 68 int5 layers, 0 int6 layers (last 0 blocks) -model_params:33317980 -XSA:[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] ws:8 gqa:8/8 -lr:embed=0.035 matrix=0.025 scalar=0.025 batch:786432 wall:600s seed:7 -warmup_step:1/20 -warmup_step:2/20 -warmup_step:3/20 -warmup_step:4/20 -warmup_step:5/20 -warmup_step:6/20 -warmup_step:7/20 -warmup_step:8/20 -warmup_step:9/20 -warmup_step:10/20 -warmup_step:11/20 -warmup_step:12/20 -warmup_step:13/20 -warmup_step:14/20 -warmup_step:15/20 -warmup_step:16/20 -warmup_step:17/20 -warmup_step:18/20 -warmup_step:19/20 -warmup_step:20/20 -step:0/20000 val_loss:6.9294 val_bpb:4.1040 train_time:0ms step_avg:0.01ms -step:1/20000 train_loss:6.9310 train_time:151ms step_avg:150.83ms -step:2/20000 train_loss:8.7138 train_time:240ms step_avg:120.17ms -step:3/20000 train_loss:7.7624 train_time:335ms step_avg:111.68ms -step:4/20000 train_loss:7.2817 train_time:430ms step_avg:107.46ms -step:5/20000 train_loss:7.1027 train_time:525ms step_avg:104.94ms -step:6/20000 train_loss:6.9241 train_time:619ms step_avg:103.24ms -step:7/20000 train_loss:6.8669 train_time:714ms step_avg:102.05ms -step:8/20000 train_loss:6.6918 train_time:809ms step_avg:101.14ms -step:9/20000 train_loss:6.3601 train_time:904ms step_avg:100.46ms -step:10/20000 train_loss:6.0035 train_time:999ms step_avg:99.88ms -step:500/20000 train_loss:2.3626 train_time:48426ms step_avg:96.85ms -step:1000/20000 train_loss:2.2381 train_time:96920ms step_avg:96.92ms -step:1500/20000 train_loss:2.1845 train_time:145469ms step_avg:96.98ms -step:2000/20000 train_loss:2.0268 train_time:194077ms step_avg:97.04ms -step:2500/20000 train_loss:2.1315 train_time:242724ms step_avg:97.09ms -step:3000/20000 train_loss:2.1123 train_time:291375ms step_avg:97.13ms -step:3500/20000 train_loss:2.1194 train_time:340018ms step_avg:97.15ms -step:4000/20000 train_loss:1.9084 train_time:388663ms step_avg:97.17ms -step:4000/20000 val_loss:1.9988 val_bpb:1.1838 train_time:388668ms step_avg:97.17ms -late_qat:enabled step:4240 scale:0.4998 -step:4500/20000 train_loss:2.0531 train_time:437292ms step_avg:97.18ms -step:5000/20000 train_loss:2.0297 train_time:485905ms step_avg:97.18ms -swa:start step:5300 -step:5500/20000 train_loss:1.9406 train_time:534723ms step_avg:97.22ms -step:5983/20000 val_loss:1.9008 val_bpb:1.1258 train_time:582062ms step_avg:97.29ms -stopping_early: wallclock_cap train_time:582062ms step:5983/20000 -peak memory allocated: 26197 MiB reserved: 26782 MiB -ema:applying EMA weights (skipping diagnostic evals) -gptq:calibrating with training data... -gptq:calibrated 68 layers in 2.0s -Serialized model: 130432585 bytes -Code size: 96428 bytes -pruning:3.0% magnitude pruning applied -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33161216 int5 params, 0 int6 params -Serialized model int6+zstd: 15331691 bytes -Total submission size int6+zstd: 15428119 bytes - ttt: pre-compiling forward+backward kernels... - ttt: pre-compile done -final_int6_sliding_window val_loss:1.8997 val_bpb:1.1251 stride:32 eval_time:165026ms -final_int6_sliding_window_exact val_loss:1.89974539 val_bpb:1.12513995 -TTT: epochs=3 lr=0.0001 freeze_first=2 chunk=131072 opt=adamw -TTT temperature: 0.98 -PPM alpha: 0.85, Byte-weighted TTT: True - Logistic context mixer enabled: eta=0.1 - Adaptive LR enabled: max_mult=3.0 -ttt:start chunks=474 chunk_tokens=131072 windows=1938176 stride=32 lr=0.0001 epochs=3 opt=adamw freeze_first=2 -ttt:params unfrozen=5780500 frozen=27537480 - Polyak averaging enabled: decay=0.998 - ttt_train [1] seqs=64 start_train... - ttt_train [1] epoch=1/3 batches=8 ... - step done ep=1 bs=0 loss=2.3476 - ttt_train [1] epoch=2/3 batches=8 ... - step done ep=2 bs=0 loss=2.3121 - ttt_train [1] epoch=3/3 batches=8 ... - step done ep=3 bs=0 loss=2.3099 - ttt_chunk [1/474] bpb=1.199612 time=1.2s - ttt_train [2] seqs=64 start_train... - ttt_train [2] epoch=1/3 batches=8 ... - step done ep=1 bs=0 loss=2.1304 - ttt_train [2] epoch=2/3 batches=8 ... - step done ep=2 bs=0 loss=2.1283 - ttt_train [2] epoch=3/3 batches=8 ... - step done ep=3 bs=0 loss=2.1239 - ttt_chunk [2/474] bpb=1.128925 time=2.4s - ttt_train [3] seqs=64 start_train... - ttt_train [3] epoch=1/3 batches=8 ... - step done ep=1 bs=0 loss=2.0552 - ttt_train [3] epoch=2/3 batches=8 ... - step done ep=2 bs=0 loss=2.0540 - ttt_train [3] epoch=3/3 batches=8 ... - step done ep=3 bs=0 loss=2.0517 - ttt_chunk [3/474] bpb=1.081615 time=3.6s - ttt_chunk [4/474] bpb=1.078973 time=4.8s - ttt_chunk [5/474] bpb=1.069501 time=5.9s - ttt_chunk [11/474] bpb=1.037589 time=13.0s - ttt_chunk [21/474] bpb=1.028394 time=24.9s - ttt_chunk [31/474] bpb=1.028114 time=36.7s - ttt_chunk [41/474] bpb=1.038811 time=48.6s - ttt_chunk [51/474] bpb=1.048907 time=60.4s - ttt_chunk [61/474] bpb=1.050285 time=72.2s - ttt_chunk [71/474] bpb=1.055243 time=84.1s - ttt_chunk [81/474] bpb=1.059893 time=95.9s - ttt_chunk [91/474] bpb=1.065566 time=107.7s - ttt_chunk [101/474] bpb=1.065416 time=119.6s - ttt_chunk [111/474] bpb=1.068729 time=131.4s - ttt_chunk [121/474] bpb=1.074869 time=143.3s - ttt_chunk [131/474] bpb=1.078175 time=155.1s - ttt_chunk [141/474] bpb=1.080151 time=166.9s - ttt_chunk [151/474] bpb=1.080074 time=178.8s - ttt_chunk [161/474] bpb=1.082475 time=190.6s - ttt_chunk [171/474] bpb=1.082532 time=202.5s - ttt_chunk [181/474] bpb=1.084595 time=214.3s - ttt_chunk [191/474] bpb=1.084726 time=226.1s - ttt_chunk [201/474] bpb=1.084927 time=238.0s - ttt_chunk [211/474] bpb=1.084565 time=249.8s - ttt_chunk [221/474] bpb=1.085631 time=261.6s - ttt_chunk [231/474] bpb=1.085931 time=273.5s - ttt_chunk [241/474] bpb=1.085693 time=285.3s - ttt_chunk [251/474] bpb=1.087461 time=297.1s - ttt_chunk [261/474] bpb=1.088614 time=309.0s - ttt_chunk [271/474] bpb=1.087539 time=320.8s - ttt_chunk [281/474] bpb=1.087260 time=332.7s - ttt_chunk [291/474] bpb=1.085792 time=344.5s - ttt_chunk [301/474] bpb=1.085870 time=356.3s - ttt_chunk [311/474] bpb=1.084865 time=368.2s - ttt_chunk [321/474] bpb=1.082823 time=380.0s - ttt_chunk [331/474] bpb=1.081088 time=391.9s - ttt_chunk [341/474] bpb=1.079459 time=403.7s - ttt_chunk [351/474] bpb=1.076696 time=415.5s - ttt_chunk [361/474] bpb=1.076032 time=427.4s - ttt_chunk [371/474] bpb=1.074447 time=439.2s - ttt_chunk [381/474] bpb=1.073773 time=451.0s - ttt_chunk [391/474] bpb=1.073248 time=462.9s - ttt_chunk [401/474] bpb=1.072241 time=474.7s - ttt_chunk [411/474] bpb=1.071242 time=486.6s - ttt_chunk [421/474] bpb=1.071481 time=498.4s - ttt_chunk [431/474] bpb=1.069344 time=510.2s - ttt_chunk [441/474] bpb=1.068440 time=522.1s - ttt_chunk [451/474] bpb=1.067601 time=533.9s - ttt_chunk [461/474] bpb=1.068029 time=545.7s - ttt_chunk [471/474] bpb=1.068628 time=557.6s - ttt_chunk [474/474] bpb=1.068643 time=560.2s -ttt:done val_loss=1.807296 val_bpb=1.070386 elapsed=560.2s -final_int6_ttt val_loss:1.8073 val_bpb:1.0704 stride:32 eval_time:561081ms -final_int6_ttt_exact val_loss:1.80729571 val_bpb:1.07038587 diff --git a/submission-2026-03-24/requirements.txt b/submission-2026-03-24/requirements.txt deleted file mode 100644 index a124ed4f5..000000000 --- a/submission-2026-03-24/requirements.txt +++ /dev/null @@ -1,5 +0,0 @@ -torch>=2.10.0 -numpy -sentencepiece -zstandard -flash-attn-3 diff --git a/submission-2026-03-24/submission.json b/submission-2026-03-24/submission.json deleted file mode 100644 index 2835282df..000000000 --- a/submission-2026-03-24/submission.json +++ /dev/null @@ -1,14 +0,0 @@ -{ - "track": "10min_16mb", - "date": "2026-03-24", - "name": "5-expert Hedge Mixer + TTT", - "author": "notapplica", - "seed_results": { - "1337": {"val_loss": 1.78307674, "val_bpb": 1.05604198, "artifact_bytes": 15484405}, - "42": {"val_loss": 1.85218082, "val_bpb": 1.09696945, "artifact_bytes": 15408595}, - "7": {"val_loss": 1.80729571, "val_bpb": 1.07038587, "artifact_bytes": 15428119} - }, - "mean_val_loss": 1.81418442, - "mean_val_bpb": 1.07446577, - "code_bytes": 96416 -} diff --git a/submission-2026-03-24/train_gpt.py b/submission-2026-03-24/train_gpt.py deleted file mode 100644 index 7460ef9be..000000000 --- a/submission-2026-03-24/train_gpt.py +++ /dev/null @@ -1,1966 +0,0 @@ -"""V25: LeakyReLU^2 + TempCal + Mixed int5/int6 + 33.6M model.""" -from __future__ import annotations -import copy -import glob -import io -import math -import os -import random -import subprocess -import sys -import time -import uuid -import zlib -from pathlib import Path -try: - import zstandard - _COMPRESSOR = "zstd" -except ImportError: - _COMPRESSOR = "zlib" -import numpy as np -import sentencepiece as spm -import torch -import torch.distributed as dist -import torch.nn.functional as F -from torch import Tensor, nn -from torch.nn.parallel import DistributedDataParallel as DDP -try: - from flash_attn_interface import flash_attn_func as flash_attn_3_func - _HAS_FA3 = True -except ImportError: - try: - from flash_attn import flash_attn_func as flash_attn_3_func - _HAS_FA3 = True - except ImportError: - _HAS_FA3 = False - flash_attn_3_func = None - -class LogisticContextMixer: - """GPU-vectorized logistic context mixing (inspired by PAQ compression). - - Maintains GPU-resident n-gram count tables and learns online mixing weights - using the Hedge/multiplicative-weights algorithm. - - Experts: - 0: Neural model (logits passed in) - 1: Unigram frequencies from scored tokens - 2: Bigram frequencies (prev_token → next_token) - 3: FastPPM (orders 0-4, CPU-side) - 4: ExactMatchCache (high-order exact matches, CPU-side) - """ - - def __init__(self, vocab_size: int = 1024, device: str = 'cuda', eta: float = 0.1): - self.V = vocab_size - self.device = device - self.eta = eta # Hedge learning rate - self.K = 5 # number of experts - - # Expert weights (log-domain for numerical stability) - self.log_weights = torch.zeros(self.K, device=device) - # Bias toward neural model initially - self.log_weights[0] = 2.0 - - # N-gram count tables (GPU-resident) - self.uni_counts = torch.zeros(vocab_size, device=device) - self.bi_counts = torch.zeros(vocab_size, vocab_size, device=device) - self.total_tokens = 0 - - # GPU Trigram: hashed table [HASH_SIZE, V] to keep memory reasonable - self.TRI_HASH = 65536 # 64K hash buckets for (prev2, prev1) pairs - self.tri_counts = torch.zeros(self.TRI_HASH, vocab_size, device=device) - self.tri_row_totals = torch.zeros(self.TRI_HASH, device=device) - - def update(self, tokens): - """Update all expert statistics with newly scored tokens.""" - if hasattr(tokens, 'cpu'): - t = tokens.to(self.device).long() - else: - t = torch.tensor(tokens, device=self.device, dtype=torch.long) - - n = t.numel() - if n == 0: - return - self.total_tokens += n - - # Unigram: in-place scatter_add - ones = torch.ones(n, device=self.device) - self.uni_counts.scatter_add_(0, t, ones) - - # Bigram: in-place scatter_add on flattened view (no temporary 1M tensor) - if n >= 2: - ctx = t[:-1] - nxt = t[1:] - bi_idx = ctx * self.V + nxt - ones_bi = torch.ones(n - 1, device=self.device) - self.bi_counts.reshape(-1).scatter_add_(0, bi_idx, ones_bi) - - # Trigram: in-place scatter_add on flattened view (no temporary 67M tensor) - if n >= 3: - prev2 = t[:-2] - prev1 = t[1:-1] - nxt3 = t[2:] - tri_ctx = ((prev2 * 36313) ^ (prev1 * 27191)) % self.TRI_HASH - tri_idx = tri_ctx * self.V + nxt3 - ones_tri = torch.ones(n - 2, device=self.device) - self.tri_counts.reshape(-1).scatter_add_(0, tri_idx, ones_tri) - self.tri_row_totals.scatter_add_(0, tri_ctx, ones_tri) - - def get_expert_log_probs(self, neural_logits, x_batch, y_batch, wlens): - """Get log-probability of targets from each expert. All GPU-vectorized. - - Args: - neural_logits: [bsz, seq_len, V] neural model logits - x_batch: [bsz, seq_len] input tokens (context) - y_batch: [bsz, seq_len] target tokens - wlens: list of actual lengths per sequence - - Returns: - expert_nll: [bsz, seq_len, K] NLL from each expert - """ - bsz, slen, V = neural_logits.shape - uniform_nll = math.log(self.V) - has_data = self.total_tokens > 0 # Python int — no GPU-CPU sync - - # Expert 0: Neural model — compute log_softmax once, reuse for entropy - neural_lp = F.log_softmax(neural_logits, dim=-1) - neural_nll = -neural_lp.gather(2, y_batch.unsqueeze(2)).squeeze(2) # [bsz, slen] - - # Expert 1: Unigram - if has_data: - uni_probs = (self.uni_counts + 0.1) / (self.total_tokens + 0.1 * self.V) - uni_nll = -uni_probs.log()[y_batch] # [bsz, slen] - else: - uni_nll = torch.full((bsz, slen), uniform_nll, device=self.device) - - # Expert 2: Bigram P(next | prev) - if has_data: - bi_total = self.bi_counts.sum(dim=1, keepdim=True) # [V, 1] - bi_probs = (self.bi_counts + 0.1) / (bi_total + 0.1 * self.V) # [V, V] - prev_flat = x_batch.reshape(-1) - next_flat = y_batch.reshape(-1) - bi_nll = -bi_probs.log()[prev_flat, next_flat].reshape(bsz, slen) - else: - bi_nll = torch.full((bsz, slen), uniform_nll, device=self.device) - - # Expert 3: GPU Trigram P(next | hash(prev2, prev1)) — vectorized - if has_data and slen >= 2: - prev2 = torch.zeros_like(x_batch) - prev2[:, 1:] = x_batch[:, :-1] - ctx_hash = ((prev2 * 36313) ^ (x_batch * 27191)) % self.TRI_HASH - ctx_flat = ctx_hash.reshape(-1).long() - next_flat = y_batch.reshape(-1).long() - tri_count = self.tri_counts[ctx_flat, next_flat] - tri_total = self.tri_row_totals[ctx_flat].clamp(min=1) - tri_prob = (tri_count + 0.01) / (tri_total + 0.01 * self.V) - tri_nll = -tri_prob.log().reshape(bsz, slen) - else: - tri_nll = torch.full((bsz, slen), uniform_nll, device=self.device) - - # Expert 4: Neural entropy — reuse neural_lp (no redundant softmax) - entropy_nll = -(neural_lp.exp() * neural_lp).sum(-1) # [bsz, slen] - - # Stack: [bsz, slen, K] - return torch.stack([neural_nll, uni_nll, bi_nll, tri_nll, entropy_nll], dim=-1) - - def mix_and_score(self, neural_logits, x_batch, y_batch, wlens): - """Compute mixed NLL using current expert weights. - - Returns (mixed_nll [bsz, slen], expert_nll [bsz, slen, K] or None). - Caller should pass expert_nll to update_weights() to avoid recomputation. - """ - if self.total_tokens < 10000: - # Not enough data for n-grams — just use neural - nll = F.cross_entropy( - neural_logits.reshape(-1, neural_logits.size(-1)), - y_batch.reshape(-1), reduction="none" - ).reshape(neural_logits.shape[0], neural_logits.shape[1]) - return nll, None - - expert_nll = self.get_expert_log_probs(neural_logits, x_batch, y_batch, wlens) # [bsz, slen, K] - - # Log-domain mixing: log(sum_k w_k * p_k) = logsumexp(log_w_k + log_p_k) - log_w = self.log_weights - self.log_weights.logsumexp(0) # normalize - mixed_lp = (-expert_nll + log_w.unsqueeze(0).unsqueeze(0)).logsumexp(dim=-1) # [bsz, slen] - - return -mixed_lp, expert_nll # mixed NLL + cached expert NLL - - def update_weights(self, expert_nll, wlens): - """Update expert weights using Hedge algorithm on pre-computed expert NLLs.""" - if expert_nll is None: - return - - with torch.no_grad(): - # Vectorized mask: compare position index against window lengths - bsz, slen = expert_nll.shape[0], expert_nll.shape[1] - wlens_t = torch.tensor(wlens, device=self.device, dtype=torch.long) - mask = torch.arange(slen, device=self.device).unsqueeze(0) < wlens_t.unsqueeze(1) # [bsz, slen] bool - - # Masked mean NLL per expert - masked_nll = expert_nll * mask.unsqueeze(-1).float() - expert_mean_loss = masked_nll.sum(dim=(0, 1)) / mask.sum().clamp(min=1) # [K] - - # Hedge update: log_w -= eta * loss - self.log_weights -= self.eta * expert_mean_loss - - -class Hyperparameters: - data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") - train_files = os.path.join(data_path, "fineweb_train_*.bin") - val_files = os.path.join(data_path, "fineweb_val_*.bin") - tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") - run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) - seed = int(os.environ.get("SEED", 1337)) - val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) - val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) - train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) - iterations = int(os.environ.get("ITERATIONS", 20000)) - warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) - warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) - train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) - train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) - eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) - max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) - qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) - vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) - num_layers = int(os.environ.get("NUM_LAYERS", 11)) - num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 8)) - model_dim = int(os.environ.get("MODEL_DIM", 512)) - num_heads = int(os.environ.get("NUM_HEADS", 8)) - mlp_mult = float(os.environ.get("MLP_MULT", 3.5)) - tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) - rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) - logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) - embed_lr = float(os.environ.get("EMBED_LR", 0.6)) - head_lr = float(os.environ.get("HEAD_LR", 0.008)) - tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) - tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) - matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) - scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) - muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) - muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) - muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) - muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) - beta1 = float(os.environ.get("BETA1", 0.9)) - beta2 = float(os.environ.get("BETA2", 0.95)) - adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) - grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) - eval_stride = int(os.environ.get("EVAL_STRIDE", 32)) - int6_last_n = int(os.environ.get("INT6_LAST_N", 0)) # all int5 (saves ~300KB vs int6 for last 2 blocks) - ttt_temperature = float(os.environ.get("TTT_TEMPERATURE", 0.98)) # post-TTT temperature calibration - muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) - swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) - swa_every = int(os.environ.get("SWA_EVERY", 50)) - - muon_wd = float(os.environ.get("MUON_WD", 0.04)) - adam_wd = float(os.environ.get("ADAM_WD", 0.04)) - qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) - bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 6144)) - bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) - xsa_last_n = int(os.environ.get("XSA_LAST_N", 11)) - rope_dims = int(os.environ.get("ROPE_DIMS", 16)) - ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) - dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) - late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.5)) - ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) - ve_dim = int(os.environ.get("VE_DIM", 128)) - ve_layers = os.environ.get("VE_LAYERS", "9,10") - prune_pct = float(os.environ.get("PRUNE_PCT", 0.03)) - -def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: - a, b, c = (3.4445, -4.7750, 2.0315) - X = G.bfloat16() - X /= X.norm() + eps - transposed = G.size(0) > G.size(1) - if transposed: - X = X.T - for _ in range(steps): - A = X @ X.T - B = b * A + c * A @ A - X = a * X + B @ X - return X.T if transposed else X - -class Muon(torch.optim.Optimizer): - def __init__(self, params, lr: float, momentum: float, backend_steps: int, - nesterov: bool = True, weight_decay: float = 0.0): - super().__init__( - params, - dict(lr=lr, momentum=momentum, backend_steps=backend_steps, - nesterov=nesterov, weight_decay=weight_decay), - ) - @torch.no_grad() - def step(self, closure=None): - loss = None - if closure is not None: - with torch.enable_grad(): - loss = closure() - distributed = dist.is_available() and dist.is_initialized() - world_size = dist.get_world_size() if distributed else 1 - rank = dist.get_rank() if distributed else 0 - for group in self.param_groups: - params = group["params"] - if not params: - continue - lr = group["lr"] - momentum = group["momentum"] - backend_steps = group["backend_steps"] - nesterov = group["nesterov"] - total_params = sum(int(p.numel()) for p in params) - updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) - curr = 0 - for i, p in enumerate(params): - if i % world_size == rank and p.grad is not None: - g = p.grad - state = self.state[p] - if "momentum_buffer" not in state: - state["momentum_buffer"] = torch.zeros_like(g) - buf = state["momentum_buffer"] - buf.mul_(momentum).add_(g) - if nesterov: - g = g.add(buf, alpha=momentum) - g = zeropower_via_newtonschulz5(g, steps=backend_steps) - g *= max(1, g.size(0) / g.size(1)) ** 0.5 - updates_flat[curr : curr + p.numel()] = g.reshape(-1) - curr += p.numel() - if distributed: - dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) - wd = group.get("weight_decay", 0.0) - curr = 0 - for p in params: - if wd > 0.0: - p.data.mul_(1.0 - lr * wd) - g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) - p.add_(g, alpha=-lr) - curr += p.numel() - return loss - -def build_sentencepiece_luts( - sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device -) -> tuple[Tensor, Tensor, Tensor]: - sp_vocab_size = int(sp.vocab_size()) - table_size = max(sp_vocab_size, vocab_size) - base_bytes_np = np.zeros((table_size,), dtype=np.int16) - has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) - is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) - for token_id in range(sp_vocab_size): - if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): - continue - is_boundary_token_np[token_id] = False - if sp.is_byte(token_id): - base_bytes_np[token_id] = 1 - continue - piece = sp.id_to_piece(token_id) - if piece.startswith("▁"): - has_leading_space_np[token_id] = True - piece = piece[1:] - base_bytes_np[token_id] = len(piece.encode("utf-8")) - return ( - torch.tensor(base_bytes_np, dtype=torch.int16, device=device), - torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), - torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), - ) - -def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: - files = [Path(p) for p in sorted(glob.glob(pattern))] - if not files: - raise FileNotFoundError(f"No files found for pattern: {pattern}") - tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() - usable = ((tokens.numel() - 1) // seq_len) * seq_len - if usable <= 0: - raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") - return tokens[: usable + 1] - -def eval_val(args: Hyperparameters, model: nn.Module, rank: int, world_size: int, - device: torch.device, grad_accum_steps: int, val_tokens: Tensor, - base_bytes_lut: Tensor, has_leading_space_lut: Tensor, - is_boundary_token_lut: Tensor, eval_seq_len: int | None = None) -> tuple[float, float]: - seq_len = eval_seq_len or args.train_seq_len - local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) - if local_batch_tokens < seq_len: - raise ValueError( - "VAL_BATCH_SIZE must provide at least one sequence per rank; " - f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " - f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" - ) - local_batch_seqs = local_batch_tokens // seq_len - total_seqs = (val_tokens.numel() - 1) // seq_len - seq_start = (total_seqs * rank) // world_size - seq_end = (total_seqs * (rank + 1)) // world_size - val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) - val_token_count = torch.zeros((), device=device, dtype=torch.float64) - val_byte_count = torch.zeros((), device=device, dtype=torch.float64) - model.eval() - with torch.inference_mode(): - for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): - batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) - raw_start = batch_seq_start * seq_len - raw_end = batch_seq_end * seq_len + 1 - local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) - x = local[:-1].reshape(-1, seq_len) - y = local[1:].reshape(-1, seq_len) - with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): - batch_loss = model(x, y).detach() - batch_token_count = float(y.numel()) - val_loss_sum += batch_loss.to(torch.float64) * batch_token_count - val_token_count += batch_token_count - prev_ids = x.reshape(-1) - tgt_ids = y.reshape(-1) - token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) - token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) - val_byte_count += token_bytes.to(torch.float64).sum() - if dist.is_available() and dist.is_initialized(): - dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) - dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) - dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) - val_loss = val_loss_sum / val_token_count - bits_per_token = val_loss.item() / math.log(2.0) - tokens_per_byte = val_token_count.item() / val_byte_count.item() - model.train() - return float(val_loss.item()), float(bits_per_token * tokens_per_byte) -CONTROL_TENSOR_NAME_PATTERNS = tuple( - pattern - for pattern in os.environ.get( - "CONTROL_TENSOR_NAME_PATTERNS", - "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale", - ).split(",") - if pattern -) -INT8_PER_ROW_SCALE_DTYPE = torch.float16 -INT8_CLIP_Q = 0.9999984 - -def tensor_nbytes(t: Tensor) -> int: - return int(t.numel()) * int(t.element_size()) - -def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: - t32 = t.float() - if t32.ndim == 2: - clip_abs = ( - torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) - if t32.numel() - else torch.empty((t32.shape[0],), dtype=torch.float32) - ) - clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) - scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) - q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() - return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() - clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 - scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) - q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() - return q, scale - -def load_data_shard(file: Path) -> Tensor: - header_bytes = 256 * np.dtype(" None: - self.file_idx = (self.file_idx + 1) % len(self.files) - self.tokens = load_data_shard(self.files[self.file_idx]) - self.pos = 0 - - def take(self, n: int) -> Tensor: - chunks: list[Tensor] = [] - remaining = n - while remaining > 0: - avail = self.tokens.numel() - self.pos - if avail <= 0: - self._advance_file() - continue - k = min(remaining, avail) - chunks.append(self.tokens[self.pos : self.pos + k]) - self.pos += k - remaining -= k - return chunks[0] if len(chunks) == 1 else torch.cat(chunks) - -class DistributedTokenLoader: - def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): - self.rank = rank - self.world_size = world_size - self.device = device - self.stream = TokenStream(pattern) - - def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: - local_tokens = global_tokens // (self.world_size * grad_accum_steps) - per_rank_span = local_tokens + 1 - chunk = self.stream.take(per_rank_span * self.world_size) - start = self.rank * per_rank_span - local = chunk[start : start + per_rank_span].to(dtype=torch.int64) - x = local[:-1].reshape(-1, seq_len) - y = local[1:].reshape(-1, seq_len) - return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) - -class RMSNorm(nn.Module): - def __init__(self, eps: float | None = None): - super().__init__() - self.eps = eps - - def forward(self, x: Tensor) -> Tensor: - return F.rms_norm(x, (x.size(-1),), eps=self.eps) - -class CastedLinear(nn.Linear): - _qat_enabled: bool = False - _soft_round_alpha: float = 1.0 # temperature for soft-round (annealed during training) - _use_soft_round: bool = False # enable soft-round QAT instead of STE - - def __init__(self, *args, **kwargs): - super().__init__(*args, **kwargs) - self._clip_range = 15 # default int5, set to 31 for int6 layers - - @staticmethod - def soft_round(y: Tensor, alpha: float) -> Tensor: - """Differentiable approximation to round() from Agustsson & Theis (NeurIPS 2020). - s_alpha(y) = floor(y) + 0.5 * tanh(alpha * r) / tanh(alpha/2) + 0.5 - where r = y - floor(y) - 0.5 (centered fractional part) - """ - fl = torch.floor(y) - r = y - fl - 0.5 - return fl + 0.5 * torch.tanh(alpha * r) / (math.tanh(alpha / 2) + 1e-10) + 0.5 - - def forward(self, x: Tensor) -> Tensor: - w = self.weight.to(x.dtype) - cr = self._clip_range - if CastedLinear._qat_enabled and self.training and w.ndim == 2: - if CastedLinear._use_soft_round: - # Soft-Round QAT: differentiable rounding with temperature annealing - w32 = self.weight.float() - row_clip = torch.quantile(w32.abs(), 0.9995, dim=1) - scale = (row_clip / float(cr)).clamp_min(1.0 / float(cr)) - w_scaled = w32 / scale[:, None] - w_rounded = CastedLinear.soft_round(w_scaled, CastedLinear._soft_round_alpha) - w_q = (torch.clamp(w_rounded, -(cr+1), cr) * scale[:, None]).to(x.dtype) - w = w_q # fully differentiable path - else: - # Original STE QAT - with torch.no_grad(): - w32 = self.weight.float() - row_clip = torch.quantile(w32.abs(), 0.9995, dim=1) - scale = (row_clip / float(cr)).clamp_min(1.0 / float(cr)) - w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -(cr+1), cr) * scale[:, None]).to(x.dtype) - w = w + (w_q - w).detach() - bias = self.bias.to(x.dtype) if self.bias is not None else None - return F.linear(x, w, bias) - -def restore_low_dim_params_to_fp32(module: nn.Module) -> None: - with torch.no_grad(): - for name, param in module.named_parameters(): - if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: - param.data = param.data.float() - -class Rotary(nn.Module): - def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): - super().__init__() - self.dim = dim - self.base = base - self.train_seq_len = train_seq_len - self.rope_dims = rope_dims if rope_dims > 0 else dim - inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) - self.register_buffer("inv_freq", inv_freq, persistent=False) - self._seq_len_cached = 0 - self._cos_cached: Tensor | None = None - self._sin_cached: Tensor | None = None - - def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: - if ( - self._cos_cached is None - or self._sin_cached is None - or self._seq_len_cached != seq_len - or self._cos_cached.device != device - ): - rd = self.rope_dims - if seq_len > self.train_seq_len: - scale = seq_len / self.train_seq_len - new_base = self.base * (scale ** (rd / (rd - 2))) - inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) - else: - inv_freq = self.inv_freq.to(device) - t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) - freqs = torch.outer(t, inv_freq) - self._cos_cached = freqs.cos()[None, :, None, :] - self._sin_cached = freqs.sin()[None, :, None, :] - self._seq_len_cached = seq_len - return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) - -def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: - if rope_dims > 0 and rope_dims < x.size(-1): - x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] - half = rope_dims // 2 - x1, x2 = x_rope[..., :half], x_rope[..., half:] - x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) - return torch.cat((x_rope, x_pass), dim=-1) - half = x.size(-1) // 2 - x1, x2 = x[..., :half], x[..., half:] - return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) - -class CausalSelfAttention(nn.Module): - def __init__(self, dim: int, num_heads: int, num_kv_heads: int, - rope_base: float, qk_gain_init: float): - super().__init__() - if dim % num_heads != 0: - raise ValueError("model_dim must be divisible by num_heads") - if num_heads % num_kv_heads != 0: - raise ValueError("num_heads must be divisible by num_kv_heads") - self.num_heads = num_heads - self.num_kv_heads = num_kv_heads - self.head_dim = dim // num_heads - if self.head_dim % 2 != 0: - raise ValueError("head_dim must be even for RoPE") - kv_dim = self.num_kv_heads * self.head_dim - self.c_q = CastedLinear(dim, dim, bias=False) - self.c_k = CastedLinear(dim, kv_dim, bias=False) - self.c_v = CastedLinear(dim, kv_dim, bias=False) - self.proj = CastedLinear(dim, dim, bias=False) - self.proj._zero_init = True - self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) - self.rope_dims = 0 - self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) - self.use_xsa = False - - def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: - B, T, H, D = y.shape - Hkv = v.size(-2) - y_g = y.reshape(B, T, Hkv, H // Hkv, D) - vn = F.normalize(v, dim=-1).unsqueeze(-2) - proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn - return (y_g - proj).reshape(B, T, H, D) - - def forward(self, x: Tensor, v_embed: Tensor | None = None) -> Tensor: - bsz, seqlen, dim = x.shape - q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim) - k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) - v = self.c_v(x) - if v_embed is not None: - v = v + v_embed - v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) - q = F.rms_norm(q, (q.size(-1),)) - k = F.rms_norm(k, (k.size(-1),)) - cos, sin = self.rotary(seqlen, x.device, q.dtype) - q = apply_rotary_emb(q, cos, sin, self.rope_dims) - k = apply_rotary_emb(k, cos, sin, self.rope_dims) - q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] - if _HAS_FA3: - y = flash_attn_3_func(q, k, v, causal=True).contiguous() - else: - y = F.scaled_dot_product_attention( - q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), - attn_mask=None, is_causal=True, - enable_gqa=(self.num_kv_heads != self.num_heads), - ).transpose(1, 2) - if self.use_xsa: - y = self._xsa_efficient(y, v) - y = y.reshape(bsz, seqlen, dim) - return self.proj(y) - -class SmearGate(nn.Module): - def __init__(self, dim: int): - super().__init__() - self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) - - def forward(self, x: Tensor) -> Tensor: - g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] - x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) - return (1 - g) * x + g * x_prev - -class BigramHashEmbedding(nn.Module): - def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): - super().__init__() - self.bigram_vocab_size = bigram_vocab_size - self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) - nn.init.zeros_(self.embed.weight) - self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None - if self.proj is not None: - nn.init.zeros_(self.proj.weight) - self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) - - def bigram_hash(self, tokens: Tensor) -> Tensor: - t = tokens.to(torch.int32) - mod = self.bigram_vocab_size - 1 - out = torch.empty_like(t) - out[..., 0] = mod - out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod - return out.long() - - def forward(self, token_ids: Tensor) -> Tensor: - h = self.embed(self.bigram_hash(token_ids)) - if self.proj is not None: - h = self.proj(h) - return h * self.scale.to(dtype=h.dtype) - -class ValueEmbedding(nn.Module): - def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): - super().__init__() - self.embed = nn.Embedding(vocab_size, ve_dim) - nn.init.normal_(self.embed.weight, std=0.01) - self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None - if self.proj is not None: - nn.init.zeros_(self.proj.weight) - self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) - - def forward(self, token_ids: Tensor) -> Tensor: - h = self.embed(token_ids) - if self.proj is not None: - h = self.proj(h) - return h * self.scale.to(dtype=h.dtype) - -class MLP(nn.Module): - def __init__(self, dim: int, mlp_mult: int): - super().__init__() - hidden = int(mlp_mult * dim) - self.fc = CastedLinear(dim, hidden, bias=False) - self.proj = CastedLinear(hidden, dim, bias=False) - self.proj._zero_init = True - - def forward(self, x: Tensor) -> Tensor: - return self.proj(F.leaky_relu(self.fc(x), negative_slope=0.5).square()) - -class Block(nn.Module): - def __init__(self, dim: int, num_heads: int, num_kv_heads: int, mlp_mult: int, - rope_base: float, qk_gain_init: float, layer_idx: int = 0, - ln_scale: bool = False, dtg: bool = False): - super().__init__() - self.attn_norm = RMSNorm() - self.mlp_norm = RMSNorm() - self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) - self.mlp = MLP(dim, mlp_mult) - self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) - self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) - self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) - self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 - if dtg: - self.dtg_gate = nn.Linear(dim, 1, bias=True) - nn.init.zeros_(self.dtg_gate.weight) - nn.init.constant_(self.dtg_gate.bias, 2.0) - else: - self.dtg_gate = None - - def forward(self, x: Tensor, x0: Tensor, v_embed: Tensor | None = None) -> Tensor: - mix = self.resid_mix.to(dtype=x.dtype) - x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 - attn_out = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, v_embed=v_embed) - x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out - x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor) - if self.dtg_gate is not None: - gate = torch.sigmoid(self.dtg_gate(x_in.detach())) - x_out = x_in + gate * (x_out - x_in) - return x_out - -class GPT(nn.Module): - def __init__(self, vocab_size: int, num_layers: int, model_dim: int, num_heads: int, - num_kv_heads: int, mlp_mult: int, tie_embeddings: bool, tied_embed_init_std: float, - logit_softcap: float, rope_base: float, qk_gain_init: float, - bigram_vocab_size: int = 0, bigram_dim: int = 128, xsa_last_n: int = 0, - rope_dims: int = 0, ln_scale: bool = False, dtg: bool = False, - ve_enabled: bool = False, ve_dim: int = 128, ve_layers: str = "9,10"): - super().__init__() - self._ve_target_dim = num_kv_heads * (model_dim // num_heads) - if logit_softcap <= 0.0: - raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") - self.tie_embeddings = tie_embeddings - self.tied_embed_init_std = tied_embed_init_std - self.logit_softcap = logit_softcap - self.tok_emb = nn.Embedding(vocab_size, model_dim) - self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None - self.smear = SmearGate(model_dim) - self.num_encoder_layers = num_layers // 2 - self.num_decoder_layers = num_layers - self.num_encoder_layers - self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) - self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) - self.blocks = nn.ModuleList([ - Block(model_dim, num_heads, num_kv_heads, mlp_mult, rope_base, - qk_gain_init, layer_idx=i, ln_scale=ln_scale, dtg=dtg) - for i in range(num_layers) - ]) - if rope_dims > 0: - head_dim = model_dim // num_heads - for block in self.blocks: - block.attn.rope_dims = rope_dims - block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) - self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] - kv_dim = self._ve_target_dim - if self.ve_layer_indices: - self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim) - self.ve_layer_scales = nn.ParameterList( - [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] - ) - else: - self.ve_shared = None - self.ve_layer_scales = nn.ParameterList() - self.value_embeds = nn.ModuleList() - self.final_norm = RMSNorm() - self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) - if self.lm_head is not None: - self.lm_head._zero_init = True - if xsa_last_n > 0: - for i in range(max(0, num_layers - xsa_last_n), num_layers): - self.blocks[i].attn.use_xsa = True - self._init_weights() - - def _init_weights(self) -> None: - if self.tie_embeddings: - nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) - num_layers = len(self.blocks) - for name, module in self.named_modules(): - if isinstance(module, nn.Linear): - if getattr(module, "_zero_init", False): - nn.init.zeros_(module.weight) - elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: - nn.init.orthogonal_(module.weight, gain=1.0) - if ".proj." in name or name.endswith(".proj"): - with torch.no_grad(): - module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) - - def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: - if self.ve_shared is None or layer_idx not in self.ve_layer_indices: - return None - if ve_cache is not None and 've' not in ve_cache: - ve_cache['ve'] = self.ve_shared(input_ids) - ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) - ve_idx = self.ve_layer_indices.index(layer_idx) - return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) - - def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: - x = self.tok_emb(input_ids) - if self.bigram is not None: - x = x + self.bigram(input_ids) - x = F.rms_norm(x, (x.size(-1),)) - x = self.smear(x) - x0 = x - skips: list[Tensor] = [] - ve_cache: dict = {} - for i in range(self.num_encoder_layers): - ve = self._get_ve(i, input_ids, ve_cache) - x = self.blocks[i](x, x0, v_embed=ve) - skips.append(x) - for i in range(self.num_decoder_layers): - bi = self.num_encoder_layers + i - if skips: - x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() - ve = self._get_ve(bi, input_ids, ve_cache) - x = self.blocks[bi](x, x0, v_embed=ve) - x = self.final_norm(x) - x_flat = x.reshape(-1, x.size(-1)) - targets = target_ids.reshape(-1) - if self.tie_embeddings: - logits_proj = F.linear(x_flat, self.tok_emb.weight) - else: - if self.lm_head is None: - raise RuntimeError("lm_head is required when tie_embeddings=False") - logits_proj = self.lm_head(x_flat) - logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) - return F.cross_entropy(logits.float(), targets, reduction="mean") - - def forward_logits(self, input_ids: Tensor) -> Tensor: - x = self.tok_emb(input_ids) - if self.bigram is not None: - x = x + self.bigram(input_ids) - x = F.rms_norm(x, (x.size(-1),)) - x = self.smear(x) - x0 = x - skips: list[Tensor] = [] - ve_cache: dict = {} - for i in range(self.num_encoder_layers): - ve = self._get_ve(i, input_ids, ve_cache) - x = self.blocks[i](x, x0, v_embed=ve) - skips.append(x) - for i in range(self.num_decoder_layers): - bi = self.num_encoder_layers + i - if skips: - x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() - ve = self._get_ve(bi, input_ids, ve_cache) - x = self.blocks[bi](x, x0, v_embed=ve) - x = self.final_norm(x) - if self.tie_embeddings: - logits_proj = F.linear(x, self.tok_emb.weight) - else: - logits_proj = self.lm_head(x) - return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) - -def eval_val_sliding(args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, - device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, - has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, - stride: int, batch_seqs: int = 32, eval_seq_len: int | None = None) -> tuple[float, float]: - seq_len = eval_seq_len or args.train_seq_len - total_tokens = val_tokens.numel() - 1 - - window_starts = [ws for ws in range(0, total_tokens, stride) - if min(ws + seq_len, total_tokens) - ws >= 1] - total_windows = len(window_starts) - my_s = (total_windows * rank) // world_size - my_e = (total_windows * (rank + 1)) // world_size - my_windows = window_starts[my_s:my_e] - loss_sum = torch.zeros((), device=device, dtype=torch.float64) - token_count = torch.zeros((), device=device, dtype=torch.float64) - byte_count = torch.zeros((), device=device, dtype=torch.float64) - base_model.eval() - compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) - - # Pre-compile: dummy forward+backward with TTT shapes to warm the compile cache - if rank == 0: - print(" ttt: pre-compiling forward+backward kernels...", flush=True) - _dummy_x = torch.zeros(1, seq_len, dtype=torch.int64, device=device) - _dummy_y = torch.zeros(1, seq_len, dtype=torch.int64, device=device) - with torch.autocast(device_type="cuda", dtype=torch.bfloat16): - _dummy_logits = base_model.forward_logits(_dummy_x) - _dummy_loss = F.cross_entropy(_dummy_logits.reshape(-1, _dummy_logits.size(-1)), _dummy_y.reshape(-1)) - _dummy_loss.backward() - base_model.zero_grad(set_to_none=True) - if rank == 0: - print(" ttt: pre-compile done", flush=True) - with torch.inference_mode(): - for bi in range(0, len(my_windows), batch_seqs): - batch_ws = my_windows[bi:bi + batch_seqs] - bsz = len(batch_ws) - x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) - y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) - wlens: list[int] = [] - for i, ws in enumerate(batch_ws): - end = min(ws + seq_len, total_tokens) - wlen = end - ws - wlens.append(wlen) - chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) - x_batch[i, :wlen] = chunk[:-1] - y_batch[i, :wlen] = chunk[1:] - with torch.autocast(device_type="cuda", dtype=torch.bfloat16): - logits = compiled_logits(x_batch) - nll = F.cross_entropy( - logits.reshape(-1, logits.size(-1)).float(), - y_batch.reshape(-1), - reduction="none", - ).reshape(bsz, seq_len) - for i, ws in enumerate(batch_ws): - wlen = wlens[i] - s = 0 if ws == 0 else max(wlen - stride, 0) - scored_nll = nll[i, s:wlen].to(torch.float64) - loss_sum += scored_nll.sum() - token_count += float(wlen - s) - tgt = y_batch[i, s:wlen] - prev = x_batch[i, s:wlen] - tb = base_bytes_lut[tgt].to(torch.float64) - tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) - byte_count += tb.sum() - if dist.is_available() and dist.is_initialized(): - dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) - dist.all_reduce(token_count, op=dist.ReduceOp.SUM) - dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) - val_loss = (loss_sum / token_count).item() - bits_per_token = val_loss / math.log(2.0) - tokens_per_byte = token_count.item() / byte_count.item() - base_model.train() - return val_loss, bits_per_token * tokens_per_byte - -def eval_val_sliding_ttt( - args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, - device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, - has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, - stride: int, ttt_epochs: int = 3, ttt_lr: float = 0.001, - ttt_momentum: float = 0.9, ttt_freeze_blocks: int = 2, - batch_seqs: int = 32, eval_seq_len: int | None = None, - ttt_chunk_tokens: int = 32768, ttt_optimizer: str = "adamw", - ttt_temp: float = 1.0, - ppm_alpha: float = 0.85, - byte_weighted_ttt: bool = True, - use_cache: bool = True, - cache_alpha: float = 0.3, - adaptive_lr: bool = True, - adaptive_lr_max_mult: float = 3.0, -) -> tuple[float, float]: - """Legal score-first TTT: score each chunk, then train on it. - Every token scored BEFORE any update that could use it.""" - seq_len = eval_seq_len or args.train_seq_len - total_tokens = val_tokens.numel() - 1 - - # Initialize GPU-vectorized logistic context mixer - use_mixer = os.environ.get("USE_MIXER", "1") == "1" - mixer = LogisticContextMixer( - vocab_size=val_tokens.to(torch.int32).max().item() + 1, - device=device, - eta=float(os.environ.get("MIXER_ETA", "0.1")), - ) if use_mixer else None - if use_mixer and rank == 0: - print(f" Logistic context mixer enabled: eta={mixer.eta}") - if adaptive_lr and rank == 0: - print(f" Adaptive LR enabled: max_mult={adaptive_lr_max_mult}") - - # Pre-compute all window starts - window_starts = [ws for ws in range(0, total_tokens, stride) - if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] - - # Assign each window to a chunk based on scored token position - num_chunks = (total_tokens + ttt_chunk_tokens - 1) // ttt_chunk_tokens - chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] - for ws in window_starts: - end = min(ws + seq_len, total_tokens) - wlen = end - ws - s = 0 if ws == 0 else max(wlen - stride, 0) - scored_start = ws + s - ci = min(scored_start // ttt_chunk_tokens, num_chunks - 1) - chunk_windows[ci].append(ws) - - if rank == 0: - print(f"ttt:start chunks={num_chunks} chunk_tokens={ttt_chunk_tokens} " - f"windows={len(window_starts)} stride={stride} " - f"lr={ttt_lr} epochs={ttt_epochs} opt={ttt_optimizer} " - f"freeze_first={ttt_freeze_blocks}") - - loss_sum = torch.zeros((), device=device, dtype=torch.float64) - token_count = torch.zeros((), device=device, dtype=torch.float64) - byte_count = torch.zeros((), device=device, dtype=torch.float64) - - # Freeze everything, then selectively unfreeze for TTT - num_blocks = len(base_model.blocks) - for p in base_model.parameters(): - p.requires_grad_(False) - ttt_params = [] - ttt_param_ids = set() - use_qttt = os.environ.get("QTTT", "0") == "1" - if use_qttt: - # qTTT: only unfreeze Q projections in last N blocks + norms + head - for i in range(max(0, num_blocks - ttt_freeze_blocks), num_blocks): - for name, p in base_model.blocks[i].named_parameters(): - if "c_q" in name: - p.requires_grad_(True) - ttt_params.append(p) - ttt_param_ids.add(id(p)) - else: - # Standard: unfreeze all params in last N blocks - for i in range(max(0, num_blocks - ttt_freeze_blocks), num_blocks): - for p in base_model.blocks[i].parameters(): - p.requires_grad_(True) - ttt_params.append(p) - ttt_param_ids.add(id(p)) - # Unfreeze norms, scales, lm_head - for name, p in base_model.named_parameters(): - if "norm" in name or "scale" in name or "lm_head" in name: - p.requires_grad_(True) - if id(p) not in ttt_param_ids: - ttt_params.append(p) - ttt_param_ids.add(id(p)) - - if rank == 0: - n_unfrozen = sum(p.numel() for p in ttt_params) - n_frozen = sum(p.numel() for p in base_model.parameters() if not p.requires_grad) - print(f"ttt:params unfrozen={n_unfrozen} frozen={n_frozen}") - - if ttt_optimizer == "adamw": - optimizer = torch.optim.AdamW(ttt_params, lr=ttt_lr, weight_decay=0.0, betas=(0.9, 0.999)) - else: - optimizer = torch.optim.SGD(ttt_params, lr=ttt_lr, momentum=ttt_momentum) - - # Polyak averaging (TTT weight EMA) for smoother scoring - use_polyak = os.environ.get("USE_POLYAK", "1") == "1" - polyak_decay = float(os.environ.get("POLYAK_DECAY", "0.998")) - if use_polyak: - polyak_state = {id(p): p.data.clone() for p in ttt_params} - if rank == 0: - print(f" Polyak averaging enabled: decay={polyak_decay}") - - t0 = time.perf_counter() - - for ci in range(num_chunks): - windows = chunk_windows[ci] - if not windows: - continue - - # --- Phase 1: SCORE this chunk (inference_mode, no grad) --- - my_s = (len(windows) * rank) // world_size - my_e = (len(windows) * (rank + 1)) // world_size - my_windows = windows[my_s:my_e] - - # Swap in Polyak-averaged weights for scoring - if use_polyak and ci > 0: - _saved_weights = {} - for p in ttt_params: - _saved_weights[id(p)] = p.data.clone() - p.data.copy_(polyak_state[id(p)]) - - base_model.eval() - with torch.inference_mode(): - for bi in range(0, len(my_windows), batch_seqs): - batch_ws = my_windows[bi:bi + batch_seqs] - bsz = len(batch_ws) - x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) - y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) - wlens: list[int] = [] - for i, ws in enumerate(batch_ws): - end = min(ws + seq_len, total_tokens) - wlen = end - ws - wlens.append(wlen) - chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) - x_batch[i, :wlen] = chunk_tok[:-1] - y_batch[i, :wlen] = chunk_tok[1:] - with torch.autocast(device_type="cuda", dtype=torch.bfloat16): - logits = base_model.forward_logits(x_batch) - logits_scaled = logits.float() / ttt_temp - - # Adaptive temperature: sharpen confident predictions more - if ttt_temp != 1.0: - with torch.no_grad(): - probs_for_entropy = F.softmax(logits.float(), dim=-1) - token_entropy = -(probs_for_entropy * (probs_for_entropy + 1e-10).log()).sum(-1) - max_ent = math.log(logits.size(-1)) - # Confident tokens (low entropy) get more sharpening - adaptive_temp = 1.0 - (1.0 - ttt_temp) * (1.0 - token_entropy / max_ent) - adaptive_temp = adaptive_temp.clamp(min=0.9, max=1.05) - logits_scaled = logits.float() / adaptive_temp.unsqueeze(-1) - - # Logistic context mixing (GPU-vectorized) or plain CE - if mixer is not None: - nll, expert_nll = mixer.mix_and_score(logits_scaled, x_batch, y_batch, wlens) - mixer.update_weights(expert_nll, wlens) - else: - nll = F.cross_entropy( - logits_scaled.reshape(-1, logits_scaled.size(-1)), - y_batch.reshape(-1), reduction="none", - ).reshape(bsz, seq_len) - for i, ws in enumerate(batch_ws): - wlen = wlens[i] - s = 0 if ws == 0 else max(wlen - stride, 0) - scored_nll = nll[i, s:wlen].to(torch.float64) - loss_sum += scored_nll.sum() - token_count += float(wlen - s) - tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] - tb = base_bytes_lut[tgt].to(torch.float64) - tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) - byte_count += tb.sum() - - # --- Update context mixer with scored chunk tokens (GPU-vectorized) --- - chunk_start_tok = ci * ttt_chunk_tokens - chunk_end_tok = min((ci + 1) * ttt_chunk_tokens, total_tokens) - if mixer is not None: - mixer.update(val_tokens[chunk_start_tok:chunk_end_tok + 1]) - - # Swap back training weights after scoring - if use_polyak and ci > 0: - for p in ttt_params: - p.data.copy_(_saved_weights[id(p)]) - - # --- Phase 2: TRAIN on this chunk (already scored = legal) --- - is_last_chunk = (ci == num_chunks - 1) - if not is_last_chunk and ttt_epochs > 0: - chunk_start = ci * ttt_chunk_tokens - chunk_end = min((ci + 1) * ttt_chunk_tokens, total_tokens) - chunk_seqs = (chunk_end - chunk_start) // seq_len - if rank == 0 and ci < 3: - print(f" ttt_train [{ci+1}] seqs={chunk_seqs} start_train...", flush=True) - if chunk_seqs > 0: - # Cosine LR across chunks + adaptive scaling - cos_lr = ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) - if adaptive_lr: - # Increase LR as we've seen more data (more confident adaptation) - progress = min(ci / max(num_chunks * 0.3, 1), 1.0) # ramp over first 30% of chunks - lr_mult = 1.0 + (adaptive_lr_max_mult - 1.0) * progress - cos_lr = cos_lr * lr_mult - for pg in optimizer.param_groups: - pg["lr"] = cos_lr - my_seq_s = (chunk_seqs * rank) // world_size - my_seq_e = (chunk_seqs * (rank + 1)) // world_size - my_chunk_seqs = my_seq_e - my_seq_s - for _ep in range(ttt_epochs): - if rank == 0 and ci < 3: - print(f" ttt_train [{ci+1}] epoch={_ep+1}/{ttt_epochs} batches={my_chunk_seqs} ...", flush=True) - for bs in range(0, my_chunk_seqs, batch_seqs): - be = min(bs + batch_seqs, my_chunk_seqs) - actual_bs = my_seq_s + bs - start_tok = chunk_start + actual_bs * seq_len - end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 - if end_tok > val_tokens.numel(): - continue - local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) - x = local[:-1].reshape(-1, seq_len) - y = local[1:].reshape(-1, seq_len) - optimizer.zero_grad(set_to_none=True) - with torch.autocast(device_type="cuda", dtype=torch.bfloat16): - if byte_weighted_ttt: - # Byte-weighted loss: tokens covering more bytes matter more - ttt_logits = base_model.forward_logits(x) - per_token_loss = F.cross_entropy( - ttt_logits.reshape(-1, ttt_logits.size(-1)), - y.reshape(-1), reduction='none' - ).reshape(y.shape) - byte_weights = base_bytes_lut[y].float() - byte_weights = byte_weights + (has_leading_space_lut[y] & ~is_boundary_token_lut[x]).float() - ttt_loss = (per_token_loss * byte_weights).sum() / byte_weights.sum() - else: - ttt_loss = base_model(x, y) - ttt_loss.backward() - if world_size > 1: - for p in ttt_params: - if p.grad is not None: - dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) - torch.nn.utils.clip_grad_norm_(ttt_params, 1.0) - optimizer.step() - # Update Polyak EMA after each step - if use_polyak: - for p in ttt_params: - polyak_state[id(p)].lerp_(p.data, 1.0 - polyak_decay) - if rank == 0 and ci < 3: - print(f" step done ep={_ep+1} bs={bs} loss={ttt_loss.item():.4f}", flush=True) - - if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1 or ci < 5): - elapsed = time.perf_counter() - t0 - rl = loss_sum.item() / max(token_count.item(), 1) - rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 - print(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s", flush=True) - - if dist.is_available() and dist.is_initialized(): - dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) - dist.all_reduce(token_count, op=dist.ReduceOp.SUM) - dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) - - val_loss = (loss_sum / token_count).item() - val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) - - for p in base_model.parameters(): - p.requires_grad_(True) - base_model.eval() - - if rank == 0: - print(f"ttt:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " - f"elapsed={time.perf_counter() - t0:.1f}s") - return val_loss, val_bpb - -def _classify_param(name: str) -> str: - if "tok_emb" in name or "lm_head" in name: - return "embed" - if ".mlp." in name: - return "mlp" - if ".attn." in name or (".proj." in name and ".mlp." not in name): - return "attn" - return "other" - -def quantize_int6_per_row(t: Tensor, clip_range: int = 15) -> tuple[Tensor, Tensor]: - t32 = t.float() - if t32.ndim == 2: - best_q, best_s, best_err = None, None, float('inf') - for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: - if pct < 1.0: - row_clip = torch.quantile(t32.abs(), pct, dim=1) - else: - row_clip = t32.abs().amax(dim=1) - s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) - q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) - recon = q.float() * s.float()[:, None] - err = (t32 - recon).pow(2).mean().item() - if err < best_err: - best_q, best_s, best_err = q, s, err - return best_q, best_s - amax = t32.abs().max().item() - scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) - q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) - return q, scale - - -def _find_best_row_scales(W: Tensor, clip_range: int = 15) -> Tensor: - """Find optimal per-row scales by searching percentile clipping thresholds.""" - t32 = W.float() - best_s = t32.abs().amax(dim=1) / clip_range - best_s = best_s.clamp_min(1.0 / clip_range) - best_err = torch.full((t32.shape[0],), float('inf')) - for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: - if pct < 1.0: - row_clip = torch.quantile(t32.abs(), pct, dim=1) - else: - row_clip = t32.abs().amax(dim=1) - s = (row_clip / clip_range).clamp_min(1.0 / clip_range) - q = torch.clamp(torch.round(t32 / s[:, None]), -clip_range, clip_range) - recon = q * s[:, None] - err = (t32 - recon).pow(2).mean(dim=1) - improved = err < best_err - best_s[improved] = s[improved] - best_err[improved] = err[improved] - return best_s - -def gptq_quantize_weight(W: Tensor, H: Tensor, clip_range: int = 15, - block_size: int = 128, percdamp: float = 0.01) -> tuple[Tensor, Tensor]: - """GPTQ: quantize weight matrix W using Hessian H = X^T X for error compensation.""" - W = W.float().clone() - rows, cols = W.shape - row_scale = _find_best_row_scales(W, clip_range) - H = H.float().clone() - damp = percdamp * H.diag().mean() - H.diagonal().add_(damp) - perm = torch.argsort(H.diag()) - invperm = torch.argsort(perm) - W = W[:, perm] - H = H[perm][:, perm] - try: - L = torch.linalg.cholesky(H) - Hinv = torch.cholesky_inverse(L) - except torch._C._LinAlgError: - Hinv = torch.diag(1.0 / H.diag().clamp_min(1e-6)) - Q = torch.zeros(rows, cols, dtype=torch.int8) - for i1 in range(0, cols, block_size): - i2 = min(i1 + block_size, cols) - W_block = W[:, i1:i2].clone() - Hinv_block = Hinv[i1:i2, i1:i2] - Err = torch.zeros_like(W_block) - for j in range(i2 - i1): - w_col = W_block[:, j] - h_inv_jj = Hinv_block[j, j].clamp_min(1e-8) - q_col = torch.clamp(torch.round(w_col / row_scale), -clip_range, clip_range) - deq_col = q_col * row_scale - Q[:, i1 + j] = q_col.to(torch.int8) - err = (w_col - deq_col) / h_inv_jj - Err[:, j] = err - if j + 1 < i2 - i1: - W_block[:, j + 1:] -= err.unsqueeze(1) * Hinv_block[j, j + 1:].unsqueeze(0) - if i2 < cols: - W[:, i2:] -= Err @ Hinv[i1:i2, i2:] - Q = Q[:, invperm] - return Q, row_scale.to(torch.float16) - -def gptq_calibrate(model: nn.Module, train_pattern: str, device: torch.device, - n_samples: int = 256, seq_len: int = 2048) -> dict[str, Tensor]: - """Collect Hessian H = X^T X for each linear layer using training data.""" - hessians: dict[str, Tensor] = {} - n_seen: dict[str, int] = {} - hooks = [] - def make_hook(name: str): - def hook_fn(module, inp, out): - x = inp[0].detach().float() - if x.ndim == 3: - x = x.reshape(-1, x.shape[-1]) - if name not in hessians: - hessians[name] = torch.zeros(x.shape[1], x.shape[1], device=x.device, dtype=torch.float32) - n_seen[name] = 0 - hessians[name].addmm_(x.t(), x) - n_seen[name] += x.shape[0] - return hook_fn - for name, module in model.named_modules(): - if isinstance(module, (nn.Linear, CastedLinear)): - hooks.append(module.register_forward_hook(make_hook(name))) - stream = TokenStream(train_pattern) - model.eval() - with torch.no_grad(): - for _ in range(n_samples): - tokens = stream.take(seq_len + 1).to(device=device, dtype=torch.int64) - x = tokens[:-1].unsqueeze(0) - with torch.autocast(device_type="cuda", dtype=torch.bfloat16): - model.forward_logits(x) - for h in hooks: - h.remove() - for name in hessians: - hessians[name] /= max(n_seen[name], 1) - return hessians - -def _get_layer_clip_range(name: str, num_layers: int, int6_last_n: int) -> int: - """Return clip_range based on which layer the param belongs to.""" - import re - m = re.search(r'blocks\.(\d+)\.', name) - if m: - layer_idx = int(m.group(1)) - if layer_idx >= num_layers - int6_last_n: - return 31 # int6 - return 15 # int5 - -def mixed_quantize_int6_gptq(state_dict: dict[str, Tensor], int6_cats: set[str], - hessians: dict[str, Tensor], - num_layers: int = 11, int6_last_n: int = 2) -> tuple[dict, dict]: - """GPTQ quantization with mixed int5/int6 precision. int6 for last int6_last_n layers, int5 for rest.""" - result: dict[str, Tensor] = {} - meta: dict[str, object] = {} - gptq_count, naive_count = 0, 0 - int5_params, int6_params = 0, 0 - for name, tensor in state_dict.items(): - t = tensor.detach().cpu().contiguous() - cat = _classify_param(name) - if not t.is_floating_point() or t.numel() <= 65536: - result[name] = t.to(torch.float16) if t.is_floating_point() else t - meta[name] = "passthrough" - continue - if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): - result[name] = t.float() - meta[name] = "passthrough_ctrl" - continue - cr = _get_layer_clip_range(name, num_layers, int6_last_n) - if cr == 31: - int6_params += t.numel() - else: - int5_params += t.numel() - if cat in int6_cats and t.ndim == 2: - module_name = name.rsplit(".weight", 1)[0] if name.endswith(".weight") else name - H = hessians.get(module_name) - if H is not None and H.shape[0] == t.shape[1]: - q, s = gptq_quantize_weight(t, H.cpu(), clip_range=cr) - gptq_count += 1 - else: - q, s = quantize_int6_per_row(t, clip_range=cr) - naive_count += 1 - result[name + ".q"] = q - result[name + ".scale"] = s - meta[name] = {"type": f"int{'6' if cr == 31 else '5'}"} - elif cat in int6_cats and t.ndim >= 1: - q, s = quantize_int6_per_row(t, clip_range=cr) - result[name + ".q"] = q - result[name + ".scale"] = s - meta[name] = {"type": f"int{'6' if cr == 31 else '5'}"} - naive_count += 1 - else: - q, s = quantize_float_tensor(t) - result[name + ".q"] = q - result[name + ".scale"] = s - meta[name] = {"type": "int8"} - print(f"gptq_quantize: {gptq_count} GPTQ layers, {naive_count} naive layers", flush=True) - print(f"mixed_precision: {int5_params} int5 params, {int6_params} int6 params", flush=True) - return result, meta - - -def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): - result: dict[str, Tensor] = {} - meta: dict[str, object] = {} - for name, tensor in state_dict.items(): - t = tensor.detach().cpu().contiguous() - cat = _classify_param(name) - if not t.is_floating_point() or t.numel() <= 65536: - result[name] = t.to(torch.float16) if t.is_floating_point() else t - meta[name] = "passthrough" - continue - if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): - result[name] = t.float() - meta[name] = "passthrough_ctrl" - continue - if cat in int6_cats and t.ndim >= 1: - q, s = quantize_int6_per_row(t) - result[name + ".q"] = q - result[name + ".scale"] = s - meta[name] = {"type": "int6"} - else: - q, s = quantize_float_tensor(t) - result[name + ".q"] = q - result[name + ".scale"] = s - meta[name] = {"type": "int8"} - return result, meta - -def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], - template_sd: dict[str, Tensor]) -> dict[str, Tensor]: - out: dict[str, Tensor] = {} - for name, orig in template_sd.items(): - info = meta.get(name) - if info is None: - continue - orig_dtype = orig.dtype - if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): - t = result[name] - if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): - t = t.to(orig_dtype) - out[name] = t - continue - q, s = result[name + ".q"], result[name + ".scale"] - if s.ndim > 0: - out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) - else: - out[name] = (q.float() * float(s.item())).to(orig_dtype) - return out - -def main() -> None: - global zeropower_via_newtonschulz5 - code = Path(__file__).read_text(encoding="utf-8") - args = Hyperparameters() - zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) - distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ - rank = int(os.environ.get("RANK", "0")) - world_size = int(os.environ.get("WORLD_SIZE", "1")) - local_rank = int(os.environ.get("LOCAL_RANK", "0")) - if world_size <= 0: - raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") - if 8 % world_size != 0: - raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") - grad_accum_steps = 8 // world_size - grad_scale = 1.0 / grad_accum_steps - if not torch.cuda.is_available(): - raise RuntimeError("CUDA is required") - device = torch.device("cuda", local_rank) - torch.cuda.set_device(device) - if distributed: - dist.init_process_group(backend="nccl", device_id=device) - dist.barrier() - master_process = rank == 0 - torch.backends.cuda.matmul.allow_tf32 = True - torch.backends.cudnn.allow_tf32 = True - from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp - enable_cudnn_sdp(False) - enable_flash_sdp(True) - enable_mem_efficient_sdp(False) - enable_math_sdp(False) - logfile = None - if master_process: - os.makedirs("logs", exist_ok=True) - logfile = f"logs/{args.run_id}.txt" - print(logfile) - - def log0(msg: str, console: bool = True) -> None: - if not master_process: - return - if console: - print(msg) - if logfile is not None: - with open(logfile, "a", encoding="utf-8") as f: - print(msg, file=f) - log0(code, console=False) - log0(f"Python {sys.version} PyTorch {torch.__version__}", console=False) - random.seed(args.seed) - np.random.seed(args.seed) - torch.manual_seed(args.seed) - torch.cuda.manual_seed_all(args.seed) - if not args.tokenizer_path.endswith(".model"): - raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") - sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) - if int(sp.vocab_size()) != args.vocab_size: - raise ValueError( - f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" - ) - dataset_dir = Path(args.data_path).resolve() - actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) - effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len - val_seq_len = max(args.train_seq_len, effective_eval_seq_len) - val_tokens = load_validation_tokens(args.val_files, val_seq_len) - base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( - sp, args.vocab_size, device - ) - log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") - log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") - log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") - CastedLinear._qat_enabled = args.qat_enabled - base_model = GPT( - vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, - num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, - tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, - logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, - bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, - xsa_last_n=args.xsa_last_n, rope_dims=args.rope_dims, ln_scale=args.ln_scale, - dtg=args.dtg_enabled, ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, - ).to(device).bfloat16() - for module in base_model.modules(): - if isinstance(module, CastedLinear): - module.float() - restore_low_dim_params_to_fp32(base_model) - compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) - model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model - block_named_params = list(base_model.blocks.named_parameters()) - matrix_params = [ - p - for name, p in block_named_params - if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) - ] - scalar_params = [ - p - for name, p in block_named_params - if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) - ] - if base_model.skip_weights.numel() > 0: - scalar_params.append(base_model.skip_weights) - scalar_params.append(base_model.smear.gate) - if base_model.bigram is not None: - scalar_params.append(base_model.bigram.scale) - token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr - tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] - if base_model.bigram is not None: - tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) - if base_model.bigram.proj is not None: - matrix_params.append(base_model.bigram.proj.weight) - if base_model.ve_shared is not None: - tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) - if base_model.ve_shared.proj is not None: - matrix_params.append(base_model.ve_shared.proj.weight) - scalar_params.append(base_model.ve_shared.scale) - for s in base_model.ve_layer_scales: - scalar_params.append(s) - optimizer_tok = torch.optim.AdamW( - tok_params, - betas=(args.beta1, args.beta2), - eps=args.adam_eps, - weight_decay=args.adam_wd, - fused=True, - ) - optimizer_muon = Muon( - matrix_params, - lr=args.matrix_lr, - momentum=args.muon_momentum, - backend_steps=args.muon_backend_steps, - weight_decay=args.muon_wd, - ) - for group in optimizer_muon.param_groups: - group["base_lr"] = args.matrix_lr - optimizer_scalar = torch.optim.AdamW( - [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], - betas=(args.beta1, args.beta2), - eps=args.adam_eps, - weight_decay=args.adam_wd, - fused=True, - ) - optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] - if base_model.lm_head is not None: - optimizer_head = torch.optim.Adam( - [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], - betas=(args.beta1, args.beta2), - eps=args.adam_eps, - fused=True, - ) - optimizers.insert(1, optimizer_head) - n_params = sum(p.numel() for p in base_model.parameters()) - # Set int6 clip_range for last N layers (mixed precision) - int6_start = args.num_layers - args.int6_last_n - for i, block in enumerate(base_model.blocks): - if i >= int6_start: - for m in block.modules(): - if isinstance(m, CastedLinear): - m._clip_range = 31 # int6 - if master_process: - int5_count = sum(1 for m in base_model.modules() if isinstance(m, CastedLinear) and m._clip_range == 15) - int6_count = sum(1 for m in base_model.modules() if isinstance(m, CastedLinear) and m._clip_range == 31) - log0(f"mixed_precision: {int5_count} int5 layers, {int6_count} int6 layers (last {args.int6_last_n} blocks)") - log0(f"model_params:{n_params}") - xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] - log0(f"XSA:{xsa_layers} ws:{world_size} gqa:{args.num_heads}/{args.num_kv_heads}") - log0(f"lr:embed={token_lr} matrix={args.matrix_lr} scalar={args.scalar_lr} batch:{args.train_batch_tokens} wall:{args.max_wallclock_seconds:.0f}s seed:{args.seed}") - train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) - def zero_grad_all() -> None: - for opt in optimizers: - opt.zero_grad(set_to_none=True) - max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None - train_reserve_ms = 18000 # reserve 18s for EMA + GPTQ calibration + quantization + save - effective_train_ms = (max_wallclock_ms - train_reserve_ms) if max_wallclock_ms is not None else None - - def lr_mul(step: int, elapsed_ms: float) -> float: - if args.warmdown_iters <= 0: - return 1.0 - if effective_train_ms is None: - warmdown_start = max(args.iterations - args.warmdown_iters, 0) - return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 - step_ms = elapsed_ms / max(step, 1) - warmdown_ms = args.warmdown_iters * step_ms - remaining_ms = max(effective_train_ms - elapsed_ms, 0.0) - return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 - # TTT_ONLY mode: skip training, load saved model, run TTT eval - if os.environ.get("TTT_ONLY", "0") == "1": - log0("TTT_ONLY mode: skipping training, loading saved model...") - sd_cpu = {k: v.cpu() for k, v in torch.load("final_model.pt", map_location="cpu").items()} - if args.prune_pct > 0: - for k, v in sd_cpu.items(): - if v.ndim == 2 and v.numel() > 65536: - thresh = torch.quantile(v.abs().float(), args.prune_pct) - v[v.abs() < thresh] = 0.0 - log0(f"pruning:{args.prune_pct*100:.1f}% magnitude pruning applied") - with open("final_model.int6.ptz", "rb") as f: - quant_blob_disk = f.read() - quant_state = torch.load( - io.BytesIO(zstandard.ZstdDecompressor().decompress(quant_blob_disk) if _COMPRESSOR == "zstd" else zlib.decompress(quant_blob_disk)), - map_location="cpu", - ) - deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) - eval_model = GPT( - vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, - num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, - tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, - logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, - bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, xsa_last_n=args.xsa_last_n, - rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, - ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, - ).to(device).bfloat16() - for m in eval_model.modules(): - if isinstance(m, CastedLinear): - m.float() - restore_low_dim_params_to_fp32(eval_model) - eval_model.load_state_dict(deq_state, strict=True) - sw_seq_len = int(os.environ.get("EVAL_SEQ_LEN", str(effective_eval_seq_len))) - log0(f"TTT_ONLY: model loaded, starting TTT eval...") - torch.cuda.synchronize() - t_ttt = time.perf_counter() - ttt_epochs = int(os.environ.get("TTT_EPOCHS", "3")) - ttt_lr = float(os.environ.get("TTT_LR", "0.0005")) - ttt_freeze = int(os.environ.get("TTT_FREEZE_BLOCKS", "2")) - ttt_chunk = int(os.environ.get("TTT_CHUNK_TOKENS", "32768")) - ttt_opt = os.environ.get("TTT_OPTIMIZER", "adamw") - log0(f"TTT: epochs={ttt_epochs} lr={ttt_lr} freeze_first={ttt_freeze} chunk={ttt_chunk} opt={ttt_opt}") - ttt_temp = args.ttt_temperature - log0(f"TTT temperature: {ttt_temp}") - ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( - args, eval_model, rank, world_size, device, - val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, - stride=args.eval_stride, ttt_epochs=ttt_epochs, ttt_lr=ttt_lr, - ttt_freeze_blocks=ttt_freeze, eval_seq_len=sw_seq_len, - ttt_chunk_tokens=ttt_chunk, ttt_optimizer=ttt_opt, - ttt_temp=ttt_temp, - ppm_alpha=float(os.environ.get("PPM_ALPHA", "0.85")), - byte_weighted_ttt=os.environ.get("BYTE_WEIGHTED_TTT", "1") == "1", - use_cache=os.environ.get("USE_CACHE", "1") == "1", - cache_alpha=float(os.environ.get("CACHE_ALPHA", "0.3")), - adaptive_lr=os.environ.get("ADAPTIVE_LR", "1") == "1", - adaptive_lr_max_mult=float(os.environ.get("ADAPTIVE_LR_MAX", "3.0")), - ) - torch.cuda.synchronize() - log0( - f"final_int6_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " - f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" - ) - log0(f"final_int6_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") - if distributed: - dist.destroy_process_group() - return - - if args.warmup_steps > 0: - initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} - initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] - model.train() - for warmup_step in range(args.warmup_steps): - zero_grad_all() - for micro_step in range(grad_accum_steps): - if distributed: - model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 - x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) - with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): - warmup_loss = model(x, y) - (warmup_loss * grad_scale).backward() - for opt in optimizers: - opt.step() - zero_grad_all() - if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: - log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") - base_model.load_state_dict(initial_model_state, strict=True) - for opt, state in zip(optimizers, initial_optimizer_states, strict=True): - opt.load_state_dict(state) - zero_grad_all() - if distributed: - model.require_backward_grad_sync = True - train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) - swa_state: dict[str, Tensor] | None = None - swa_count = 0 - ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} - ema_decay = 0.997 - training_time_ms = 0.0 - stop_after_step: int | None = None - torch.cuda.synchronize() - t0 = time.perf_counter() - step = 0 - while True: - last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) - should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) - if should_validate: - torch.cuda.synchronize() - training_time_ms += 1000.0 * (time.perf_counter() - t0) - val_loss, val_bpb = eval_val( - args, - model, - rank, - world_size, - device, - grad_accum_steps, - val_tokens, - base_bytes_lut, - has_leading_space_lut, - is_boundary_token_lut, - ) - log0( - f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " - f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" - ) - torch.cuda.synchronize() - t0 = time.perf_counter() - if last_step: - if stop_after_step is not None and step < args.iterations: - log0( - f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " - f"step:{step}/{args.iterations}" - ) - break - elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) - scale = lr_mul(step, elapsed_ms) - # Anneal soft-round alpha based on QAT progress - if CastedLinear._use_soft_round and CastedLinear._qat_enabled: - qat_progress = max(0.0, 1.0 - scale / max(args.late_qat_threshold, 0.01)) - CastedLinear._soft_round_alpha = 1.0 + 15.0 * qat_progress - if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: - CastedLinear._qat_enabled = True - CastedLinear._use_soft_round = os.environ.get("SOFT_ROUND_QAT", "0") == "1" - if CastedLinear._use_soft_round and master_process: - log0(f"soft_round_qat:enabled initial_alpha=1.0") - log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") - zero_grad_all() - train_loss = torch.zeros((), device=device) - for micro_step in range(grad_accum_steps): - if distributed: - model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 - x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) - with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): - loss = model(x, y) - train_loss += loss.detach() - (loss * grad_scale).backward() - train_loss /= grad_accum_steps - frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 - muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum - for group in optimizer_muon.param_groups: - group["momentum"] = muon_momentum - for opt in optimizers: - for group in opt.param_groups: - group["lr"] = group["base_lr"] * scale - if args.grad_clip_norm > 0: - torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) - for opt in optimizers: - opt.step() - zero_grad_all() - with torch.no_grad(): - for name, t in base_model.state_dict().items(): - ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) - step += 1 - approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) - if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: - if swa_state is None: - swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} - swa_count = 1 - log0(f"swa:start step:{step}") - else: - for name, t in base_model.state_dict().items(): - swa_state[name] += t.detach().cpu() - swa_count += 1 - should_log_train = ( - args.train_log_every > 0 - and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) - ) - if should_log_train: - log0( - f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " - f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" - ) - reached_cap = effective_train_ms is not None and approx_training_time_ms >= effective_train_ms - if distributed and max_wallclock_ms is not None: - reached_cap_tensor = torch.tensor(int(reached_cap), device=device) - dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) - reached_cap = bool(reached_cap_tensor.item()) - if stop_after_step is None and reached_cap: - stop_after_step = step - log0( - f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " - f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" - ) - # Apply EMA weights directly (skip diagnostic evals to save ~5s of reserve) - log0("ema:applying EMA weights (skipping diagnostic evals)") - current_state = base_model.state_dict() - ema_sd = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} - base_model.load_state_dict(ema_sd, strict=True) - # GPTQ calibration on final model (within reserved training budget) - log0("gptq:calibrating with training data...") - t_gptq = time.perf_counter() - gptq_hessians = gptq_calibrate(base_model, args.train_files, device, n_samples=128, seq_len=args.train_seq_len) - log0(f"gptq:calibrated {len(gptq_hessians)} layers in {time.perf_counter()-t_gptq:.1f}s") - export_sd = base_model.state_dict() - if master_process: - torch.save(export_sd, "final_model.pt") - model_bytes = os.path.getsize("final_model.pt") - code_bytes = len(code.encode("utf-8")) - log0(f"Serialized model: {model_bytes} bytes") - log0(f"Code size: {code_bytes} bytes") - sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} - if args.prune_pct > 0: - for k, v in sd_cpu.items(): - if v.ndim == 2 and v.numel() > 65536: - thresh = torch.quantile(v.abs().float(), args.prune_pct) - v[v.abs() < thresh] = 0.0 - if master_process: - log0(f"pruning:{args.prune_pct*100:.1f}% magnitude pruning applied") - quant_result, quant_meta = mixed_quantize_int6_gptq(sd_cpu, {"mlp", "attn"}, gptq_hessians, num_layers=args.num_layers, int6_last_n=args.int6_last_n) - quant_buf = io.BytesIO() - torch.save({"w": quant_result, "m": quant_meta}, quant_buf) - quant_raw = quant_buf.getvalue() - quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) if _COMPRESSOR == "zstd" else zlib.compress(quant_raw, 9) - if master_process: - with open("final_model.int6.ptz", "wb") as f: - f.write(quant_blob) - quant_file_bytes = len(quant_blob) - code_bytes = len(code.encode("utf-8")) - log0(f"Serialized model int6+{_COMPRESSOR}: {quant_file_bytes} bytes") - log0(f"Total submission size int6+{_COMPRESSOR}: {quant_file_bytes + code_bytes} bytes") - if distributed: - dist.barrier() - with open("final_model.int6.ptz", "rb") as f: - quant_blob_disk = f.read() - quant_state = torch.load( - io.BytesIO(zstandard.ZstdDecompressor().decompress(quant_blob_disk) if _COMPRESSOR == "zstd" else zlib.decompress(quant_blob_disk)), - map_location="cpu", - ) - deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) - eval_model = GPT( - vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, - num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, - tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, - logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, - bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, xsa_last_n=args.xsa_last_n, - rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, - ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, - ).to(device).bfloat16() - for m in eval_model.modules(): - if isinstance(m, CastedLinear): - m.float() - restore_low_dim_params_to_fp32(eval_model) - eval_model.load_state_dict(deq_state, strict=True) - sw_seq_len = int(os.environ.get("EVAL_SEQ_LEN", str(effective_eval_seq_len))) - if sw_seq_len != effective_eval_seq_len and rank == 0: - log0(f"Eval seq_len override: {effective_eval_seq_len} -> {sw_seq_len}") - if args.eval_stride > 0 and args.eval_stride < sw_seq_len and not os.environ.get("SKIP_SLIDING"): - torch.cuda.synchronize() - t_slide = time.perf_counter() - sw_val_loss, sw_val_bpb = eval_val_sliding( - args, eval_model, rank, world_size, device, - val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, - stride=args.eval_stride, - eval_seq_len=sw_seq_len, - ) - torch.cuda.synchronize() - log0( - f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " - f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" - ) - log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") - torch.cuda.synchronize() - t_ttt = time.perf_counter() - ttt_epochs = int(os.environ.get("TTT_EPOCHS", "3")) - ttt_lr = float(os.environ.get("TTT_LR", "0.0005")) - ttt_freeze = int(os.environ.get("TTT_FREEZE_BLOCKS", "2")) - ttt_chunk = int(os.environ.get("TTT_CHUNK_TOKENS", "32768")) - ttt_opt = os.environ.get("TTT_OPTIMIZER", "adamw") - log0(f"TTT: epochs={ttt_epochs} lr={ttt_lr} freeze_first={ttt_freeze} chunk={ttt_chunk} opt={ttt_opt}") - ttt_temp = args.ttt_temperature - log0(f"TTT temperature: {ttt_temp}") - ppm_alpha_val = float(os.environ.get("PPM_ALPHA", "0.85")) - bw_ttt = os.environ.get("BYTE_WEIGHTED_TTT", "1") == "1" - log0(f"PPM alpha: {ppm_alpha_val}, Byte-weighted TTT: {bw_ttt}") - ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( - args, eval_model, rank, world_size, device, - val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, - stride=args.eval_stride, ttt_epochs=ttt_epochs, ttt_lr=ttt_lr, - ttt_freeze_blocks=ttt_freeze, eval_seq_len=sw_seq_len, - ttt_chunk_tokens=ttt_chunk, ttt_optimizer=ttt_opt, - ttt_temp=ttt_temp, - ppm_alpha=float(os.environ.get("PPM_ALPHA", "0.85")), - byte_weighted_ttt=os.environ.get("BYTE_WEIGHTED_TTT", "1") == "1", - use_cache=os.environ.get("USE_CACHE", "1") == "1", - cache_alpha=float(os.environ.get("CACHE_ALPHA", "0.3")), - adaptive_lr=os.environ.get("ADAPTIVE_LR", "1") == "1", - adaptive_lr_max_mult=float(os.environ.get("ADAPTIVE_LR_MAX", "3.0")), - ) - torch.cuda.synchronize() - log0( - f"final_int6_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " - f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" - ) - log0(f"final_int6_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") - if distributed: - dist.destroy_process_group() -if __name__ == "__main__": - main() From a366675fbcda83df0f634a8967d6f041b868a651 Mon Sep 17 00:00:00 2001 From: RoyiRa <37896343+RoyiRa@users.noreply.github.com> Date: Wed, 25 Mar 2026 12:41:36 +0200 Subject: [PATCH 4/4] Update author and add github_id in submission.json --- .../track_10min_16mb/2026-03-24_HedgeMixer_TTT/submission.json | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/submission.json b/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/submission.json index 2835282df..3bc12ffd4 100644 --- a/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/submission.json +++ b/records/track_10min_16mb/2026-03-24_HedgeMixer_TTT/submission.json @@ -1,8 +1,9 @@ { "track": "10min_16mb", "date": "2026-03-24", + "github_id": "royira", "name": "5-expert Hedge Mixer + TTT", - "author": "notapplica", + "author": "royira", "seed_results": { "1337": {"val_loss": 1.78307674, "val_bpb": 1.05604198, "artifact_bytes": 15484405}, "42": {"val_loss": 1.85218082, "val_bpb": 1.09696945, "artifact_bytes": 15408595},