contrib: add MiMo-V2.5 (FP8 on Trn2)#148

Open

whn09 wants to merge 15 commits intoaws-neuron:mainfrom

whn09:contrib/MiMo-V2.5

Commits on Apr 27, 2026

Add contrib MiMo-V2.5 as a rename of MiMo-V2-Flash

whn09
and
claude
committed
preprocess: rebuild weight_map from actual shard files for V2.5

whn09
and
claude
committed
preprocess: split MiMo-V2.5 fused qkv_proj back into q/k/v for NxDI

whn09
and
claude
committed

Commits on Apr 28, 2026

preprocess: fix V2.5 fused qkv split to use 4 interleaved groups

whn09
and
claude
committed
README/smoke/bench: V2.5-specific content + move scratch off /tmp

whn09
and
claude
committed
README: add 'Testing the vLLM server' subsection + fix NC count

whn09
and
claude
committed
0_setup: resolve CONTRIB_SRC before cd-ing into vllm-neuron

whn09
and
claude
committed
README: fix maintainer name (Henan Wang, not Wan)

whn09
and
claude
committed
bench scripts: source vllm venv, not the nxd-training one

whn09
and
claude
committed
vllm: reuse MiMoV2Flash arch name to satisfy vLLM's builtin validator

whn09
and
claude
committed
Consolidate env var docs + point scratch dirs at persistent storage

whn09
and
claude
committed
bench: drop the BS=128 config, keep only BS=32

whn09
and
claude
committed
sanity_check + run_bench_single: default to Neuron-FP8 checkpoint

whn09
and
claude
committed
perf_test: split out start_vllm_server.sh; make bench_mimo_v2_5 a thin wrapper

whn09
and
claude
committed
README: record actual bench_mimo_v2_5 results (FP8 BS=32, c=1/16/32)

whn09
and
claude
committed