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

Open

whn09 wants to merge 28 commits into

aws-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
05b1e5e
preprocess: rebuild weight_map from actual shard files for V2.5

whn09
and
claude
committed
b830a78
preprocess: split MiMo-V2.5 fused qkv_proj back into q/k/v for NxDI

whn09
and
claude
committed
d381601

Commits on Apr 28, 2026

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

whn09
and
claude
committed
eb6b2aa
README/smoke/bench: V2.5-specific content + move scratch off /tmp

whn09
and
claude
committed
94d001d
README: add 'Testing the vLLM server' subsection + fix NC count

whn09
and
claude
committed
1516d7a
0_setup: resolve CONTRIB_SRC before cd-ing into vllm-neuron

whn09
and
claude
committed
2149480
README: fix maintainer name (Henan Wang, not Wan)

whn09
and
claude
committed
60c17a9
bench scripts: source vllm venv, not the nxd-training one

whn09
and
claude
committed
89e04cd
vllm: reuse MiMoV2Flash arch name to satisfy vLLM's builtin validator

whn09
and
claude
committed
4e788f2
Consolidate env var docs + point scratch dirs at persistent storage

whn09
and
claude
committed
d724a94
bench: drop the BS=128 config, keep only BS=32

whn09
and
claude
committed
1469d15
sanity_check + run_bench_single: default to Neuron-FP8 checkpoint

whn09
and
claude
committed
9b44f57
perf_test: split out start_vllm_server.sh; make bench_mimo_v2_5 a thin wrapper

whn09
and
claude
committed
2e5348b
README: record actual bench_mimo_v2_5 results (FP8 BS=32, c=1/16/32)

whn09
and
claude
committed
1ef5ef4

Commits on Jul 21, 2026

Merge branch 'aws-neuron:main' into contrib/MiMo-V2.5
whn09
authored
d06ab78
[contrib] MiMo-V2.5: fix decode sliding-window mask to use per-batch-slot position

whn09
and
claude
committed
d47d9e8
Add H100 GPU baseline for MiMo-V2.5 (SGLang single-node)

whn09
and
claude
committed
b2488f6
Fix batched-decode garble: per-row causal upper bound on prior KV

whn09
and
claude
committed
8a64b5f
h100/MiMo-V2.5: use cookbook DP=2 as SGLang default (no DeepEP needed)

whn09
and
claude
committed
4e69aee
h100/MiMo-V2.5: vLLM works with PR #42270 (verified) + add vLLM benchmarks

whn09
and
claude
committed
6ea3156
h100/MiMo-V2.5: enable chunked prefill for vLLM (default) + add its bench

whn09
and
claude
committed
d87a41a

Commits on Jul 22, 2026

h100/MiMo-V2.5: add 32K long-context concurrency exploration (SGLang)

whn09
and
claude
committed
8ebc732
Add data-parallel + context-parallel attention for long context (512->16K)

whn09
and
claude
committed
163a7c6
h100/MiMo-V2.5: add 15K-input / 1K-output bench (Trn2 16K-context point)

whn09
and
claude
committed
94ce6b2
docs: document long-context (16K) support and benchmarks

whn09
and
claude
committed
b078f51
h100/MiMo-V2.5: add prefill benchmark (16K in / 10 out) + root-cause vLLM gap

whn09
and
claude
committed
0063b00
h100/MiMo-V2.5: add H100-vs-Trn2 cross-platform table to 15K/1K results

whn09
and
claude
committed
a07e2cb