Skip to content

MiniMax-M3.1-Flash MI355X SGLang AgentX: opt into full-precision custom all-reduce / MiniMax-M3.1-Flash MI355X SGLang AgentX：启用全精度 custom all-reduce #4745

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
ZhengGong-amd wants to merge 2 commits into from
Closed

MiniMax-M3.1-Flash MI355X SGLang AgentX: opt into full-precision custom all-reduce / MiniMax-M3.1-Flash MI355X SGLang AgentX：启用全精度 custom all-reduce #4745

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
2d1c39b
perf(minimaxm31flash-mi355x-sglang): opt into full-precision custom a…
ZhengGong-amd Nov 2, 2026
8708b7b
perf(minimaxm31flash-mi355x-sglang): drop redundant INT4 quick-reduce…
ZhengGong-amd Nov 2, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 4 additions & 0 deletions ...nchmarks/single_node/srt-slurm-recipes/minimaxm31flash/sglang/mi355x-fp4-mtp/agentic.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,6 +49,10 @@ base:
# Recompute the sparse index top-k in every layer (SGLang's ROCm default 2 shares
# one selection across each pair of sparse layers).
SGLANG_MINIMAX_M3_INDEX_TOPK_FREQ: '1'
# MiniMax-M3 on ROCm disables custom all-reduce unless opted in.
SGLANG_M3_ALLOW_CUSTOM_AR: '1'
# The image otherwise quantizes the all-reduce to INT8, or INT4 for M3.1.
ROCM_QUICK_REDUCE_QUANTIZATION: 'NONE'
benchmark:
env:
# Warmup can leave bytes unacknowledged past AIPerf's 30 s default.
Expand Down
7 changes: 7 additions & 0 deletions inferencex-e2e/perf-changelog.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11800,3 +11800,10 @@
- "Add aggregated SGLang behind Dynamo for MiniMax-M3.1-Flash AgentX on GB300 with lmsysorg/sglang:nightly-cu134-20261030-e371881: TP4 at conc 1-16; TP4 with HiCache 180 GB per rank, shortest-prefill-first and prefill-decode interval 32 at conc 24; TP2 with HiCache 180 GB per rank at conc 40 (with and without shortest-prefill-first) and 44."
- "Disagg prefill chunk 16384 (8192 per DP rank); add 4 DEP2 prefill workers with a DEP8 decode at conc 224 and 256, 8 DEP2 prefill workers with a DEP8 decode at conc 440 and 480, conc 340 on the 6 DEP2 + DEP8 shape, and conc 360 and 400 on the 6 DEP2 + DEP16 shape."
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4698

- config-keys:
- minimaxm31flash-fp4-mi355x-sglang-agentic-dspark
scenario-type: [agentic-coding]
description:
- "Opt MI355X SGLang MiniMax-M3.1-Flash AgentX into custom all-reduce (SGLANG_M3_ALLOW_CUSTOM_AR=1), which MiniMax-M3 on ROCm otherwise disables, at full precision (ROCM_QUICK_REDUCE_QUANTIZATION=NONE)."
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4745
Loading