[ZhengGong-amd]

Summary

SGLang disables custom all-reduce for MiniMax-M3 on ROCm unless SGLANG_M3_ALLOW_CUSTOM_AR=1 (sgl-project/sglang#32230). The pinned image enables it automatically only for the Quark exports, so this recipe runs its TP4 decode all-reduces on RCCL. This PR sets the opt-in, plus ROCM_QUICK_REDUCE_QUANTIZATION=NONE to keep the all-reduce full precision. Without that, the image quantizes it to INT8, and its MiniMax-M3.1 override raises that to INT4, which #4510 would forbid.

Evidence comes from a Hyperloom AgentX run on 4x MI355X at conc 8, with amd/MiniMax-M3.1-Flash-Quark-MXFP4 and SGLang 0.5.22.dev20261028. Both sides of that comparison set SGLANG_MINIMAX_M3_INDEX_TOPK_FREQ=4 and quick reduce NONE. The opt-in raised p50 normalized interactivity from 121.1 to 134.6 tok/s/user (+11.2%) and output throughput from 345.7 to 367.6 tok/s. That run used a different checkpoint and image from this recipe, so the sweep here is the measurement.

AI model disclosure

• Model/version: Claude Opus 5.5 in Cursor; the runtime does not expose the exact model identifier, so it could not be verified.
  • Role: analysis, the change and this description.
• Model/version: claude-sonnet-5.5, as configured for the Hyperloom optimizer.
  • Role: proposed and measured the opt-in in the source run.

PR checklist and change type

Type of Change

• Bug fix
• New feature
• Configuration change
• Documentation update
• Other (please describe)

Checklist

• I have completed the AI model disclosure and kept it current
• I have tested my changes locally
• I have updated documentation if necessary
• If python -m infx.matrix.changed --base origin/main lists affected config keys, I have appended one new entry covering them to the physical end of inferencex-e2e/perf-changelog.yaml; if it lists none, I have added none. I have not edited historical entries
• If this PR appends an inferencex-e2e/perf-changelog.yaml entry, it carries exactly one primary sweep label (a maintainer applies it on fork PRs): full-sweep-fail-fast (recommended), full-sweep-enabled, or non-canary-full-sweep-enabled. Optional modifiers all-evals, evals-only, and agentx-fast require a primary label; the last two block reuse while applied.
• Before merging via reuse, an authorized maintainer (OWNER/MEMBER/COLLABORATOR) has commented /use <run_id> (or the legacy /reuse-sweep-run) on this PR. Do this only once there is a final full sweep that is all green with evals passing, since after this comment the primary sweep label will no longer automatically kick off new sweeps. Remove and re-add the primary sweep label to force a new sweep.

中文

SGLang 默认在 ROCm 上为 MiniMax-M3 关闭 custom all-reduce，除非设置 SGLANG_M3_ALLOW_CUSTOM_AR=1（sgl-project/sglang#32230）。固定镜像只对 Quark 导出版本自动开启它，所以本配方 TP4 下的 decode all-reduce 走的是 RCCL。本 PR 设置该开关，并设置 ROCM_QUICK_REDUCE_QUANTIZATION=NONE 让 all-reduce 保持全精度。如果不设置，镜像会把 all-reduce 量化为 INT8，它对 MiniMax-M3.1 的 override 还会再提升为 INT4，而 #4510 会禁止 INT4。

依据来自一次 Hyperloom AgentX 运行：4 张 MI355X，conc 8，使用 amd/MiniMax-M3.1-Flash-Quark-MXFP4 和 SGLang 0.5.22.dev20261028。对比的两组都设置了 SGLANG_MINIMAX_M3_INDEX_TOPK_FREQ=4，quick reduce 都为 NONE。开启该开关后，p50 归一化交互性从 121.1 提升到 134.6 tok/s/user（+11.2%），输出吞吐从 345.7 提升到 367.6 tok/s。那次运行的 checkpoint 和镜像都与本配方不同，以本 PR 的 sweep 为准。

AI 模型使用说明：Claude Opus 5.5（Cursor；运行环境没有提供精确的模型标识，无法确认）负责分析、改动和撰写本说明；Hyperloom 优化器配置的 claude-sonnet-5.5 在来源运行中提出并测量了该开关。

改动类型：配置改动。

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[ZhengGong-amd]

Closing: this is a no-op on the pinned image. MiniMaxAI/MiniMax-M3.1-Flash resolves to minimax_m3_nvfp4, and its gfx950 path (_hip.nvfp4_overrides) returns before the MiniMax-M3 override that disables custom all-reduce, so SGLANG_M3_ALLOW_CUSTOM_AR is never read. The conc 8 server log of the #4694 sweep already shows:

• serving as minimax_m3_nvfp4 and disable_custom_all_reduce: False;
• [AR] Using AiterCustomAllreduce on every TP rank;
• MiniMax-M3.1: quick all-reduce INT4 (from INT8 or unset).

The measured gain came from upstream SGLang serving the Quark MXFP4 checkpoint, where custom all-reduce is off by default; this recipe already has it on.

中文

关闭原因：在固定镜像上这项改动不起作用。MiniMaxAI/MiniMax-M3.1-Flash 会被解析为 minimax_m3_nvfp4，它在 gfx950 上的路径（_hip.nvfp4_overrides）在 MiniMax-M3 override 关闭 custom all-reduce 之前就已返回，所以 SGLANG_M3_ALLOW_CUSTOM_AR 根本不会被读取。#4694 的 sweep 中 conc 8 的 server 日志已经显示：

• serving as minimax_m3_nvfp4 和 disable_custom_all_reduce: False；
• 每个 TP rank 都使用 [AR] Using AiterCustomAllreduce；
• MiniMax-M3.1: quick all-reduce INT4 (from INT8 or unset)。

之前测得的收益来自上游 SGLang 加 Quark MXFP4 checkpoint 的组合，那里 custom all-reduce 默认关闭；本配方已经默认开启。