Skip to content

[VideoGenX] add MiniMax-H3 MI355X SGLang video sweep / [VideoGenX] 新增 MiniMax-H3 MI355X SGLang 视频扫描 - #4809

Merged
edwingao28 merged 125 commits into
mainfrom
feat/h3-mi355x
Nov 6, 2026
Merged

edwingao28 merged 125 commits into
mainfrom
feat/h3-mi355x

Conversation

@edwingao28

Copy link
Copy Markdown
Collaborator

Summary

Add a MiniMax-H3 C1 video sweep on four MI355X GPUs with SGLang Ulysses4 on the h3-v2.yaml workload of the H100 and B200 rows. The image is the SGLang v0.5.23 ROCm 7.2.4 MI35x build. On ROCm the server names each GPU by its board serial, matching AMD power samples; NVIDIA keeps CUDA UUIDs.

Testing: CPU checks pass; the matrix plans one MI355X point. No GPU run yet: a trusted e2e-tests.yml dispatch follows once H3 weights are staged on the MI355X shared cache.

Stacked on #2916.

AI model disclosure
  • Model/version: claude-opus-5-5.
  • Role: implementation and PR text.
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

新增 MiniMax-H3 在 4 张 MI355X 上的 C1 视频扫描,使用 SGLang Ulysses4,工作负载与 H100、B200 两行相同,都是 h3-v2.yaml。镜像为 SGLang v0.5.23 的 ROCm 7.2.4 MI35x 版本。在 ROCm 上,服务端用板卡序列号标识每张 GPU,与 AMD 功耗采样的标识方式一致;NVIDIA 仍使用 CUDA UUID。

测试: CPU 检查通过,矩阵规划出一个 MI355X 数据点。尚未跑 GPU:等 H3 权重放到 MI355X 共享缓存后,再用 trusted 模式手动触发 e2e-tests.yml。

基于 #2916。

AI 模型:claude-opus-5-5,负责实现和 PR 文案。

@edwingao28
edwingao28 requested a review from zijiecode as a code owner November 5, 2026 06:56
@github-actions

github-actions Bot commented Nov 5, 2026 •

Copy link
Copy Markdown
Contributor

inferencex-e2e/infx diff: +3 -1

Generally most InferenceX submissions don't need to add changes to the infx package. On the off chance they do, generally the diff is small. Please prompt your agent to clean your PR to keep your diff to only necessary & minimal changes too. We have an increased code quality bar for changes to the infx package.

中文

大多数 InferenceX 提交通常不需要修改 infx 包。即使确实需要,改动通常也很小。请让你的 agent 清理 PR,使改动只包含必要且最小的内容。infx 包的改动有更高的代码质量要求。

@claude

claude Bot commented Nov 5, 2026 •

Copy link
Copy Markdown
Contributor

Claude finished @edwingao28's task in 49s —— View job


Review: no blocking issues found

I read the full diff against origin/feat/h3-cross-hardware and found nothing that blocks merge.

  • Config key: the only new key is h3-bf16-mi355x-sglang-video. The new perf-changelog.yaml entry covers it and is appended at the end of the file.
  • Image: public lmsysorg/sglang ROCm image, pinned by digest.
  • Quick-reduce setting: ROCM_QUICK_REDUCE_QUANTIZATION: NONE is set explicitly, as MODELS.md requires for SGLang on ROCm. The server passes it on to the engine through os.environ.copy().
  • Shared HF cache rule: the new MiniMaxAI/MiniMax-H3 match applies only to the mi355x-amds cluster, so the existing H100 and B200 points are unaffected.
  • NVIDIA GPU naming: gpu_devices() uses the ROCm path only when /dev/kfd exists. NVIDIA hosts keep using cuda_devices(), and the existing tests that monkeypatch cuda_devices still apply.

Not verified here:

  • The sandbox blocked uv run, so I didn't run infx.matrix.changed or pytest.
  • The ROCm board-serial naming needs the first GPU run to confirm that each GPU's /sys/bus/pci/devices/<bus>/serial_number matches the serial_number label on the AMD exporter's power samples. If they don't match, the native power binding will reject the run.

@claude claude Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nothing blocking. The comments below are optional suggestions. There is no need to push a fix for them before merging.

Base automatically changed from feat/h3-cross-hardware to main November 5, 2026 21:45
[skip-sweep]

# Conflicts:
#	inferencex-e2e/docs/MODELS.md
#	inferencex-e2e/docs/video-generation.md
#	inferencex-e2e/perf-changelog.yaml
#	video-generation/README.md
#	video-generation/README_zh.md
#	video-generation/evaluator/srt_selected.py
#	video-generation/tests/test_srt_selected.py
@edwingao28

Copy link
Copy Markdown
Collaborator Author

/use 42118905701

@github-actions

github-actions Bot commented Nov 6, 2026 •

Copy link
Copy Markdown
Contributor

@edwingao28 staged run 42118905701.

This run has no staged performance curves.

This run remains available across future /stage-results requests. Staging the same run ID again updates its staged data. Staging workflow

@edwingao28
edwingao28 merged commit e738157 into main Nov 6, 2026
31 checks passed
@edwingao28
edwingao28 deleted the feat/h3-mi355x branch November 6, 2026 00:10
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

Status: Done

Development

Successfully merging this pull request may close these issues.

1 participant