[Klaud Cold] Update minimaxm3-fp8-mi300x-vllm-agentic-mtp vLLM ROCm image to nightly-d9105ea8001e0a6d77a96327d17515bb5791fb36 - #2873
Conversation
…mage to nightly-d9105ea8001e0a6d77a96327d17515bb5791fb36 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
4feda6b to
4ef1fc1
Compare
|
Thanks for the contribution! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase For PR verification, add the PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. See GitHub's docs on re-running failed jobs 感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 如需进行 PR 验证，请为此 PR 添加 PR 作者有责任确保合并后所有 GitHub Action 任务完全通过。 很多时候失败只是偶发抖动（flake），重新运行失败的任务即可解决。参见 GitHub 关于重新运行失败任务的文档 |
2 similar comments
|
Thanks for the contribution! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase For PR verification, add the PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. See GitHub's docs on re-running failed jobs 感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 如需进行 PR 验证，请为此 PR 添加 PR 作者有责任确保合并后所有 GitHub Action 任务完全通过。 很多时候失败只是偶发抖动（flake），重新运行失败的任务即可解决。参见 GitHub 关于重新运行失败任务的文档 |
|
Thanks for the contribution! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase For PR verification, add the PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. See GitHub's docs on re-running failed jobs 感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 如需进行 PR 验证，请为此 PR 添加 PR 作者有责任确保合并后所有 GitHub Action 任务完全通过。 很多时候失败只是偶发抖动（flake），重新运行失败的任务即可解决。参见 GitHub 关于重新运行失败任务的文档 |
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Looks good, straightforward image-tag bump with a matching changelog entry.
What was reviewed: the single-line image bump in configs/amd-master.yaml for the minimaxm3-fp8-mi300x-vllm-agentic-mtp recipe (single-node vLLM, so no model.container mirror is needed); the appended perf-changelog.yaml entry, confirmed it only adds bytes at the physical tail without touching prior entries; and the pr-link: PRLINK_PLACEHOLDER value against utils/validate_perf_changelog.py — the PR-time gate (run-sweep.yml) never calls compare_entries/validate_added_pr_link, so it won't block this PR, but the merge-time script (merge_with_reuse.sh → prepare_perf_changelog_merge.py) does call compare_entries, so the author will need to replace the placeholder with a real/XXX value before merge succeeds.
Extended reasoning...
Overview
The diff is two files: a one-line image tag bump in configs/amd-master.yaml for the minimaxm3-fp8-mi300x-vllm-agentic-mtp recipe (pinned release v0.27.1 → commit-pinned ROCm nightly nightly-d9105ea8...), and a single appended entry in perf-changelog.yaml describing that bump. No benchmark scripts, gating logic, or other recipes are touched.
Security risks
None. This is a config value change (a Docker image tag string) and a documentation/changelog append; no code execution paths, auth, or data handling are affected.
Level of scrutiny
Low. This matches the "simple/mechanical config change" profile the approval guidelines describe — a version-bump-style change with a corresponding changelog entry, no CODEOWNERS-sensitive paths, and no design decision for a human to weigh in on beyond the already-documented risk (ROCm nightly tags expiring from Docker Hub) that the changelog entry itself explicitly calls out.
Other factors
I traced the pr-link: PRLINK_PLACEHOLDER value (flagged by earlier bug-hunter passes and ruled out) through utils/validate_perf_changelog.py and utils/prepare_perf_changelog_merge.py: the CLI path invoked by run-sweep.yml (validate_matrix_compatible_change → validate_generated_config) never calls compare_entries/validate_added_pr_link, so this PR's CI will not fail on the placeholder. That validation only runs later, at merge time, via merge_with_reuse.sh → prepare_perf_changelog_merge.py, which does call compare_entries and would reject PRLINK_PLACEHOLDER (it's neither XXX, .../pull/XXX, nor the real PR URL) — so the author will simply need to fix it before merge, rather than it silently landing in perf-changelog.yaml. This makes it a non-blocking, self-correcting issue rather than a bug worth delaying approval for.
This review covers commit 4feda6b, which is no longer the latest commit on this pull request; later commits are not covered by it.
…with breakable graphs off Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34174124043 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34175049433 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34175049433 |
|
/reuse-sweep-run |
# Conflicts: # perf-changelog.yaml
1 participant
Summary
Update vLLM ROCm image for the MI300X MiniMax-M3 MXFP8 AgentX EAGLE3 recipe from
vllm/vllm-openai-rocm:v0.27.1(v0.27.1 release) tovllm/vllm-openai-rocm:nightly-d9105ea8001e0a6d77a96327d17515bb5791fb36(2026-09-07 upstream ROCm nightly).sha256:74d4a95f3ae672ecddf9acb7296917def82d9eca51687fa2862ae72b03ff1907; tag commit vllm-project/vllm@d9105ea8, the same commit the B200 MiniMax-M3 AgentX recipe moved to in [Klaud Cold] Update minimaxm3-fp4-b200-vllm-agentic-mtp vLLM image to nightly-8a728663c1c3eeace834a95f5654fa653cc1998c #2860.benchmarks/single_node/agentic/minimaxm3_fp8_mi300x_mtp.shis unchanged and pins nothing image-specific: TRITON_ATTN attention, fp8 KV, block-size 128, EAGLE3 with the Inferact EAGLE3-GQA draft and committed golden synthetic AL, minimax_m3 parsers. Search space unchanged.vllm-openai-rocmcommit-nightly tags have expired from Docker Hub within days before. Node squash caches keep a merged config running, but a re-pin to a durable tag may be needed later.Recipes touched:
minimaxm3-fp8-mi300x-vllm-agentic-mtpTest plan
🤖 Generated with Claude Code
Note
Low Risk
Benchmark-only recipe image pin and vLLM serve flag for MI300X AgentX; no application auth or data-path changes, though nightly image availability may affect future sweeps.
Overview
Updates
minimaxm3-fp8-mi300x-vllm-agentic-mtpfromvllm/vllm-openai-rocm:v0.27.1to commit-pinned ROCm nightlynightly-d9105ea8001e0a6d77a96327d17515bb5791fb36(same vLLM commit as the B200 MiniMax-M3 AgentX bump in #2860). Agentic search space and EAGLE3/MTP settings are unchanged.After the image bump, engine init failed on the nightly because
MiniMaxM3SparseForConditionalGenerationis not torch-compiled and withVLLM_USE_BREAKABLE_CUDAGRAPH=0the defaultFULL_AND_PIECEWISEmode errors on missing piecewise CUDA graphs (run 34174124043).minimaxm3_fp8_mi300x_mtp.shnow passes--compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY"}', aligned with the MI355X MiniMax-M3 nightly path (#2825).perf-changelog.yamldocuments the image change, the cudagraph workaround, and the ROCm nightly tag expiry caveat.Reviewed by Cursor Bugbot for commit 1ba8e60. Bugbot is set up for automated code reviews on this repo. Configure here.
Update
First sweep (run 34174124043 on the MI300X sibling) failed at engine init: the nightly does not torch-compile
MiniMaxM3SparseForConditionalGeneration, and withVLLM_USE_BREAKABLE_CUDAGRAPH=0the defaultFULL_AND_PIECEWISEgraph mode aborts ("piecewise CUDA graphs unavailable ... Set VLLM_USE_BREAKABLE_CUDAGRAPH=1 or cudagraph_mode=NONE/FULL"). Added--compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY"}'to the serve command, matching the MI355X MiniMax-M3 sibling on its nightly (#2825). Script is otherwise unchanged.