[Klaud Cold] Update minimaxm3-fp8-mi325x-vllm-agentic-mtp vLLM ROCm image to nightly-d9105ea8001e0a6d77a96327d17515bb5791fb36 - #2872
Conversation
…mage to nightly-d9105ea8001e0a6d77a96327d17515bb5791fb36 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
91a463a to
6532842
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}
| - agentic-coding | ||
| description: | ||
| - "Update vLLM ROCm image from vllm/vllm-openai-rocm:v0.27.1 (v0.27.1 release) to vllm/vllm-openai-rocm:nightly-d9105ea8001e0a6d77a96327d17515bb5791fb36 (2026-09-07 upstream ROCm nightly, digest sha256:74d4a95f3ae672ecddf9acb7296917def82d9eca51687fa2862ae72b03ff1907, tag commit vllm-project/vllm@d9105ea8; Docker Hub last pushed 2026-09-07T05:26:48Z). benchmarks/single_node/agentic/minimaxm3_fp8_mi325x_mtp.sh is unchanged: TRITON_ATTN attention, fp8 KV, block-size 128, EAGLE3 speculative decoding with the Inferact MiniMax-M3 EAGLE3-GQA draft and the committed golden synthetic acceptance length, minimax_m3 tool-call and reasoning parsers; the search space is unchanged. The upstream commit-pinned ROCm nightly is the same vllm commit the B200 MiniMax-M3 AgentX recipe moved to in #2860. Note that vllm-openai-rocm commit-nightly tags have expired from Docker Hub within days in the past; node squash caches keep merged configs running, but a re-pin to a durable tag may be needed later." | ||
| pr-link: PRLINK_PLACEHOLDER |
There was a problem hiding this comment.
🔴 The new entry's pr-link: PRLINK_PLACEHOLDER is not in validate_perf_changelog.py's PR_LINK_PLACEHOLDERS set ({"XXX", the canonical PR URL}). When this full-sweep-enabled PR is merged via utils/merge_with_reuse.sh, prepare_perf_changelog_merge.py's canonicalize step raises ChangelogValidationError("appended entry N has unexpected pr-link 'PRLINK_PLACEHOLDER'"), and set -euo pipefail aborts the whole reuse-merge script, blocking the squash-merge until someone manually edits the changelog. Fix: the appended entry's pr-link must be exactly XXX (the only pre-PR placeholder the merge tooling accepts) or the real PR URL — note TBD (the docs-suggested placeholder) would also fail this same guard.
Extended reasoning...
utils/merge_with_reuse.sh runs prepare_perf_changelog_merge.py canonicalize --pr-number PR --base-ref origin/main after merging main into the PR branch (used to land full-sweep-enabled PRs with sweep-run reuse, matching this PR's label). canonicalize_appended_links() (prepare_perf_changelog_merge.py) iterates the newly appended entries, computes link = entry['pr-link'], and does if link != expected_link: if link not in PR_LINK_PLACEHOLDERS: raise ChangelogValidationError(...). PR_LINK_PLACEHOLDERS (validate_perf_changelog.py) is only {'XXX', 'https://.../pull/XXX'}. 'PRLINK_PLACEHOLDER' matches neither, so the raise fires, the python process exits non-zero, and the bash script (set -euo pipefail) dies before pushing/squash-merging — the PR cannot be merged through this tool until the placeholder is corrected to 'XXX' or a real link.
Verification: normal. The appended entry uses pr-link: PRLINK_PLACEHOLDER (perf-changelog.yaml, final diff line). This value is not in PR_LINK_PLACEHOLDERS = {"XXX", "https://github.com/SemiAnalysisAI/InferenceX/pull/XXX"} (utils/validate_perf_changelog.py:24-27) and is not the canonical expected URL. Two reachable failures result: (1) The reuse-merge path — utils/merge_with_reuse.sh runs… | Severity:…
…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=34174118730 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34175043014 |
1 participant
Summary
Update vLLM ROCm image for the MI325X 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_mi325x_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-mi325x-vllm-agentic-mtpTest plan
🤖 Generated with Claude Code
Note
Low Risk
Scoped benchmark/config and vLLM serve flags only; no application auth or data-path changes, with a targeted fix for a known init failure on the new image.
Overview
Updates the minimaxm3-fp8-mi325x-vllm-agentic-mtp recipe to vllm/vllm-openai-rocm:nightly-d9105ea8001e0a6d77a96327d17515bb5791fb36 (2026-09-07 ROCm nightly, same vLLM commit as the B200 AgentX move in #2860), replacing v0.27.1.
After the first sweep failed at engine init (run 34174124043), minimaxm3_fp8_mi325x_mtp.sh now passes
--compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY"}'so serve starts when the nightly skips torch-compile forMiniMaxM3SparseForConditionalGenerationandVLLM_USE_BREAKABLE_CUDAGRAPH=0blocks default FULL_AND_PIECEWISE graphs—aligned with the MI355X MiniMax-M3 nightly approach (#2825). Agentic search space and other serve flags are unchanged.perf-changelog.yaml records the image bump and compilation-config fix for the agentic-coding scenario.
Reviewed by Cursor Bugbot for commit 5f4e8e6. 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.