[Klaud Cold] Update minimaxm3-fp4-b200-vllm-agentic-mtp vLLM image to nightly-8a728663c1c3eeace834a95f5654fa653cc1998c

[functionstackx]

Summary

Update vLLM image for the MiniMax-M3 NVFP4 B200 AgentX MTP recipe from vllm/vllm-openai:nightly-1dc464d42681d22f38caf1fdc1eb632dc4421c45 (2026-08-30 nightly) to vllm/vllm-openai:nightly-d9105ea8001e0a6d77a96327d17515bb5791fb36 (2026-09-07 nightly).

• Tag verified on Docker Hub: last pushed 2026-09-07T06:16:01Z, digest sha256:254eebf919e8b7b0d530d97fccc606c36f380ff6724d64bc190951bec1aee838
• Tag commit: vllm-project/vllm@d9105ea8 ("[Qwen3.8-Flash-Next] Remove torch.compile for NVIDIA implementation")
• Recipe script, search space, and engine flags are unchanged; image-only bump.

Recipes touched: minimaxm3-fp4-b200-vllm-agentic-mtp

Test plan

• full-sweep-enabled sweep passes on cluster:b200-nscale.

🤖 Generated with Claude Code

---

Note

Low Risk
Config-only Docker tag change with no application code; risk is limited to B200 AgentX sweep behavior if the chosen nightly regresses elsewhere.

Overview
Re-pins the minimaxm3-fp4-b200-vllm-agentic-mtp recipe’s vLLM container from the 2026-08-30 nightly (1dc464d…) to nightly-8a728663c1c3eeace834a95f5654fa653cc1998c (2026-09-04). Script, search space, and engine flags are unchanged—image only.

The 2026-09-07 nightly was skipped after B200 engine init failed during CUDA-graph profiling: post–Flash-Attention sync nightlies (2026-09-05+) hit an EAGLE3 FA4 CuTe + fp8 KV descale stride error on Blackwell. 2026-09-04 is the newest nightly still on the prior vllm-flash-attn pin, aligned with the MI355X MiniMax-M3 AgentX image.

Adds a matching perf-changelog.yaml entry for the agentic-coding scenario (PR #2860).

^{Reviewed by Cursor Bugbot for commit 50e127a. Bugbot is set up for automated code reviews on this repo. Configure here.}

Update: re-pinned to the 2026-09-04 nightly

The 2026-09-07 nightly failed at engine init on B200 (run 34168437161, TP4 vllm-simple c36): during CUDA-graph memory profiling the EAGLE3 draft's FLASH_ATTN backend takes the FA4 CuTe path on Blackwell and, with the fp8 KV cache, its descale tensors fail to_cute_tensor with RuntimeError: Expected strides[leading_dim] == 1, but got 0. Cause: vllm-project/vllm@4ee259551 ("Sync FA with upstream", #54819, 2026-09-05) moved vllm-flash-attn from 06bdd47c to 506341a1; every nightly from 2026-09-05 on carries it, and no fix has landed on vllm main as of 2026-09-08T03:00Z. Re-pinned to vllm/vllm-openai:nightly-8a728663c1c3eeace834a95f5654fa653cc1998c (2026-09-04, pushed 2026-09-04T06:18:14Z, digest sha256:f5df5cc3302b5f404848c4eca88d7bf7ed5226e151c056da22816d7734644d67): the newest nightly still on FA 06bdd47c, five days newer than the recipe's previous pin, and the same vllm commit the MI355X MiniMax-M3 vLLM AgentX recipe already runs. Recipe otherwise unchanged. The Hopper (#2874/#2875) and ROCm (#2872/#2873) MiniMax siblings do not take the FA4 CuTe path and stay on the 09-07 nightly.

[github-actions]

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 As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

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 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

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]

Beyond the inline finding, I checked utils/validate_perf_changelog.py: it only accepts the literal placeholders XXX or https://github.com/SemiAnalysisAI/InferenceX/pull/XXX (or a real .../pull/\d+ link) as pr-link. The value used here, PRLINK_PLACEHOLDER, matches none of those, so this entry should actually fail changelog validation as-is, not just be a style nit.

Extended reasoning...

The inline finding already flags that pr-link: PRLINK_PLACEHOLDER is not a real PR URL. I additionally verified the enforcement mechanism: utils/validate_perf_changelog.py defines PR_LINK_PLACEHOLDERS = {"XXX", "https://github.com/SemiAnalysisAI/InferenceX/pull/XXX"} and a CANONICAL_PR_LINK regex requiring https://github.com/SemiAnalysisAI/InferenceX/pull/\d+. The string PRLINK_PLACEHOLDER matches neither the accepted placeholder set nor the canonical link pattern, so this is not merely a cosmetic gap that a script would tolerate — it should trip the validator's checks before merge. This is concrete information beyond the already-posted inline comment, so it's worth a short defer note rather than staying silent.

[claude]

🔴 The new changelog entry's pr-link is the literal placeholder "PRLINK_PLACEHOLDER", not a real PR URL, unlike every other entry in the file. Fix: replace with the actual GitHub PR URL for this change before merge, consistent with all prior entries (e.g. #2792 on the line above).

Extended reasoning...

perf-changelog.yaml is append-only and consumed by tooling/CODEOWNER review to trace each performance-affecting change back to its PR; every prior entry has a resolvable github.com/.../pull/N link. If this entry merges with PRLINK_PLACEHOLDER still in place, anyone or any script auditing the changelog for provenance of the minimaxm3-fp4-b200-vllm-agentic-mtp image bump gets a dead placeholder instead of the source PR, breaking traceability and the CODEOWNER sign-off checklist requirement to verify image provenance via the linked PR.

Verification: normal. The new entry at perf-changelog.yaml:6931 has pr-link: PRLINK_PLACEHOLDER, whereas the entry above uses canonical https://github.com/SemiAnalysisAI/InferenceX/pull/2792. This is a real, unintended defect: (1) the generating workflow (.claude/commands/nuke.md:103,124-127) writes PRLINK_PLACEHOLDER only as a temporary marker and mandates replacing it with the real PR URL via… | normal.…

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34168437161
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=34168437161

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34182370413
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=34182370413