config(minimaxm3): update GB200 Dynamo vLLM AgentX recipes / 更新 GB200 Dynamo vLLM AgentX 配方

[RohitNagraj]

Summary

• Update the MiniMax-M3 GB200 aggregate matrix with TP4 SimpleCPU recipe variants that use separate scheduler caps.
• Replace the disaggregated matrix with DEP4 prefill and TP2 decode pools for the submitted concurrency points.
• Add five multi-node recipes and pin each recipe to its declared vLLM, Dynamo, and Mooncake runtime.
• Keep AgentX evaluation on real target verification while using the EAGLE3 draft model at its shipped precision.

Validation

• Parsed the changed YAML files successfully.
• Generated the two exact config keys and the affected GB200 family: 9 configurations total.
• Ran srtctl dry-run -f successfully for all 5 new recipes with the repository-pinned srt-slurm version.
• Passed 370 focused matrix, changelog-gate, and SRT cluster tests.
• Verified the EAGLE3 draft artifacts are BF16 and the vLLM configuration does not request draft-weight quantization.
• Passed git diff --check.

Review notes

• This is exclusively a multi-node configuration change, so no separate single-node serving recipe is applicable.
• The disaggregated matrix contains four submitted points and therefore needs core-maintainer review under the current Pareto-coverage guidance.
• No sweep-triggering label has been added.

AI assistance

OpenAI GPT-5 (Codex; exact runtime version not exposed) prepared the configuration translation, validation, and pull request text. The submitter reviewed the resulting files and metadata before publication.

简体中文

概要

• 更新 MiniMax-M3 GB200 聚合部署矩阵，新增使用不同调度上限的 TP4 SimpleCPU 配方变体。
• 将分离部署矩阵调整为 DEP4 prefill 与 TP2 decode 池，并覆盖本次提交的并发点。
• 新增 5 个多节点配方，并将每个配方固定到其声明的 vLLM、Dynamo 与 Mooncake 运行时。
• AgentX 评测继续使用目标模型进行真实验证，同时按 EAGLE3 草稿模型随附精度加载权重。

验证

• 已成功解析本次修改的 YAML 文件。
• 已生成两个精确配置键及受影响的 GB200 配置族，共 9 个配置。
• 使用仓库固定的 srt-slurm 版本，对 5 个新增配方执行 srtctl dry-run -f，全部通过。
• 通过 370 项针对矩阵、变更日志门禁和 SRT 集群配置的测试。
• 已确认 EAGLE3 草稿模型工件为 BF16，且 vLLM 配置未请求对草稿权重进行量化。
• git diff --check 通过。

审阅说明

• 本次仅修改多节点配置，因此不适用单独的单节点服务配方。
• 分离部署矩阵包含 4 个提交点，依据当前 Pareto 覆盖指引，需要核心维护者审阅。
• 未添加任何触发 sweep 的标签。

AI 协助

OpenAI GPT-5（Codex；运行环境未公开更精确的版本标识）协助完成配置整理、验证与拉取请求文本。提交者会在发布前审阅最终文件和元数据。

[github-actions]

Thanks for the contribution!

• Review: If this PR changes files owned by someone other than a repository admin or @SemiAnalysisAI/core, ask one eligible CODEOWNER to complete the latest PR_REVIEW_CHECKLIST.md before contacting a core maintainer on Slack. Follow the template exactly, including As a PR reviewer and CODEOWNER, I have reviewed this and have, so sign-off verification triggers.
• PR verification: Sweeps only run on labeled PRs. Add full-sweep-fail-fast (strongly recommended); use full-sweep-enabled only when matrix jobs should continue after a failure.
• After merging: PR authors must ensure all GitHub Actions jobs pass. Transient failures often pass on rerun; see how to rerun failed jobs.

中文

感谢你的贡献！

• **审阅：**如果 PR 修改的文件归属于仓库管理员及 @SemiAnalysisAI/core 之外的 CODEOWNER，请先联系一位有资格的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，再通过 Slack 联系核心维护者。必须严格遵循模板，并保留 As a PR reviewer and CODEOWNER, I have reviewed this and have，才能触发签核验证。
• **PR 验证：**扫描仅在带有标签的 PR 上运行。强烈建议添加 full-sweep-fail-fast；仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。
• **合并后：**PR 作者必须确保所有 GitHub Actions 任务通过。临时性失败通常可以通过重新运行恢复；参见重新运行失败任务的说明。

[claude]

Beyond the inline pr-link finding, I also checked whether the decode max-num-seqs cap being far below each recipe's benchmarked concurrency was a bug (e.g. disagg-1p5d-dep4-tp2-mns4-c50.yaml: max-num-seqs 4 x 5 workers = 20 cap vs c50). Ruled out as unintentional: the same undersized-cap-vs-concurrency pattern appears consistently across all three new disagg recipes, and each filename literally encodes the chosen mns (max-num-seqs) value alongside its concurrency point(s) (mns1-c15, mns2-c20-c30, mns4-c50), which reads as a deliberate queuing/backpressure scheduler-cap design rather than a mismatch.

Extended reasoning...

This is a config/recipe-only PR (5 new srt-slurm YAML recipes plus matching master-config and changelog entries for the MiniMax-M3 GB200 FP4 benchmark matrix); it touches no application code and no auth/crypto/permissions surface. A confirmed inline finding already flags that the new perf-changelog.yaml entry's pr-link: TBD is not in the tool's accepted placeholder set (only "XXX" and its full URL form), so it will fail validate_added_pr_link in infx/workflows/validate_perf_changelog.py and block the changelog gate rather than pass as the PR description claims. I additionally checked the decode scheduler-cap-vs-concurrency pattern across all three new disagg recipes and found it consistent and filename-encoded, so I ruled that out as an intentional design choice rather than a separate bug. Given the confirmed blocking finding, this should not be approved.

[claude]

🔴 The new changelog entry's pr-link: TBD will make the merge-automation bot reject this PR's own changelog entry instead of canonicalizing it, unlike every other entry in the file. validate_added_pr_link in infx/workflows/validate_perf_changelog.py:134-143 only accepts a real https://github.com/SemiAnalysisAI/InferenceX/pull/<n> link or a placeholder in PR_LINK_PLACEHOLDERS = {"XXX", ".../pull/XXX"}; TBD matches neither and raises ChangelogValidationError("new PR entry must use ... or an XXX placeholder; found 'TBD'") when compare_entries runs (invoked from infx/workflows/prepare_perf_changelog_merge.py at merge time). Fix: use the documented XXX (or full pull/XXX URL) placeholder per CONTRIBUTING.md's example, not a free-form string like TBD.

Why this was flagged

The new perf-changelog.yaml entry at line 8901 sets pr-link: TBD. When the repo's merge-prep tooling (infx/workflows/prepare_perf_changelog_merge.py, via compare_entries/validate_added_pr_link) processes this PR to canonicalize appended entries' pr-link into the real PR URL, it calls validate_added_pr_link('TBD', pr_number). Since 'TBD' is not in PR_LINK_PLACEHOLDERS = {'XXX', 'https://github.com/SemiAnalysisAI/InferenceX/pull/XXX'} and does not match CANONICAL_PR_LINK, this raises ChangelogValidationError. CONTRIBUTING.md:190 documents the required placeholder as the literal 'XXX' form; grep across perf-changelog.yaml and the changelog_gate_tests fixtures shows every other placeholder entry uses 'XXX', never 'TBD'. This blocks the automated pr-link canonicalization/merge step for this PR, requiring a manual fix, whereas on the base branch every appended entry's placeholder is recognized and processed automatically.

Verification: normal. The new changelog entry at perf-changelog.yaml:8901 sets pr-link: TBD. The supported merge path utils/merge_with_reuse.sh:169-170 runs python3 -m infx.workflows.prepare_perf_changelog_merge canonicalize, which calls canonicalize_appended_links (prepare_perf_changelog_merge.py:77), which calls compare_entries(base, head, pr_number) at line 86. compare_entries iterates additions and…