[AgentX] add Muse Spark Open NVFP4 on B300 / [AgentX] 新增 B300 Muse Spark Open NVFP4 支持

[edwingao28]

Summary

Add TP4/EP4 Muse Spark Open NVFP4 AgentX on B300 with DFlash2 and the upstream recipe.

Testing: Sweep, a23546bea, attempt 1: ten canonical throughput points and GSM8K 97.65% passed; raw/power verified. Retains one request error and 383 cutoff cancellations, including C64/128 acknowledgement timeouts.

Blocker: Main sync/reuse, cookbook coverage, CODEOWNER/Core review, and an admin exception for three P90 Pareto points.

AI model disclosure

• Model/version: Exact model identifiers were not exposed and could not be verified.
• Role: AI-assisted implementation and review, including delegated agents.

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 B300 TP4/EP4 Muse Spark Open NVFP4 AgentX，使用 DFlash2 和上方上游配方。

验证：a23546bea 的 sweep 第1次运行：10个标准吞吐点及 GSM8K 97.65% 通过，原始产物和功耗已核验。保留1次请求错误及383次截止取消，其中C64/128取消确认超时。

**待办：**同步 main/复用、cookbook 覆盖、CODEOWNER/Core 审查，以及三个 P90 Pareto 前沿点的管理员例外。

改动类型：新功能、配置变更。模型/版本：运行环境未提供可核实的精确标识。AI 用于实现和审查，包括委派给其他 agent 的工作。

[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 run only when the PR appends an inferencex-e2e/perf-changelog.yaml entry and carries exactly one primary label: full-sweep-fail-fast (strongly recommended; canary plus per-matrix fail-fast), full-sweep-enabled (canary; matrix jobs continue after a failure), or non-canary-full-sweep-enabled (no canary or fail-fast). The modifiers all-evals, evals-only, and agentx-fast require a primary label. On fork PRs, a maintainer applies the label. See sweep labels and reuse.
• 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：**只有当 PR 在 inferencex-e2e/perf-changelog.yaml 末尾追加了条目，并且恰好带有一个主标签时，才会运行扫描：full-sweep-fail-fast（强烈推荐；canary 加逐矩阵 fail-fast）、full-sweep-enabled（有 canary；矩阵任务在失败后继续运行）或 non-canary-full-sweep-enabled（无 canary，也无 fail-fast）。修饰标签 all-evals、evals-only 和 agentx-fast 必须与主标签一起使用。fork PR 的标签由维护者添加。参见扫描标签与复用。
• **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。

[github-actions]

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=38825818513

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=38825818513

[nvpohanh]

[by Claude Code] The SGLang cookbook integration of this PR's AgentX config (musespark-fp4-b300-sglang-agentic-dflash) will happen in sgl-project/sglang#44654, in the Agentic Long-Context Deployment section of the Muse Spark Open cookbook page.

This PR has no authorized /use <run_id> comment yet, so the cookbook shows the recipe at this PR's head (a23546bea) as "in progress". It switches to the measured run at the next cookbook sync after a run is pinned or this PR merges.

[functionstackx]

/use 38825818513

[github-actions]

@functionstackx staged run 38825818513.

• Muse-Spark-Open · agentic-traces · FP4 · b300_sglang · 2026-10-15

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

[functionstackx]

/reuse-sweep-run 38825818513

[claude]

Claude finished @edwingao28's task in 27s —— View job

---

LGTM - no blocking issues found

What I checked

• I checked zip_override_tp4 in b300-fp4-mtp/agentic.yaml. Its list lengths match the 10-point conc-list in nvidia-master.yaml, and max-running-requests and cuda-graph-max-bs-decode are 2× concurrency, the same as the B200 and GB200 Muse Spark recipes.
• In nvidia-master.yaml, the image is a public Docker Hub reference pinned to the same digest as the B200 and GB200 configs. The musespark prefix is documented in docs/MODELS.md.
• The runners.yaml model entry points to the existing data-models root on b300-dsxe.
• The perf-changelog.yaml entry is appended at the end of the file and covers the new config key.
• I couldn't run infx.matrix.changed because the command needed approval in this environment. The changelog entry covers the one new key either way.

[claude]

I reviewed this PR and didn't find any bugs. It adds a new benchmark config to CODEOWNER-owned NVIDIA paths, so a code-owner review is still worthwhile.

What was reviewed:

• The new single-node agentic.yaml recipe: TP4/EP4 agg role on 4 GPUs, DFLASH draft settings, the health-check and AIPerf timeout env, and zip_override_tp4. All three swept lists have 10 entries, so they line up.
• The nvidia-master.yaml entry: srt-recipe-dir and srt-recipe point to the new file, and the conc-list matches the recipe's CONC list.
• The runners.yaml models mapping and the perf-changelog.yaml block. The block is appended at the end of the file and covers the new config key.

Extended reasoning...

The change adds one new srt-slurm recipe, one nvidia-master.yaml entry, one runners.yaml model mapping and one perf-changelog.yaml entry, 89 added lines in total. It is benchmark configuration only and touches no auth, crypto or data-exposure surface. The recipe hard-codes no synthetic acceptance length, and the new entry follows existing master and recipe patterns. nvidia-master.yaml and the b300-* recipe trees are covered by CODEOWNERS entries for NVIDIA reviewers and core, so code-owner sign-off is expected.

[nvpohanh]

[by Claude Code] sgl-project/sglang#44654 has been merged. This PR's AgentX config (musespark-fp4-b300-sglang-agentic-dflash) is now in the Agentic Long-Context Deployment section of the SGLang Muse Spark Open cookbook: https://docs.sglang.io/cookbook/autoregressive/Meta/MuseSparkOpen#agentic-long-context-deployment

This PR has no authorized /use <run_id> comment yet, so the cookbook shows the recipe at this PR's head (a23546bea) as "in progress". It switches to the measured run at the next cookbook sync after a run is pinned or this PR merges.