[AgentX] add Muse Spark Open NVFP4 on H200 / [AgentX] 新增 H200 Muse Spark Open NVFP4 支持

[edwingao28]

Description

Add TP8/EP8 Muse Spark Open NVFP4 on H200 with Marlin W4A16 and DFlash2.

Testing: Sweep on adc40a25d has 5/9 throughput points accepted (C1/C2/C4/C24/C48); C64 GSM8K passed 97.57% in attempt 1. C24/C48 retain 79 acknowledged drain cancellations and high latency. Strict power is invalid from clock-sync failures; energy metrics are withheld.

Blocker: Four throughput points, cookbook, main sync/reuse and human review.

中文

新增 H200 TP8/EP8 Muse Spark Open NVFP4，使用 Marlin W4A16 和 DFlash2。

测试：adc40a25d 的扫描已验收 5/9 个吞吐点（C1/C2/C4/C24/C48）；attempt 1 的 C64 GSM8K 通过率为 97.57%。C24/C48 保留 79 个已确认的收尾取消和高延迟记录。时钟同步失败使严格功耗无效，能耗指标不予发布。扫描链接见上方。

**阻塞：**剩余四个吞吐点、上游 cookbook、主分支同步/复用及人工审查。

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.

[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=38826180170

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=38826180170

[Oseltamivir]

/use 38826180170

[github-actions]

@Oseltamivir staged run 38826180170.

• Muse-Spark-Open · agentic-traces · FP4 · h200_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

[github-actions]

inferencex-e2e/infx diff: +14 -1

Generally most InferenceX submissions don't need to add changes to the infx package. On the off chance they do, generally the diff is small. Please prompt your agent to clean your PR to keep your diff to only necessary & minimal changes too. We have an increased code quality bar for changes to the infx package.

中文

大多数 InferenceX 提交通常不需要修改 infx 包。即使确实需要，改动通常也很小。请让你的 agent 清理 PR，使改动只包含必要且最小的内容。infx 包的改动有更高的代码质量要求。

[claude]

I reviewed this PR and found no bugs. A human should still look before merge: it adds a new per-workload staging flag to the shared SRT launch policy, and the description says four of the nine throughput points are still unaccepted.

What was reviewed:

• Override.single_node_staged in drivers/srt/models.py: it switches single_node_model_path to the staged checkpoint only for the Muse Spark glob on h200-dgxc. Hub-mode behavior for every other model is unchanged.
• The new dflash branch in spec_parameters: draft length is speculative-num-draft-tokens - 1, and a bool or non-int width gives None, which the new tests check is rejected. Also the matching dflash entry in the single-node method allowlist.
• The recipe's zip sweep: the 9 CONC values line up with the 9 max-running-requests and cuda-graph-max-bs-decode values. The perf-changelog block is a single entry appended at the end of the file.

Extended reasoning...

The change adds a new H200 Muse Spark NVFP4 AgentX recipe, master-config, runner and doc entries, and small code changes to the SRT model-path policy and to how speculative parameters are extracted for DFlash, with tests. It touches no security-sensitive surface such as auth, secrets handling or injection. I deferred for two reasons. First, the new override flag changes shared launch-policy behavior. Second, by the PR's own account the sweep has only 5 of 9 points accepted, and remaining blockers include an upstream cookbook and a final sweep.

This review covers commit adc40a2, which is no longer the latest commit on this pull request; later commits are not covered by it.

[Oseltamivir]

ok