perf(gb300): add AgentX aggregate TP4 points / 添加 GB300 AgentX 聚合式 TP4 数据点

[ivanium]

Summary / 摘要

• Expand the GB300 AgentX aggregate TP4 sweep to concurrency 1, 2, 4, 6, and 8.
• 将 GB300 AgentX 聚合式 TP4 扫描扩展到并发 1、2、4、6 和 8。

Validation / 验证

Exact-key generation and append-only validation passed; 36 focused tests passed. / 精确配置生成和仅追加校验通过；36 项专项测试通过。

[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]

Claude finished @ivanium's task in 2m 4s —— View job

---

PR Review

• Gather context and read PR diff
• Verify perf-changelog entry is appended at end of file with append-only: true
• Verify master-config change is behaviorally isolated to appended points (base matrix ⊆ head matrix)
• Validate append-only semantics (by direct diff inspection — see note)
• Post review

LGTM - no blocking issues found

Verification details:

• The only benchmark-affecting change is configs/nvidia-master.yaml:7790, expanding the TP4 MTP variant's conc-list from [8] to [1, 2, 4, 6, 8]. The existing concurrency-8 point is retained with an identical recipe, image, and settings; the TP8 variants are untouched. Base matrix is a strict subset of the head matrix, and the config keeps its single non-null image (vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-426e59f, public on Docker Hub).
• The new perf-changelog.yaml entry is appended at the very end of the file, carries append-only: true, has no eval modifiers, and is the only new entry.
• No benchmark scripts, launchers, or generation logic changed — the diff is exactly these two declarative YAML edits, so the benchmark effect is exclusive to the newly appended concurrency points by construction.

Note: I could not execute the exact-key generation script in this sandbox (Python execution requires approval), so the base⊆head check was done by inspecting the diff rather than by re-running the generator. Since the change is a pure conc-list expansion with no logic changes, the inspection is conclusive.
· Branch

[claude]

LGTM — straightforward sweep expansion, no bugs found.

What was reviewed: the diff only widens conc-list from [8] to [1, 2, 4, 6, 8] on the existing GB300 AgentX aggregate TP4 recipe entry, with the referenced recipe file (agg-gb300-tp4-mtp-agentic.yaml) left untouched, plus a matching append-only changelog entry. Confirmed the TP4 block sits alongside the untouched TP8 block in nvidia-master.yaml and that no other config or launcher logic was touched.

Extended reasoning...

Overview

This PR touches only configs/nvidia-master.yaml and perf-changelog.yaml. It expands the conc-list for the existing dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-agg TP4 scenario entry from [8] to [1, 2, 4, 6, 8], and appends a corresponding changelog entry describing the change with an append-only: true marker. No recipe files, launcher code, or scoring logic were modified.

Security risks

None. This is a pure data change to a benchmark sweep configuration — no new code paths, no external inputs, no auth/crypto/permission surfaces touched.

Level of scrutiny

Low. This is a mechanical, additive change to an existing, already-validated recipe (same image, same CONFIG_FILE, same TP/EP/dp-attn settings) — only the concurrency sweep points are widened. The changelog entry is consistent with the diff and marked append-only, matching repository convention for this file.

Other factors

The bug-hunting system found no issues, and the one candidate concern raised (a stale 'concurrency 8' comment in the referenced TP4 recipe) was investigated and ruled out. The PR description states 36 focused tests passed for exact-key generation and append-only validation. No unresolved reviewer comments exist in the timeline.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32242794988
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=32242794988

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32242794988
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=32242794988