[edwingao28]

Summary

Qwen3.5 FP8 B300 Dynamo-SGLang AgentX publishes no measured power. Its curve (qwen3.5-fp8-b300-dynamo-sglang-agentic-disagg, 2026-10-21) has valid power on 0 / 9 points. The curve predates default-on multinode power collection (#3670).

This PR originally updated the Qwen3.5 FP8 B300 SGLang 8k1k config. #4697 retired every Qwen3.5 8k1k config, so the PR now reruns this AgentX curve to collect power.

Changes:

• inferencex-e2e/perf-changelog.yaml: append one block listing qwen3.5-fp8-b300-dynamo-sglang-agentic-disagg. It is the only key of its curve, so the rerun replaces the curve whole.
• The 8k1k image bump and its changelog block are dropped, because main deleted that config.
• Configs and recipes are unchanged, so infx.matrix.changed lists no keys.
• The FP4 curve (qwen3.5-fp4-b300-dynamo-sglang-agentic-pareto) is not rerun, because the Qwen3.5 NVFP4 weights are missing on b300-dsxe.

Testing:

• Sweep run 42278372774, attempt 1: 28 jobs passed, 0 failed. All 9 AgentX points carry valid GPU power, and the AgentX evals passed.
• curve_coverage.py: no dashboard point lost (9 published points, 9 in the run).
• validate_perf_changelog --base-ref origin/main passes after the main sync.

Curve change after publishing:

• Output throughput per GPU moves by -3.8% to +5.8% per point.
• The frontier drops 22% to 33% between interactivity 196 and 256 tok/s/user. The published c32 TP2/EP2 point (2026-10-21) showed 245 tok/s/user, well above c40 at 150; this run measures 176 at 3.8% lower throughput. The rest of the frontier stays within 4%.

Limits:

• GPU power only: b300-dsxe declares no CPU power exporter.

PR checklist and change type

Type of Change

• Bug fix
• New feature
• Configuration change
• Documentation update
• Other (please describe): rerun of an unchanged config to collect power

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。

[claude]

I reviewed this PR and didn't find any bugs. Because nvidia-master.yaml has designated code owners, a human look is still worthwhile. One small thing before merge: the new changelog entry's pr-link still ends in pull/XXX, so it should be changed to this PR's number.

What was reviewed:

• The new tag and digest for qwen3.5-fp8-b300-sglang are identical to the v0.5.22-cu130 pin already used by qwen3.5-fp8-b300-sglang-mtp and about 15 other entries in the file.
• The changelog block is appended at the physical end, and no earlier bytes changed. It covers the one affected key and is in English.
• The missing scenario-type is consistent with earlier entries for this key (PRs 1048, 1451 and 4059), none of which set it.

Extended reasoning...

This change bumps one container image (tag plus sha256 digest) for the qwen3.5-fp8-b300-sglang benchmark config and appends one matching perf-changelog entry. It touches no security-sensitive surface. The new digest matches pins that already run elsewhere in the same file. The PR still needs human review because CODEOWNERS assigns named reviewers to inferencex-e2e/configs/nvidia-master.yaml, and the changelog pr-link is still a placeholder.

This review covers commit f6cbf01, which is no longer the latest commit on this pull request; later commits are not covered by it.

[github-actions]

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=42278372774

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=42278372774

[edwingao28]

/use 42278372774

[github-actions]

@edwingao28 staging run 42278372774. InferenceX-app staging workflow is now running. Existing staged runs will be preserved; this comment will be updated when staging completes.

[github-actions]

Ingest report — ✅ Results ingested into the production database.

Ingest workflow · Merge Ingest run · Benchmark source run

Category	new	duplicate	skipped	unmapped	failed
benchmark	9	0	9	0	0
eval	7	7	0	0	0
trace	9	0	0	0	0
power	9	0	0	0	0

_{benchmarkFiles 10, rawBenchmarkRows 18, coveredByPerJobFiles 9, purgedPoints 0, failedRunRows 0, runStatsNew 0, runStatsDuplicate 0, evalSamplesNew 9233, gpuMetricSamplesNew 266196, powerPublicationPoints 9, changelogEntriesWritten 1}

Counts cover only this run. Rows: benchmark_results, eval_results, agentic rows linked to a trace, gpu_metrics series. This comment updates if the ingest is rerun.

中文

Ingest 报告 — ✅ 结果已写入 production 数据库。

上表仅统计本次运行写入的数据，各行单位依次为 benchmark_results 行、eval_results 行、已关联 trace 的 agentic 行和 gpu_metrics 序列。重新运行 ingest 时本评论会原地更新。