Repository navigation
[PowerX] backfill Qwen3.5 FP4 B300 SGLang AgentX power / [PowerX] 补采 Qwen3.5 FP4 B300 SGLang AgentX 功耗 - #4704
edwingao28 wants to merge 1 commit into
Conversation
中文：重新运行 Qwen3.5 FP4 B300 SGLang AgentX 以采集功耗
|
Thanks for the contribution!
中文感谢你的贡献！
|
2f00523 to
2f2a68f
Compare
edwingao28
added
the
full-sweep-fail-fast
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
| - qwen3.5-fp4-b300-sglang-agentic-mtp | ||
| description: | ||
| - "Rerun the unchanged config to collect GPU-board power. Its published curve (2026-08-03) predates B300 single-node power collection and carries no power fields." | ||
| pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/PENDING |
There was a problem hiding this comment.
🔴 The PR's changelog validation will fail, and the reuse merge cannot fill in the real link, so this entry blocks the sweep and the merge. The placeholder pull/PENDING is not in PR_LINK_PLACEHOLDERS (validate_perf_changelog.py:20-23), which only accepts XXX and .../pull/XXX. Fix: set pr-link to https://github.com/SemiAnalysisAI/InferenceX/pull/XXX or to this PR's real URL. Both pass validate_added_pr_link and are rewritten by prepare_perf_changelog_merge.
Why this was flagged
The new entry at inferencex-e2e/perf-changelog.yaml:11777 sets pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/PENDING. On a PR run, validate_added_pr_link (inferencex-e2e/infx/workflows/validate_perf_changelog.py:105-117) accepts only https://github.com/SemiAnalysisAI/InferenceX/pull/{pr_number} or a member of PR_LINK_PLACEHOLDERS. That set is {"XXX", "https://github.com/SemiAnalysisAI/InferenceX/pull/XXX"}, so this link raises "new PR entry must use ... or an XXX placeholder". The merge path fails too: prepare_perf_changelog_merge.py:108-111 raises "appended entry N has unexpected pr-link" and never swaps in the real link. As a result the power-rerun entry cannot pass changelog validation or merge through reuse, so its power results are never ingested. Before this PR the changelog validated cleanly. No other placeholder handling rescues PENDING.
Verification: PR_LINK_PLACEHOLDERS in validate_perf_changelog.py:21-24 holds only "XXX" and "https://github.com/SemiAnalysisAI/InferenceX/pull/XXX". In merge_pr, merge_with_reuse.py:688 always calls canonicalize_changelog. The placeholder is not rewritten, and the PR cannot be merged with reuse until it changes to .../pull/XXX or the real PR URL. The PR-time CI check does not block the sweep.
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=41290415206 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=41290415206 |
Summary
Qwen3.5 FP4 B300 SGLang AgentX publishes no measured power. Its 28-point curve (2026-08-03) predates B300 single-node power collection, which later B300 curves report as valid. This reruns the unchanged config; only the changelog changes, so
infx.matrix.changedlists no keys. FP8 B300 is refreshed separately by #4482.Testing: the changelog plan generates the same 28 points plus 1 eval, and the image still resolves. The PR sweep is the GPU proof.
AI model disclosure
PR checklist and change type
Type of Change
Checklist
python -m infx.matrix.changed --base origin/mainlists affected config keys, I have appended one new entry covering them to the physical end ofinferencex-e2e/perf-changelog.yaml; if it lists none, I have added none. I have not edited historical entriesinferencex-e2e/perf-changelog.yamlentry, it carries exactly one primary sweep label (a maintainer applies it on fork PRs):full-sweep-fail-fast(recommended),full-sweep-enabled, ornon-canary-full-sweep-enabled. Optional modifiersall-evals,evals-only, andagentx-fastrequire a primary label; the last two block reuse while applied.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
Qwen3.5 FP4 B300 SGLang AgentX 目前没有实测功耗。其 28 个点的曲线（2026-08-03）早于 B300 单节点功耗采集上线，而之后的 B300 曲线已带有有效功耗。本 PR 重新运行这个未改动的配置，只新增 changelog 条目，因此
infx.matrix.changed不会列出任何配置。FP8 B300 由 #4482 另行更新。测试： changelog 生成的矩阵与原曲线相同（28 个点，外加 1 个评估），镜像仍可拉取。GPU 验证以本 PR 的 sweep 为准。
AI 模型：claude-opus-5-5（Claude Code），负责功耗覆盖率审计和 PR 准备。