Repository navigation
[PowerX] collect GLM-5.2 B200 Dynamo-SGLang AgentX power / [PowerX] 采集 GLM-5.2 B200 Dynamo-SGLang AgentX 功耗 - #4656
Conversation
中文：采集 GLM-5.2 B200 Dynamo-SGLang AgentX 功耗。
中文：将 GLM-5.2 B200 功耗 changelog 关联到 PR。
|
Thanks for the contribution!
中文感谢你的贡献！
|
中文：恢复历史 changelog 条目中的 pr-link 占位符。
edwingao28
added
the
full-sweep-fail-fast
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
| - dram-utilization: 1.0 | ||
| search-space: | ||
| - spec-decoding: mtp | ||
| power: true |
There was a problem hiding this comment.
🔴 All eight GLM-5.2 B200 AgentX points will now fail at launch with PowerPolicyError instead of running. Today they run without power. Neither b200-nscale lane in POWER_LANES (power.py:84-114) has a rule for glm5.2 fp4 dynamo-sglang agentic, so decide_power rejects every power: true point. Fix: add a matching PowerRule to the b200-nscale lane (SRT_MULTI, and SRT_NATIVE if these single-node agg points launch there) in this PR, or drop power: true. Same pattern at 8 sites (nvidia-master.yaml:7479, 7491, 7503, 7515, 7527, 7539, 7569, 7586).
Why this was flagged
Both config keys use runner: cluster:b200-nscale, framework: dynamo-sglang, model-prefix: glm5.2, precision: fp4 under the agentic-coding scenario (nvidia-master.yaml:7463-7473, 7551-7562). With power: true, decide_power (inferencex-e2e/infx/launch/drivers/srt/power.py:146-162) looks up the lane and walks its rules. The ('b200-nscale', SRT_MULTI) lane only allows qwen3.5 fp8 dynamo-sglang agentic and dsv4 fp4 dynamo-vllm fixed-seq. The ('b200-nscale', SRT_NATIVE) lane only allows kimik3 fp4 dynamo-vllm agentic and dsv4 fp4 fixed-seq. No rule matches, so it raises PowerPolicyError 'B200 Nscale dcgm-power requires fixed-sequence DSV4 FP4 dynamo-vllm or Qwen3.5 FP8 AgentX dynamo-sglang'. Callers are infx/launch/drivers/srt/init.py:171 and infx/srt_slurm/generate.py:128. On base, request.power is false, decide_power returns NO_POWER and the points run. The PR claims matrix generation passes, but that step does not call decide_power, so the full sweep is where every one of the eight points fails.
Verification: launch_path (infx/launch/policy.py:80-85) routes them to SRT_NATIVE. When request.power is set, decide_power (power.py:146-162) finds no matching rule and raises PowerPolicyError(lane.error). The diff touches only nvidia-master.yaml and perf-changelog.yaml and adds no PowerRule, so all 8 power: true sites (6 agg, 2 disagg) are affected.
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=41003299710 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=41003299710 |
中文：允许 B200 native 通道采集 GLM-5.2 AgentX SGLang 功耗。
|
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。 |
中文：将 GLM-5.2 B200 分离部署固定到可拉取的 SGLang 镜像。
中文：将 main 合并到 GLM-5.2 B200 功耗分支。
|
/use 41003299710 |
|
@edwingao28 staged run 41003299710. This run remains available across future |
functionstackx
left a comment
functionstackx
left a comment
There was a problem hiding this comment.
Approved. Reviewed the eight power-enabled points, the B200 native-lane policy change and its positive/negative test cases, and the disaggregated image pin. Sweep 41003299710 passed at the reviewed head, including all eight benchmark points and three agentic eval jobs; /use 41003299710 is recorded. Strict power enforcement remains follow-up work in #4552. Resolve the merge conflict and obtain the required NVIDIA config CODEOWNER checklist before merging through the reuse path; this approval does not replace that sign-off.
中文
批准。已审阅 8 个点的功耗启用改动、B200 native 通道策略及其正反测试用例，以及分离部署的镜像固定。所审阅提交的 sweep 41003299710 已通过，包括全部 8 个 benchmark 点和 3 个 agentic eval 任务；已记录 /use 41003299710。严格功耗校验仍由 #4552 跟进。请先解决合并冲突并取得所需的 NVIDIA 配置 CODEOWNER 检查清单签核，再通过复用流程合并；本次批准不替代该签核。
…power # Conflicts: # inferencex-e2e/perf-changelog.yaml
|
/reuse-sweep-run 41003299710 |
# Conflicts: # inferencex-e2e/perf-changelog.yaml
|
Ingest report — ✅ Results ingested into the Ingest workflow · Merge Ingest run · Benchmark source run
benchmarkFiles 9, rawBenchmarkRows 16, coveredByPerJobFiles 8, purgedPoints 0, failedRunRows 0, runStatsNew 0, runStatsDuplicate 0, evalSamplesNew 3957, gpuMetricSamplesNew 393524, powerPublicationPoints 8, 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 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。 |
Summary
GLM-5.2 B200 Dynamo-SGLang AgentX publishes no measured power for its six aggregated and two disaggregated points. Enable GPU-board power on all eight and allow it in the B200 native srt lane's power policy. The disaggregated key's
nightly-dev-20260910image was pruned (Docker Hub 404), so it moves to the aggregated key's digest-pinned v0.5.22-cu130.Testing: the new launch-policy case fails before and passes after; changelog and matrix checks pass. The PR sweep is the GPU proof. Strict enforcement follows #4552.
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
GLM-5.2 B200 Dynamo-SGLang AgentX 的 6 个聚合部署点和 2 个分离部署点目前没有发布实测功耗。本改动为全部 8 个点启用 GPU 板卡功耗采集，并在 B200 native srt 通道的功耗策略中允许该路由。分离部署配置原用的
nightly-dev-20260910镜像已从 Docker Hub 删除（404），因此改用聚合部署配置已在使用的、按 digest 固定的 v0.5.22-cu130 镜像。测试： launch 策略测试覆盖新规则（修复前失败、修复后通过）；changelog 校验与矩阵生成通过。GPU 验证以本 PR 的完整 sweep 为准。严格校验将在 #4552 的 AgentX
require-power传递合入后跟进。AI 模型：claude-opus-5-5（Claude Code），负责配置与策略改动、覆盖率审计和 PR 准备。