[AgentX H200] enable Kimi-K3 balanced power / 启用 balanced 实测功耗 - #3053
[AgentX H200] enable Kimi-K3 balanced power / 启用 balanced 实测功耗#3053edwingao28 wants to merge 2 commits into
Conversation
|
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 For PR verification, add the 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 如需进行 PR 验证，请为此 PR 添加 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 |
为 H200 Kimi-K3 balanced 配方的完整 12 点范围启用实测功耗。
128e5c0 to
89967b5
Compare
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Looks good, straightforward config change. Reviewed the telemetry block and concurrency list added to the H200 Kimi-K3 balanced recipe, and the appended perf-changelog entry.
- Telemetry block is byte-identical to the already-merged dcgm-power block used by the sibling Kimi-K3 H200 latency recipe (PR #3044), so it follows established precedent rather than introducing a new pattern.
- The new
concurrencies: [1,2,3,4,5,6,7,8,10,12,14,16]matches theconc-listalready present inconfigs/nvidia-master.yamlfor thekimik3-fp4-h200-vllm-agentic-balancedconfig key, so recipe and master config stay in sync. - perf-changelog.yaml entry is correctly appended at the physical tail, prior bytes untouched, config-keys match the master config entry; the placeholder
pull/XXXpr-link matches the documented placeholder pattern in CONTRIBUTING.md and is expected to be filled in before merge (already called out as unchecked by the author).
Extended reasoning...
Overview
This PR touches exactly two files: it adds a telemetry block (dcgm-power provider, required=true, standard timeouts, dcgm_exporter image/port) and a benchmark.concurrencies list to one srt-slurm recipe YAML (H200 Kimi-K3 balanced, vLLM), and appends one new entry to perf-changelog.yaml. No code paths, launchers, or CI logic are modified — purely declarative config.
Security risks
None identified. There is no auth, crypto, or user-input handling involved; the change only toggles a benchmarking/telemetry feature flag and a list of integers in a YAML recipe, plus an append-only changelog entry.
Level of scrutiny
This warrants light scrutiny consistent with a mechanical config change. I verified it against the actual precedent in the repo rather than trusting the PR description: the telemetry block added here is byte-for-byte identical to the block already merged and running in production for the sibling H200 Kimi-K3 latency recipe (from PR #3044, whose changelog entry is directly above this one in perf-changelog.yaml). The new concurrency list [1,2,3,4,5,6,7,8,10,12,14,16] matches exactly the conc-list already present in configs/nvidia-master.yaml for the kimik3-fp4-h200-vllm-agentic-balanced config key — so master config and recipe config are already consistent (the AGENTS.md rule to update master config alongside the recipe is satisfied; that master-config entry was apparently added previously and needed no changes here since there is no image bump). The perf-changelog.yaml diff is append-only at the tail, all prior bytes preserved, and the config-keys/scenario-type match the master config. The pull/XXX placeholder pr-link is not a bug — it is the exact documented placeholder shown in CONTRIBUTING.md's example changelog entry, and the author's own PR checklist leaves that box unchecked pending the real PR number.
Other factors
The AGENT_OPERATIONS.md note that "only eligible recipe-gated dynamo-sglang dcgm-power lanes are validated" today refers to a pre-existing scope limitation of the local test suite that predates this PR (the same caveat already applied to the vLLM latency recipe merged in #3044); this PR does not change that validation surface or introduce new risk beyond what was already accepted for the sibling recipe. No CODEOWNERS entries restrict the two changed paths. Given the change is a narrow, mechanical copy of an already-validated pattern plus a correctly-appended changelog entry, I'm confident this does not need further human scrutiny beyond the author filling in the PR number.
This review covers commit 128e5c0, which is no longer the latest commit on this pull request; later commits are not covered by it.
同步 latency 父分支的主线合并，保留 balanced 完整矩阵与实测功耗配置；代码树不变。
1 participant
Description
Enable required measured power for the complete H200 Kimi-K3 balanced recipe (12 points), using the recipe-gated launcher from #3044.
Testing: Changelog validation and exact matrix checks passed; shared routing tests passed.
Pending: Depends on #3044. The full 12-point sweep and 12 tool-schema evals remain unrun; current capacity cannot complete them overnight.
中文
为 H200 Kimi-K3 balanced 完整配方（12 点）启用必需实测功耗，复用 #3044 按配方控制的启动路径。
测试： Changelog 校验及精确矩阵检查通过，公共路由测试通过。
待完成： 依赖 #3044；完整 12 点扫描和 12 项工具 schema 评估尚未运行，当前容量无法在本夜间窗口完成。
Related Issue
Parent and shared evidence / 前置 PR 与公共证据: #3044.
Type of Change
Checklist
perf-changelog.yamland have not edited historical entriesOWNER/MEMBER/COLLABORATOR) has commented/reuse-sweep-runon this PR. Do this only once there is a final full sweep that is all green with evals passing, since after this comment the sweep label will no longer automatically kick off new sweeps. Remove and re-add the label to force one.Note
Low Risk
Benchmark recipe and changelog-only changes; no application auth or data-path logic, with behavior gated on the shared launcher from #3044.
Overview
Enables required DCGM power telemetry on the H200 Kimi-K3 balanced agentic recipe (
agg-h200-tp8dp4ep32-balanced-agentic.yaml), matching the pattern from the latency recipe (#3044):dcgm-powercollector,required: true, and exporter settings.Defines an explicit 12-point concurrency sweep (
1–8,10,12,14,16) on the customagentic_srt.shbenchmark so balanced runs use the full matrix with measured power.Documents the change in
perf-changelog.yamlunder config keykimik3-fp4-h200-vllm-agentic-balanced.Reviewed by Cursor Bugbot for commit 71d7670. Bugbot is set up for automated code reviews on this repo. Configure here.