Repository navigation
[GLM-5.3 GB300] add Dynamo-SGLang AgentX agg and disagg recipes with power / [GLM-5.3 GB300] 新增 Dynamo-SGLang AgentX 聚合与分离部署配方并采集功耗 - #4706
edwingao28 wants to merge 8 commits into
Conversation
中文：采集 GLM-5.2 GB300 Dynamo-SGLang AgentX 功耗
|
Thanks for the contribution!
中文感谢你的贡献！
|
|
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。 |
1238f20 to
1a66b16
Compare
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Besides the inline finding, I checked the new GB300 GLM-5.2 PowerRule in infx/launch/drivers/srt/power.py. Its Match arguments are in the right order (prefixes, precisions, frameworks, then agentic=True). The rule also matches the existing GB200 GLM-5.2 rule exactly, which doesn't set require=True either, so leaving that flag off follows the existing pattern.
Extended reasoning...
I compared the new rule with Match's field order in infx/launch/policy.py and with the GB200 srt-multi lane, which already has an identical GLM-5.2 FP4 dynamo-sglang AgentX rule that doesn't set require.
| description: | ||
| - "Collect GPU-board power for the two aggregated and six disaggregated GLM-5.2 GB300 Dynamo-SGLang AgentX points, which currently publish no power." | ||
| - "Route GLM-5.2 FP4 AgentX dynamo-sglang on the GB300 srt-multi lane to AgentX power instead of DCGM-only power." | ||
| pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/PENDING |
There was a problem hiding this comment.
🔴 The changelog check will reject this PR, and merge prep cannot canonicalize the link, because pull/PENDING is neither an accepted placeholder nor a PR number. validate_added_pr_link only accepts links in PR_LINK_PLACEHOLDERS (XXX or .../pull/XXX) or the exact .../pull/<pr_number> URL. Fix: set the link to https://github.com/SemiAnalysisAI/InferenceX/pull/XXX or the real PR URL. XXX is what the merge tooling rewrites.
Why this was flagged
The new tail entry at inferencex-e2e/perf-changelog.yaml:11779 sets pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/PENDING. On PR runs, validate_added_pr_link (infx/workflows/validate_perf_changelog.py:105-116) accepts only a link in PR_LINK_PLACEHOLDERS (lines 21-24: XXX and https://github.com/SemiAnalysisAI/InferenceX/pull/XXX) or the exact https://github.com/SemiAnalysisAI/InferenceX/pull/{pr_number}. Any other value raises "new PR entry must use ... or an XXX placeholder". Without a PR number, CANONICAL_PR_LINK requires \d+, so the check fails there too. prepare_perf_changelog_merge.py:108 also rejects links outside PR_LINK_PLACEHOLDERS with "unexpected pr-link". As a result the changelog validation job fails and the reuse/merge path cannot rewrite the link. No other entry in the file uses PENDING.
Verification: Trigger: an authorized maintainer runs the reuse merge (/use <run_id>) for this PR. validate_added_pr_link (validate_perf_changelog.py:112-116) raises new PR entry must use ... or an XXX placeholder for pull/PENDING, unless the value is exactly .../pull/{pr_number}. Result: the reuse merge errors out and the placeholder is never rewritten to the real PR URL.
edwingao28
added
the
full-sweep-fail-fast
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=41407333903 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=41407333903 |
中文：将 main 合并到 GLM-5.3 GB300 Dynamo-SGLang AgentX
中文：新增 GLM-5.3 FP4 GB300 Dynamo-SGLang AgentX 聚合配置
中文：新增 GLM-5.3 FP4 GB300 Dynamo-SGLang AgentX 分离式配置
中文：将 main 合并到 GLM-5.3 GB300 Dynamo-SGLang AgentX
edwingao28
changed the title
|
/use 41407333903 |
|
@edwingao28 staged run 41407333903. This run remains available across future |
中文：更正 GLM-5.3 GB300 changelog 中的功耗规则描述。
中文：将 main 合并到 GLM-5.3 GB300 Dynamo-SGLang AgentX。
edwingao28
removed
the
full-sweep-fail-fast
Summary
Add GLM-5.3 NVFP4 GB300 Dynamo-SGLang AgentX: two aggregated points (TP4 and TP8 at conc 1) and six disaggregated PP4-prefill points (conc 30 to 264). The recipes port GLM-5.2 GB300, whose FP4 weights left the cluster, with NEXTN K=3 MTP and pinned weights. A new GB300 power rule routes these points to AgentX power.
Testing: sweep 41407333903 at 26d2411: 8/8 points strict power, 8/8 gsm8k 0.973 to 0.977, at most 0.7% request errors. Against GLM-5.2 GB300, output throughput per GPU is within 4% and interactivity up to 13% lower.
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
新增 GLM-5.3 NVFP4 GB300 Dynamo-SGLang AgentX 配置：2 个聚合部署点（TP4 和 TP8，并发 1）和 6 个 PP4 prefill 分离部署点（并发 30 到 264）。配方移植自 GLM-5.2 GB300（其 FP4 权重已从集群移除），改用 NEXTN K=3 MTP，并固定权重版本。新增的 GB300 功耗规则将这些点路由到 AgentX 功耗采集。
测试： sweep 41407333903（26d2411f3）：8/8 个点通过严格功耗校验，8/8 个 gsm8k 评估得分 0.973 到 0.977，请求错误率最高 0.7%。与 GLM-5.2 GB300 相比，每 GPU 输出吞吐量相差不超过 4%，interactivity 最多低 13%。
AI 模型：claude-opus-5-5（Claude Code），负责配方移植、launch 策略改动、sweep 验证和 PR 准备。