[GLM-5.2 GB300] Preserve draft precision and extend scheduler watchdog / [GLM-5.2 GB300] 保留 draft 精度并延长调度 watchdog - #3402
Conversation
|
Thanks for the contribution!
中文感谢你的贡献！
|
0079a3a to
66a5f68
Compare
edwingao28
added
the
full-sweep-fail-fast
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=36538942770 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=36538942770 |
edwingao28
removed
the
full-sweep-fail-fast
edwingao28
changed the title
edwingao28
changed the title
edwingao28
added
the
full-sweep-fail-fast
581f91f to
d8b13ce
Compare
edwingao28
changed the title
d8b13ce to
945377b
Compare
edwingao28
changed the title
edwingao28
removed
the
full-sweep-fail-fast
945377b to
9bced94
Compare
edwingao28
added
full-sweep-enabled
9bced94 to
3cf29c6
Compare
edwingao28
changed the title
|
Sorry, over the weekend, there was 2 major refactors to clean up the technical debt accumalated over the past 11 months of moving at the speed of light. We don't see any major refactors in the forthseeable future besides cleaning up AMD multinode AgentX pile of bash. As much, due to the refactors, u would need to ask your agent to rebase from remote main@latest. Thank you in advance for ur understanding |
4d4d437 to
1903073
Compare
Set SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE=0 so the NextN/MTP draft keeps its shipped precision, and use the 1800 s scheduler watchdog of the GLM-5.2 B200/GB200 recipes for disaggregated prefill and decode. 设置 SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE=0，使 NextN/MTP draft 保持原始发布精度； 分离式 prefill 和 decode 采用 GLM-5.2 B200/GB200 配方的 1800 秒调度 watchdog。 Co-authored-by: Wenyao Gao <[email protected]>
1903073 to
fc03e75
Compare
edwingao28
changed the title
中文：同步 main 并保留 GLM-5.2 GB300 的改动及验证历史。[skip-sweep]
|
Claude finished @edwingao28's task in 35s —— View job PR review
LGTM - no blocking issues found |
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Nothing blocking. The comments below are optional suggestions. There is no need to push a fix for them before merging.
Beyond the inline finding, I also checked whether the GB300 agg.yaml recipe needed the same 1800s watchdog-timeout bump as its disagg sibling and the GB200/B200 agg recipes (which already carry it) — this was already examined and is not raised as a separate issue here.
Extended reasoning...
The diff flips SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE from '1' to '0' in two GLM-5.2 GB300-FP4 SGLang recipes (moving toward compliance with the repo's draft-precision rule) and adds watchdog-timeout: 1800 to disagg-mtp-variants.yaml's prefill/decode blocks, plus one append-only perf-changelog entry; none of this touches auth, crypto, or permission logic. A confirmed inline finding (Chinese description lines violating the English-only perf-changelog rule in AGENTS.md) will be posted, which by itself means this should not be approved. I independently verified the diff's precision flips and watchdog addition, and separately confirmed the GB300 agg.yaml does not receive the same watchdog-timeout bump that GB200-fp4's agg.yaml already has, but that gap was already investigated and not raised as a new finding.
| - "关闭 GB300 GLM-5.2 AgentX 对 NextN/MTP draft MoE 的 FP8 转换，使 draft 保持原始发布精度；镜像、拓扑、工作负载和验收不变。" | ||
| - "分离式 prefill 和 decode 采用 GLM-5.2 B200/GB200 配方的 1800 秒调度 watchdog，替代默认 300 秒。" | ||
| pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3402 |
There was a problem hiding this comment.
🟡 (optional) New perf-changelog entry mixes two Simplified-Chinese description lines with the English ones, violating the repo's English-only rule for this file.
Why this was flagged
AGENTS.md:126 states new inferencex-e2e/perf-changelog.yaml entries must be English-only and must not add Chinese translations or bilingual descriptions. The entry appended at inferencex-e2e/perf-changelog.yaml:9238-9250 (config-keys glm5.2-fp4-gb300-dynamo-sglang-agentic-agg/-disagg) includes two Chinese description lines at 9249-9250 alongside the English ones at 9247-9248. Any changelog consumer or maintainer relying on English-only entries (e.g. tooling that parses/greps descriptions, or CODEOWNER review per the PR_REVIEW_CHECKLIST rule) gets a bilingual entry the base branch's convention forbids. Fix: remove the two Chinese description lines so the entry is English-only, consistent with every other entry in the file.
Verification: The new perf-changelog entry appended at inferencex-e2e/perf-changelog.yaml:9238-9250 contains two Simplified-Chinese description lines (9249-9250) alongside the English lines at 9247-9248. This contradicts AGENTS.md:126, which requires new entries to be English-only with no Chinese translations or bilingual descriptions. The violation is introduced by this PR; the base branch has no such entry.
|
/reuse-sweep-run 36538942770 |
中文：同步 main 后重新追加 GB300 更新记录。[skip-sweep]
Resolve dirty mergeable_state from main changelog appends (#3402/#3652/#3653 et al.). Keep origin/main perf-changelog bytes as prefix and re-append all #3088 tip entries including CONC 32+ max-num-seqs 1x. Recipe tip unchanged. 解决 main changelog 追加（#3402/#3652/#3653 等）导致的 dirty。保留 origin/main 的 perf-changelog 字节为前缀，并重新追加全部 #3088 tip 条目（含 CONC 32+ max-num-seqs 1×）。配方 tip 不变。 Co-authored-by: Wenyao Gao <[email protected]>
2 participants
Description
Preserve GLM-5.2 GB300 draft precision and the 1800-second disaggregated watchdog. Synchronize main while preserving the tested recipe bytes and sweep ancestry.
Testing: Sweep 36538942770/a3, head
fc03e75c: seven performance points and five evals passed with per-cell artifact/runtime review. Current changelog/matrix validation passes.Review limits: Identity, power, cancellation and shutdown limitations remain. Main's newer launcher/runtime has not been GPU-qualified by that sweep; reuse compatibility requires review.
中文
保留 GLM-5.2 GB300 draft 原始发布精度及分离式 1800 秒 watchdog。同步 main，并保留已测试配方的原始字节及 sweep 提交历史。
测试： sweep
36538942770/a3（fc03e75c）的 7 个性能点和 5 个 eval 通过，已逐格审查 artifact/runtime；当前 changelog/矩阵验证通过。审阅限制： 身份比对、功耗、取消及关闭限制保留。历史 sweep 未验证 main 的新 launcher/runtime；仍需审查结果复用兼容性。
AI 模型： Claude Opus 5.5 (
claude-opus-5-5) 负责实施及起草；GPT-6（具体变体不可确认）负责恢复、验证、冲突解决及委派复核。AI model disclosure
claude-opus-5-5): implementation/drafting.Related Issue
Related to #3228. / 关联 #3228。
Type of Change
Checklist
inferencex-e2e/perf-changelog.yamland have not edited historical entriesOWNER/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 sweep label will no longer automatically kick off new sweeps. Remove and re-add the label to force one.