[Klaud Cold] Update glm5.2-fp4-b200-sglang-agentic-mtp SGLang image to nightly-dev-cu13-20260908-20ca564b

[functionstackx]

Summary

Update SGLang image for the GLM-5.2 NVFP4 B200 AgentX MTP recipe from lmsysorg/sglang:nightly-dev-cu13-20260901-07c8f729 (2026-09-01 cu13 dev nightly) to lmsysorg/sglang:nightly-dev-cu13-20260907-30705c00 (2026-09-07 cu13 dev nightly).

• Tag verified on Docker Hub: last pushed 2026-09-07T01:43:42Z, digest sha256:19b8fa1223cc339c1eae7a5b703f1a8c2543b5b119155bf3d7efaef18f77f007
• Tag commit: sgl-project/sglang@30705c00 ("[Deepseek V4] Keep fp32 routing weights in the mxfp4 trtllm MoE")
• Carries forward the FlashInfer 0.6.18 BF16 TRTLLM MoE allocation fix the 2026-09-01 pin was taken for (perf(agentx): refresh GLM-5.2 B200 SGLang curve / 性能：刷新 GLM-5.2 B200 SGLang 曲线 #2808); the master-config comment is updated to say so.
• Same tag as the Qwen3.5 B200 SGLang bumps ([Klaud Cold] Update qwen3.5-fp8-b200-sglang-agentic-mtp SGLang image to nightly-dev-cu13-20260907-30705c00 #2861, [Klaud Cold] Update qwen3.5-fp4-b200-sglang-agentic-mtp SGLang image to nightly-dev-cu13-20260907-30705c00 #2862) and the new GLM-5.2 FP8 B200 sibling ([Klaud Cold] glm5.2-fp8-b200-sglang-agentic-mtp: day-zero GLM-5.2 FP8 SGLang AgentX MTP on B200 / B200 上 GLM-5.2 FP8 SGLang AgentX MTP 首发配方 #2863), so all four B200 SGLang AgentX curves share one SGLang commit.
• Recipe script, EAGLE MTP settings, golden AL 2.99, HiCache sizing, and the TP8 conc [1, 4, 8, 12, 16] grid are unchanged; image-only bump.

Recipes touched: glm5.2-fp4-b200-sglang-agentic-mtp

Test plan

• full-sweep-enabled sweep passes on cluster:b200-nscale.

🤖 Generated with Claude Code

---

Note

Low Risk
Benchmark config and changelog only; no engine flags or search space changes, though serving stability depends on the newer SGLang tag.

Overview
Image-only update for glm5.2-fp4-b200-sglang-agentic-mtp: the SGLang container moves from nightly-dev-cu13-20260901-07c8f729 to nightly-dev-cu13-20260908-20ca564b, aligning with other B200 SGLang AgentX recipes while keeping EAGLE MTP, golden AL 2.99, HiCache, and the TP8 concurrency grid unchanged.

The master-config comment is refreshed to note the 2026-09-08 pin still includes the FlashInfer 0.6.18 BF16 TRTLLM MoE fix needed for GLM-5.2’s unquantized EAGLE draft at CUDA-graph capture.

perf-changelog.yaml documents the path: an intermediate 2026-09-07 nightly was skipped because an unguarded kv_index_translator.translate_dcp_read_ids on the DSA fp8 KV path caused intermittent AttributeError on MTP runs; 2026-09-08 includes sgl-project/sglang#38318’s None guard.

^{Reviewed by Cursor Bugbot for commit 57d2ed5. Bugbot is set up for automated code reviews on this repo. Configure here.}

Update: re-pinned to the 2026-09-08 nightly

The 2026-09-07 cu13 nightly (nightly-dev-cu13-20260907-30705c00) carries an unguarded kv_index_translator.translate_dcp_read_ids call on the DSA fp8 KV read path; the EAGLE draft backend never binds a translator, so GLM-5.2 MTP runs crash intermittently with AttributeError: 'NoneType' object has no attribute 'translate_dcp_read_ids' (seen on the MI355X FP8 sibling, run 34173459478, after 74 minutes of clean serving). sgl-project/sglang#38318 (merged 2026-09-07T20:03Z) adds the None guard. Re-pinned to lmsysorg/sglang:nightly-dev-cu13-20260908-20ca564b (pushed 2026-09-08T01:40:59Z, digest sha256:9a352a35c973a2357372e85f3bcb5388b6b3c46c1329165987260f3b089647dc), which is six commits past that fix. Recipe otherwise unchanged.

[github-actions]

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 As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

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 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

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

[github-actions]

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 As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

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 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

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

[claude]

Nothing blocking. The comments below are optional suggestions. There is no need to push a fix for them before merging.

[claude]

🟡 (optional) The new changelog entry's pr-link is left as the literal placeholder string "PRLINK_PLACEHOLDER" even though this PR already exists (opened 2026-09-07), so the append-only changelog record for this recipe permanently points nowhere unless someone catches and edits it before merge. Fix: before merge, replace pr-link with the real PR URL (https://github.com/SemiAnalysisAI/InferenceX/pull/) per docs/configuration-procedures.md's "Append the changelog safely" step 3 — note the docs only sanction "TBD" pre-PR, not this custom placeholder.

Extended reasoning...

docs/configuration-procedures.md step 3 states pr-link may be "TBD" only before the PR exists and must become the real URL immediately after PR creation. This diff's entry uses neither a real URL nor the sanctioned TBD value but an ad-hoc placeholder token, and because perf-changelog.yaml is append-only and byte-sensitive (never edited in place after merge per rule 4-6), if this slips through, the historical record for this image bump permanently lacks provenance back to its PR, undermining the file's stated purpose of tracking why each performance-affecting change was made.

Verification: normal. The appended changelog entry at perf-changelog.yaml:6931 sets pr-link: PRLINK_PLACEHOLDER. This is neither a canonical PR URL nor a sanctioned placeholder. utils/validate_perf_changelog.py:24-27 defines the only allowed placeholders as PR_LINK_PLACEHOLDERS = {"XXX", "https://github.com/SemiAnalysisAI/InferenceX/pull/XXX"}, and CANONICAL_PR_LINK (lines 21-23) is… | nit. The new…

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34169357752
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=34169357752

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34179014799
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=34179014799