[Klaud Cold] Update glm5.2-fp4-b300-sglang-agentic-mtp SGLang image to nightly-dev-cu13-20260908-20ca564b and move to cluster:b300-dsxe

[functionstackx]

Summary

Update the SGLang image for glm5.2-fp4-b300-sglang-agentic-mtp from lmsysorg/sglang:v0.5.16-cu130 to lmsysorg/sglang:nightly-dev-cu13-20260907-30705c00, and move the recipe from the retired cluster:b300-nv runner to cluster:b300-dsxe.

• Tag verified on Docker Hub: last pushed 2026-09-07T01:43:42Z, digest sha256:19b8fa1223cc339c1eae7a5b703f1a8c2543b5b119155bf3d7efaef18f77f007; same tag as the B200 GLM-5.2 SGLang bumps ([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, [Klaud Cold] Update glm5.2-fp4-b200-sglang-agentic-mtp SGLang image to nightly-dev-cu13-20260908-20ca564b #2864).
• Overlap: the open [Klaud Cold] glm5.2-fp4-b300-sglang-agentic-mtp: extend the AgentX curve to concurrency 48 and move to cluster:b300-dsxe / 将 GLM-5.2 B300 SGLang AgentX 曲线扩展至并发度 48 并迁移至 cluster:b300-dsxe #2829 also repoints this key to DSXE and extends its grid to concurrency 48. This PR is image-only on the grid; whichever lands second rebases the one-line conflict.
• Runner: cluster:b300-nv was retired in feat(runners): add the B300 DSXE cluster and retire the B300 NV launcher / 添加 B300 DSXE 集群并下线 B300 NV 启动脚本 #2826 (launcher and runner labels removed), so a sweep on it can never be scheduled. Repointed to cluster:b300-dsxe, the same move [Klaud Cold] glm5.2-fp4-b300-sglang-agentic-mtp: extend the AgentX curve to concurrency 48 and move to cluster:b300-dsxe / 将 GLM-5.2 B300 SGLang AgentX 曲线扩展至并发度 48 并迁移至 cluster:b300-dsxe #2829 makes for the GLM-5.2 FP4 B300 sibling. Because the runner changed this is not an append-only bump; the whole curve reruns on DSXE.
• Recipe script and concurrency grid unchanged.

Recipes touched: glm5.2-fp4-b300-sglang-agentic-mtp

Test plan

• full-sweep-enabled sweep passes on cluster:b300-dsxe.

🤖 Generated with Claude Code

---

Note

Low Risk
Benchmark recipe and changelog only; no application or auth logic. Runner change triggers a full benchmark rerun on DSXE.

Overview
Updates the glm5.2-fp4-b300-sglang-agentic-mtp recipe so it can run on current hardware and a stable SGLang build.

Runner moves from retired cluster:b300-nv to cluster:b300-dsxe, so sweeps can be scheduled; this is not an append-only image bump—the agentic curve is expected to rerun on DSXE.

Container image moves from lmsysorg/sglang:v0.5.16-cu130 to lmsysorg/sglang:nightly-dev-cu13-20260908-20ca564b, aligning with other GLM-5.2 SGLang AgentX bumps and picking up a fix for intermittent MTP AttributeError on an unguarded KV index translator path in the prior 2026-09-07 nightly. The TP8 + HiCache + MTP search space and concurrency grid [1, 4, 8, 12, 16] are unchanged.

perf-changelog.yaml documents the change and notes overlap with open work that would extend the same key’s grid to concurrency 48.

^{Reviewed by Cursor Bugbot for commit f7ab351. 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

[claude]

Beyond the inline pr-link finding, I checked two other things that panned out fine: cluster:b300-dsxe is a valid, already-defined runner label in configs/runners.yaml, and cluster:b300-nv is indeed absent from that file (consistent with the claimed retirement). The new nightly-dev-cu13-<date>-<hash> image tag also matches the format of several other existing entries in nvidia-master.yaml, so it's not an invented tag pattern.

Extended reasoning...

A confirmed finding (perf-changelog.yaml pr-link uses the unaccepted placeholder "PRLINK_PLACEHOLDER") is already being posted as an inline comment, so this is a findings-present run. Beyond that, I independently verified two claims embedded in the PR description that were plausible risk points: (1) that cluster:b300-nv was actually retired and cluster:b300-dsxe is a real, valid runner label — confirmed by grepping configs/runners.yaml, which defines cluster:b300-dsxe with 18 nodes and contains no cluster:b300-nv entry; and (2) that the new nightly-dev-cu13-20260907-30705c00 tag follows an established naming convention rather than an invented one — confirmed by finding multiple pre-existing lmsysorg/sglang:nightly-dev-cu13-- entries elsewhere in nvidia-master.yaml. Both checks ruled out plausible additional concerns beyond the inline finding, so per the narrow exception I'm posting a brief defer note recording that, without restating the inline finding itself.

[claude]

🔴 pr-link uses the literal string "PRLINK_PLACEHOLDER", which is not an accepted placeholder for utils/validate_perf_changelog.py: it only allows "XXX", the full ".../pull/XXX" URL, or the exact expected ".../pull/" URL. This new entry will fail changelog validation on this PR's CI run. Fix: use "XXX" (or the canonical pull/XXX URL) as the placeholder, or fill in the real PR number/link before merge.

Extended reasoning...

validate_added_pr_link() in utils/validate_perf_changelog.py:144-159 checks, when pr_number is not None (a PR run), that link is in PR_LINK_PLACEHOLDERS = {"XXX", ".../pull/XXX"} or equals the expected ".../pull/<pr_number>" URL. "PRLINK_PLACEHOLDER" is neither, so it raises ChangelogValidationError("new PR entry must use ... or an XXX placeholder; found 'PRLINK_PLACEHOLDER'"), failing the perf-changelog validation check that CI/CONTRIBUTING.md require to pass before merge.

Verification: normal (stage mislabeled): The defect is real — "PRLINK_PLACEHOLDER" is not a recognized placeholder. PR_LINK_PLACEHOLDERS = {"XXX", "https://github.com/SemiAnalysisAI/InferenceX/pull/XXX"} (validate_perf_changelog.py:24-27), and validate_added_pr_link (lines 144-160) raises ChangelogValidationError for any link that is neither a placeholder nor the exact expected pull/ URL when…

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34177481357
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=34177481357

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34179022786
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=34179022786

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34179022786
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=34179022786