[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
Conversation
Add concurrency points 20, 24, 28, 32, 40, and 48 to the TP8 HiCache MTP arm of glm5.2-fp4-b300-sglang-agentic-mtp, and move the recipe from the retired cluster:b300-nv fleet to cluster:b300-dsxe so the sweep can be generated and scheduled again. 为 glm5.2-fp4-b300-sglang-agentic-mtp 的 TP8 HiCache MTP 分支新增并发度 20、24、28、32、40、48，并将配方从已下线的 cluster:b300-nv 迁移到 cluster:b300-dsxe，使 sweep 能重新生成与调度。 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
|
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 |
补全 perf-changelog 条目中的 PR 链接（#2829）。 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
![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 that cluster:b300-dsxe is a real, already-registered runner (configs/runners.yaml, runners/launch_b300-dsxe.sh) and that the recipe script benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang_mtp.sh it depends on exists — the runner migration itself is not fabricated.
Extended reasoning...
This is a small, mechanical config change (two lines in configs/nvidia-master.yaml plus one appended perf-changelog.yaml entry), and a CONFIRMED finding was already reported inline about the doc comment above the recipe going stale (it still says "hard stop at conc 16" and cites 48 as "well above that cap" as justification for omitting a DEP arm, which no longer holds once conc-list itself reaches 48). Given that finding, I did not restate it and only add what I separately verified: cluster:b300-dsxe is a genuine, pre-existing runner fleet defined in configs/runners.yaml and referenced by runners/launch_b300-dsxe.sh and other benchmark scripts, and the target recipe script glm5.2_fp4_b300_sglang_mtp.sh exists on disk, so the runner migration itself is legitimate infrastructure and not a broken reference. This does not resolve the outstanding stale-comment finding, so a human should still take a final look before merge.
| image: lmsysorg/sglang:v0.5.16-cu130 | ||
| model: nvidia/GLM-5.2-NVFP4 | ||
| model-prefix: glm5.2 | ||
| runner: cluster:b300-nv | ||
| runner: cluster:b300-dsxe | ||
| precision: fp4 | ||
| framework: sglang | ||
| multinode: false | ||
| scenarios: |
There was a problem hiding this comment.
🟡 (optional) The doc comment above this recipe still says the concurrency grid has 'a hard stop at conc 16' and justifies omitting a DEP throughput arm because its frontier peak (48) is 'well above that cap' — but this diff raises conc-list to 48, so the stated cap is now exactly the DEP arm's peak and the omission rationale no longer holds. Fix: update the comment to reflect the new [1..48] grid and re-justify (or reconsider) excluding the DEP arm now that 48 is in-range, not just the conc-list value.
Extended reasoning...
Lines 8980-8989 explain the design rationale for this recipe; line 8991-8998 is the recipe itself. The diff only touches the runner and conc-list (lines 8994, 8998) but leaves the rationale comment unchanged, so a reader (or future engineer using the comment to decide whether to add a DEP arm) is told the grid stops at 16 and that DEP's peak (48) is safely above the cap, when in fact the recipe now measures exactly conc 48 without a DEP baseline for comparison — misleading future changes to this recipe.
Verification: nit. The finding is factually accurate but is a stale-comment issue, not a functional defect. The recipe's design-rationale comment (configs/nvidia-master.yaml:8980-8989) still reads "on the AgentX MTP concurrency grid [1, 4, 8, 12, 16]" (8981), "a hard stop at conc 16" (8983), and justifies omitting a DEP arm because "its measured frontier peak is conc 48, well above that cap"…
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33924247919 |
1 participant
Summary / 摘要
Extend the GLM-5.2 NVFP4 B300 SGLang AgentX MTP curve (
glm5.2-fp4-b300-sglang-agentic-mtp) from concurrency 16 to 48. The TP8 HiCache arm grows from[1, 4, 8, 12, 16]to[1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48], following the grid the Qwen3.5 B200/B300 SGLang AgentX recipes already use (steps of 4 through 32, then 8).将 GLM-5.2 NVFP4 B300 SGLang AgentX MTP 曲线（
glm5.2-fp4-b300-sglang-agentic-mtp）的并发度从 16 扩展到 48。TP8 HiCache 分支由[1, 4, 8, 12, 16]扩展为[1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48]，网格与 Qwen3.5 B200/B300 SGLang AgentX 配方一致（32 以内步长 4，之后步长 8）。Runner move / 运行集群迁移
#2826 retired
launch_b300-nv.shand its runner labels, so every config still pinned tocluster:b300-nv(13 onmain, this one included) fails matrix generation withRunner 'cluster:b300-nv' requires 'available-cpu-dram-mib' in runner hardware metadata. This PR moves the recipe tocluster:b300-dsxe, the replacement fleet registered in #2826 and already used by #2814 for Kimi K3 B300 AgentX. Because the runner is a config-wide field, this entry is a regular full-sweep entry rather thanappend-only, so the whole 11-point curve is re-measured on DSXE.agentic/glm5.2_fp4_b300_sglang_mtp.shthrough its existing<prefix>_<precision>_b300_<framework>_mtp.shrule and exportsMODEL_PATH, which the script already honors.#2826 已下线
launch_b300-nv.sh及其 runner 标签，因此所有仍指向cluster:b300-nv的配置（main上共 13 个，包括本配置）在矩阵生成时报错Runner 'cluster:b300-nv' requires 'available-cpu-dram-mib' in runner hardware metadata。本 PR 将配方迁移到cluster:b300-dsxe，即 #2826 注册的替代集群，#2814 的 Kimi K3 B300 AgentX 也已使用。由于 runner 是配置级字段，本条目为常规全量 sweep 条目而非append-only，整条 11 点曲线将在 DSXE 上重新测量。<prefix>_<precision>_b300_<framework>_mtp.sh规则解析到agentic/glm5.2_fp4_b300_sglang_mtp.sh，并导出脚本已支持的MODEL_PATH。Changes / 改动
configs/nvidia-master.yaml—runner: cluster:b300-dsxe;conc-listextended to 48.perf-changelog.yaml— appended entry.configs/nvidia-master.yaml：runner: cluster:b300-dsxe；conc-list扩展至 48。perf-changelog.yaml：追加条目。Verification / 验证
generate_sweep_configs.py test-configrenders 11 jobs at[1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48]oncluster:b300-dsxe, all steps ≥ 2.utils/process_changelog.py --base-ref origin/main --head-ref HEADaccepts the entry and schedules the 11 benchmark jobs plus one agentic eval.generate_sweep_configs.py test-config在cluster:b300-dsxe上生成 11 个任务，并发度[1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48]，步长均 ≥ 2。utils/process_changelog.py --base-ref origin/main --head-ref HEAD接受该条目，调度 11 个基准任务和 1 个 agentic eval。Sibling PR extends the B200 curve to 48 on the same grid. This is the first single-node config on
cluster:b300-dsxe, so the sweep also exercises that launcher's single-node path for the first time.同系列 PR 将 B200 曲线以相同网格扩展至 48。本 PR 是
cluster:b300-dsxe上首个单节点配置，因此本次 sweep 也将首次验证该启动脚本的单节点路径。🤖 Generated with Claude Code
Note
Low Risk
YAML-only benchmark and changelog updates with no application, auth, or data-path changes; main operational effect is more GPU jobs and a cluster migration for one recipe.
Overview
Updates the
glm5.2-fp4-b300-sglang-agentic-mtpbenchmark recipe so the agentic-coding TP8 HiCache arm sweeps concurrency through 48 instead of stopping at 16, adding points at 20, 24, 28, 32, 40, and 48 (aligned with other B200/B300 SGLang AgentX grids).Because
cluster:b300-nvwas retired, the reciperunneris switched tocluster:b300-dsxeso matrix generation and scheduling work again; that runner change implies a full re-measure of the expanded 11-point curve on the new fleet, not an append-only delta.perf-changelog.yamlrecords the config key, scenario, and PR link for the sweep.Reviewed by Cursor Bugbot for commit 778a004. Bugbot is set up for automated code reviews on this repo. Configure here.