[Klaud Cold] Add conc 10/12/14 to kimik3-fp4-b300-vllm-agentic-dspark and move it to cluster:b300-dsxe / 为 kimik3-fp4-b300-vllm-agentic-dspark 新增并发 10/12/14 并迁移至 cluster:b300-dsxe - #2846
Conversation
|
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 |
1 similar comment
|
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 |
Extend the TP8 x DCP8 Mooncake sweep with concurrency 10, 12 and 14. They land in the recipe's conc 9-16 band and draft with DSpark level 3 at the committed golden AL 3.00. Repoint the runner from the retired cluster:b300-nv (#2826) to cluster:b300-dsxe so the sweep can schedule. 为 kimik3-fp4-b300-vllm-agentic-dspark 的 TP8 x DCP8 Mooncake 扫描新增并发 10、12、14。这些点落在 recipe 现有的 conc 9-16 区间，以 DSpark level 3 和已提交的 golden AL 3.00 起草。同时将 runner 从已在 #2826 退役的 cluster:b300-nv 改为 cluster:b300-dsxe，使扫描可以调度。 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
25685b3 to
43b7be4
Compare
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Found a concrete issue: the new perf-changelog.yaml entry uses pr-link: PLACEHOLDER, but PR_LINK_PLACEHOLDERS in utils/validate_perf_changelog.py only recognizes "XXX" or "https://github.com/SemiAnalysisAI/InferenceX/pull/XXX" as placeholders (utils/validate_perf_changelog.py:24-27). Because main() never calls compare_entries/validate_added_pr_link, utils/validate_perf_changelog.py --base-ref ... --head-ref ... (what the author ran) does not catch this. But utils/merge_with_reuse.sh unconditionally runs prepare_perf_changelog_merge.py canonicalize at merge time (line ~186-190, not guarded by if, under set -euo pipefail), which calls compare_entries -> validate_added_pr_link("PLACEHOLDER", pr_number). Since "PLACEHOLDER" doesn't match either recognized token, this raises ChangelogValidationError and aborts the merge script. The pr-link field needs to be XXX (matching the pattern used everywhere else in the file, e.g. utils/prepare_perf_changelog_merge.py's own header comment: "re-appending the PR's entry at the bottom with XXX -> the canonical PR URL"), not the literal word PLACEHOLDER.
Extended reasoning...
Overview
This PR makes a narrow, mechanical change: it adds concurrency points 10/12/14 to an existing DSpark-level-3 arm in configs/nvidia-master.yaml (verified against benchmarks/single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh's actual CONC -le 16 branch, which already covered these values — only the comment was stale), repoints the recipe's runner label from a retired cluster (cluster:b300-nv, confirmed absent from configs/runners.yaml) to a live one (cluster:b300-dsxe, confirmed present), and appends one perf-changelog.yaml entry.
Security risks
None. This is a config/benchmark-recipe change with no new code paths, no secrets, and no auth/permission logic touched.
Level of scrutiny
Config-only PRs like this are normally straightforward-approve candidates, and the concurrency/runner changes here check out cleanly against the actual shell script logic and configs/runners.yaml. However, I traced the pr-link: PLACEHOLDER value through utils/validate_perf_changelog.py and utils/prepare_perf_changelog_merge.py and found it will break the automated merge-with-reuse tooling (utils/merge_with_reuse.sh's unconditional canonicalize step, under set -euo pipefail), since "PLACEHOLDER" is not one of the two literal tokens ("XXX" or the full pull/XXX URL) the validator's PR_LINK_PLACEHOLDERS set recognizes. This is a concrete, mechanically-verified defect that the author's own claimed validation command does not exercise (since main() in validate_perf_changelog.py never calls compare_entries), so it would only surface at merge time. That's enough to warrant a human/author fix-up before merging, even though the rest of the diff is clean.
Other factors
No test coverage gap in the traditional sense — the existing utils/changelog_gate_tests/test_validate_perf_changelog.py and test_prepare_perf_changelog_merge.py already assert this exact placeholder-matching behavior (e.g. test_compare_entries_rejects_wrong_pr_link_on_append), it's just that this specific PR's local validation command doesn't invoke that code path, so the mismatch wasn't caught before submission.
This review covers commit 25685b3, which is no longer the latest commit on this pull request; later commits are not covered by it.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33996641472 |
1 participant
Summary / 摘要
Adds concurrency 10, 12 and 14 to the
kimik3-fp4-b300-vllm-agentic-dsparkTP8 x DCP8 Mooncake sweep and repoints the config to the live B300 cluster.为
kimik3-fp4-b300-vllm-agentic-dspark的 TP8 x DCP8 Mooncake 扫描新增并发 10、12、14，并将该配置指向当前在线的 B300 集群。Changes / 变更
configs/nvidia-master.yaml: conc-list[1, 2, 4, 8, 16, 24, 32, 40, 48, 56, 70]->[1, 2, 4, 8, 10, 12, 14, 16, 24, 32, 40, 48, 56, 70]on the{ tp: 8, ep: 1, dcp-size: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" } }entry. Runnercluster:b300-nv->cluster:b300-dsxe: the NV launcher and its runner labels were 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, so the sweep could not schedule on the old label.benchmarks/single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh: comment-only update of the arm table. No serve-flag or logic change. The new points fall in the existingCONC <= 16branch and draft with DSpark level 3 at the committed golden AL 3.00, the same arm as concurrency 16.perf-changelog.yaml: appended entry forkimik3-fp4-b300-vllm-agentic-dspark.configs/nvidia-master.yaml：在{ tp: 8, ep: 1, dcp-size: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" } }条目上，conc-list 从[1, 2, 4, 8, 16, 24, 32, 40, 48, 56, 70]改为[1, 2, 4, 8, 10, 12, 14, 16, 24, 32, 40, 48, 56, 70]。runner 从cluster:b300-nv改为cluster:b300-dsxe：NV launcher 及其 runner 标签已在 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 退役，旧标签上的扫描无法调度。benchmarks/single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh：仅更新注释中的 arm 表，不改动 serve 参数或逻辑。新增的并发点落在现有CONC <= 16分支，以 DSpark level 3 和已提交的 golden AL 3.00 起草，与并发 16 同一 arm。perf-changelog.yaml：追加kimik3-fp4-b300-vllm-agentic-dspark条目。Validation / 验证
utils/validate_perf_changelog.py --base-ref origin/main --head-ref HEADpasses.generate_sweep_configs.py test-config --config-keys kimik3-fp4-b300-vllm-agentic-dsparkyields 14 rows oncluster:b300-dsxe, includingconc10,conc12andconc14.Full sweep requested via the
full-sweep-enabledlabel.utils/validate_perf_changelog.py --base-ref origin/main --head-ref HEAD通过。generate_sweep_configs.py test-config --config-keys kimik3-fp4-b300-vllm-agentic-dspark生成 14 行，全部在cluster:b300-dsxe上，包含conc10、conc12、conc14。通过
full-sweep-enabled标签触发完整扫描。🤖 Generated with Claude Code
Note
Low Risk
Benchmark sweep and scheduling metadata only; no serving, auth, or runtime logic changes beyond comments.
Overview
Extends the kimik3-fp4-b300-vllm-agentic-dspark agentic-coding sweep with mid-band concurrency points and points it at a schedulable B300 cluster.
The TP8 × DCP8 Mooncake entry in
configs/nvidia-master.yamlnow uses conc-list[1, 2, 4, 8, 10, 12, 14, 16, 24, …, 70](adds 10, 12, 14) and runnercluster:b300-dsxeinstead of retiredcluster:b300-nv. Comments in the config andkimik3_fp4_b300_vllm_mtp.shdescribe the DSpark level-3 arm as conc 10–16; serve logic is unchanged and still mapsCONC <= 16to DSpark level 3 / golden AL 3.00.perf-changelog.yamlrecords the sweep expansion and runner repoint for this config key.Reviewed by Cursor Bugbot for commit 43b7be4. Bugbot is set up for automated code reviews on this repo. Configure here.