Update Kimi K3 GB300 Agentx

[wzhao18]

Note

Low Risk
Changes are benchmark/Slurm recipe and sweep metadata only; they affect how large GB300 jobs are launched and compared, not application runtime code paths in this repo.

Overview
Refreshes Kimi-K3 GB300 agentic-coding Slurm recipes and the matching nvidia-master sweep entries to track the measured global p90-ITL Pareto front.

All affected recipes move to vllm/vllm-openai:nightly-dev-arm64-cu13-3696c77, switch Dynamo routing from random to least-loaded, and align Mooncake with 160GB segments (dram-utilization 0.1775), default_kv_lease_ttl=60000, and richer MooncakeStoreConnector extras (compact_group_io, max_load_batch_keys). DSpark speculative JSON drops draft_sample_method: probabilistic across agg and disagg configs.

Disaggregated recipes share a tightened vLLM profile: explicit dcp-comm-backend: a2a, gpu-memory-utilization: 0.92, decode max-cudagraph-capture-size: 512, and removal of several decode-only limits (e.g. max-num-seqs, compilation-config, prefill max_num_batched_tokens). The 1p3d topology is corrected to 1 prefill / 3 decode workers (6 decode nodes); disagg-gb300-1p3d-dcp8-dcp8-dspark4 replaces a misnamed 1p1d DSpark-7 recipe, and a new disagg-gb300-1p3d-dcp8-dcp8-dspark7 recipe is added.

nvidia-master expands disagg concurrency grids (e.g. 1P1D 48–56, 1P2D 32–64), adds 1P3D points for DSpark-4 and DSpark-7 with synthetic acceptance lengths, trims the no-spec agg list (drops c74), and records the change in perf-changelog.yaml.

^{Reviewed by Cursor Bugbot for commit f6984c1. Bugbot is set up for automated code reviews on this repo. Configure here.}

[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]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33699200986
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=33699200986

[claude]

Additional findings (no inline location):

• 🔴 perf-changelog.yaml — New entry uses pr-link: TBD, which is not a valid placeholder: the CI gate (utils/validate_perf_changelog.py PR_LINK_PLACEHOLDERS = {"XXX", ".../pull/XXX"}) only accepts "XXX" or the exact PR URL, so this fails validate_added_pr_link and the changelog gate blocks the PR merge, unlike prior entries that all use the accepted convention. Fix: use "XXX" as the placeholder (or the actual PR URL) so it matches PR_LINK_PLACEHOLDERS / the expected pull/ link.

  Extended reasoning...

  utils/validate_perf_changelog.py:validate_added_pr_link() checks link not in PR_LINK_PLACEHOLDERS and link != expected where PR_LINK_PLACEHOLDERS = {"XXX", "https://github.com/SemiAnalysisAI/InferenceX/pull/XXX"} (lines 25-26, 154-159). The new entry appended at perf-changelog.yaml:1082 sets pr-link: TBD, which matches neither placeholder nor the expected pull/<pr_number> URL, so validate_added_pr_link raises ChangelogValidationError('new PR entry must use ... or an XXX placeholder; found 'TBD''), causing the changelog CI gate to fail on this PR until the value is corrected.

  Verification: normal. The diff appends a new changelog entry ending in pr-link: TBD (perf-changelog.yaml diff hunk after line 6831). utils/validate_perf_changelog.py:205-206 iterates every appended entry and calls validate_added_pr_link(str(entry.get("pr-link") or ""), pr_number). In validate_added_pr_link (lines 144-160): on a PR run, link not in PR_LINK_PLACEHOLDERS and link != expected is true for…

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33699230099
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=33699230099

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33700071162
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=33700071162

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit bce55d5. Configure here.}

[cursor]

Changelog key missing from master

High Severity

The new changelog entry lists kimik3-fp4-gb300-dynamo-vllm-agentic-dspark-mooncake-dcp8-agg, but that key is not in nvidia-master.yaml. process_changelog.py raises when a changelog key is missing from master configs, so sweep generation for this PR fails.

 

^{Reviewed by Cursor Bugbot for commit bce55d5. Configure here.}