[AgentX] Refresh Kimi K3 DSpark on B300 / [AgentX] 刷新 B300 上的 Kimi K3 DSpark - #2476
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 |
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Looks good — a straightforward, correctly-formatted perf-changelog entry refreshing an existing AgentX submission. The only finding is a non-functional metadata nit (missing bilingual PR title), which doesn't block this.
Extended reasoning...
Overview
The PR adds a single 5-line entry to perf-changelog.yaml, referencing the existing kimik3-fp4-b300-vllm-agentic-dspark config key (confirmed present in configs/nvidia-master.yaml) to record a refreshed AgentX submission for B300 Kimi K3 DSpark. No search-space, code, or config-key changes — purely a changelog/metadata append.
Security risks
None. The change is a static YAML changelog entry with no executable logic, no new config keys, no credentials, and no changes to auth/permissions paths.
Level of scrutiny
Minimal scrutiny is appropriate here — this is a mechanical, additive changelog entry following the exact structural pattern of the immediately preceding entries in the file (config-keys/description/pr-link), referencing a pre-existing, already-defined config key.
Other factors
The bug hunting system's only finding is a nit about the PR title lacking a required Simplified Chinese translation per AGENTS.md convention — a process/metadata gap, not a functional defect, and consistent with several recent adjacent commits in the log. The other candidate issue raised (placeholder pr-link) was already ruled out as expected/consistent with existing entries. No outstanding unaddressed reviewer comments exist in the timeline.
中文：将 main 合并到 B300 分支
中文：使用新版 Kimi nightly 镜像测试 B300
中文：补齐 Kimi 基准测试变更日志条目的末尾换行。
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31407692525 |
中文：将 Kimi GPU 常驻基准测试的并发上限设为 8，并保留 DRAM 卸载在并发 16 的容量测试点。
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31409490573 |
中文：合并 main 并解决冲突
中文：合并 main 后保留性能变更日志末尾换行
中文：排除发生不可纠正 NVLink 错误的 B300 节点，避免基准测试继续调度到故障硬件。
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31444781885 |
中文：让 B300 启动器把 SALLOC_EXCLUDE 传递给 Slurm，避免任务调度到已知故障节点。
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31446383426 |
中文：合并最新 main，并按追加规则保留 B300 Kimi K3 基准测试变更日志条目。
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31447976375 |
中文：将 Kimi K3 B300 的 MLA 预填充后端对齐到上游 FlashInfer 配置，并保留真实 block 验证的评估路径。
adibarra
changed the title
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31509160691 |
1 similar comment
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31509160691 |
中文：将 main 合并到 B300 分支，并保留 Kimi K3 DSpark 配置及规范的性能变更日志顺序。
|
/stage-results 31509160691 |
|
@cquil11 staged run 31509160691: https://inferencemax-app-git-staging-semianalysisai.vercel.app/inference?i_dates=2026-08-11~r31509160691 This run remains available across future @cquil11 已将运行 31509160691 发布到预发布环境：https://inferencemax-app-git-staging-semianalysisai.vercel.app/inference?i_dates=2026-08-11~r31509160691 后续的 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31534110529 |
|
/reuse-sweep-run 31509160691 |
Remove the PR-specific Slurm exclusion changes and restore the unrelated benchmark comment so the PR stays focused on the Kimi K3 refresh.\n\n中文：移除本 PR 中与 Slurm 节点排除相关的改动，并恢复无关的基准测试注释，使 PR 仅聚焦于 Kimi K3 刷新。
2 participants
Summary
eval-onlyrow uses realrejection_sample_method=blockverification and no synthetic acceptance.Validation
eval-onlyDRAM-offload concurrency-16 row.pytest -q utils/matrix_logic/test_generate_sweep_configs.py -k 'multinode or agentic': 27 passed, 78 deselected.bash -n benchmarks/single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh,git diff --check, andutils/validate_perf_changelog.py.中文说明
eval-only配置使用真实rejection_sample_method=block验证，不使用合成接受率。验证
eval-only配置。pytest -q utils/matrix_logic/test_generate_sweep_configs.py -k 'multinode or agentic'：27 passed、78 deselected。bash -n benchmarks/single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh、git diff --check和utils/validate_perf_changelog.py。