[cquil11]

Adds three points to Kimi-K3.1-Pro GB200 SGLang 1P1D AgentX from #4804 at conc 64, 76 and 96, using an EP16 decode on DeepGEMM MegaMoE. This is append-only: the image and the existing c16, c32 and c48 points are unchanged.

• Decode: EP16 with moe-a2a-backend: megamoe (MXFP8 dispatch, BF16 combine). Sharding the experts frees about 5 GB of weights per GPU, so the decode KV pool grows from 878k to 1.18M tokens per rank.
• Batch limit: without the megakernel (SM103/107 only), DCP16 merges indexer top-k through an fp32 buffer of 4 MiB per verify row. That caps decode at 96 running requests and CUDA graphs at batch 96. The MegaMoE token cap is sized to 96 × 8 draft tokens.
• Prefill: the new points use page size 64 and 8 draft tokens to match the decode.

Test run

Run 42201408790 ran the same recipe at c76. Compared with #4804's sweep:

Point	Total tput/GPU	Interactivity p50 (tok/s/user)	TTFT p90
#4804 agg c32	8,891	69.7	1.95 s
#4804 agg c56	10,567	36.9	4.05 s
#4804 1P1D c48	7,868	99.9	3.72 s
1P1D EP16 MegaMoE c76	11,324	69.9	3.10 s

[github-actions]

Thanks for the contribution!

• Review: If this PR changes files owned by someone other than a repository admin or @SemiAnalysisAI/core, ask one eligible CODEOWNER to complete the latest PR_REVIEW_CHECKLIST.md before contacting a core maintainer on Slack. Follow the template exactly, including As a PR reviewer and CODEOWNER, I have reviewed this and have, so sign-off verification triggers.
• PR verification: Sweeps run only when the PR appends an inferencex-e2e/perf-changelog.yaml entry and carries exactly one primary label: full-sweep-fail-fast (strongly recommended; canary plus per-matrix fail-fast), full-sweep-enabled (canary; matrix jobs continue after a failure), or non-canary-full-sweep-enabled (no canary or fail-fast). The modifiers all-evals, evals-only, and agentx-fast require a primary label. On fork PRs, a maintainer applies the label. See sweep labels and reuse.
• After merging: PR authors must ensure all GitHub Actions jobs pass. Transient failures often pass on rerun; see how to rerun failed jobs.

中文

感谢你的贡献！

• **审阅：**如果 PR 修改的文件归属于仓库管理员及 @SemiAnalysisAI/core 之外的 CODEOWNER，请先联系一位有资格的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，再通过 Slack 联系核心维护者。必须严格遵循模板，并保留 As a PR reviewer and CODEOWNER, I have reviewed this and have，才能触发签核验证。
• **PR 验证：**只有当 PR 在 inferencex-e2e/perf-changelog.yaml 末尾追加了条目，并且恰好带有一个主标签时，才会运行扫描：full-sweep-fail-fast（强烈推荐；canary 加逐矩阵 fail-fast）、full-sweep-enabled（有 canary；矩阵任务在失败后继续运行）或 non-canary-full-sweep-enabled（无 canary，也无 fail-fast）。修饰标签 all-evals、evals-only 和 agentx-fast 必须与主标签一起使用。fork PR 的标签由维护者添加。参见扫描标签与复用。
• **合并后：**PR 作者必须确保所有 GitHub Actions 任务通过。临时性失败通常可以通过重新运行恢复；参见重新运行失败任务的说明。

[claude]

Nothing blocking. The comments below are optional suggestions. There is no need to push a fix for them before merging.

[claude]

🟡 nit (optional): The comment above the new override is a header with past results, which AGENTS.md comment rules forbid. It describes the whole override and quotes the measured KV pool change "from 878k to 1.18M tokens". Those numbers are history that belongs in the PR or perf-changelog.yaml. Fix: drop the header, or keep one short line on the line it explains (for example, why expert-parallel-size: 16 is set) without the measured before/after figures.

Why this was flagged

AGENTS.md:56-58, which CLAUDE.md delegates to, says "No header comments describing a file, topology, config, recipe, override, or section" and "No history: ... or past results". The new lines at disagg-variants.yaml:216-217 sit above override_1p1d_tp16_ep16 and describe the override as a whole. They also record a measured result, the decode KV pool growing "from 878k to 1.18M tokens per rank". That figure will go stale as soon as the image or memory settings change. The other overrides in this file (c16, c32, c48) have no header comment. The base branch has no such comment.

Verification: disagg-variants.yaml:216-217 adds a top-level comment directly above override_1p1d_tp16_ep16: (line 218). AGENTS.md:56 forbids "header comments describing a file, topology, config, recipe, override, or section", and AGENTS.md:57 forbids "past results". The comment describes the whole override and records measured KV pool figures (878k to 1.18M). Nothing breaks at runtime.

[github-actions]

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=42530422785

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=42530422785

[cquil11]

/use 42530422785

[github-actions]

@cquil11 staged run 42530422785.

• Kimi-K3.1-Pro · agentic-traces · FP4 · gb200_dynamo-sglang · 2026-11-08

This run remains available across future /stage-results requests. Staging the same run ID again updates its staged data. Staging workflow

[cquil11]

/reuse-sweep-run 42530422785

[github-actions]

Ingest report — ✅ Results ingested into the production database.

Ingest workflow · Merge Ingest run · Benchmark source run

Category	new	duplicate	skipped	unmapped	failed
benchmark	3	0	0	0	0
eval	0	0	0	0	0
trace	3	0	0	0	0
power	24	0	0	0	0

_{benchmarkFiles 3, rawBenchmarkRows 3, coveredByPerJobFiles 0, purgedPoints 0, failedRunRows 0, runStatsNew 0, runStatsDuplicate 0, evalSamplesNew 0, gpuMetricSamplesNew 587952, powerPublicationPoints 3, changelogEntriesWritten 1}

Counts cover only this run. Rows: benchmark_results, eval_results, agentic rows linked to a trace, gpu_metrics series. This comment updates if the ingest is rerun.

中文

Ingest 报告 — ✅ 结果已写入 production 数据库。

上表仅统计本次运行写入的数据，各行单位依次为 benchmark_results 行、eval_results 行、已关联 trace 的 agentic 行和 gpu_metrics 序列。重新运行 ingest 时本评论会原地更新。