[Klaud Cold] kimik2.5-fp4-b200-vllm: nightly image, TP/DEP/TEP sweep / Kimi K2.5 NVFP4 B200 nightly 镜像，TP/DEP/TEP 扫描

[xinli-sw]

Summary

Kimi K2.5 NVFP4 B200 vLLM with nightly image: extended sweep space with TP8, TP4, TEP8, TEP4, and DEP8 arms.

中文说明

Kimi K2.5 NVFP4 B200 vLLM nightly 镜像：扩展 TP8、TP4、TEP8、TEP4 及 DEP8 配置扫描空间。

[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]

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

[claude]

🟡 The new perf-changelog.yaml entry for kimik2.5-fp4-b200-vllm links to pull/2443, but this change is actually part of PR #2512. Every other entry in the file links to the PR that introduced it, so this looks like a copy-paste/typo (2443 vs 2447 from the preceding entry). Please update the pr-link to point to #2512.

Extended reasoning...

The newly added perf-changelog.yaml entry (lines 5521-5523) for kimik2.5-fp4-b200-vllm sets:

pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2443

But the PR metadata for this change unambiguously identifies it as PR #2512 ([Klaud Cold] kimik2.5-fp4-b200-vllm: nightly image, TP/DEP/TEP sweep).

Convention in this file: every prior entry links to the PR that actually introduced it, and this is verifiable against git history. The immediately preceding entry (GLM-5.2 B300 AgentX MTP) links to #2447, which matches commit bb8e26d ... (#2447). Entries above it link to #2492, #2390, and #2099, which likewise match commits d4363bd (#2492), 4689d9b (#2390), and 7667f21 (#2099) in the recent git log. This is a consistent, load-bearing convention — the changelog is used to trace which PR/commit introduced a given config change.

Root cause: this is almost certainly a mistyped digit — 2443 vs the correct 2512, and notably very close to 2447 (the preceding entry's PR number), suggesting a copy/paste-and-edit slip where only part of the number was updated.

Step-by-step proof:

1. PR metadata states <pr number="2512">.
2. The diff adds a new changelog block for kimik2.5-fp4-b200-vllm with pr-link: .../pull/2443.
3. Cross-checking the file's own convention: the entry directly above it (GLM-5.2 B300 AgentX) has pr-link: .../pull/2447, which matches the commit that introduced it (bb8e26d ... (#2447) in git log).
4. Therefore the new entry's pr-link should read .../pull/2512, not .../pull/2443 — 2443 refers to an unrelated, already-merged PR and misattributes this config change to it.

Impact: This is purely a metadata/attribution issue in a changelog file — it does not affect config resolution, the benchmark sweep, or any runtime behavior. Nothing will break at merge time. It's a one-line fix (2443 -> 2512) but doesn't block merging, since it degrades traceability rather than functionality.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31147925932
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=31147925932

[xinli-sw]

/reuse-sweep-run

[Ankur-singh]

/reuse-sweep-run 31147925932

[Ankur-singh]

As a PR reviewer and CODEOWNER, I have reviewed this and have:

• Verified that as of the moment of typing this, this is the latest version of PR_REVIEW_CHECKLIST.md
• Verified that the general code quality meets the InferenceX standard and does not make the code quality any worse.
• Verified that this PR has passed PR validation. Please link to GitHub Action workflow that shows this. — https://github.com/SemiAnalysisAI/InferenceX/actions/runs/31147925932
• Verified that this PR passes evals. Please link to GitHub Action workflow that shows this. — https://github.com/SemiAnalysisAI/InferenceX/actions/runs/31147925932
• Verified that speculative decoding PRs uses chat templates to align the AL distribution to real world
• For agentic workloads: verified that speculative-decoding configs (EAGLE / MTP / draft models) run with simulated synthetic acceptance, with the acceptance-length value taken from the committed golden AL curve in golden_al_distribution/ for that model, thinking mode, and draft length. A submission may choose any supported draft length, but it may not substitute a different acceptance target.
• Verified against the current MODELS.md that this PR does not submit a deprecated model, scenario, or model-scenario combination.
• Verified that the model architecture isn't changed with benchmark hacks like using --hf-overrides to skipping indexer for every x layers on models that don't natively support this. As a general rule, we won't accept optimizations that reduces the number of model architecture FLOPs. Anything that makes that same computation run faster is fair game; FLOPs at lower precisions is fine, given that the config passes private evals. As an general north star princple, we should only use optimizations which is used in production by customers that care about accuracy
• If an company claims that they support vLLM/SGLang as first class LLM inference engines on their hardware, I have verified that the respective vLLM submission made using upstream https://hub.docker.com/u/vllm docker repo, upstream SGLang https://hub.docker.com/u/lmsysorg docker repo. The only exceptions are for new hardware, such as MI455X UALoE72, Vera Rubin NVL72, Rubin NVL8, etc., and for new model architectures where there is an actual reason why vLLM/SGLang does not fundamentally support them yet as supported by vLLM/SGLang community maintainers
• If an company claims that they support vLLM/SGLang as first class upstream in-tree LLM inference engines on their hardware, I have have verified that the respective vLLM/SGLang submission has been made before additional frameworks (TRT-LLM, ATOM, etc.). The only exceptions are for new hardware, such as MI455X UALoE72, Vera Rubin NVL72, Rubin NVL8, etc., and for new model architectures where there is an actual reason why vLLM/SGLang does not fundamentally support them yet.
• Verified that every single-node vLLM/SGLang recipe in this PR is documented in the official vLLM recipes and/or the SGLang cookbook:
  • I linked the corresponding upstream PR in the vLLM recipe repo or SGLang repo and verified that it is MERGED before this InferenceX PR merges. An opened, draft, or closed-without-merge upstream PR does not satisfy this requirement. If the matching recipe was already published, I linked the published recipe/cookbook page in the additional detail section below.
• Verified that this PR does not patch the inference engine or serving stack — the pinned image must run as shipped. This covers .patch files / git apply / patch, inline patches embedded in benchmark scripts (e.g. a python3/sed heredoc that rewrites installed engine sources before serving), in-place edits of site-packages, monkey-patching, overwriting container files, and installing forked/rebuilt engine wheels on top of the pinned image. The only exception is a patch covered by a filled-out waiver at docs/waiver/<PR_NUMBER>.md — named after the PR that introduces the patch and filed in that same PR, stating what is patched, why the unmodified upstream image cannot run this benchmark, the upstream PR/issue link, and the removal plan — which I have linked below in the additional detail section.
• If any of the above criteria cannot reasonably be satisfied, I have provided additional reasoning below.

Additional detail section:

• Scope. kimik2.5-fp4-b200-vllm moves to vllm/vllm-openai:nightly-e2fa28594f7baad142a426b0b6a2cfe2c79201c7 and replaces the two-point search space with five arms: TP8/EP1 and TP4/EP1 at conc 1–512, TP8/EP8 and TP4/EP4 tensor-EP arms at conc 1–512, and a TP8/EP8 attention-DP arm at conc [512, 1024]. Reviewed at head 8d034aa96d340b569de06f4e45886179210f00ed.
• Validation and evals. Run 31147925932 executed at in-PR commit 11b75f153 and is a real sweep, not a reuse no-op: setup = success with reuse-sweep-gate skipped. 42/42 single-node 8k1k and 9/9 eval lanes concluded success, with collect-results, collect-evals, compare-results and calc-success-rate all green and zero failures.
• Why the evidence is linked at an in-PR commit rather than the tip. The head moved after that run because I fixed two things in this PR (below). The submission itself — the benchmark script and the kimik2.5-fp4-b200-vllm entry — is byte-identical to what run 31147925932 measured; the only later changes are the changelog pr-link correction and a merge of main. /reuse-sweep-run 31147925932 is pinned on this PR, and the resulting head run 31204430170 consumed that authorization (reuse-sweep-gate = success, setup skipped) with check-changelog passing on the corrected file.
• Two defects I fixed in this PR while reviewing.
  1. The perf-changelog.yaml entry carried pr-link: .../pull/2443 — a different PR. Corrected to 2512.
  2. perf-changelog.yaml had drifted: main gained several entries after this branch was cut, so the file was in conflict. Resolved by taking main's copy verbatim and re-appending this PR's entry last, which keeps the diff against main additions-only (6/0) and makes it impossible to drop an entry that main owns. The PR is now MERGEABLE.
  3. A bare /reuse-sweep-run had been posted. That command is not one-shot — utils/find_reusable_sweep_run.py re-reads it on every synchronize, so it would have set skip-pr-sweep=true for every future push on this PR. It is now superseded by the explicit pin above.
• MODELS.md (unchecked, and the reason stated plainly). This PR submits kimik2.5 Single-turn 8k1k. The current MODELS.md on main still reads "Single-turn 8k1k (until 2026-08-06) — fully retired after 2026-08-06", and today is 2026-08-07. I am told by the maintainers that the kimik2.5 deadline has been extended, and I am signing off on that basis — but the extension has not yet landed in MODELS.md, so I cannot honestly tick a box that says the current document permits this combination. This item should be re-read once the MODELS.md update merges; if the extension is not in fact granted, this submission should not merge.
• Upstream recipe (unchecked). This is a single-node vLLM recipe, so the requirement applies. The published Kimi-K2.5 vLLM recipe does document the NVFP4-on-Blackwell serving shape this config builds on, including vllm serve nvidia/Kimi-K2.5-NVFP4 with tensor parallelism, --enable-expert-parallel, and --data-parallel-size. It does not document the backend flags this PR introduces — --attention-config '{"mla_prefill_backend":"FLASHINFER","use_prefill_query_quantization":true}', --linear-backend flashinfer_cutlass, and VLLM_USE_V2_MODEL_RUNNER=0 — nor a nightly image pin. Those are MAJOR serving arguments, so I am leaving this item unchecked rather than arguing the published page covers them. An upstream recipe PR adding the nightly/FlashInfer arm would close it.
• Speculative decoding and agentic golden AL (unchecked). Both not applicable: there is no speculative decoding anywhere in kimik2.5_fp4_b200.sh (no draft model, no speculative-* flags, no SIMULATE_ACC knobs), and this is a single-turn fixed-seq-len 8k1k submission, not an agentic workload.
• Model architecture. No --hf-overrides, model-override arguments, or layer/expert trimming. The changes are parallelism selection (TP vs attention-DP), expert parallelism, kernel backends, GPU-memory utilisation (0.90, or 0.80 on the DP-attention arm), CUDA-graph capture sizing and stream interval — none of which reduce model-architecture FLOPs.
• Upstream image and framework ordering. vllm/vllm-openai:nightly-e2fa28594f7baad142a426b0b6a2cfe2c79201c7 comes from the official https://hub.docker.com/u/vllm org — a nightly tag, but not a fork or vendor rebuild. vLLM is the serving engine here, so engine-first ordering is satisfied by construction.
• No engine patching. No .patch files, git apply, sed -i rewrites, site-packages edits, monkey-patching, or forked/rebuilt engine wheels; the pinned image runs as shipped and no waiver is needed. The only pip install in the script is datasets pandas for the client-side eval path, not an engine wheel.

Signed: Ankur-singh

[Klaud-Cold]

❌❌❌ REJECTED ❌❌❌

@Ankur-singh — two blockers: the kimik2.5 Single-turn 8k1k scenario is retired as of today per the current MODELS.md (last day was 2026-08-06; the claimed extension has not landed), and the PR's major FlashInfer backend args are not documented in any merged/published upstream recipe. The sign-off itself flagged both honestly, but neither meets the merge standard yet.

✅ Check 0 (CODEOWNER): PASS — Ankur-singh owns configs/nvidia-master.yaml in CODEOWNERS; catch-all paths covered by any recognized CODEOWNER.
✅ Check 1 (sweep on in-PR commit): PASS — in-PR commit 11b75f153 has 42/42 single-node 8k1k / and 9/9 eval / check-runs with conclusion success from executed run 31147925932; head 8d034aa matches the pinned sign-off SHA.
✅ Check 2 (evals pass): PASS — downloaded agg_eval_all.json from run 31147925932: 9 GSM8K results (all five arms incl. TP8/EP8-DPA c1024) scoring 0.969–0.974, run on the same vllm/vllm-openai:nightly-e2fa28594f… image as this PR's config.
❌ Check 3 (upstream recipe): FAIL — the linked published Kimi-K2.5 recipe covers the NVFP4 serve shape (model, TP, EP, fuse_allreduce_rms) but does NOT document this PR's major kernel-selection args: --attention-config '{"mla_prefill_backend":"FLASHINFER","use_prefill_query_quantization":true}', --linear-backend flashinfer_cutlass, and VLLM_USE_V2_MODEL_RUNNER=0; no merged upstream recipe PR adding them is linked. A merged vllm-project/recipes PR documenting the nightly/FlashInfer arm closes this. (GMU, --stream-interval, cudagraph sizing are InferenceX-specific tuning — informational only.)
✅ Check 4 (reuse command): PASS — /reuse-sweep-run 31147925932 posted by Ankur-singh (COLLABORATOR); head run 31204430170 consumed it.
✅ Check 5 (latest checklist): PASS — sign-off contains every item of the current docs/PR_REVIEW_CHECKLIST.md template; each unchecked item is explained in the additional detail section.
✅ Check 6 (upstream image / engine-first): PASS — vllm/vllm-openai:nightly-e2fa28594f… is from the upstream vLLM Docker Hub org; framework is vLLM itself, so ordering is satisfied.
❌ Check 7 (deprecated models): FAIL — this PR re-tunes kimik2.5-fp4-b200-vllm (kimik2.5, Single-turn 8k1k), but current MODELS.md states Single-turn 8k1k on kimik2.5 ends 2026-08-06 and the model is "fully retired after August 6, 2026"; review date is 2026-08-07. The extension the sign-off relies on is not in MODELS.md — land that update first, then re-verify.
✅ Check 8 (no architecture hacks): PASS — no --hf-overrides or FLOPs-reducing overrides; changes are parallelism, kernel backends, and memory/scheduling tuning (prefill query quantization is precision, not architecture, and evals pass).
➖ Check 9 (spec-decode chat templates): N/A — no speculative-decoding flags anywhere in this PR.
✅ Check 10 (no engine patches): PASS — no patch files, heredoc rewrites, site-packages edits, or engine wheel installs; the only pip install is client-side datasets pandas.
➖ Check 11 (agentic golden AL): N/A — no agentic speculative-decoding changes; no synthetic-acceptance knobs on this non-agentic config.