[chore] bump kimi-k2.5 int4 b200/b300 vllm images to v0.25.0 / [chore] 将 kimi-k2.5 int4 b200/b300 vLLM 镜像升级至 v0.25.0

[xinli-sw]

Bump the B200 and B300 Kimi-K2.5 INT4 vLLM images from v0.24.0 to v0.25.0.

---

中文说明

将 B200 和 B300 上的 Kimi-K2.5 INT4 vLLM 镜像从 v0.24.0 升级至 v0.25.0。

[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]

Additional findings (outside current diff — PR may have been updated during review):

• 🟡 configs/nvidia-master.yaml:3126-3132 — The PR title and description don't follow AGENTS.md's mandatory bilingual convention (<English title> / <中文标题>, plus a body with a Chinese translation). Sibling PRs in the git log (e.g. #2224, #2244) follow this format; this one is English-only with an empty body. Not blocking merge, but should be brought into compliance.

  Extended reasoning...

  AGENTS.md (top of file, line 7) states in a blockquote-emphasized rule that overrides default behavior: "PR and GitHub-issue titles & descriptions must be bilingual — include a Simplified Chinese version in addition to English." The required title format is <English title> / <中文标题>, and the body should follow the English content with a Chinese translation (e.g. a ## 中文说明 section). The rule explicitly states "This applies to every PR and every issue."

  PR #2256's title is [chore] bump kimi-k2.5 int4 b200/b300 vllm images to v0.25.0 — purely English, with no / <中文标题> suffix — and the PR body is empty (no description, no ## 中文说明 section at all).

  This is easy to verify by comparing against sibling PRs merged into the same repo just before this one: commit 4bbb2ac is titled [AgentX] vLLM DeepSeek-V4 B200 aggregate / vLLM DeepSeek-V4 B200 聚合配置 (#2224), and commit 1e44aa8 is titled Simplify reusable sweep ingestion with direct source dispatch / 通过直接分发源运行简化复用扫描摄取 (#2244). Both follow the mandated bilingual title format; PR #2256 does not. This confirms the bilingual title/description requirement is an actively-enforced convention in this repo, not a dead rule.

  Step-by-step proof:

  1. Open AGENTS.md line 7 — the bilingual title/description rule, marked as applying to every PR.
  2. Open PR #2256's metadata — title field contains only English text, no / <中文> segment; body field is empty.
  3. Compare against #2224 and #2244 titles in the git log — both contain the bilingual / <中文> suffix, showing the convention is followed elsewhere in the same repo history.
  4. Conclusion: PR #2256 violates the explicit, repo-mandated bilingual title/description rule.

  Why this doesn't block merge: the change itself (bumping two vLLM image tags from v0.24.0 to v0.25.0, and removing a now-stale NOTE comment about B300 lacking a dedicated recipe) is functionally correct and low-risk — nothing in the actual YAML diff is affected by this issue. Unlike the perf-changelog check, this bilingual title/description rule isn't wired into a blocking CI gate in this repo, so it's a metadata/policy compliance gap rather than a functional defect. The fix is trivial: the author (or a maintainer editing the PR) should update the title to [chore] bump kimi-k2.5 int4 b200/b300 vllm images to v0.25.0 / [chore] 将 kimi-k2.5 int4 b200/b300 vllm 镜像升级到 v0.25.0 and add a short bilingual description to the PR body.

• 🔴 configs/nvidia-master.yaml:3129 — This PR bumps the vLLM image tag for kimik2.5-int4-b200-vllm and kimik2.5-int4-b300-vllm (v0.24.0 → v0.25.0) but does not append the required perf-changelog.yaml entry. Per AGENTS.md's 'Updating Docker images' section, this is required and is what triggers the benchmark sweep for the changed configs — please add an entry (e.g. mirroring the prior bump's entry) before merging.

  Extended reasoning...

  What's missing: configs/nvidia-master.yaml is the only file changed in this PR — it bumps the image: tag for both kimik2.5-int4-b200-vllm and kimik2.5-int4-b300-vllm from vllm/vllm-openai:v0.24.0 to vllm/vllm-openai:v0.25.0. There is no corresponding addition to perf-changelog.yaml.

  Why this matters (not just style): AGENTS.md's 'Updating Docker images' section is explicit: 'Update the image tag ... and append a perf-changelog.yaml entry (required - triggers benchmarks).' This isn't a documentation nicety — the changelog is the mechanism that drives which configs actually get swept/validated. The repo's own .github/workflows/claude-pr-review.yml codifies this exact check as a blocking review item: if a master config file is edited and perf-changelog.yaml is not edited in the same PR, reviewers should flag that the changelog must also be updated.

  Proof / verification steps:

  1. git show --stat on this PR's commit shows only configs/nvidia-master.yaml changed (2 insertions, 6 deletions) — perf-changelog.yaml is untouched.
  2. Tailing perf-changelog.yaml at HEAD shows the most recent entry references PR #2224 (an unrelated DSv4 AgentX change); there is no entry mentioning kimik2.5-int4-b200-vllm, kimik2.5-int4-b300-vllm, or v0.25.0 anywhere near the tail.
  3. Every previous image bump for these same two config keys (e.g. the v0.21.0 → v0.24.0 bump) has a matching perf-changelog.yaml entry with config-keys, a description line noting the image update, and a pr-link. This PR breaks that established pattern.

  Impact if merged as-is: Because the changelog entry is what triggers the benchmark sweep for the touched configs, the v0.25.0 image bump may not actually get benchmarked/validated through the normal pipeline, defeating the purpose of the version bump (there's also a risk that whatever mechanism ingests the changelog to select configs for sweeping simply skips these two configs).

  Fix: Append a perf-changelog.yaml entry at the end of the file following the existing format, e.g.:

  - config-keys:
      - kimik2.5-int4-b200-vllm
      - kimik2.5-int4-b300-vllm
    description:
      - "Bump vLLM image from v0.24.0 to v0.25.0"
    pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2256

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29553384191
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29553384191

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29553384191
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29553384191

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29650514816
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29650514816

[xinli-sw]

/reuse-sweep-run

[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. GitHub Actions workflow
• Verified that this PR passes evals. GitHub Actions workflow
• 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 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:

• Exact-head validation and evals passed: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/29650514816
• The published Kimi-K2.5 vLLM recipe is backed by merged vllm-project/recipes #204 and #258.
• This PR changes only the upstream vLLM image tag; it has no speculative-decoding or AgentX configuration.

Signed: Ankur-singh

[Klaud-Cold]

✅✅✅ Verdict: PASS ✅✅✅

✅ Check 0 (CODEOWNER): PASS — configs/nvidia-master.yaml is owned by ankur-singh (last-matching CODEOWNERS line); perf-changelog.yaml falls to the catch-all, satisfied by a recognized CODEOWNER.
✅ Check 1 (sweep on in-PR commit): PASS — head 62a18cd has a green pull_request sweep run 29650514816 with executed (non-skipped) single-node 8k1k / and eval / jobs all success; both configs define only the 8k1k lane.
✅ Check 2 (eval accuracy): PASS — GSM8K em_strict 0.9727–0.9750 across all 6 lanes (B200 TP8 c32/c64, B300 TP8+TP4 c32/c64) vs the 0.90 kimik2.5 threshold in utils/evals/thresholds.yaml; eval jobs ran vllm/vllm-openai:v0.25.0, the exact image this PR pins.
✅ Check 3 (recipe merged + args match): PASS — published recipe recipes.vllm.ai/moonshotai/Kimi-K2.5 backed by MERGED vllm-project/recipes#204 and #258; major args match (model moonshotai/Kimi-K2.5, TP, kimi_k2 reasoning/tool parsers, --compilation_config.pass_config.fuse_allreduce_rms true added upstream by #258, INT4 auto-detected from the native checkpoint — no quant flag in either). This PR changes no server args (image bump only). Informational: VLLM_USE_FLASHINFER_MOE_INT4=1 follows the recipe's documented Blackwell FlashInfer-MoE pattern; --max-num-seqs / --no-enable-prefix-caching / mem-util are InferenceX harness tuning.
✅ Check 4 (reuse command): PASS — /reuse-sweep-run posted by xinli-sw (COLLABORATOR).
✅ Check 5 (latest template): PASS — every item of the current docs/PR_REVIEW_CHECKLIST.md is present and checked.
✅ Check 6 (upstream image / engine-first): PASS — both changed entries pin vllm/vllm-openai:v0.25.0 from the upstream vLLM Docker Hub org on established B200/B300; ordering N/A (submission is vLLM itself).
✅ Check 7 (no architecture hacks): PASS — diff touches only the image tag and perf-changelog; no --hf-overrides or FLOPs-reducing changes.
➖ Check 8 (spec-decode chat template): N/A — no speculative-decoding changes in this PR.
✅ Check 9 (no engine patches): PASS — no patch files, heredoc rewrites, or forked wheels; the pinned image runs as shipped.
➖ Check 10 (agentic golden AL): N/A — no agentic speculative-decoding changes in this PR.