[chore] Bump dsv4-fp4 vLLM image to v0.25.0 on B200 and B300

[xinli-sw]

Bump vLLM to v0.25.0 for DSv4 FP4 B200/B300 agg + B200 tuning

Bumps dsv4-fp4-b200-vllm, dsv4-fp4-b200-vllm-mtp, dsv4-fp4-b300-vllm, and dsv4-fp4-b300-vllm-mtp to vllm/vllm-openai:v0.25.0.

B200 config/script tuning for this image:

• Reduce gpu-memory-utilization to 0.9 for DEP configs in the MTP script
• Add --prefill-schedule-interval 4 for DEP configs in both base and MTP scripts
• Extend TP+EP (ep:8, no DEP) sweep from conc=128 to conc=256 for 1k1k and add it to 8k1k
• Extend 8k1k TP-only sweep from conc=32/16 to conc=64

[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]

🔴 This PR bumps the vLLM image tag on four dsv4-fp4 configs but does not append a corresponding perf-changelog.yaml entry. Per AGENTS.md's 'Updating Docker images' section, this entry is "required — triggers benchmarks", so without it the sweep won't validate v0.24.0 on these configs and the bump ships un-benchmarked. Please add a perf-changelog.yaml entry listing dsv4-fp4-b200-vllm, dsv4-fp4-b200-vllm-mtp, dsv4-fp4-b300-vllm, dsv4-fp4-b300-vllm-mtp before merging.

Extended reasoning...

What the bug is. This PR bumps the image: field on four configs in configs/nvidia-master.yaml from various older tags to vllm/vllm-openai:v0.24.0 — dsv4-fp4-b200-vllm (line 1742), dsv4-fp4-b200-vllm-mtp (line 1830), dsv4-fp4-b300-vllm (line 3184), and dsv4-fp4-b300-vllm-mtp (line 3279). The changed-files list on this PR is only configs/nvidia-master.yaml — no companion entry was appended to perf-changelog.yaml.\n\nWhy this is required. AGENTS.md § 'Updating Docker images' is explicit: "Update the image tag in the relevant configs/*-master.yaml and/or benchmarks/*.sh, update any related env vars / config params, and append a perf-changelog.yaml entry (required — triggers benchmarks)". The parenthetical spells out the consequence: perf-changelog.yaml is what selects which config-keys the sweep actually runs. Elsewhere in AGENTS.md it's described as 'the benchmark trigger log'. Without an entry naming these four config-keys, the full-sweep does not pick them up for validation against the new image.\n\nWhy existing code doesn't prevent it. There is no CI check that cross-references image-tag diffs in configs/*-master.yaml against perf-changelog.yaml additions — the convention is documented but not enforced. A PR that only touches the image tag will pass lint and typecheck; the sweep simply won't cover the changed configs.\n\nImpact. The stated purpose of this PR (per its title, '[chore] Bump dsv4-fp4 vLLM image to v0.24.0 on B200 and B300') is to validate v0.24.0 on the four dsv4 FP4 vLLM configs. Merging without the changelog entry means the image change ships to main without any of those four configs being benchmarked against v0.24.0. Any regression in v0.24.0 (accuracy, throughput, or an outright crash on DeepSeek-V4-Pro) would not be caught until a downstream benchmark or a real user runs it.\n\nHow to fix. Append an entry to perf-changelog.yaml naming all four bumped config-keys. Following the format used by every recent analogous PR (#2073 for kimik2.5-int4-b200-vllm v0.24.0, #2070 for kimik2.5-int4-b300-vllm v0.24.0, #2074, #2077 — each of these mirrors this PR's shape and each appended such an entry):\n\nyaml\n- config-keys:\n - dsv4-fp4-b200-vllm\n - dsv4-fp4-b200-vllm-mtp\n - dsv4-fp4-b300-vllm\n - dsv4-fp4-b300-vllm-mtp\n description:\n - "Bump vLLM image to v0.24.0"\n pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2169\n\n\nStep-by-step proof this manifests.\n1. PR author changes image: on dsv4-fp4-b200-vllm from a nightly tag to v0.24.0 and does the same for the three siblings.\n2. PR is labeled full-sweep-fail-fast per the recipe-reminder bot.\n3. The sweep launcher reads perf-changelog.yaml to determine which config-keys to run in the sweep for this PR.\n4. Grep of the current perf-changelog.yaml for '2169' or 'v0.24.0' on any of dsv4-fp4-b200-vllm/-mtp/dsv4-fp4-b300-vllm/-mtp: no match — the four bumped configs are not in the trigger set for this PR.\n5. Sweep completes green because it did not run any of the four configs whose image actually changed. The bump is merged as 'validated' when in fact no dsv4-fp4 vLLM benchmark exercised v0.24.0.\n\nThis is a required convention that gates the actual validation the PR is intended to perform, so it should be resolved before merge.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29138257980
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29138257980

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29140720192
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29140720192

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29140720192
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29140720192

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29177379558
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29177379558

[xinli-sw]

/reuse-sweep-run

[xinli-sw]

recipes PR: vllm-project/recipes#629

cc @Ankur-singh

[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. Run Sweep
• Verified that this PR passes evals. Please link to GitHub Action workflow that shows this. Run Sweep
• 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 the single-node recipes are similar to the official vLLM recipes and/or theSGLang cookbook:
  • If they are not, I have verified that a PR has been opened in vLLM recipe repo or SGLang repo and linked it below in the additional detail section:
• 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:

• Validation + evals: full Run Sweep on head SHA 238dc77 is green: actions/runs/29177379558
• All four touched configs (dsv4-fp4-b200-vllm, dsv4-fp4-b200-vllm-mtp, dsv4-fp4-b300-vllm, dsv4-fp4-b300-vllm-mtp) are single-node agg submissions (multinode: false).
• Spec-decoding (MTP) configs route benchmark prompts through chat-formatted encoding via --dsv4 in the benchmark scripts, as required for realistic acceptance lengths.
• Image is upstream vllm/vllm-openai:v0.25.0 from the official vLLM Docker Hub repo; no engine patching (only benchmark-client pip install datasets pandas).
• Recipe: the configs follow the official vLLM recipe models/deepseek-ai/DeepSeek-V4-Pro.yaml; the newly added --prefill-schedule-interval 4 (DEP prefill scheduling) is covered by open recipe PR docs: add DeepSeek V4 Pro DEP prefill scheduling vllm-project/recipes#629 (same author): docs: add DeepSeek V4 Pro DEP prefill scheduling vllm-project/recipes#629
• No agentic-workload configs are touched by this PR (the agentic golden-AL item is vacuously satisfied).

Signed: ankur-singh

[Klaud-Cold]

✅✅✅ Verdict: PASS ✅✅✅

✅ Check 0 (CODEOWNER): PASS — @Ankur-singh is a listed owner of configs/nvidia-master.yaml; remaining paths are catch-all-only.
✅ Check 1 (sweep on in-PR commit): PASS — head 238dc77 has 129 executed single-node */ + eval / check-runs, all success (none skipped), from run 29177379558.
✅ Check 2 (evals pass): PASS — 16 gsm8k results in that run's eval_results_all artifact, em_strict 0.959–0.967, all above the 0.91 dsv4 bar; all lanes ran vllm/vllm-openai:v0.25.0, matching this PR's configs.
✅ Check 3 (recipe): PASS — single-node vLLM configs match the official DeepSeek-V4-Pro recipe on all major args (model, fp8 KV cache, TP/TEP/DEP parallelism, deep_gemm_mega_moe, FP4 indexer cache, MTP spec-config); the new --prefill-schedule-interval 4 is covered by linked open recipes#629. Informational: pre-existing --enable-eplb (B200 DEP lane) and per-lane GMU/conc tuning are InferenceX-side tuning, not recipe blockers.
✅ Check 4 (reuse command): PASS — /reuse-sweep-run posted by xinli-sw (COLLABORATOR).
✅ Check 5 (latest template): PASS — sign-off contains every item of the current docs/PR_REVIEW_CHECKLIST.md, all checked.
✅ Check 6 (upstream image / engine-first): PASS — all four changed entries use upstream vllm/vllm-openai:v0.25.0; no new non-vLLM/SGLang framework entries.
✅ Check 7 (no architecture hacks): PASS — diff only adds prefill scheduling / GMU / concurrency tuning; no --hf-overrides or FLOPs-reducing knobs.
✅ Check 8 (spec-decode chat template): PASS — both MTP scripts benchmark via --dsv4, which routes prompts through the DeepSeek-V4 chat template (auto-enables --use-chat-template).
✅ Check 9 (no engine patches): PASS — no patching of the serving stack; only client-side pip install datasets pandas.
➖ Check 10 (agentic golden AL): N/A — no agentic spec-decode configs touched, and no synthetic-acceptance knobs added to the (non-agentic) MTP configs.

[xinli-sw]

Hi @adibarra can you also review? Thanks!