[NV] Kimi K2.5 NVFP4 B200 disaggregated TRT-LLM (Dynamo) recipe / Kimi K2.5 NVFP4 B200 TRT-LLM（Dynamo）分离式推理配置

[richardhuo-nv]

[NV] kimik2.5-fp4-b200-dynamo-trt: Kimi K2.5 NVFP4 B200 disaggregated TRT-LLM (Dynamo) recipe

Summary

Adds a new multi-node benchmark config kimik2.5-fp4-b200-dynamo-trt — Kimi K2.5 NVFP4 on B200, disaggregated prefill/decode served via Dynamo + TensorRT-LLM. This is the B200 sibling of the existing kimik2.5-fp4-gb200-dynamo-trt entry, with topologies and engine configs sourced from the NVIDIA srt-slurm B200 recipes.

Scenario: ISL 8192 / OSL 1024, 11 STP (non-MTP) Pareto points.

Topologies

Config	Prefill	Decode	Concurrency
1p4d-dep4-tep8	1×DEP4	4×TEP8	4, 24
1p5d-dep4-tep4	1×DEP4	5×TEP4	5, 30, 60, 115, 195
1p1d-dep4-dep8	1×DEP4	1×DEP8	308
2p1d-dep4-dep8	2×DEP4	1×DEP8	615
3p1d-dep4-dep8	3×DEP4	1×DEP8	1127
4p1d-dep4-dep8	4×DEP4	1×DEP8	2151

• Container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc21
• Dynamo: installed from source (hash 4025cb6a8992d19fd19eead36065e08ea5301f35), nixl KV transfer, TCP request plane
• Runner: b200-multinode (b200-dgxc-slurm)

Changes

• benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/ — 11 flat recipe YAMLs copied from srt-slurm, with model.path/model.container rewritten to the cluster-resolved aliases (kimik2.5-fp4 / dynamo-trtllm)
• configs/nvidia-master.yaml — new kimik2.5-fp4-b200-dynamo-trt entry (all 11 configs, each pointing at its local CONFIG_FILE)
• runners/launch_b200-dgxc.sh — added a dynamo-trt + kimik2.5 + fp4 branch that clones srt-slurm and copies the local B200 recipes into recipes/trtllm/kimi-k25-nvfp4/b200-fp4/
• perf-changelog.yaml — new changelog entry (selects this config for the sweep)

中文说明

概要

新增多节点基准测试配置 kimik2.5-fp4-b200-dynamo-trt：在 B200 上运行 Kimi K2.5 NVFP4，并通过 Dynamo + TensorRT-LLM 提供预填充/解码分离式推理服务。该配置是现有 kimik2.5-fp4-gb200-dynamo-trt 的 B200 对应版本，拓扑和引擎配置来自 NVIDIA srt-slurm 的 B200 recipe。

测试场景：ISL 8192 / OSL 1024，包含 11 个 STP（非 MTP）Pareto 数据点。

拓扑

配置	预填充	解码	并发数
1p4d-dep4-tep8	1×DEP4	4×TEP8	4, 24
1p5d-dep4-tep4	1×DEP4	5×TEP4	5, 30, 60, 115, 195
1p1d-dep4-dep8	1×DEP4	1×DEP8	308
2p1d-dep4-dep8	2×DEP4	1×DEP8	615
3p1d-dep4-dep8	3×DEP4	1×DEP8	1127
4p1d-dep4-dep8	4×DEP4	1×DEP8	2151

• 镜像： nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc21
• Dynamo： 从源码安装（哈希 4025cb6a8992d19fd19eead36065e08ea5301f35），使用 nixl 传输 KV 缓存，并采用 TCP 请求平面
• 运行器： b200-multinode（b200-dgxc-slurm）

变更

• benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/ — 从 srt-slurm 复制 11 个扁平化 recipe YAML，并将 model.path/model.container 改为集群解析后的别名（kimik2.5-fp4 / dynamo-trtllm）
• configs/nvidia-master.yaml — 新增 kimik2.5-fp4-b200-dynamo-trt 配置（包含全部 11 个配置，每个配置均通过 CONFIG_FILE 指向对应的本地 recipe）
• runners/launch_b200-dgxc.sh — 新增 dynamo-trt + kimik2.5 + fp4 分支，用于克隆 srt-slurm，并将本地 B200 recipe 复制到 recipes/trtllm/kimi-k25-nvfp4/b200-fp4/
• perf-changelog.yaml — 新增变更日志条目，以便扫描流程选择该配置

[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

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29528243378
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29528243378

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29528755773
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29528755773

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29528755773
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29528755773

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29528755773
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29528755773

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29528755773
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29528755773

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29528755773
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29528755773

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29528755773
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29528755773

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29528755773
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29528755773

[xinli-sw]

/reuse-sweep-run

[Ankur-singh]

/reuse-sweep-run 29528755773

[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.
• Verified that this PR passes evals. Please link to GitHub Action workflow that shows this.
• 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:

• Run Sweep 29528755773, attempt 7 passed all 11 applicable multi-node 8k1k benchmark jobs and all 9 applicable multi-node eval jobs for head 0869a4a41161642cf59dc6e9d2dd157494a6c985.
• This is a multi-node disaggregated TRT-LLM STP submission. Speculative-decoding, AgentX, and single-node recipe requirements are not applicable. The existing upstream vLLM configuration kimik2.5-fp4-b200-vllm satisfies the upstream-engine-first requirement.
• The CODEOWNER reviewed and accepted the srt-slurm provenance wording difference: the PR description and changelog mention sa-submission-q2-2026, while the launcher pins v1.0.29.
• The PR is currently CONFLICTING / DIRTY; the user explicitly accepted resolving the conflict later. This approval does not resolve mergeability.

Signed: Ankur-singh

[Klaud-Cold]

✅✅✅ Verdict: PASS ✅✅✅

✅ Check 0 (CODEOWNER): PASS — @Ankur-singh is a listed owner of configs/nvidia-master.yaml; all other changed paths carry only the catch-all, which a recognized CODEOWNER satisfies.
✅ Check 1 (sweep on in-PR commit): PASS — head 0869a4a (still the PR tip) has 11 green multi-node 8k1k / and 9 green multi-node eval / executed check-runs on run 29528755773.
✅ Check 2 (evals pass): PASS — downloaded agg_eval_all.json from that run: 9/9 configs, GSM8K em_strict 0.970–0.974 (n_eff 1319 each), framework dynamo-trt fp4 on image nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc21, matching this PR's config.
➖ Check 3 (recipe link): N/A — disaggregated/multi-node submission (multinode: true, disagg: true, files under benchmarks/multi_node/**); the recipe-link requirement applies to single-node recipes only.
✅ Check 4 (reuse command): PASS — /reuse-sweep-run 29528755773 posted by COLLABORATOR Ankur-singh (plus a bare /reuse-sweep-run from COLLABORATOR xinli-sw).
✅ Check 5 (latest checklist): PASS — sign-off contains every item of the current docs/PR_REVIEW_CHECKLIST.md template, all checked.
✅ Check 6 (upstream image / engine-first): PASS — framework: dynamo-trt (upstream-image rule targets vLLM/SGLang entries); engine-first satisfied by existing kimik2.5-fp4-b200-vllm (vllm/vllm-openai:v0.22.0, runner b200) in configs/nvidia-master.yaml.
✅ Check 7 (no architecture hacks): PASS — no --hf-overrides/model-config edits; diff only adds recipes, config entries, and launcher routing.
➖ Check 8 (spec-decode chat template): N/A — STP (non-MTP) configs only, no speculative decoding in this PR.
✅ Check 9 (no engine patches): PASS — no patch/sed/heredoc edits of engine sources; TRT-LLM runs as shipped in the pinned image (Dynamo from source is the declared dynamo-trt framework layer, same as the merged GB200 sibling).
➖ Check 10 (agentic golden AL): N/A — no agentic speculative-decoding changes.

Note (informational, not a check failure): the PR is currently in a conflicting state against main per the sign-off; conflicts must be resolved before merge, and the sign-off itself flags this.