Add Qwen3.5-397B-A17B-NVFP4-V2 GB300 TRT-LLM disaggregated AgentX - #2612
Conversation
|
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 For PR verification, add the 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 如需进行 PR 验证，请为此 PR 添加 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 |
2 similar comments
|
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 For PR verification, add the 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 如需进行 PR 验证，请为此 PR 添加 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 |
|
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 For PR verification, add the 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 如需进行 PR 验证，请为此 PR 添加 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[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31833296173 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31835246631 |
…b300-qwen3.5-nvfp4-agentx # Conflicts: # perf-changelog.yaml
Run 22805 showed the dynamo-trt frontend registers the qwen3.5 model
as "Qwen3.5-397B-A17B-NVFP4-V2" (no nvidia/ prefix) regardless of
DYN_TRTLLM_SERVED_MODEL_NAME, while build_replay_cmd passed the raw
$MODEL HF id ("nvidia/Qwen3.5-397B-A17B-NVFP4-V2") to aiperf --model,
404ing on every warmup request before profiling could start (same
root cause as the DSv4 GB300 dynamo-trt case in PR #2595).
Add SERVED_MODEL_NAME to all 6 qwen3.5 agentx recipes and make
build_replay_cmd prefer it over $MODEL, consistent with PR #2595.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31835246631 |
Run 22866 (qwen3.5 GB300 trtllm) died at Configure Profiling:
aiperf's dataset manager resolves the tokenizer from --model, and
--model is now SERVED_MODEL_NAME ("Qwen3.5-397B-A17B-NVFP4-V2"), which
isn't a valid HF repo id — only "nvidia/Qwen3.5-397B-A17B-NVFP4-V2" is.
Pass --tokenizer $MODEL explicitly so routing (--model) and tokenizer
resolution (--tokenizer) use the correct name for each.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31904594706 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31910374232 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31927376673 |
|
/reuse-sweep-run |
kedarpotdar-nv
left a comment
kedarpotdar-nv
left a comment
There was a problem hiding this comment.
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 run 31927376673
- Verified that this PR passes evals. GitHub Actions run 31927376673
- 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 this PR uses
append-only: true, verified that it only adds generated points or recipe variants inside a selected existing config/scenario and existing same-image visual curve: every previously generated point remains present with the same recipe, no prior point is removed or rerun, and every benchmark-affecting change in the complete diff can affect only the corresponding newly appended points (never an existing point), regardless of which file contains it. - If any of the above criteria cannot reasonably be satisfied, I have provided additional reasoning below.
Additional detail section:
- Validation and evals passed in run 31927376673. The synchronized head retained that passing commit in the PR and the reuse gate recognized the authorized
/reuse-sweep-run. - This is exclusively a disaggregated/multi-node TRT-LLM submission; the single-node upstream recipe-link requirement is N/A.
- Engine-first ordering is satisfied by the existing Qwen3.5 FP4 GB300 Dynamo-SGLang AgentX aggregated and disaggregated submissions, using an upstream
lmsysorg/sglangimage. - TRT-LLM simulated accepted-draft-token values are
3.80for draft length 6 and4.04for draft length 7, corresponding to committed golden AL values4.80and5.04. Eval-only runs remove simulation and use real verification. - The eval-only recipe edit removes only the synthetic-acceptance setting from copied recipe YAML; it does not patch the inference engine.
- This PR does not use
append-only: true.
Signed: @kedarpotdar-nv
✅✅✅ Verdict: PASS ✅✅✅✅ Check 0 (CODEOWNER): PASS — @kedarpotdar-nv is a named owner of |
|
/stage-results 31927376673 |
|
@cquil11 staged run 31927376673: https://inferencemax-app-git-staging-semianalysisai.vercel.app/inference?i_dates=2026-08-16~r31927376673 This run remains available across future |
6 participants
Add Qwen3.5 FP4 GB300 TRT-LLM Disaggregated AgentX / 新增 Qwen3.5 FP4 GB300 TRT-LLM 分离式 AgentX
Summary / 摘要
Registers six disaggregated Dynamo-TRT agentic benchmark recipes for Qwen3.5-397B-A17B-NVFP4-V2 on GB300, spanning low-latency small-scale to high-throughput large-scale topologies, with MTP speculative decoding throughout.
新增六个 Qwen3.5-397B-A17B-NVFP4-V2 GB300 分离式 Dynamo-TRT AgentX Benchmark 配置，涵盖从低延迟小规模到高吞吐量大规模的拓扑，全部采用 MTP 推测解码。
Changes / 变更内容
configs/nvidia-master.yamlNew config key
qwen3.5-fp4-gb300-dynamo-trt-agentic-disagg:nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24nvidia/Qwen3.5-397B-A17B-NVFP4-V2dynamo-router 1.4.0.dev20260807, KV transfer: NIXLagentic-codingsearch-space entries:benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/Six recipe YAMLs added (with trailing newlines matching sibling directory convention).
runners/launch_gb300-nv.shNew branch for
IS_AGENTIC=1 && dynamo-trt && qwen3.5:NVIDIA/srt-slurmatv1.0.50and overlays the agentx recipe directory.EVAL_ONLY=true, stripsTLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENSfrom all agentx recipe YAMLs beforesrtctlis invoked so EVAL_ONLY runs switch to real block verification by removing synthetic acceptance.perf-changelog.yamlNew entry for
qwen3.5-fp4-gb300-dynamo-trt-agentic-disagg.Validation / 验证
Matrix generator produces all 6 expected
agentic-codingrows: