[NV]Add GLM-5 NVFP4 GB300 disagg-mtp TRT-LLM benchmarks via Dynamo

[xinli-sw]

opened on behalf of @richardhuo-nv

---

Note

Low Risk
Benchmark and launcher configuration only; no application runtime or security-sensitive logic changes.

Overview
Adds glm5-fp4-gb300-dynamo-trt-mtp to the NVIDIA master benchmark matrix: GLM-5 NVFP4 on GB300 with disaggregated prefill/decode, MTP spec-decoding, and Dynamo + TensorRT-LLM (tensorrtllm-runtime:1.3.0-dev.1-cuda13). Coverage is 23 MTP search points—13 at 1K/1K and 10 at 8K/1K—each wired to an CONFIG_FILE recipe on NVIDIA srt-slurm sa-submission-q2-2026.

runners/launch_gb300-nv.sh gains a glm5 + fp4 + dynamo-trt branch that sets SERVED_MODEL_NAME, MODEL_PATH, and SRT_SLURM_MODEL_PREFIX for nvidia/GLM-5-NVFP4. perf-changelog.yaml documents the new config key.

^{Reviewed by Cursor Bugbot for commit 4dc40fe. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-actions]

Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook

• https://docs.sglang.io/cookbook/intro
• https://recipes.vllm.ai/

If it is not, please create a PR first before we can merge your single node PR into the master branch. Let's ensure that the documentation is first class such that the entire ML community can benefit from your hard work! Thank you

• https://github.com/vllm-project/recipes
• https://github.com/sgl-project/sglang/tree/main/docs_new

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. If re-running failed jobs is attempted, PR authors are responsible for ensuring it passes. See GitHub's docs on re-running failed jobs: https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs#re-running-failed-jobs-in-a-workflow

As a rule of thumb, generally, PR authors should request a review & get a PR approval from the respective companies' CODEOWNERS before requesting a review from core maintainers.

If additional help is needed, PR authors can reach out to core maintainers over Slack.

[github-actions]

Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook

• https://docs.sglang.io/cookbook/intro
• https://recipes.vllm.ai/

If it is not, please create a PR first before we can merge your single node PR into the master branch. Let's ensure that the documentation is first class such that the entire ML community can benefit from your hard work! Thank you

• https://github.com/vllm-project/recipes
• https://github.com/sgl-project/sglang/tree/main/docs_new

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. If re-running failed jobs is attempted, PR authors are responsible for ensuring it passes. See GitHub's docs on re-running failed jobs: https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs#re-running-failed-jobs-in-a-workflow

As a rule of thumb, generally, PR authors should request a review & get a PR approval from the respective companies' CODEOWNERS before requesting a review from core maintainers.

If additional help is needed, PR authors can reach out to core maintainers over Slack.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 6598e48. Configure here.}

[cursor]

Duplicate non-monotonic conc-list 615 suggests misplaced entry

Medium Severity

In the ISL 8192 section, conc-list: [615] appears twice — once at line 2668 (with 10 prefill workers, decode tp=16) and again at line 2698 (with 11 prefill workers, decode tp=32) placed after the conc-list: [1229] entry. This breaks the otherwise strictly ascending concurrency order (5, 15, 30, 84, 180, 333, 615, 1229, 615, 2253). The second 615 likely has an incorrect concurrency value or is misplaced, potentially causing the benchmark to run two different hardware configurations at the same concurrency target unintentionally, or missing a distinct concurrency sweep point.

Additional Locations (1)

• .github/configs/nvidia-master.yaml#L2666-L2681

 

^{Reviewed by Cursor Bugbot for commit 6598e48. Configure here.}

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27621692220
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27621692220

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27622387448
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27622387448

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27634400529
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27634400529

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27634553033
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27634553033

[functionstackx]

waiting on glm5 nvfp4 gb300 mtp sgl PR first