[NVIDIA][GB300] update DSR1 FP8 GB300 TRTLLM image to latest

[xinli-sw]

The release of nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13 contains a fix for NVIDIA/srt-slurm#51, where accuracy for one MTP config drops to 88% on GSM8k. While the point still passes evals on InferenceX, 95% is usually expected.

Dropped these configs

gen_worker_env_var: ' TRTLLM_FORCE_ALLTOALL_METHOD=NVLinkTwoSided ENABLE_CONFIGURABLE_MOE=1
    TRTLLM_FORCE_COMM_METHOD=NVLINK_TWO_SIDED'

In the latest container, NVLinkOneSided is the preferred backend and also the default, removing these to recover perf.

---

Note

Medium Risk
Changes published inference benchmark image and hardware runner for a flagship multinode TRT config, which directly affects accuracy and throughput numbers but does not alter application runtime code paths.

Overview
Updates dsr1-fp8-gb300-dynamo-trt to use tensorrtllm-runtime:1.3.0-dev.1-cuda13 (from 0.8.1.post2) and switches the runner from gb300 to gb300-nv, aligning the GB300 DeepSeek-R1 FP8 Dynamo/TRT benchmark with the latest container that fixes GSM8K accuracy and MTP concurrency issues.

Documents the change in perf-changelog.yaml: restored ~95% GSM8K accuracy (vs 88% on one point), resolution of numeric issues affecting some MTP points, and note that DSR1 TRTLLM FP8 configs should use the new image.

^{Reviewed by Cursor Bugbot for commit 4b8d282. 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]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27523259199
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27523259199

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27523630547
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27523630547

[xinli-sw]

might need to do a quick update to 1k1k config , please do not enable sweep yet

[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 9031353. Configure here.}

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27590368246
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27590368246

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27590368246
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27590368246

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27590368246
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27590368246

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27595478969
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27595478969

[functionstackx]

overall lgtm besides more detailed perfchange plz, explaining the onesidealltoall error vs twosidealltoall

[functionstackx]

@xinli-sw can u add an new more of an description of what was broken and what was the fix?

[xinli-sw]

@functionstackx done

[cquil11]

/reuse-sweep-run