Add DSV4 GB300 wide-EP sweep configs (EP=12/16/24/32/40)

[yhyang201]

Summary

• Add 5 new search-space entries and recipe files for DSV4 GB300 non-MTP wide-EP sweep, matching srt-slurm PR#173 topology (18 nodes total).
• EP sizes: 12/16/24/32/40, decode nodes: 3/4/6/8/10, concurrencies: 12000/8192/3000/2500/2048.
• Uses InferenceX env vars and sglang_config (megamoe, W4A4, nightly-20260520 image), with Weiliang's tuned decode params (swa-full-tokens-ratio=0.20, max-running-requests=18432).

---

Note

Medium Risk
Changes multinode SLURM benchmark topology and serving flags for expensive GB300 runs; mistakes could schedule wrong node/GPU layouts or skew published perf, but scope is config/recipes only.

Overview
Replaces the DSV4 GB300 dynamo-sglang 8k/1k disaggregated search space with a wide expert-parallel sweep aligned to srt-slurm PR#173: five 18-node points at decode EP 12/16/24/32/40 with matching concurrencies (12000 down to 2048), plus a retained low-concurrency 1p1d tp4/tp4 case. Four older dominated topology entries (mixed node counts and DEP16-focused decode) are removed from nvidia-master.yaml.

The linked srt-slurm recipe YAMLs are retargeted or added (disagg-gb300-*-18-c*.yaml): node counts, decode TP/EP, multi-frontend routing, benchmark concurrency, and tuned SGLang decode settings (swa-full-tokens-ratio=0.20, max-running-requests=18432, prefill enable-dp-lm-head / cuda-graph-max-bs=512). One new recipe file is added for the EP=24 topology.

Smaller benchmark pins: dsv4-fp4-b200-vllm moves from a nightly image digest to vllm/vllm-openai:v0.22.0; gptoss-fp4-mi355x-vllm on AMD master switches the model to amd/gpt-oss-120b-w-mxfp4-a-fp8. The B200 DSv4 vLLM launch script drops EPLB flags when DP attention is enabled. perf-changelog.yaml documents the GB300 sweep.

^{Reviewed by Cursor Bugbot for commit 74f1a45. 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=26642196095
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26642196095

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26642196095
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26642196095

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26824028702
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26824028702

[weireweire]

there should be bugs on the visualizer, the result looks normal but the plot make it's interactivity too big

dsv4 | deepseek-ai/DeepSeek-V4-Pro | GB300-CW | DYNAMO-SGLANG | FP4 | 8192 | 1024 | 4 | 4 | true | 15 | 60 | 12 | 12 | true | 1 | 12 | 12000 | 147376 | 0 | 0 | 0 | 18.4703 | 0 | 54.141 | 0 | 0 | 0 | 164.389 | 0 | 0
-- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | --

[weireweire]

@kedarpotdar-nv @jgangani could you have a look?

[weireweire]

actually, the median TPOT is showed as 0 in the table(while mean is goo), so issue is not visualize, there are real issue for tpot.

Comparing to NVIDIA/srt-slurm#173, the decode side added load-balance-method which is buggy, can be the issue.

[weireweire]

and enable_multiple_frontends: true and set num 8 is also missing

[cursor]

EP40 max-running-requests typo

Medium Severity

The EP=40 wide-EP recipe sets decode max-running-requests to 18400, while the other four new sweep recipes and perf-changelog.yaml specify 18432 for Weiliang-aligned decode tuning.

 

^{Reviewed by Cursor Bugbot for commit 4b2448d. Configure here.}

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26974513741
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26974513741

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27032160756
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27032160756

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27032160756
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27032160756

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27032160756
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27032160756

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27033619607
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27033619607

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27033666549
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27033666549

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

There are 3 total unresolved issues (including 1 from previous review).

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

^{Reviewed by Cursor Bugbot for commit d02ccd7. Configure here.}

[cursor]

B200 vLLM drops EPLB flags

Medium Severity

The B200 DSV4 vLLM launch script no longer passes --enable-eplb / --eplb-config when DP_ATTENTION is true, while the same commit bumps dsv4-fp4-b200-vllm to v0.22.0 with no perf-changelog entry for either change.

Additional Locations (1)

• .github/configs/nvidia-master.yaml#L1758-L1759

 

^{Reviewed by Cursor Bugbot for commit d02ccd7. Configure here.}

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27033807266
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27033807266

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27033910325
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27033910325

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27033992200
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27033992200

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27034082957
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27034082957