dsv4-fp4-b300-sglang: update image to nightly

[yhyang201]

Summary

• Update image from deepseek-v4-b300@sha256:2fec8d... to nightly-dev-cu13-20260518-c67b2870
• Refactor benchmark script to dispatch by CONC instead of nested DP_ATTENTION/CONC/EP_SIZE
• Switch high-concurrency profiles (CONC 2048/4096/8192) from --moe-a2a-backend deepep to megamoe
• Remove env vars deleted from sglang main or redundant with defaults
• Remove --deepep-config (not needed by megamoe)
• Fix CONC=512 yaml ep: 4 → ep: 1 (flashinfer_mxfp4 doesn't set ep=tp)

---

Note

Low Risk
Benchmark and CI config only—no production serving or auth paths; risk is mis-tuned launch flags affecting perf numbers.

Overview
Updates dsv4-fp4-b300-sglang to a newer SGLang nightly image and aligns the B300 DeepSeek-V4-Pro benchmark with megamoe for high concurrency.

The nvidia-master config drops the B200-recipe note, documents CONC-based recipes (TP-only / DP+flashinfer / DP+megamoe), and fixes the CONC=512 search point from ep: 4 to ep: 1 so result names match flashinfer_mxfp4 (no implicit ep=tp).

dsv4_fp4_b300_sglang.sh is refactored to choose launch settings by CONC only (no nested DP_ATTENTION / EP_SIZE branches). 2048 / 4096 / 8192 now use --moe-a2a-backend megamoe instead of deepep; 512 stays DP-attn + flashinfer_mxfp4. Several SGLANG_OPT_* env vars and --deepep-config are removed as obsolete or megamoe-handled. pip install --upgrade transformers is added before serving benchmarks. perf-changelog.yaml records the change.

^{Reviewed by Cursor Bugbot for commit 0ba92fd. 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.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26109529858
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26109529858

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26109534591
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26109534591

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26221509538
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26221509538

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26221509538
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26221509538

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26658560606
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26658560606

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26658745339
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26658745339

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26658745339
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26658745339