dsv4-fp4-b200-vllm: bump to vllm v0.20.0, deep_gemm_mega_moe MoE

[functionstackx]

Summary

• Pin dsv4-fp4-b200-vllm image to vllm/vllm-openai:v0.20.0-cu130 (canonical v0.20.0 tag) — replaces the floating deepseekv4-cu130 tag. DeepGEMM is preinstalled in this image, so no install_deepgemm.sh step is needed at benchmark time.
• All configs share --compilation-config '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}' and --attention_config.use_fp4_indexer_cache=True.
• Gate --moe-backend deep_gemm_mega_moe and --gpu-memory-utilization 0.85 on DP_ATTENTION=true per the v0.20.0 recipe. The two flags travel together via MOE_ARGS / GMU_ARGS.
• Drop --pipeline-parallel-size 1 (no longer in the v0.20.0 recipe). Retain --no-enable-prefix-caching (so cross-request cache hits don't skew steady-state throughput) and --max-cudagraph-capture-size 2048 (preserve cudagraph coverage at the top of the search space).

Resulting flag matrix

search-space entry	DP_ATTENTION	EP	compile-config	moe-backend	gpu-mem-util
{tp:8}	false	–	FULL_AND_PIECEWISE,custom_ops:all	(default)	(default)
{tp:8, ep:8} (1k1k @ 128)	false	enabled	FULL_AND_PIECEWISE,custom_ops:all	(default)	(default)
{tp:8, ep:8, dp-attn:true}	true	enabled	FULL_AND_PIECEWISE,custom_ops:all	deep_gemm_mega_moe	0.85

PARALLEL_ARGS / EP_ARGS / GMU_ARGS / MOE_ARGS arrays are preserved so the existing DP_ATTENTION / EP_SIZE search-space branching drives the right flags cleanly.

Adds a perf-changelog.yaml entry to trigger the affected configs.

Test plan

• Trigger the dsv4-fp4-b200-vllm benchmark workflow on a B200 runner and confirm the engine starts and the sweep completes for at least one cell in each row of the matrix above.
• Confirm the v0.20.0-cu130 image pulls and DeepGEMM is already importable inside the container (no install step at runtime).
• Spot-check server.log to verify the right flags are logged per branch:
  • DP_ATTENTION=false (any CONC, any EP): no --moe-backend, no --gpu-memory-utilization.
  • DP_ATTENTION=true: --moe-backend deep_gemm_mega_moe --gpu-memory-utilization 0.85.
  • All branches: --compilation-config '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}', --no-enable-prefix-caching, --max-cudagraph-capture-size 2048.

🤖 Generated with Claude Code

[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://cookbook.sglang.io/
• https://docs.vllm.ai/projects/recipes/en/latest/

If it is not, please create a PR first before we can merge your 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/sgl-cookbook

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://cookbook.sglang.io/
• https://docs.vllm.ai/projects/recipes/en/latest/

If it is not, please create a PR first before we can merge your 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/sgl-cookbook

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.

[functionstackx]

talked to vllm maintainer esmeetu and he said that deepgemm extra install only needed for pip wheels and docker already has deepgemm megamoes kernels

[functionstackx]

https://github.com/vllm-project/recipes/pull/409/changes

[functionstackx]