[NV] minimaxm2.5 fp8 b300 vllm update

[hshrivastava-droid]

Summary

Updates the MiniMax-M2.5 FP8 B300 vLLM benchmark configuration and environment variables:

• Environment variable change: Replace VLLM_FLASHINFER_ALLREDUCE_BACKEND=mnnvl with VLLM_FLOAT32_MATMUL_PRECISION=high in the benchmark script
• Search-space overhaul (1k/1k): Restructure TP/EP/concurrency sweep — use TP=4 at low concurrency (4–128), TP=4 EP=4 at mid concurrency (256–512), TP=2 EP=2 at high concurrency (512–1024), and add TP=2 EP=2 with dp-attn at very high concurrency (1024–2048)
• Search-space overhaul (8k/1k): Replace broad TP=2/TP=4 sweeps with targeted configs — TP=1 at low concurrency (4–16), TP=2 at mid concurrency (64–256), and TP=4 at low concurrency (4–8)
• Changelog: Add perf-changelog.yaml entry documenting the env var changes

Changed Files

File	Change
benchmarks/single_node/minimaxm2.5_fp8_b300.sh	Swap env var from VLLM_FLASHINFER_ALLREDUCE_BACKEND=mnnvl to VLLM_FLOAT32_MATMUL_PRECISION=high
.github/configs/nvidia-master.yaml	Revise search-space for minimaxm2.5-fp8-b300-vllm (both 1k/1k and 8k/1k)
perf-changelog.yaml	Add changelog entry for this PR

Test Plan

• Run e2e benchmarks for minimaxm2.5-fp8-b300-vllm to validate new search-space configurations
• Verify the VLLM_FLOAT32_MATMUL_PRECISION=high env var produces expected performance results
• Confirm dp-attn config at high concurrency (1024–2048) works correctly

[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

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

[hshrivastava-droid]

@claude - add PR description

[Klaud-Cold]

Claude finished @hshrivastava-droid's task in 44s —— View job

---

Adding PR Description

• Gather context on PR changes
• Write and update PR description

PR description has been added. It covers:

• The environment variable swap (VLLM_FLASHINFER_ALLREDUCE_BACKEND=mnnvl → VLLM_FLOAT32_MATMUL_PRECISION=high)
• The restructured search-space configurations for both 1k/1k and 8k/1k sequence lengths
• A test plan checklist for validation

[functionstackx]

can u update the vllm recipe?

[hshrivastava-droid]

vllm recipe- vllm-project/recipes#353

[hshrivastava-droid]

@functionstackx - could you please review this?

[hshrivastava-droid]

@functionstackx - could you please review this?

[functionstackx]

lgtm! thank you for this PR!

[jgangani]

LGTM