Update MiniMax M2.5 FP8 H200 vLLM agg recipes

[anish-shanbhag]

(Identical to #1298 except the source branch is no longer from a fork so that CI can run)

Set vLLM serving knobs in benchmarks/single_node/minimaxm2.5_fp8_h200.sh: generated benchmark max-model-len, previous eval max-model-len handling, fp8 KV cache, FlashInfer attention/autotune, Triton MoE, and MiniMax QK norm fusion.

---

Note

Low Risk
Benchmark and CI config only—no production auth or app logic; changes affect how perf jobs run and what hardware knobs are swept.

Overview
Tightens MiniMax-M2.5 FP8 on H200 (vLLM) benchmark recipes by changing how the server is launched and how CI sweeps concurrency/tensor parallelism.

benchmarks/single_node/minimaxm2.5_fp8_h200.sh now sets runtime env (e.g. disabling Deep GEMM / blockscale FlashInfer FP8 GEMM), defaults for max-num-seqs, max-num-batched-tokens, and a compilation-config with MiniMax QK-norm fusion, and passes new vLLM flags: fp8 KV cache, Triton MoE, FlashInfer attention with autotune, plus safer quoting and array-style expert-parallel args.

In .github/configs/nvidia-master.yaml, fixed-seq-len search for minimaxm2.5-fp8-h200-vllm shifts from tp 8 / conc 4–128 to tp 4 / conc 1–256 (two ISL/OSL points). perf-changelog.yaml documents the recipe update (image v0.20.1-ubuntu2404 and serving knobs).

^{Reviewed by Cursor Bugbot for commit 111df5b. 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=25772346949
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=25772346949

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=25814571481
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=25814571481

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=25814571481
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=25814571481

[kedarpotdar-nv]

@claude please review

[Klaud-Cold]

Claude finished @kedarpotdar-nv's task in 2m 38s —— View job

---

Review: Update MiniMax M2.5 FP8 H200 vLLM agg recipes

• Gather context and read changed files
• Review .github/configs/nvidia-master.yaml changes
• Review benchmarks/single_node/minimaxm2.5_fp8_h200.sh changes
• Review perf-changelog.yaml changes
• Post review feedback

---

Overall: Looks good. The PR modernizes the MiniMax M2.5 FP8 H200 vLLM benchmark script with proper quoting, FP8 KV cache, FlashInfer attention with autotune, Triton MoE backend, and MiniMax QK norm fusion via compilation config. CI is green.

.github/configs/nvidia-master.yaml

• TP 8 → TP 4: Makes sense — TP=4 on 8×H200 leaves more memory per shard and allows wider concurrency sweeps. The H100 config (minimaxm2.5-fp8-h100-vllm) also uses TP=4 with EP=4.
• Concurrency range widened from 4–128 to 1–256: Provides broader coverage, consistent with other configs that use conc-start: 1.
• No ep field: Without ep in the search-space, EP_SIZE defaults to 1, so the script's if [ "$EP_SIZE" -gt 1 ] guard correctly skips --enable-expert-parallel. However, the H100 config uses ep: 4 with TP=4 — was EP intentionally omitted here, or should it be { tp: 4, ep: 4, conc-start: 1, conc-end: 256 } to match? With TP=4 on an 8-GPU node, you have 4 unused GPUs unless EP fills them. Worth confirming this is intentional.

benchmarks/single_node/minimaxm2.5_fp8_h200.sh

• Proper quoting: All variable expansions are now quoted ("$MODEL", "$PORT", etc.) — good shell hygiene.
• EP guard fix: Changed from -ge 1 (always true since EP_SIZE defaults to 1) to -gt 1 and switched from string concatenation (EP=" --enable-expert-parallel") to array syntax (EP=(--enable-expert-parallel) / "${EP[@]}"). This matches the canonical pattern from the repo guidelines.
• New vLLM flags:
  • --kv-cache-dtype fp8 — FP8 KV cache for memory savings
  • --moe-backend triton — Triton MoE kernels
  • --attention-backend FLASHINFER + --enable-flashinfer-autotune — FlashInfer attention with autotuning
  • --compilation-config with fuse_minimax_qk_norm — model-specific QK norm fusion
  • --max-num-seqs 512 and --max-num-batched-tokens 32768 — configurable via env with reasonable defaults
• New env vars: PYTHONNOUSERSITE=1, SAFETENSORS_FAST_GPU=1, VLLM_USE_DEEP_GEMM=0, VLLM_BLOCKSCALE_FP8_GEMM_FLASHINFER=0, VLLM_FLOAT32_MATMUL_PRECISION=high — these align with typical vLLM tuning knobs for FP8 MoE workloads.

perf-changelog.yaml

• Entry looks well-formed with the correct config key and PR link.
• Minor note: description mentions updating to v0.20.1-ubuntu2404 but the actual image in the config is v0.20.2. This was likely true in an earlier iteration (PR (Replaced by #1354) Update MiniMax M2.5 FP8 H200 vLLM agg recipes #1298) but is now stale.

Questions / Nits

1. Missing ep in H200 config (see above): The H100 MiniMax FP8 config uses { tp: 4, ep: 4 }. If TP=4 without EP on 8×H200, that means only 4 of 8 GPUs are utilized. Is this the intended configuration, or should EP=4 be added? Fix this →

2. Stale perf-changelog description: The changelog mentions v0.20.1-ubuntu2404 but the config uses v0.20.2. Fix this →

---

| CI run (green)

[functionstackx]

update recieps docs plz

[anish-shanbhag]

Opened a recipes PR here: vllm-project/recipes#474

[functionstackx]

lgtm

[anish-shanbhag]

@functionstackx vllm-project/recipes#474 was merged, could you please approve this PR? thanks

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26137455891
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26137455891

[functionstackx]

lgtm

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26468137754
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26468137754