Add MiniMax-M2.5 FP8 vLLM benchmark for B200

[functionstackx]

Add MiniMax-M2.5 FP8 vLLM benchmark configuration for B200 GPUs.

• New benchmark script benchmarks/single_node/minimaxm2.5_fp8_b200.sh
• Config entry minimaxm2.5-fp8-b200-vllm in nvidia-master.yaml
• Image: vllm/vllm-openai:v0.16.0-cu130
• TP=2 and TP=4, concurrency 4-64
• Sequence lengths: 1k1k, 1k8k, 8k1k

Closes #756

Generated with Claude Code

[kedarpotdar-nv]

@claude please review. i made a few changes

[Klaud-Cold]

Claude finished @kedarpotdar-nv's task in 2m 40s —— View job

---

Review of PR #757

• Gather context and read changed files
• Review benchmark script
• Review nvidia-master.yaml config
• Review perf-changelog.yaml
• Post review feedback

---

Bug: Wrong benchmark_lib.sh source path

benchmarks/single_node/minimaxm2.5_fp8_b200.sh:3 sources benchmark_lib.sh with:

source "$(dirname "$0")/benchmark_lib.sh"

This resolves to benchmarks/single_node/benchmark_lib.sh, which does not exist. The file is at benchmarks/benchmark_lib.sh. Every other script in single_node/ uses /../:

source "$(dirname "$0")/../benchmark_lib.sh"

This will cause the script to fail immediately at runtime.

Fix this →

---

Benchmark Script (minimaxm2.5_fp8_b200.sh)

Good changes:

• --enable-expert-parallel (line 32) is appropriate — MiniMax-M2.5 is a MoE model (456B total, ~47B active params). The MI355X version (minimaxm2.5_fp8_mi355x.sh) does NOT have this flag, so this is a deliberate addition for B200.
• --trust-remote-code on both the vLLM serve and benchmark client side is correct for this model.
• --block-size=32, --gpu-memory-utilization 0.95, and the overall script structure match existing patterns well.

No explicit FP8 quantization flags: The script doesn't pass --quantization fp8 or --kv-cache-dtype fp8. This is fine since MiniMax-M2.5 appears to ship with native FP8 weights, matching the approach in the MI355X version.

---

Config (nvidia-master.yaml)

The config entry at line 3018 looks correct:

• runner: b200-nvs — matches the appropriate B200 runner
• image: vllm/vllm-openai:v0.15.1-cu130 — new vLLM version, first config to use this tag
• Search space (TP=2 and TP=4, conc 4-64) across all three seq-len configs is reasonable

No issues found here.

---

Changelog (perf-changelog.yaml)

Content is correct. Minor: trailing whitespace on line 690 (pr-link: line) and trailing blank line at 692. Not a blocker.

---

Summary

One blocking issue: the benchmark_lib.sh source path bug will prevent the script from running. Everything else looks good — the --enable-expert-parallel addition and overall config are solid.
| Branch

[functionstackx]

@claude implement this fix in this PR's runners/launch_b200-dgxc-slurm.sh too

[Klaud-Cold]

Claude finished @functionstackx's task in 1m 5s —— View job

---

Implementing fix in runners/launch_b200-dgxc-slurm.sh

• Gather context and understand the fix from the screenshot
• Implement the fix — change HF_HUB_CACHE_MOUNT path from /home/sa-shared/gharunners/hf-hub-cache/ to /scratch/fsw/gharunners/hf-hub-cache
• Commit and push changes

Updated HF_HUB_CACHE_MOUNT path in runners/launch_b200-dgxc-slurm.sh:216 from /home/sa-shared/gharunners/hf-hub-cache/ to /scratch/fsw/gharunners/hf-hub-cache to match the correct mount point.

View job | Branch

[cquil11]

Updated sweep: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/22411684877

[functionstackx]

@ankursingh-nv i cancelled the workflow as it seems to be hard erroring out

[functionstackx]

@ankursingh-nv can u set the ep setting in nvidia-master.yaml and then in this script do if ep is set then enable-experrt-parallel

[functionstackx]

@claude create an new PR from remote main that talks situtations where vllm serve doesnt use expert-parallel-size directly --enable-expert-parallel in CLAUDE.md, claude.yml, claude-review.yml

[Klaud-Cold]

Claude finished @functionstackx's task in 2m 7s —— View job

---

Creating PR for --enable-expert-parallel documentation

• Gather context: read CLAUDE.md, claude.yml, claude-pr-review.yml
• Create new branch from remote main
• Add documentation about --enable-expert-parallel vs --expert-parallel-size usage
• Commit and push changes
• Create PR

---

Created PR #864: Document expert parallelism flag differences in Claude configs

Changes made to claude.yml and claude-pr-review.yml:

• vLLM uses --enable-expert-parallel (boolean flag) — does NOT accept --expert-parallel-size
• SGLang uses --expert-parallel-size N (explicit integer)
• ATOM (AMD vLLM fork) uses --enable-expert-parallel (same as vLLM)
• Scripts must conditionally enable --enable-expert-parallel based on EP_SIZE env var, not hardcode it
• Added PR review validation rule to flag hardcoded --enable-expert-parallel

Note: There is no CLAUDE.md file in this repo — the Claude instructions live in the workflow prompt: fields in claude.yml and claude-pr-review.yml.
| Branch

[functionstackx]

@ankursingh-nv @kedarpotdar-nv can u also link vllm recipes for minimax and kimi k2.5 in the official nvidia docs too

https://docs.nvidia.com/llm-inference-quick-start-recipes/index.html

[ankursingh-nv]

Sure, let me address the ep setting.

[functionstackx]

@claude update PR description, seems like the image is now 0.16 and not 0.15.1

[Klaud-Cold]

Claude finished @functionstackx's task in 32s —— View job

---

Updating PR Description

• Check current image version in config
• Update PR description

Updated the PR description to reflect the correct image version: vllm/vllm-openai:v0.16.0-cu130 (was v0.15.1-cu130).

| Branch

[functionstackx]

document this too please in the vllm recipes