b200 MiniMax M2.5 FP8

[functionstackx]

@claude use vllm/vllm-openai:v0.15.1-cu130 for TP=2 & TP=4 from conc=4 to 64 on minimax model MiniMaxAI/MiniMax-M2.5

[Klaud-Cold]

Claude finished @functionstackx's task in 69m 32s —— View job • claude/issue-756-20260219-0235 • Create PR ➔

---

Task: Add B200 MiniMax M2.5 FP8 vLLM Benchmarks

• Gather context and understand repo structure
• Identify existing MiniMax and B200 vLLM configs
• Create benchmark script minimaxm2.5_fp8_b200.sh
• Add config entry minimaxm2.5-fp8-b200-vllm to nvidia-master.yaml
• Update perf-changelog.yaml
• Commit and push changes
• Trigger e2e smoke test

Changes

New files:

• benchmarks/minimaxm2.5_fp8_b200.sh — vLLM benchmark script for MiniMax-M2.5 FP8 on B200 (adapted from MI355X script, removing ROCm-specific config)

Modified files:

• .github/configs/nvidia-master.yaml — Added minimaxm2.5-fp8-b200-vllm config:
  • Image: vllm/vllm-openai:v0.15.1-cu130
  • Model: MiniMaxAI/MiniMax-M2.5
  • TP=2 and TP=4, concurrency 4–64
  • Sequence lengths: 1k1k, 1k8k, 8k1k
• perf-changelog.yaml — Added changelog entry

E2E Test Status

Smoke test running (TP=2,4 conc=4 1k1k): Run #22166363309

• TP=4 — in progress on b200-nb_1 (model download + server startup for 456B MoE model)
• TP=2 — queued (waiting for B200 runner)

|