MiniMaxAI

/

MiniMax-M2.1

like 415

Follow

MiniMax 3.05k

Hi MiniMax team,

I am trying to follow the official vLLM deployment guide for MiniMaxAI/MiniMax-M2.1 on a 4× H100 node, but the engine fails to start with a CUDA error, even though all GPUs are healthy and visible.

Environment

• GPUs: 4× NVIDIA H100 PCIe 80GB (same node)
• nvidia-smi: all 4 GPUs visible, idle, almost no memory used, no running processes
• PyTorch CUDA view:
  • torch.cuda.device_count() == 4
  • torch.cuda.is_available() == True
  • Devices: 0, 1, 2, 3 → all NVIDIA H100 PCIe
• CUDA runtime: 12.8
• Driver: 570.195.03
• Python: 3.12
• vLLM: installed from wheels.vllm.ai/nightly as recommended in the guide
• Model: MiniMaxAI/MiniMax-M2.1

Command (from your vLLM guide)

SAFETENSORS_FAST_GPU=1 vllm serve \
  MiniMaxAI/MiniMax-M2.1 --trust-remote-code \
  --tensor-parallel-size 4 \
  --enable-auto-tool-choice --tool-call-parser minimax_m2 \
  --reasoning-parser minimax_m2_append_think

I also tried adding the suggested flag from the guide:

--compilation-config "{\"cudagraph_mode\": \"PIECEWISE\"}"

but the behavior did not change.

Error

The engine starts initializing, logs the expected config (model, TP=4, bf16, fp8 quant, etc.), then fails during worker startup with:

torch.AcceleratorError: CUDA error: CUDA-capable device(s) is/are busy or unavailable
...
WorkerProc initialization failed due to an exception in a background process.
Engine core initialization failed. See root cause above.

This happens inside torch.cuda.set_device(device) in vllm/platforms/cuda.py, i.e., when vLLM is assigning devices to workers. There are no other processes on the GPUs at that time (verified via nvidia-smi), and TP=4 should be valid because all four H100s are visible and empty.

What I already tried

• Verified GPUs and visibility:
  • nvidia-smi: 4 GPUs, no processes
  • Python check: torch.cuda.device_count() == 4, names printed correctly
  • CUDA_VISIBLE_DEVICES is empty (all GPUs available)
• Tried:
  • VLLM_SKIP_CUDA_P2P_CHECK=1 / VLLM_SKIP_NCCL_P2P_CHECK=1
  • tensor-parallel-size 1 (TP=1 works, TP=4 fails with the error above)
  • Fresh virtualenv with nightly vLLM only, no older vLLM versions

Given that:

• The command line matches your guide,
• Hardware matches your “4× high-end GPUs” recommendation,
• TP=1 works but TP>1 fails at device init on otherwise healthy H100s,

this looks like a specific interaction between MiniMax-M2.1’s fp8/bf16 setup, the vLLM nightly recommended in the docs, and H100 on CUDA 12.8.

Questions

1. Is there a known-good vLLM version / nightly build (commit or version pin) that you recommend for running MiniMaxAI/MiniMax-M2.1 on 4× H100 with TP=4?
2. Are there any additional flags you recommend (e.g. disabling custom all-reduce, specific compilation-config, env vars) for H100 deployments beyond what’s in the vLLM guide?
3. Could you provide an example of a full, working vllm serve invocation + environment spec (driver, CUDA, torch, vLLM) that you have verified on 4× H100 with M2.1?

Thanks in advance for any guidance, and for publishing this model and the vLLM deployment recipe.