vLLM + openai/gpt-oss-20b on 3× RTX 3090 (CUDA 12.8) — FlashAttention Error

1

Hi,

Trying to run openai/gpt-oss-20b on vLLM with:

  • 3 × RTX 3090 (24 GB each)

  • CUDA 12.8, Driver 570.xx

  • Python 3, PyTorch 2.7.0

I want to split the model across all GPUs, but I’m getting a FlashAttention error (likely needs FA3).

Any tips for multi-GPU vLLM setup and installing FlashAttention 3 for this environment?

Thanks!