B200/B300 support: current FA4 attention backend underperforms and lacks tuned SM10x kernel path #113

Trying to benchmark Helios-Distilled T2V on NVIDIA B300 SXM6 AC, but the current attention backend path does not seem to provide a fully optimized B200/B300 path.

Environment

• GPU: NVIDIA B300 SXM6 AC
• Compute capability: sm_103
• CUDA: 13.0
• PyTorch: tested with 2.11.0+cu130 and 2.12.0+cu130
• diffusers: 0.38.0
• kernels: 0.14.1
• flash-attn-4:
  • kernels-community/flash-attn4 v0
  • also tested latest GitHub source from Dao-AILab/flash-attention main

But results arount 14 fps on single card

What we observed

The current Helios code selects _flash_3_hub for cuda_major >= 9:

cuda_major = torch.cuda.get_device_capability()[0]
if cuda_major >= 9:
    transformer.set_attention_backend("_flash_3_hub")
else:
    transformer.set_attention_backend("flash_hub")