Issue evidence · Run · kijai/ComfyUI-WanVideoWrapper

I need help with my A100 GPU; I can't use sageattn or flash-attn, only sdpa works

GitHub issue: https://github.com/kijai/ComfyUI-WanVideoWrapper/issues/1829

Not run

On an A100, SageAttention and flash-attn fail in the Wan wrapper and only SDPA works.

What Badgr ran

Nothing was run; read the reporter's log only.

What came back

  • SageAttention logs "Input tensors must be in dtype of torch.float16 or torch.bfloat16" repeatedly, so the tensors it receives are another dtype.
  • flash-attn fails with "TypeError: only integer tensors of a single element can be converted to an index".
  • The reporter notes A100 has no fp8, so all models use float16 and bfloat16.

The first message points to the model or attention inputs being in a dtype the kernels reject, which looks like a configuration or dtype issue to check rather than a missing install. Not confirmed by a run.

Checked 2026-10-14 in Badgr’s local development environment.

← All issue evidence