Summary

Cuddn supports fused_attention with FP8 inputs.
See: https://docs.nvidia.com/deeplearning/cudnn/developer-guide/index.html#flash-fused-multi-head-att-fp8

There already exists a PR on PyTorch to add support for the Cudnn Backend to SDPA:pytorch/pytorch#101916

This seems like a potential path attention support for FP8