[QST] B200 FP8 Prefill Kernels

[yigithanyigit]

Hey!

Do flashinfer has B200 fp8 single prefill kernels different than CUTLASS implementation? as far as I can see there is not but wanted to double check. I am planning to use mainly on Diffusion models, since batched prefill kernels are mostly designed for streaming, I believe only options is the cutlass backend. I wonder your thoughts and am I missing something?

Thanks

[yzh119]

#1238 It should be supported, let me finalizing this PR.

[yigithanyigit]

Thanks for the response, is there any roadmap for cudnn fp8 backend? CUTLASS fp8 throughput seems not high compared to cuDNN. 🥲