Is FP8 tensor supported in GEMM?

[akaitsuki-ii]

In the code of GemmSm90 it said:

Constraints:

• Supported input data types: fp16, fp8 (e4m3fn, e5m2)

However in the GemmWrapperBase.validate_tensor method, fp8 tensor will break the assertion.

If fp8 tensor is not actually right now, how to adapt it? How to convert the fp8 tensor to CuTe tensor?

Thanks a lot!

[tridao]

You can add fp8 to the GemmWrapperBase

[tridao]

It's just about the wrapper. The kernel does support fp8. You can run the benchmark her
https://github.com/Dao-AILab/quack/blob/main/benchmarks/benchmark_gemm_sm90.py