[QST] I am writing a mix-dtype input gemm for blackwell but no example

[haowen-han]

What is your question?
I am developing a GEMM operator for the Blackwell GPU where the B matrix uses MXFP4 and the A matrix uses MXFP8. However, there are no relevant examples for either CUTE C++ or CUTE DSL. What's more, CUTE DSL even requires the data types of the A and B matrices input to MMA to be identical, but in reality, the Blackwell architecture supports mismatched data types. Could anyone provide relevant examples?

[CalebDu]

Hi @haowen-han, cutlass c++ has mxf8* mxf4 example in examples/72_blackwell_narrow_precision_gemm/72c_blackwell_mixed_mxfp8_bf16_gemm.cu. Cute DSL can not mix dtype in current, will support it in future release soon.