Does cublasLt support rowwise/blockwise scaled Matmul?

[ghostplant]

I only see point-wise Matmul examples in the tutorial.

[rsdubtso]

Hi @ghostplant. cuBLAS does not support rowwise/blockwise yet. For the record, on Blackwell it supports block-scaling https://docs.nvidia.com/cuda/cublas/#d-block-scaling-for-fp8-and-fp4-data-types, which provide a similar level of accuracy.

[ghostplant]

Thanks, does Blackwell support FlashMLA and DeepGemm? I found plenty of dependencies incompatible with B200, so I have to use H100 if there is no similar solutions on Blackwell.

[rsdubtso]

My understanding is that DeepSeek did not implement FlashMLA and DeepGemm for Blackwell at least yet.

[ghostplant]

Sure, I'll use H100 instead.