Closed yelite closed 1 year ago
@yelite Continuing dense
vs matmul
terminology issue, so far CUTLASS BYOC has only supported dense, so dense
terminology is used everywhere. For example, in the code below we should rename DenseOp
to MatmulOp
and replace the condition with func_name.find("dense") != std::string::npos || func_name.find("matmul") != std::string::npos
.
This PR adds support of matmul, including the variants with bias and with relu/gelu, to relax cutlass BYOC. This PR also makes cutlass profiler and codegen support column major RHS matrix, which is the default layout for
R.matmul
.There will be a followup PR to add batched matmul.
cc: @vinx13 @masahi @mbaret