Closed vatai closed 3 years ago
mmmmm now torch kernell spends a bi of time converting tensors while in cblas they were simply ignored
the current state is clearly not good, but we need some smarter soluon like converting or generating data depending on framework bun not in call method
This is why we do reviews... I might have a solution for that...
cblas
(and other non python)gemm
and friends were importing torch. As a (very undesirable) result you needed to have pytorch installed, even if you wanted to measure some non-pytorch, C/C++ gemm benchmark. Hopefully this helps with the issue.The original error message was: