Hi,
I was wondering about the reason behind torch.matmul being only supported for float in PyTorch? I would really appreciate if someone can describe the best work around for it without losing performance.
Thanks
Hi,
I was wondering about the reason behind torch.matmul being only supported for float in PyTorch? I would really appreciate if someone can describe the best work around for it without losing performance.
Thanks
This is a known upstream issue: Add support for integer matrix multiplication (particularly for dtype = torch.int8 ) · Issue #29961 · pytorch/pytorch (github.com)
As for a workaround if you know the dynamic range of your integers you can do exact integer accumulation in float16 from to -2**11
to 2**11
and float32 from -2**24
to 2**24
.