I updated my pytorch version from 2.0.1 to 2.1.0 today and noticed that torch.backends
no longer has opt_einsum as an attribute. The error can be simply reproduced by running
torch.backends.opt_einsum
When I go into the source code of torch.backends
, I found the following block of code
from torch.backends import (
cpu as cpu,
cuda as cuda,
cudnn as cudnn,
mkl as mkl,
mkldnn as mkldnn,
mps as mps,
openmp as openmp,
quantized as quantized,
)
which seems to suggest that opt_einsum has removed from the imports. However, the directory opt_einsum
still exists in torch/backends
.
I’m just curious if this change is intentional, and if so what is the reason?