Closed kevinstephano closed 1 year ago
Interesting, it seems to be tested in PyTorch: https://github.com/pytorch/pytorch/blob/c6d8d10b3e974019dae7ec91a85c6192c6d511fa/torch/testing/_internal/common_methods_invocations.py#L19823-L19826 Maybe it was working in the previous version of nvFuser.
Yes, this was a relatively recent bug.
🐛 Describe the bug
This bug with
var_mean
is very specific to the case where the reductions occur across all dimensions of the input tensor. If you remove 1, it will be fine.Error :
Versions
ToT