Open
Description
I run it on my Mac mini m1, but the loss is always nan.
Then, I tried it on the Ubuntu and It worked.
It seems that pytorch doesn't work well on m1.
{{ message }}
I run it on my Mac mini m1, but the loss is always nan.
Then, I tried it on the Ubuntu and It worked.
It seems that pytorch doesn't work well on m1.
Activity
xshaun commentedon May 27, 2021
Is the random seed same one in your settings? It would be helpful if you have a try on other models and give more runtime information.
shreyaspadhy commentedon Sep 6, 2021
Hi Dongqi, did you have any followups to this issue? I am running into a closely related issue on an M1 Mac Mini where the default MNIST example does not seem to have good test time evaluation (accuracy is always chance ~10%), but the training loss goes down a significant amount. However, on a 2017 x86 Macbook Pro, the test accuracy reaches ~99% within 2 epochs, with similar training loss as the M1 machine. I'm happy to provide any further information to get to the bottom of this.
On both machines, I set up a Python, Conda and Pytorch Environment from scratch, following this tutorial pretty much exactly - https://towardsdatascience.com/yes-you-can-run-pytorch-natively-on-m1-macbooks-and-heres-how-35d2eaa07a83
I'm on Python 3.9.6 on both machines.
msaroufim commentedon Mar 9, 2022
Super strange, that said for M1 updates I'd suggest following this thread instead pytorch/pytorch#47702 (comment)