Is there a way to set M1 mps fallback environment variable in-code?

1

I want to automatically set 'PYTORCH_ENABLE_MPS_FALLBACK=1' environment variable.

I tried

import os
os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1'

It didn’t work.
Is there a direct way to set it in Pytorch, e.g. torch.backends.mps.enable_fallback().

2

I would recommend exporting environment variables directly to your terminal as they are often set too late in the Python script and this won’t be read again.

3

I’m unsure what you mean. What won’t be read again?

4

Certain environment variables are read only during the import of a library and often users set them too late, which can cause confusion.
For this reason, I would not recommend setting any environment variables inside the Python script but to properly export them in your terminal before launching the Python process.

1 Like
5

This is helpful, thanks!