Older version of PyTorch: with torch.autocast('cuda'): AttributeError: module 'torch' has no attribute 'autocast'

autograd
1

I have two machines that I need to check my code across one is Ubuntu 18.04 and the other is Ubuntu 20.04.

In Ubuntu 18.04 I have:

$ python
Python 3.8.0 (default, Dec  9 2021, 17:53:27)
[GCC 8.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.__version__
'1.9.1+cu102'

However, the code that works in Ubuntu 20.04, throws this error:

    with torch.autocast('cuda'):
AttributeError: module 'torch' has no attribute 'autocast'

I have this version of PyTorch on Ubuntu 20.04:

 python
Python 3.8.10 (default, Nov 26 2021, 20:14:08)
[GCC 9.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.__version__
'1.10.1+cu113'

Here’s the GPU info on Ubuntu 20.04:

Screen Shot 2022-02-01 at 2.45.51 AM
Screen Shot 2022-02-01 at 2.45.51 AM814Ă—566 59.7 KB

and here’s the GPU info on Ubuntu 18.04:

Screen Shot 2022-02-01 at 2.46.59 AM
Screen Shot 2022-02-01 at 2.46.59 AM822Ă—551 59.4 KB

Ideally I want the same code to run across two machines.

2

The best approach would be to use the same PyTorch release on both machines. If that’s not possible, and assuming you are using the GPU, use torch.cuda.amp.autocast.

1 Like
3

Thank you. I will spend some more time digging into this but

    with torch.cuda.amp.autocast:
AttributeError: __enter__

didn’t work as well. The error is unfortunately not super descriptive or guiding me how to fix it.

4

You have to “call” the decorator as given in the docs and examples:

with torch.cuda.amp.autocast():
    ...
2 Likes