[kit1980]

Before this PR, TorchFix in standalone mode crashes on

from torch import ger
just_name = ger(pos_seq, inv_freq)

With this PR it works fine, replacing the code with

from torch import outer, ger
just_name = outer(pos_seq, inv_freq)

Note that the replacement code still imports no longer needed ger, but that can be caught and fixed by existing linters like ruff.
Nevertheless, I'll create an issue to clean this up later in TorchFix.

[kit1980]

Issue to clean up uneeded imports: #24

[izaitsevfb]

LGTM

looks like it should work correctly with nested qualified names, but a test for that may be warranted.

[kit1980]

LGTM

looks like it should work correctly with nested qualified names, but a test for that may be warranted.

What do you mean exactly, could you give an example?

[izaitsevfb]

LGTM
looks like it should work correctly with nested qualified names, but a test for that may be warranted.

What do you mean exactly, could you give an example?

Something like this:

import torch

tp_mesh = torch.distributed.device_mesh.init_device_mesh("cuda", (8,))

[kit1980]

Something like this:

import torch

tp_mesh = torch.distributed.device_mesh.init_device_mesh("cuda", (8,))

I don't understand. init_device_mesh is not a known deprecated function for TorchFix, so nothing will happen with the code.

Do you mean a case when there is more than 1 dot in the qualified name?
There is no deprecated functions with replacements like this currently, but there is a tested functionality to replace, for example, functorch.vmap() with torch.func.vmap().
And that's a different code path from the bug this PR addresses.

[izaitsevfb]

I don't understand. init_device_mesh is not a known deprecated function for TorchFix, so nothing will happen with the code.

Do you mean a case when there is more than 1 dot in the qualified name? There is no deprecated functions with replacements like this currently, but there is a tested functionality to replace, for example, functorch.vmap() with torch.func.vmap(). And that's a different code path from the bug this PR addresses.

Sorry for not being clear. In my example I was indeed referring to the hypothetical case when init_device_mesh is deprecated, and is used in the code with multiple nested qualifying names like torch.distributed.device_mesh.init_device_mesh.

If that case is supported and tested, then awesome.