torch.nn.functional.log_softmax parameter '_stacklevel' undocumented

[cheyennee]

📚 The doc issue

The documentation of the very popular torch.nn.functional.log_softmax doesn't explain what _stacklevel is, why it's set to 3 and what to do with it.

Suggest a potential alternative/fix

No response

cc @albanD @mruberry @jbschlosser @walterddr @kshitij12345 @saketh-are

[vadimkantorov]

Related: #36524 #64038. As discussed there, this _stacklevel is a remnant of migrating to explicit-only dim and to control warnings about not setting dim and the default value used. I think since a long time ago, this _stacklevel argument should be deleted.

[cheyennee]

Related: #36524. As discussed there, this _stacklevel is a remnant of migrating to explicit-only dim and to control warnings about not setting dim and the default value used. I think since a long time ago, this _stacklevel argument should be deleted.

I want to know whether the value of '_stacklevel' can set to string❓ Even value in documentation is set to 3, I found that when it's set to string, such as "max", these two API works normally.

import torch
results={}
arg_1 = torch.rand([4, 12, 141, 141], dtype=torch.float32)
arg_2 = -1
arg_3 = "max"
results['res'] = torch.nn.functional.softmax(arg_1,arg_2,_stacklevel=arg_3,)

[vadimkantorov]

I think the essence of the story is that _stacklevel is passed to some warnings.warn(...) function or even is unused now / cast implicitly. So it's not very interesting to inquire about its possible values. If I'm correct, this argument should be dropped entirely from arguments list, and this is the right course of action about _stacklevel

But hopefully we can hear from devs @albanD @ezyang

[cheyennee]

I think the essence of the story is that _stacklevel is passed to some warnings.warn(...) function or even is unused now / cast implicitly. So it's not very interesting to inquire about its possible values. If I'm correct, this argument should be dropped entirely from arguments list, and this is the right course of action about _stacklevel

But hopefully we can hear from devs @albanD @ezyang

Get it. I think you're right, because just now I found that bool / list or other value could be assigned to _stacklevel. Thank u🥰

[ezyang]

Let's drop it

[albanD]

This argument is only used for the depraction warning raised when dim=None. Otherwise it is not used and its type is not checked indeed.

Since dim=None has been deprecated for 5+ years now (98e6744), I think it is ok to remove it (and remove the _stacklevel arg)

[vadimkantorov]

After dropping, probably even functional python wrapper can be dropped or auto-generated instead?