"No source for code" error when dealing with PyTorch's Dynamo

[BurnzZ]

Describe the bug

It would seem that the following error is encountered when importing torchvision or transformers.models.distilbert.DistilBertModel (which basically imports torch._dynamo):

No source for code: '/private/var/folders/lh/yz1yvyvx5nd2vqq6t9t3qbmw0000gn/T/tmpavruueea/_remote_module_non_scriptable.py'.

To Reproduce
How can we reproduce the problem? Please be specific. Don't link to a failing CI job. Answer the questions below:

1. What version of Python are you using? Python 3.13.1
2. What version of coverage.py shows the problem? 7.6.10
3. What versions of what packages do you have installed? torch==2.6.0, torchvision==0.2.1, transformers==4.48.2

Install the latest versions of these packages: pip install torch torchvision transformers.

Create 2 files with these contents:

foo.py

def say_hello():
    return "hello"

test_foo.py

from foo import say_hello

def test_foo():
    assert say_hello() == "hello"

Now, if we run the following code, it's all good ✅:

coverage run test_foo.py
coverage report

However, when adding any the following imports test_foo.py:

import torchvision
from transformers.models.distilbert import DistilBertModel

import torch._dynamo

... then coverage report results in the "No source for code: ..." error (status code of 1). ❌

Expected behavior

In this example, the mere imports of the given modules are enough to cause the issue.

Additional context

It was discovered in bazel-contrib/rules_python#2575.

It can be solved by having the --ignore-errors flag as was recommended by the author in https://stackoverflow.com/a/2401206.

However, it's still worth documenting this weird behavior.

[nedbat]

Thanks for the reproducible report. It seems that torch creates and executes a file in your temp directory.

Some other ways to avoid the message:

coverage run --source=. test_foo.py
# or:
coverage run --omit="$TMPDIR/*" test_foo.py

I'm wondering if there's something coverage could do for this as a default.

[nedbat]

However, it's still worth documenting this weird behavior.

I'm not sure what you want added to the documentation. This behavior is due to an intricacy of PyTorch.

[BurnzZ]

Thanks for checking it out @nedbat 🎉

I'm not sure what you want added to the documentation. This behavior is due to an intricacy of PyTorch.

Indeed. Though I think it might be useful if there was a hint about the error, since No source for code: ... isn't clear upfront what it means. Perhaps another log msg pointing to https://coverage.readthedocs.io/en/7.6.10/cmd.html#reporting?

[nedbat]

I see what you mean. The message could be longer: "Couldn't find source for code observed at ...." or something? I'm not coming up with a one-line message that explains it.

[BurnzZ]

Yeah, expanding it would be great. I think it would be clearer if we change " ... code observed at ... " into "... code executed ...".

Would this technically still be correct?

[nedbat]

"Source file no longer available. Was executed at ..." ?

[BurnzZ]

+1 much better.