Extend import utils to cover "editable" torch versions

[bhack]

What does this PR do?

Fixes #28999

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[bhack]

Why formatting error we have? it isn't clear from the CI log

[khipp]

The repository uses double quotes for string literals. You can format your code by running 'make style' (see contributor guideline).

[bhack]

Done, who can review this?

[amyeroberts]

@bhack Thanks for opening this PR!

For anyone who's coming to this PR in the future, could you share the torch version in your environment i.e. when running pip list | grep torch?

I've requested a review from @ydshieh as he is the king of all things version and package handling. I'd like to get his opinion before we merge. He's off for a for days, so would next week when this can get merged in.

[bhack]

Gently ping

[amyeroberts]

Gentle ping @ydshieh

@bhack Could you also share the torch version listed in your environment, as requested?

[bhack]

It is in the mentioned ticket in the description. This is for the devpytorch versions.

[amyeroberts]

@bhack Yes, I understand it's for the dev version, what I'm asking for is how that appears in your env when running pip list | grep torch

[bhack]

These are the nightly releases.. you can find the patterns at https://download.pytorch.org/whl/nightly/torch/

[ydshieh]

Hi @bhack Thank you for the PR.

I am up to have this, but I am worried about the fact that the dev version might cause problem in the ecosystem.
For example, what would happen

version.parse(_torch_version)
and
version.parse(_torch_version) >= version.parse("2.1.1")

Would they fail? Would they give the correct results? ect.

[bhack]

Are we not working with this on dev only when the import fail?

[ydshieh]

Well, I might lack some context, but my understanding is

if importlib.metadata.version(pkg_name) gives importlib.metadata.PackageNotFoundError (which happens with dev), we use importlib.import_module(pkg_name) to grab the version (and this means the import will work).

So dev could be imported, and we can grab its version. But it is unclear to me if the usages (that I listed in the previous comment) will work or give the correct results.

Hope the above make my concern a bit more clear

[bhack]

Yes but I don't understand your example. Where we are using this?

version.parse(_torch_version)
and
version.parse(_torch_version) >= version.parse("2.1.1")

[bhack]

If you meant here:

transformers/src/transformers/utils/import_utils.py

Line 292 in 6957c09

return version.parse(_torch_version) >= version.parse("2.1.1")

We can add the 'dev` version as it has the date as a fake version and control the date.

[bhack]

E.g just tested the PR code today and the assert is ok

``` 2.3.0.dev20240314`

import importlib
from importlib import metadata
package_exists = importlib.util.find_spec("torch") is not None
pkg_name="torch"
try:
    # Primary method to get the package version
    package_version = importlib.metadata.version(pkg_name)
    package_exists = True
except importlib.metadata.PackageNotFoundError:
    # Fallback method: Only for "torch" and versions containing "dev"
    if pkg_name == "torch":
        try:
            package = importlib.import_module(pkg_name)
            temp_version = getattr(package, "__version__", "N/A")
            # Check if the version contains "dev"
            if "dev" in temp_version:
                package_version = temp_version
                package_exists = True
            else:
                package_exists = False
        except ImportError:
            # If the package can't be imported, it's not available
            package_exists = False
    else:
        # For packages other than "torch", don't attempt the fallback and set as not available
        package_exists = False
print(f"Detected  version {package_version}")
print(f"Detected  version: {package_version}")
from packaging import version
assert(version.parse(package_version) >= version.parse("2.1.1"))

[ydshieh]

Very nice! Thank you for checking, that's exactly what I am looking for.

@amyeroberts I tried to do the same (install torch editable version, but it took forever). The changes in this PR LGTM however.

[ydshieh]

Thank you

[ydshieh]

BTW, @bhack could you rebase the PR on the latest main branch? Thanks.

[bhack]

Done

[bhack]

I tried to do the same (install torch editable version, but it took forever). The changes in this PR LGTM however.

Strange, follow the official mentiond pytorch contributing markdown for rightly

It is quite fast to prepare.

[ydshieh]

pytorch contributing

well, I was trying to use pip -e and found probably an old doc (in pytorch forum) that is probably not working anymore

[bhack]

No it is https://github.com/pytorch/pytorch/blob/main/CONTRIBUTING.md#nightly-checkout--pull

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[amyeroberts]

Thanks for working on this!

Just one comment

[ydshieh]

Still good to me, but we don't need the last commit Restore package_exists.

I will leave @amyeroberts to make the final decision however.

[bhack]

I think the same but it was explicitly requested. So, as you like.

[ydshieh]

I think the same but it was explicitly requested. So, as you like.

Sure :-)

(but you can still express you thought instead of just following the request 😃 )

[amyeroberts]

Thanks for enabling this @bhack!