Pin torchao==0.12.0 to avoid PyTorch ABI warnings, also pin numpydoc>=1.6.0 and ty==0.0.1a21 for compatibility.

[ParagEkbote]

Description

As described in the issue, this PR pins the torchao version to 0.12.0. If we want to support a version higher, we will need to bump the torch version as well. After pinning to this version, the warning does not appear. Could you please review?

cc: @davidberenstein1957

Related Issue

Fixes #411

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update

How Has This Been Tested?

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

Additional Notes

---

Note

Pins torchao to version 0.12.0 to avoid the PyTorch ABI warning.

^{Written by Cursor Bugbot for commit 474202c. This will update automatically on new commits. Configure here.}

[github-actions]

This PR has been inactive for 10 days and is now marked as stale.

[ParagEkbote]

Not stale.

[ParagEkbote]

I've pinned the sphinx version to avoid a conflict with jinja. Currently, the janus test is failing since the version supporting it is not being installed while testing. Can I add a step like pip install --upgrade transformers in the CI?

Secondly, the linting with ty seems to be failing due to updates in the newer version of ty, could we pin this package as well. WDYT?

cc: @davidberenstein1957

[ParagEkbote]

Could you please review?

cc: @davidberenstein1957

[davidberenstein1957]

@ParagEkbote, could you rebase on main_ Also, it seems one of the tests is failing because the Janus architecture is not being recognised? Perhaps we should see if this is caused by wrong transformer versions.

[davidberenstein1957]

why are we doing this?

[ParagEkbote]

We need to pin the sphinx error to prevent a version conflict with jinja. The GH Action logs for the same:https://github.com/PrunaAI/pruna/actions/runs/18781274712/job/53587851292

[ParagEkbote]

Furthermore, I have also pinned the ty version and transformers version, so the test and linting checks are passing. Could you please review?

cc: @davidberenstein1957

[gsprochette]

Sphinx is only a dev dependency so we shouldn't pin it here. I don't understand why we're pinning it in the first place and the logs are not available anymore (sorry for the delay), could you re-explain / re-run the logs please?

[davidberenstein1957]

hi @ParagEkbote sorry for the outcome of this PR. I feel that the resctrictions w.r.t. the versioning don't outweigh the benefits of skipping the warning message in this case. WDYT @johannaSommer ?

[ParagEkbote]

hi @ParagEkbote sorry for the outcome of this PR. I feel that the resctrictions w.r.t. the versioning don't outweigh the benefits of skipping the warning message in this case. WDYT @johannaSommer ?

Can we consider updating the torch version so that a newer version of torchao can be used?

cc: @davidberenstein1957

[johannaSommer]

hi @ParagEkbote sorry for the outcome of this PR. I feel that the resctrictions w.r.t. the versioning don't outweigh the benefits of skipping the warning message in this case. WDYT @johannaSommer ?

Can we consider updating the torch version so that a newer version of torchao can be used?

cc: @davidberenstein1957

Hey @ParagEkbote @davidberenstein1957 thanks for the progress so far! Only catching up now but definitely yes, we can upgrade the torch version and then we should be able to support also newer versions of torchao. Infact @gsprochette is currently working on this torch update, maybe he can give a quick heads up as to when this will be merged.

[github-actions]

This PR has been inactive for 10 days and is now marked as stale.

[ParagEkbote]

Not Stale.

[gsprochette]

Thanks a lot for fixing the torchao version. I think all the pinned dependencies requires a justification. We should try not to restrict the environment too much, and especially for transformers, >=4.55.0 seems restrictive.

[gsprochette]

Is there a reason for pinning the specific version of ty?

[ParagEkbote]

Since ty is a pre-release package, the number of rules and their enforcement levels vary. The pinned version helps to pass the majority of the linting checks in line with the configuration defined in pyproject.toml .

[gsprochette]

Makes sense, can we pin it to 0.0.1a21 then? This is the version already in the current uv.lock

[gsprochette]

why are we restraining the transformer version this much?

[ParagEkbote]

As seen in this pipdeptree output, older versions of gliner had a minimum requirement of transformers>=4.51.0 , but newer versions support a higher version, should I update it and omit transformers v5 for now?

[gsprochette]

I understand that pruna is compatible with transformers>=4.55.0 but here this is not the side that is restrained, you're declaring it's not compatible with transformers<4.55.0 and I don't understand why.
In the issue you linked it sounds like you want to use more recent model architecture, but this is a dependency external to pruna, right? I don't think this should justify restraining the compatibility of pruna with other packages.

[gsprochette]

Sphinx is only a dev dependency so we shouldn't pin it here. I don't understand why we're pinning it in the first place and the logs are not available anymore (sorry for the delay), could you re-explain / re-run the logs please?

[ParagEkbote]

The logs for tests if we don't pin Sphinx are as follows:

=========================== short test summary info ============================
ERROR tests/algorithms/test_algorithms.py - sphinx.errors.VersionRequirementError: 
Sphinx<4.0.2 is incompatible with Jinja2>=3.1.
If you wish to continue using sphinx<4.0.2 you need to pin Jinja2<3.1.
ERROR tests/algorithms/test_algorithms.py - sphinx.errors.VersionRequirementError: 
Sphinx<4.0.2 is incompatible with Jinja2>=3.1.
If you wish to continue using sphinx<4.0.2 you need to pin Jinja2<3.1.
ERROR tests/algorithms/test_algorithms.py - sphinx.errors.VersionRequirementError: 
Sphinx<4.0.2 is incompatible with Jinja2>=3.1.
If you wish to continue using sphinx<4.0.2 you need to pin Jinja2<3.1.
ERROR tests/algorithms/test_algorithms.py - sphinx.errors.VersionRequirementError: 
Sphinx<4.0.2 is incompatible with Jinja2>=3.1.
If you wish to continue using sphinx<4.0.2 you need to pin Jinja2<3.1.
ERROR tests/algorithms/test_algorithms.py - sphinx.errors.VersionRequirementError: 
Sphinx<4.0.2 is incompatible with Jinja2>=3.1.
If you wish to continue using sphinx<4.0.2 you need to pin Jinja2<3.1.
ERROR tests/algorithms/test_algorithms.py - sphinx.errors.VersionRequirementError: 
Sphinx<4.0.2 is incompatible with Jinja2>=3.1.
If you wish to continue using sphinx<4.0.2 you need to pin Jinja2<3.1.
ERROR tests/algorithms/test_algorithms.py - sphinx.errors.VersionRequirementError: 
Sphinx<4.0.2 is incompatible with Jinja2>=3.1.
If you wish to continue using sphinx<4.0.2 you need to pin Jinja2<3.1.
!!!!!!!!!!!!!!!!!!!!!!!!!! stopping after 7 failures !!!!!!!!!!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!! xdist.dsession.Interrupted: stopping after 1 failures !!!!!!!!!!!!!
======================== 8 warnings, 7 errors in 28.94s ========================

Could we pin sphinx as a dev dependency instead? WDYT?

cc: @gsprochette

[gsprochette]

For the Sphinx and Jinja2 issue, I would be in favor of pinning numpydoc>=1.6.0 which seems to be the first version with a pyproject.toml, the CI should already be using something above that in the uv.lock and numpydoc>=1.6.0 relies on Sphinx>=5. If you can leave a quick comment stating that this is fixing dependencies in numpydoc-validation that would be great :) Thanks a lot!

[gsprochette]

I understand that pruna is compatible with transformers>=4.55.0 but here this is not the side that is restrained, you're declaring it's not compatible with transformers<4.55.0 and I don't understand why.
In the issue you linked it sounds like you want to use more recent model architecture, but this is a dependency external to pruna, right? I don't think this should justify restraining the compatibility of pruna with other packages.

[gsprochette]

Makes sense, can we pin it to 0.0.1a21 then? This is the version already in the current uv.lock

[ParagEkbote]

For the Sphinx and Jinja2 issue, I would be in favor of pinning numpydoc>=1.6.0 which seems to be the first version with a pyproject.toml, the CI should already be using something above that in the uv.lock and numpydoc>=1.6.0 relies on Sphinx>=5. If you can leave a quick comment stating that this is fixing dependencies in numpydoc-validation that would be great :) Thanks a lot!

Pinning the numpydoc version has fixed the jinja version conflict and CI, thanks for getting back to me quickly 👍

cc: @gsprochette

[ParagEkbote]

I understand that pruna is compatible with transformers>=4.55.0 but here this is not the side that is restrained, you're declaring it's not compatible with transformers<4.55.0 and I don't understand why.
In the issue you linked it sounds like you want to use more recent model architecture, but this is a dependency external to pruna, right? I don't think this should justify restraining the compatibility of pruna with other packages.

Based on the pipdeptree graph for the given deps, we could bump the min. version to 4.57.3 based on the minimum requirement of gliner and restrict installation of v5, something like:

transformers >= 4.57.3, <5.0

pipdeptree graph for minimum version constraints :

transformers==4.57.3
├── llmcompressor==0.6.0 [requires: transformers>4.0,<5.0]
│   └── pruna==0.3.0 [requires: llmcompressor]
├── hqq==0.2.7.post1 [requires: transformers>=4.36.1]
│   └── pruna==0.3.0 [requires: hqq==0.2.7.post1]
├── compressed-tensors==0.10.2 [requires: transformers]
│   └── llmcompressor==0.6.0 [requires: compressed-tensors==0.10.2]
│       └── pruna==0.3.0 [requires: llmcompressor]
├── whisper-s2t==1.3.1 [requires: transformers]
│   └── pruna==0.3.0 [requires: whisper-s2t==1.3.1]
├── gliner==0.2.24 [requires: transformers>=4.57.3]
│   └── pruna==0.3.0 [requires: gliner]
├── pruna==0.3.0 [requires: transformers]
├── DeepCache==0.1.1 [requires: transformers]
│   └── pruna==0.3.0 [requires: DeepCache]
├── optimum==2.0.0 [requires: transformers>=4.29]
│   └── whisper-s2t==1.3.1 [requires: optimum]
│       └── pruna==0.3.0 [requires: whisper-s2t==1.3.1]
└── vbench-pruna==0.0.1 [requires: transformers]
    └── pruna==0.3.0 [requires: vbench-pruna]

WDYT?

cc: @gsprochette

[gsprochette]

I think you only see this requirement of transformer>=4.57.3 because gliner resolved to 0.2.24 on your side, but lower versions are compatible with the requirements of pruna. Running uv pip install --dry-run transformers==4.37 pruna to check compatibility resolves, meaning the transformers version can go much lower than you expect.

We should not constraint the transformers version unless the code in pruna is incompatible with these anterior versions, so I'm in favor of not touching the transformers version in this PR.

[ParagEkbote]

I think you only see this requirement of transformer>=4.57.3 because gliner resolved to 0.2.24 on your side, but lower versions are compatible with the requirements of pruna. Running uv pip install --dry-run transformers==4.37 pruna to check compatibility resolves, meaning the transformers version can go much lower than you expect.

We should not constraint the transformers version unless the code in pruna is incompatible with these anterior versions, so I'm in favor of not touching the transformers version in this PR.

Do we need to guard against a future transformers v5 for breaking changes or not touch that as well?

cc: @gsprochette

[gsprochette]

I don't think it's necessary for now to guard against the future transformers versions, we'll restrict compatibility if necessary when the version comes out

[ParagEkbote]

I have unpinned transformers as suggested. Could you please review?

cc: @gsprochette

[gsprochette]

Looks good to me, thanks for your work :) @johannaSommer it would be cool to merge this soon to stabilize the CI