feature: added import complexity checker

[tarekziade]

What does this PR do?

• Adds a new checker in make chek-repo that will import transformers and count the number of imported module.
• Lazy import of torch when doing import transformers

The change will reduce the import time from ~1.5s to ~0.7s (+3000 modules to 619 modules)
The checker will error out when it goes over a threshold - for instance if our code imports a lib like torch not lazily
This will prevent imports slowdowns regression.

Can be used directly to display the import counts and import tree:

✗ python utils/check_import_complexity.py
Import complexity regression: `import transformers` triggered 1890 module imports (maximum allowed: 1500).

Run the following command to display the full import tree and identify the cause:

    python utils/check_import_complexity.py --display

✗ python utils/check_import_complexity.py --display
...
    │   │   │   │   ├── tqdm.contrib
    │   │   │   │   ├── tqdm.contrib.concurrent
    │   │   │   │   ├── huggingface_hub._buckets
    │   │   │   │   ├── huggingface_hub._commit_api
    │   │   │   │   │   └── huggingface_hub.lfs
    │   │   │   │   ├── huggingface_hub._dataset_viewer
    │   │   │   │   ├── huggingface_hub._eval_results
    │   │   │   │   ├── huggingface_hub._inference_endpoints
    │   │   │   │   ├── huggingface_hub._jobs_api
    │   │   │   │   │   └── huggingface_hub._space_api
    │   │   │   │   ├── huggingface_hub._upload_large_folder
    │   │   │   │   ├── huggingface_hub.community
    │   │   │   │   ├── huggingface_hub.repocard_data
    │   │   │   │   ├── huggingface_hub.utils._deprecation
    │   │   │   │   ├── huggingface_hub.utils._verification
    │   │   │   │   └── huggingface_hub.utils.endpoint_helpers
    │   │   │   ├── huggingface_hub.repocard
    │   │   │   └── huggingface_hub._snapshot_download
    │   │   ├── transformers.utils.kernel_config
    │   │   └── transformers.utils.peft_utils
    │   ├── transformers.utils.versions
    │   ├── tokenizers
    │   └── accelerate
    ├── sentencepiece
    ├── mistral_common
    ├── torchvision
    ├── torchaudio
    ├── librosa
    ├── timm
    └── scipy

Total modules imported: 1890

[tarekziade]

The CI fails on purpose to demo it:

https://app.circleci.com/pipelines/gh/huggingface/transformers/169580/workflows/c1bc1013-980f-4895-8a28-e56a2be6e882/jobs/2238716?utm_campaign=vcs-integration-link&utm_medium=referral&utm_source=github-checks-link&utm_content=summary

[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.

[ydshieh]

Not super clear too me, this depends on the environment, no? I mean, say in our .[dev] environment which has almost everything, import transformers will import a lot of modules, then this check will fail (?)

[tarekziade]

Not super clear too me, this depends on the environment, no? I mean, say in our .[dev] environment which has almost everything, import transformers will import a lot of modules, then this check will fail (?)

The test will do an import transformers on a fresh Python interpreter and will detect what modules get pulled in when you do this.

when you install .[dev] you install a lot of extra packages used for developing. In theory those should not be pulled under a vanilla import transformers unless dev installs specific versions of packages that are also in production. But in practice you are absolutely right! we should run this checker in a prod-like environment if we want to make sure we exercise what people out there do.

[LysandreJik]

Love it, and aligned with the direction. Thanks Tarek!

[LysandreJik]

sounds good

[LysandreJik]

quite surprised, there are fewer offending torch import statements than I would have thought!

[github-actions]

View the CircleCI Test Summary for this PR:

https://huggingface.co/spaces/transformers-community/circle-ci-viz?pr=45013&sha=b7fd8f