Generate inference types + start using output types

[Wauplin]

cc @SBrandeis @julien-c for visibility.

This PR adds scripts to generate inference types + start to use the output types in the InferenceClient.

The broader goal is to use all the spec-ed types in the InferenceClient but also in transformers and diffusers pipelines.

How it works:

• generating script in huggingface.js/tasks => done separately
• utils/generate_inference_types.py => clean Python generated code + generate imports + generate package reference
• all generated classes inherit from src/huggingface_hub/inference/_generated/types/base.py
  • dataclass + inherit from dict for backward compatibility => goal is to make it dataclass-only in the future (1.0?)
  • adds parsing methods to avoid pydantic
  • all fields are optional E.g. don't raise if server "forget" a required field
  • extra fields are set as attributes for future compatibility
• output types are now used in src/huggingface_hub/inference/_client.py

Docs:

• Inference types
• Inference client

TODO:

• add types to inference package reference in docs
• how to handle parsing audio or images in output? currently done "manually", e.g. not from the json specs
• fix tests
• fix type annotations in _client.py...
• make dataclasses first class citizen in huggingface_hub?
• better auto-generated docstrings => following HF docs format => will be done later

Breaking changes in outputs 💔:

• automatic_speech_recognition: str => AutomaticSpeechRecognitionOutput
• summarization: str => SummarizationOutput
• translation: str => TranslationOutput

Left out of scope for this PR:

• text_generation => requires more typing for stream=True
• deal with conversational => will be deprecated/removed soon
• use input parameters generated from specs

[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.

[codecov]

Codecov Report

Attention: Patch coverage is 97.85047% with 23 lines in your changes are missing coverage. Please review.

Project coverage is 82.71%. Comparing base (e7f243c) to head (e25bcaf).

❗ Current head e25bcaf differs from pull request most recent head 56955a8. Consider uploading reports for the commit 56955a8 to get more accurate results

Files	Patch %	Lines
...gingface_hub/inference/_generated/_async_client.py	21.42%	22 Missing ⚠️
src/huggingface_hub/inference/_client.py	96.42%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2036      +/-   ##
==========================================
+ Coverage   80.70%   82.71%   +2.01%     
==========================================
  Files          71      102      +31     
  Lines        8519     9490     +971     
==========================================
+ Hits         6875     7850     +975     
+ Misses       1644     1640       -4     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[julien-c]

as discussed irl, looks quite cool already!

[Wauplin]

This PR is now ready for review :) 🎉
I've updated the PR description accordingly. This PR will introduce breaking changes for some tasks but hopefully not too problematic. I'm planning to make a release of huggingface_hub early next week but no rush to merge this PR. It can wait the next one which would allow to implement a few more TODOs before shipping.

[LysandreJik]

Looks good! Clear implementation, awesome to have the types so neatly defined within the runtime.

It would be awesome to gradually move towards the same input/output definitions within transformers & diffusers pipelines.

(No need to review but FYI @sanchit-gandhi @Rocketknight1 @amyeroberts @molbap @yiyixuxu as we'll want to adopt this soon)

[SBrandeis]

Output is actually a List of those if I'm not mistaken :(

I had to manually parse the TS code to add that array definition in huggingface.js (bug also reported in quicktype)

[Wauplin]

Yeah I already had a script to handle the shared classes (typically ClassificationOutput) but now I pushed 56955a8 so that classes are named AudioClassificationOutputElement, to emphasize that the expected output will be a list of those.

Does it look better for you like this?

[SBrandeis]

Very cool that the inference types are used in the Inference client right away!

[Wauplin]

Let's merge this PR now that everything's addressed! I have listed remaining todos in a follow-up issue: #2063.