(misc) Deprecate some hf-inference specific features (wait-for-model header, can't override model's task, get_model_status, list_deployed_models)

[Wauplin]

In this PR:

• deprecate sentence-similarity (use feature-extraction instead) => not deprecated after all (see comment)
• kill ability to override model's default task ([inference] Kill the ability to override a model's task huggingface.js#1199). This means a model will be either sentence-similarity, either feature-extraction but not both
• deprecate get_model_status (will be replaced by HfApi.model_info
• deprecate list_deployed_models (will be replaced by HfApi.list_models)
• do not use X-wait-for-model anymore
• raise on HTTP 503 => previously we were waiting for 1s and then retrying. Now that all models will be warm-only, if shouldn't get a 503 (and if we do so, we should raise and let the user define what they want to do)
• remove Make sure '{request_parameters.task}' task is supported by the model. msg in error => since now task is necessarily supported otherwise it would have failed before

=> the http 503 thingy is a breaking change but mostly due to the fact that forbidding cold starts in Inference API is a breaking change => so not a problem

Related:

• [inference] Kill the ability to override a model's task huggingface.js#1199
• Inference: remove the hf-only options now that the API is warm-only huggingface.js#1192

[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.

[hanouticelina]

thank you for taking care of this! @Wauplin

[julien-c]

I'm not sure about

deprecate sentence-similarity (use feature-extraction instead)

given it'll still be a task on hf.co/tasks for the foreseeable future.

Personally i wouldn't remove it 🤷

Everything else LGTM

[Wauplin]

given it'll still be a task on hf.co/tasks for the foreseeable future.
Personally i wouldn't remove it 🤷

Ok I added it back in 4218f8d but removed the ability to call f"{self.base_url}/pipeline/{self.task}/{mapped_model}". This means a model can be used either with feature-extraction or sentence-similarity but not both

[julien-c]

yes, perfect imo

[SpecLad]

Does this change mean that other clients should also stop using X-Wait-For-Model? I see that it's still documented at https://huggingface.co/docs/api-inference/parameters#wait-for-the-model.

[Wauplin]

Hey @SpecLad ,

Does this change mean that other clients should also stop using X-Wait-For-Model?

Yes it does. You can send it but it won't have any effect anymore

I see that it's still documented at https://huggingface.co/docs/api-inference/parameters#wait-for-the-model.

Thanks for reporting. We are currently revamping these docs (see huggingface/hub-docs#1652) and will remove it from there.