Add an implementation of HF model and an example sentiment analysis a…

[mhawasly]

No description provided.

[fdalvi]

Thanks a lot for starting this, will be very good to have in the package; I left a few comments, and have some more higher level questions:

• Is this tied to a specific class of models? I see in the Inference docs https://huggingface.co/docs/api-inference/detailed_parameters that the parameters differ per task type? If not lets leave a comment (perhaps module level) that tells the users what they are expected to put in the "prompt fn"
• Currently, there is no "input" key in the payload we send to the API, while all samples in their docs have the key - maybe "text" was an older API that is not official anymore?
• Perhaps we can make things even clearer by changing the name of the model to HuggingFaceInferenceAPIModel; I know thats quite a mouthful but there are HuggingFace Inference Endpoints that are paid, so we should disambiguate from that (+ also disambiguate from general HuggingFace models that run locally)

Let me know what you think and if something is unclear!

[mhawasly]

Thanks a lot for starting this, will be very good to have in the package; I left a few comments, and have some more higher level questions:

* Is this tied to a specific class of models? I see in the Inference docs `https://huggingface.co/docs/api-inference/detailed_parameters` that the parameters differ per task type? If not lets leave a comment (perhaps module level) that tells the users what they are expected to put in the "prompt fn"

Indeed, different task types have different input/output formats and types. I tried only an example from the classification type.

* Currently, there is no "input" key in the payload we send to the API, while all samples in their docs have the key - maybe "text" was an older API that is not official anymore?

Interesting. I did try that but "inputs" does not work, at least for the model I used in the asset. It returns {"error":"You need to specify either textortext_target.","warnings":["There was an inference error: You need to specify either textortext_target."]}'

* Perhaps we can make things even clearer by changing the name of the model to `HuggingFaceInferenceAPIModel`; I know thats quite a mouthful but there are HuggingFace Inference Endpoints that are paid, so we should disambiguate from that (+ also disambiguate from general HuggingFace models that run locally)

Sure!

Let me know what you think and if something is unclear!

[fdalvi]

Thanks a lot for the changes; The only big thing I'm worried about is the mismatching APIs ("inputs" in the docs vs "text" that seems to work for us) - we should get this committed nevertheless, but if you have some time lets try to dig into this and see why this is happening

[fdalvi]

Any particular reason for the sleep here?

[mhawasly]

Hoping to give the model some time to load before retrying?

[fdalvi]

The retry mechanism has an inherent random delay (that gets sampled from an exponentially increasing range), so we don't need to worry about it here