[InferenceClient] Add trust_env parameter in InferenceClient

[morgandiverrez]

Description

Add trust_env parameter in InferenceClient

Issue
[InferenceClient / langchain-HuggingFaceEndpoint] impossible to used Langchain HuggingFaceEndpoint whith trust_env=True #2522

[morgandiverrez]

hello, @Wauplin what do you think about this PR ?

[Wauplin]

Hi @morgandiverrez, sorry about the long delay. I wanted to check where exactly this is an issue in langchain integration. Can you pin-point the breaking line on GitHub? I just want to check the context of it

[morgandiverrez]

hello, yes, in this line : https://github.com/langchain-ai/langchain/blob/398718e1cb555e526bbf76ae163ded4c8d3cd1f8/libs/partners/huggingface/langchain_huggingface/llms/huggingface_endpoint.py#L216

if we add in kwargs trust_env=true for set in asyncClient, we have an error by inferenceClient

[Wauplin]

Hi @morgandiverrez, thanks for the reference. First of all, sorry for the delay, I first wanted to check how to deal with this. I just opened a PR to fix things on the langchain side rather than here. See langchain-ai/langchain#26378.

While this PR is correct, I prefer not to add parameters when it's not necessary. As mentioned in langchain-ai/langchain#26378, I think that langchain integration should not assume a perfect match in supported kwargs for both clients. trust_env is a first issue but we might add new ones in the future that are only relevant to one or the other client but not both.

If the PR gets accepted, I would prefer to close this one without merging. Advanced users already have a way to configure trust_env in huggingface_hub globally (see docs) for the sync client and giving the possibility here for a new argument might be misleading (+ create side effects since we are mutating the global session, not used only in InferenceClient). If you have any remarks and questions, please let me know :)

[Wauplin]

As explained above, I'll close this PR without merging in favor of langchain-ai/langchain#26378 that just got merged.