[1.0] Httpx migration

[Wauplin]

Draft to migrate from requests to httpx.

Opening this PR to trigger CI. Mid-term goal is to merge the branch into a "v1-milestone" branch.

Breaking changes so far:

• "per method" proxies ignored => should be set globally
• configure_http_backend => set_client_factory/set_async_client_factory (and all this low-level stuff in general). Still keeping get_session the same.
• no more requests.HTTPError raised, only httpx.HTTPError. Good practice is to catch HfHubHttpError (both backward and forward compatible)
• httpx doesn't have the concept of SSLError (it's just a generic ConnectError). In the download process we used to raise on SSLError (which is a subclass of ConnectionError) but not other ConnectionErrors. Let's reassess this later.
• InferenceClient can now be used as a context manager with InferenceClient(...) as client:. Benefit is to be sure connections are closed (especially useful when streaming tokens from an LLM)
• trust_env removed from AsyncInferenceClient init parameters
• LocalEntryNotFoundError do not derive from HTTPError anymore. We now have EntryNotFoundError(Exception), LocalEntryNotFoundError (file not found + entry not found errors) and RemoteEntryNotFoundError (http error + entry not found error)

Resources:

• https://www.python-httpx.org/compatibility

[HuggingFaceDocBuilderDev]

It looks like you've updated code related to model or dataset cards in this PR.

Some content is duplicated among the following files. Please make sure that everything stays consistent.

• src/.../repocard.py
• src/.../datasetcard_template.md
• src/.../modelcard_template.md
• modelcard.md (hub-docs repo)
• docs/hub/model-cards.md (hub-docs repo)
• docs/hub/model-card-annotated.md (hub-docs repo)
• datasetcard.md (hub-docs repo)
• docs/hub/datasets-cards.md (hub-docs repo)

[Wauplin]

Script to test inference locally:

import asyncio

from huggingface_hub import AsyncInferenceClient, InferenceClient


CHAT_MODEL = "openai/gpt-oss-20b"
CHAT_MESSAGE = {"role": "user", "content": "What is the capital of France?"}

# chat completion - sync

print("\nChat completion - sync\n")

client = InferenceClient(provider="groq", model=CHAT_MODEL)
stream = client.chat.completions.create(messages=[CHAT_MESSAGE], stream=True)
for chunk in stream:
    print(chunk.choices[0].delta.reasoning or chunk.choices[0].delta.content, end="")


# chat completion - async

print("\n\nChat completion - async\n")


async def main():
    client = AsyncInferenceClient(provider="groq", model=CHAT_MODEL)
    stream = await client.chat.completions.create(messages=[CHAT_MESSAGE], stream=True)
    async for chunk in stream:
        print(chunk.choices[0].delta.reasoning or chunk.choices[0].delta.content, end="")


asyncio.run(main())


# image to image - sync

print("\n\nImage to image - sync\n")

client = InferenceClient(provider="fal-ai", model="Qwen/Qwen-Image-Edit")
image = client.image_to_image("https://cataas.com/cat", prompt="turn the cat into a tiger")
image.save("tiger_sync.jpg")

# image to image - async

print("\nImage to image - async\n")


async def main():
    client = AsyncInferenceClient(provider="fal-ai", model="Qwen/Qwen-Image-Edit")
    image = await client.image_to_image("https://cataas.com/cat", prompt="turn the cat into a tiger")
    image.save("tiger_async.jpg")


asyncio.run(main())

[hanouticelina]

Very excited about this! 🔥
the number of loc changes and file changed might seems scary but it's quite easy to review! I didn't review the tests yet though, I will do that in a second round of review 😄

[hanouticelina]

is it necessary to make this function async since it makes no I/O and doesn't await anything?

[Wauplin]

good point. Not needed indeed 👍 (I changed the logic several times TBH and at first it was required)

[hanouticelina]

Reviewed all the changes made to the tests! all good

[Wauplin]

I've updated the target branch to v1.0-release. Possible to re-review/approve @hanouticelina

I'll handle the documentation part separately.

[Wauplin]

Merging on v1.0-release for now but feel free to comment on this PR.