True OpenAI drop-in replacement by InferenceClient

[Wauplin]

Closes #2369 @lappemic

Goal is to be able to use InferenceClient exactly the same way as OpenAI client. To do so we need to:

• rename model to base_url => added an alias for it
• rename model_id to model
• rename token to api_key => added an alias for it
• add alias for client.chat.completions.create

@philschmid could you have a look at it and confirm it meets your expectations? See tests for real example (here and here)

Sync + stream=False

client = InferenceClient(
    base_url="https://api-inference.huggingface.co/models/meta-llama/Meta-Llama-3-8B-Instruct",
    api_key="my-api-key",
)
output = client.chat.completions.create(
    model="meta-llama/Meta-Llama-3-8B-Instruct",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Count to 10"},
    ],
    stream=False,
    max_tokens=1024,
)
assert output.choices[0].message.content == "1, 2, 3, 4, 5, 6, 7, 8, 9, 10!"

Sync + stream=True

client = InferenceClient()
output = client.chat.completions.create(
    model="meta-llama/Meta-Llama-3-8B-Instruct",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Count to 10"},
    ],
    stream=True,
    max_tokens=1024,
)

for chunk in output:
    print(chunk.choices[0].delta.content)

Async + stream=False

    client = AsyncInferenceClient(
        base_url="https://api-inference.huggingface.co/models/meta-llama/Meta-Llama-3-8B-Instruct",
        api_key="my-api-key",
    )
    output = await client.chat.completions.create(
        model="meta-llama/Meta-Llama-3-8B-Instruct",
        messages=[
            {"role": "system", "content": "You are a helpful assistant."},
            {"role": "user", "content": "Count to 10"},
        ],
        stream=False,
        max_tokens=1024,
    )
    assert output.choices[0].message.content == "1, 2, 3, 4, 5, 6, 7, 8, 9, 10!"

Async + stream=True

client = AsyncInferenceClient()
output = await client.chat.completions.create(
    model="meta-llama/Meta-Llama-3-8B-Instruct",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Count to 10"},
    ],
    stream=True,
    max_tokens=1024,
)

chunked_text = [chunk.choices[0].delta.content async for chunk in output]

TODO:

• document this properly (see inference.md)

[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.

[LysandreJik]

Awesome, very clear docs! Thanks @Wauplin

[julien-c]

@philschmid could you have a look at it and confirm it meets your expectations?

from experience, a very tough test to pass usually 🤣

EDIT: PR looks cool!

[Wauplin]

Thanks for the reviews!