Support Azure OpenAI

[joeddav]

Adds a wrapper around the OpenAI class, creating an AsyncAzureOpenAI client instead of an AsyncOpenAI client.

Also required one change to openai.py: I put the tiktoken logic inside a computed tokenizer() property so that I could override this in the subclass. This is necessary because Azure OpenAI has a distinction between a deployment and a model, the former of which is used when making calls to the API, but the latter is what must be passed to tiktoken.

Usage:

model = outlines.models.azure_openai(
    model_name="gpt-4",
    deployment_name="gpt4-1106",   # my custom deployment name
    azure_endpoint="https://<myendpoint>.openai.azure.com/",
    api_key="<mykey>"
)
model.generate_choice(
    "What is the capital of the country where DaVinci is from?",
    choices=["Milan", "Rome", "Paris", "Italy"]
)
# 'Rome'

vs.

model = outlines.models.openai("gpt-4")
model.generate_choice(
    "What is the capital of the country where DaVinci is from?",
    choices=["Milan", "Rome", "Paris", "Italy"]
)
# 'Rome'

I coded this up for my own local use and but thought I'd share here in case you'd like to incorporate it, or in case someone else wanted to use it.

Resolves #271

[rlouf]

This is a great addition, thank you so much for your contribution!

[pamelafox]

Thanks for adding Azure support! When I try to use azure_openai, it fails at this line of code, since the OpenAI client attempts to authenticate. Is there a different way we can validate the model name, besides instantiating the client?

[pamelafox]

Just in case it's useful, here's a mapping that we maintain for an Azure repo of Azure model names to OpenAI names: https://github.com/Azure-Samples/azure-search-openai-demo/blob/3424475e6a61388c114f4a4db1de246734c37728/app/backend/core/modelhelper.py#L20

[rlouf]

I am now leaning towards "letting it crash" i.e. let users enter anything for the model name and then handle a potential error gracefully, letting them know that the model name is incorrect.

[rlouf]

I don’t have the time to fix it right now, but happy to review a PR.