Fireworks support for tool_choice=required

[aidando73]

🚀 Describe the new functionality needed

Support for tool_choice=required (in inference).

When testing against fireworks, they support tool_choice=any (which semantically means the same thing) [1].

💡 Why is this needed? What if we don't build it?

There are use-cases where you rely on tool calls being generated but llama-stack generates 0 tool calls. I've created a repro here: aidando73/llama-code-review#3.

In that example, Llama-stack gives me 0 tool calls. But hitting the fireworks api directly I get pretty solid tool calls.

If we don't build this then users would be forced to hit the API directly instead of using llama-stack

[aidando73]

Here's a POC pr: #657

This uses fireworks chat completion api - so we're relying on their implementation.

[aidando73]

Oh - I realize now that the problem I was having is because we haven't updated prompt_adapter.py for 3.3 models yet. The prompt format for 3.3 works the same as 3.1: https://www.llama.com/docs/model-cards-and-prompt-formats/llama3_3. So we need to update prompt_adapter.py to:

    if model.model_family == ModelFamily.llama3_1 or (
        model.model_family == ModelFamily.llama3_2
        and is_multimodal(model.core_model_id)
    ) or model.model_family == ModelFamily.llama3_3:
        # llama3.1, llama3.2 multimodal and llama3.3 models follow the same tool prompt format
        messages = augment_messages_for_tools_llama_3_1(request)
    elif model.model_family == ModelFamily.llama3_2:
        messages = augment_messages_for_tools_llama_3_2(request)
    else:
        messages = request.messages

But is it just me or is fireworks a bit more consistent with their tool calls when tool_choice=any - are they doing something special under the hood 🤔?