Add tool calling to ``OpenAIToMessages`` Transform

[joecummings]

Currently, we only support text and images from the OpenAI format to the torchtune Messages format.

We should incorporate tool calling as it is supported by the OpenAI format and therefore can be found in many datasets: https://platform.openai.com/docs/guides/function-calling?api-mode=chat&example=get-weather.

[SlinkoIgor]

I think this other bug might be related:

When experimenting with the OpenAI sample from https://platform.openai.com/docs/guides/fine-tuning (Tool calling section), I encountered that the assistant content, if absent (as in the example) or null, will be copy-pasted from the previous message.

Also, you can see that the tool calls in assistant messages do not appear in the output.

Code to reproduce:

from torchtune.models.qwen2_5 import qwen2_5_tokenizer
from torchtune.datasets import chat_dataset

tokenizer=qwen2_5_tokenizer(
    path="/tmp/Qwen2.5-0.5B-Instruct/vocab.json",
    merges_file="/tmp/Qwen2.5-0.5B-Instruct/merges.txt",
    special_tokens_path="/tmp/Qwen2.5-0.5B-Instruct/tokenizer.json",
)

ds = chat_dataset(
    source="json",
    tokenizer=tokenizer,
    conversation_column="messages",
    conversation_style="openai",
    data_files="sample.json",
    split="train",
)

for i in range(len(ds)):
    decoded_input = tokenizer.decode(ds[i]['tokens'])
    print(decoded_input)

>>>
<|im_start|>user
What is the weather in San Francisco?<|im_end|>
<|im_start|>assistant
What is the weather in San Francisco?<|im_end|>
<tool_call>

</tool_call><|im_end|>
<|im_start|>tool
21.0<|im_end|>
<|im_end|>

sample.json:

{
  "messages": [
    {
      "role": "user",
      "content": "What is the weather in San Francisco?"
    },
    {
      "role": "assistant",
      "tool_calls": [
        {
          "id": "call_id",
          "type": "function",
          "function": {
            "name": "get_current_weather",
            "arguments": "{\"location\": \"San Francisco, USA\", \"format\": \"celsius\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "tool_call_id": "call_id",
      "content": "21.0"
    },
  ],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_current_weather",
        "description": "Get the current weather",
        "parameters": {
          "type": "object",
          "properties": {
            "location": {
              "type": "string",
              "description": "The city and country, eg. San Francisco, USA"
            },
            "format": {
              "type": "string",
              "enum": [
                "celsius",
                "fahrenheit"
              ]
            }
          },
          "required": [
            "location",
            "format"
          ]
        }
      }
    }
  ]
}

[SlinkoIgor]

Also found incorrect format in the tests:

torchtune/tests/torchtune/data/test_messages.py

Line 785 in 1075d9c

def test_call_tool_messages(self):

Moreover, there are no separate fields for tool_call, and it is assumed here that the assistant cannot return both content and tool_call. However, it can and does. This often proves to be very useful, for example in ReAct agents, when we ask the LLM to provide thought along with action. In fact, thought generation comes first to ensure the action is of higher quality. This does not contradict the chat template, for example, of the Qwen-2.5 model. In the template the <tool_call></tool_call> can be placed anywhere in the text.

I also discovered that this same assumption exists in the tokenizer for Qwen:

torchtune/torchtune/models/qwen2_5/_tokenizer.py

Line 223 in e3d95cb

if message.role == "assistant" and message.ipython:

[krammnic]

I don't see someone assigned on this and I've promised that I will add this after the ICLR, so let me take this. Furthermore, it is definitely related to the #2723