fix(vLLM): Add tool calling support to VLLMClient.chat()

Fixes #4871

Previously, using GRPOTrainer with `vllm_mode="server"` raised a
`NotImplementedError` when tools were passed to `VLLMClient.chat()`.
This prevented users from using tool calling features with the vLLM
server mode.

Changes:
- Remove the NotImplementedError check in VLLMClient.chat()
- Add `tools` parameter to the HTTP request payload
- Add `tools` field to ChatRequest model in vllm_serve.py
- Pass tools to vLLM's chat() method on the server side

Author: kansalaman

trl/extras/vllm_client.py

@@ -352,8 +352,6 @@ def chat(
                 - `logprobs` (`list[list[float]]`):
                     List of lists of log probabilities for each generated token.
         """
-        if tools is not None:
-            raise NotImplementedError("Tool calling is not yet implemented in VLLMClient.chat().")
         if chat_template is not None:
             raise NotImplementedError("Custom chat templates are not yet implemented in VLLMClient.chat().")
 
@@ -383,6 +381,7 @@ def chat(
                 "structured_outputs_regex": structured_outputs_regex,
                 "generation_kwargs": generation_kwargs or {},
                 "chat_template_kwargs": chat_template_kwargs or {},
+                "tools": tools,
             },
         )
         if response.status_code == 200:

trl/scripts/vllm_serve.py

@@ -644,6 +644,7 @@ class ChatRequest(BaseModel):
         structured_outputs_regex: str | None = None
         generation_kwargs: dict = field(default_factory=dict)
         chat_template_kwargs: dict = field(default_factory=dict)
+        tools: list | None = None
 
     class ChatResponse(BaseModel):
         prompt_ids: list[list[int]]
@@ -756,6 +757,7 @@ async def chat(request: ChatRequest):
                 "messages": messages,
                 "sampling_params": sampling_params,
                 "chat_template_kwargs": request.chat_template_kwargs,
+                "tools": request.tools,
             }
             connection.send({"type": "call", "method": "chat", "kwargs": kwargs})