fix: multiple tool calls in remote-vllm chat_completion (#2161)

# What does this PR do?

This fixes an issue in how we used the tool_call_buf from streaming tool
calls in the remote-vllm provider where it would end up concatenating
parameters from multiple different tool call results instead of
aggregating the results from each tool call separately.

It also fixes an issue found while digging into that where we were
accidentally mixing the json string form of tool call parameters with
the string representation of the python form, which mean we'd end up
with single quotes in what should be double-quoted json strings.

Closes #1120

## Test Plan

The following tests are now passing 100% for the remote-vllm provider,
where some of the test_text_inference were failing before this change:

```
VLLM_URL="http://localhost:8000/v1" INFERENCE_MODEL="RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic" LLAMA_STACK_CONFIG=remote-vllm python -m pytest -v tests/integration/inference/test_text_inference.py --text-model "RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic"

VLLM_URL="http://localhost:8000/v1" INFERENCE_MODEL="RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic" LLAMA_STACK_CONFIG=remote-vllm python -m pytest -v tests/integration/inference/test_vision_inference.py --vision-model "RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic"

```

All but one of the agent tests are passing (including the multi-tool
one). See the PR at https://github.com/vllm-project/vllm/pull/17917 and
a gist at
https://gist.github.com/bbrowning/4734240ce96b4264340caa9584e47c9e for
changes needed there, which will have to get made upstream in vLLM.

Agent tests:

```
VLLM_URL="http://localhost:8000/v1" INFERENCE_MODEL="RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic" LLAMA_STACK_CONFIG=remote-vllm python -m pytest -v tests/integration/agents/test_agents.py --text-model "RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic"
````

---------

Signed-off-by: Ben Browning <bbrownin@redhat.com>

llama_stack/providers/remote/inference/vllm/vllm.py

@@ -162,7 +162,7 @@ def _process_vllm_chat_completion_end_of_stream(
     finish_reason: str | None,
     last_chunk_content: str | None,
     current_event_type: ChatCompletionResponseEventType,
-    tool_call_buf: UnparseableToolCall,
+    tool_call_bufs: dict[str, UnparseableToolCall] | None = None,
 ) -> list[OpenAIChatCompletionChunk]:
     chunks = []
 
@@ -171,9 +171,8 @@ def _process_vllm_chat_completion_end_of_stream(
     else:
         stop_reason = StopReason.end_of_message
 
-    if tool_call_buf.tool_name:
-        # at least one tool call request is received
-
+    tool_call_bufs = tool_call_bufs or {}
+    for _index, tool_call_buf in sorted(tool_call_bufs.items()):
         args_str = tool_call_buf.arguments or "{}"
         try:
             args = json.loads(args_str)
@@ -225,8 +224,14 @@ def _process_vllm_chat_completion_end_of_stream(
 async def _process_vllm_chat_completion_stream_response(
     stream: AsyncGenerator[OpenAIChatCompletionChunk, None],
 ) -> AsyncGenerator:
-    event_type = ChatCompletionResponseEventType.start
-    tool_call_buf = UnparseableToolCall()
+    yield ChatCompletionResponseStreamChunk(
+        event=ChatCompletionResponseEvent(
+            event_type=ChatCompletionResponseEventType.start,
+            delta=TextDelta(text=""),
+        )
+    )
+    event_type = ChatCompletionResponseEventType.progress
+    tool_call_bufs: dict[str, UnparseableToolCall] = {}
     end_of_stream_processed = False
 
     async for chunk in stream:
@@ -235,17 +240,22 @@ async def _process_vllm_chat_completion_stream_response(
             return
         choice = chunk.choices[0]
         if choice.delta.tool_calls:
-            tool_call = convert_tool_call(choice.delta.tool_calls[0])
-            tool_call_buf.tool_name += str(tool_call.tool_name)
-            tool_call_buf.call_id += tool_call.call_id
-            # TODO: remove str() when dict type for 'arguments' is no longer allowed
-            tool_call_buf.arguments += str(tool_call.arguments)
+            for delta_tool_call in choice.delta.tool_calls:
+                tool_call = convert_tool_call(delta_tool_call)
+                if delta_tool_call.index not in tool_call_bufs:
+                    tool_call_bufs[delta_tool_call.index] = UnparseableToolCall()
+                tool_call_buf = tool_call_bufs[delta_tool_call.index]
+                tool_call_buf.tool_name += str(tool_call.tool_name)
+                tool_call_buf.call_id += tool_call.call_id
+                tool_call_buf.arguments += (
+                    tool_call.arguments if isinstance(tool_call.arguments, str) else json.dumps(tool_call.arguments)
+                )
         if choice.finish_reason:
             chunks = _process_vllm_chat_completion_end_of_stream(
                 finish_reason=choice.finish_reason,
                 last_chunk_content=choice.delta.content,
                 current_event_type=event_type,
-                tool_call_buf=tool_call_buf,
+                tool_call_bufs=tool_call_bufs,
             )
             for c in chunks:
                 yield c
@@ -266,7 +276,7 @@ async def _process_vllm_chat_completion_stream_response(
     # the stream ended without a chunk containing finish_reason - we have to generate the
     # respective completion chunks manually
     chunks = _process_vllm_chat_completion_end_of_stream(
-        finish_reason=None, last_chunk_content=None, current_event_type=event_type, tool_call_buf=tool_call_buf
+        finish_reason=None, last_chunk_content=None, current_event_type=event_type, tool_call_bufs=tool_call_bufs
     )
     for c in chunks:
         yield c

llama_stack/providers/utils/inference/openai_compat.py

@@ -531,13 +531,19 @@ async def convert_message_to_openai_dict(message: Message, download: bool = Fals
             tool_name = tc.tool_name
             if isinstance(tool_name, BuiltinTool):
                 tool_name = tool_name.value
+
+            # arguments_json can be None, so attempt it first and fall back to arguments
+            if hasattr(tc, "arguments_json") and tc.arguments_json:
+                arguments = tc.arguments_json
+            else:
+                arguments = json.dumps(tc.arguments)
             result["tool_calls"].append(
                 {
                     "id": tc.call_id,
                     "type": "function",
                     "function": {
                         "name": tool_name,
-                        "arguments": tc.arguments_json if hasattr(tc, "arguments_json") else json.dumps(tc.arguments),
+                        "arguments": arguments,
                     },
                 }
             )