A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Bugfix] fix tool call arguments is empty (#25223)
Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com> Co-authored-by: xin.li <xin.li@daocloud.io>
Chauncey committed
Sep 20, 2025 at 13:29 UTC
f91480b2d44c263fb600b5cba5b0e6c7a195f742
2 files changed
+65
-3
tests/entrypoints/openai/test_completion_with_function_calling.py
+60
@@ -1,6 +1,7 @@
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
+import datetime
5
from typing import Union
6
7
import openai # use the official client for correctness check
@@ -284,3 +285,62 @@ async def test_tool_id_kimi_k2(k2_client: openai.AsyncOpenAI, model_name: str,
285
output.extend(chunk.choices[0].delta.tool_calls)
286
for o in output:
287
assert o.id is None or o.id == 'functions.get_current_weather:0'
288
+
289
+
290
+@pytest.mark.asyncio
291
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
292
+@pytest.mark.parametrize("arguments", ["{}", ''])
293
+async def test_no_args_tool_call(client: openai.AsyncOpenAI, model_name: str,
294
+ arguments: str):
295
+ # Step 1: Define a tool that requires no parameters
296
+ tools = [{
297
+ "type": "function",
298
+ "function": {
299
+ "name": "get_current_time",
300
+ "description":
301
+ "Get the current date and time. No parameters needed.",
302
+ "parameters": {
303
+ "type": "object",
304
+ "properties": {}, # No parameters
305
+ "required": [] # No required fields
306
+ }
307
+ }
308
+ }]
309
+ messages = [{"role": "user", "content": "What time is it now?"}]
310
+ # Step 2: Send user message and let model decide whether to call the tool
311
+ response = await client.chat.completions.create(
312
+ model=model_name,
313
+ messages=messages,
314
+ tools=tools,
315
+ tool_choice="auto" # Let model choose automatically
316
+ )
317
+
318
+ # Step 3: Check if model wants to call a tool
319
+ message = response.choices[0].message
320
+ if message.tool_calls:
321
+ # Get the first tool call
322
+ tool_call = message.tool_calls[0]
323
+ tool_name = tool_call.function.name
324
+ # Step 4: Execute the tool locally (no parameters)
325
+ if tool_name == "get_current_time":
326
+ # Test both empty string and "{}" for no-arg tool calls
327
+ tool_call.function.arguments = arguments
328
+ messages.append(message)
329
+ current_time = datetime.datetime.now()
330
+ result = current_time.isoformat()
331
+ messages.append({
332
+ "role": "tool",
333
+ "tool_call_id": tool_call.id,
334
+ "content": result,
335
+ })
336
+ # Step 5: Send tool result back to model to continue conversation
337
+ final_response = await client.chat.completions.create(
338
+ model=model_name,
339
+ messages=messages,
340
+ )
341
+ # Output final natural language response
342
+ assert final_response.choices[0].message.content is not None
343
+
344
+ else:
345
+ # No tool called — just print model's direct reply
346
+ assert message.content is not None
vllm/entrypoints/chat_utils.py
+5
-3
@@ -1450,9 +1450,11 @@ def _postprocess_messages(messages: list[ConversationMessage]) -> None:
1450
and isinstance(message["tool_calls"], list)
1451
):
1452
for item in message["tool_calls"]:
1453
- item["function"]["arguments"] = json.loads(
1454
- item["function"]["arguments"]
1455
- )
1453
+ # if arguments is None or empty string, set to {}
1454
+ if content := item["function"].get("arguments"):
1455
+ item["function"]["arguments"] = json.loads(content)
1456
+ else:
1457
+ item["function"]["arguments"] = {}
1458
1459
1460
def parse_chat_messages(