Qwen3 response_schema might not be correct #42220
Description
Qwen3 response schema taken from
transformers/tests/utils/test_chat_parsing_utils.py
Lines 149 to 181 in 8598421
from transformers import AutoModelForCausalLM, AutoTokenizer
checkpoint = "Qwen/Qwen3-0.6B"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint, dtype="auto", device_map="auto")
# taken from test
qwen3_schema = {
"x-regex": r"^(?:(?:<think>)?\s*(?P<thinking>.+?)\s*</think>)?\s*(?:<tool_call>(?P<tool_calls>.*?)\s*</tool_call>)?\s*(?P<content>.+?)?\s*$",
"type": "object",
"properties": {
"role": {"const": "assistant"},
"content": {"type": "string"},
"thinking": {"type": "string"},
"tool_calls": {
"x-regex-iterator": r"^(.*)$", # We have already extracted tool calls and there can only be one, so just make it a list
"type": "array",
"items": {
"type": "object",
"properties": {
"type": {"const": "function"},
"function": {
"type": "object",
"properties": {
"name": {"type": "string", "x-regex": r"<function=(\w+)>"},
"arguments": {
"type": "object",
"x-regex-key-value": r"<parameter=(?P<key>\w+)>\n(?P<value>.*?)\n</parameter>",
"additionalProperties": {
"x-parser": "json",
"x-parser-args": {"allow_non_json": True},
},
},
},
},
},
},
},
},
}
tokenizer.response_schema = qwen3_schema
def multiply(a: int, b: int) -> int:
"""
Multiplies two integers.
Args:
a: The first integer.
b: The second integer.
Returns:
The product of the two integers.
"""
return a * b
messages = [{"role": "user", "content": "Use the tool to multiply 3 and 4."}]
processed = tokenizer.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=True,
return_tensors="pt",
tools=[multiply],
enable_thinking=False,
)
input_ids = processed["input_ids"].to(model.device)
outputs = model.generate(input_ids, max_new_tokens=1024)[0, input_ids.shape[1] :]
out_text = tokenizer.decode(outputs)
print(out_text)<tool_call>
{"name": "multiply", "arguments": {"a": 3, "b": 4}}
</tool_call><|im_end|>
parsed = tokenizer.parse_response(out_text)
print(parsed){'role': 'assistant', 'content': '<|im_end|>', 'tool_calls': [{'type': 'function', 'function': {}}]}
Interestingly, if I use smollm_schema instead, it seems to work fine:
tokenizer.response_schema = smollm_schema
parsed = tokenizer.parse_response(out_text)
print(parsed){'role': 'assistant', 'content': '<|im_end|>', 'tool_calls': [{'type': 'function', 'function': {'name': 'multiply', 'arguments': {'a': 3, 'b': 4}}}]}