Qwen

/

Qwen3-1.7B

like 213

Follow

Qwen 40.3k

Previously when messages contained multiple assistant messages, applying tokenizer template with enable_thinking=False would result in applying no thinking tokens to the first assistant message, but applying them to the second assistant message.

For example,

messages = [
    {'role': 'user', 'content': 'i am user 1'},
    {'role': 'assistant', 'content': 'i am assistant 1'},
    {'role': 'user', 'content': 'i am user 2'},
    {'role': 'assistant', 'content': 'i am assistant 2'},
]

input_ids = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=False,
    truncate=True,
    return_tensors='pt',
    enable_thinking=False

).squeeze(0)

This input_ids would result in the following decoded output:
<|im_start|>user\ni am user 1<|im_end|>\n<|im_start|>assistant\ni am assistant 1<|im_end|>\n<|im_start|>user\ni am user 2<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\ni am assistant 2<|im_end|>\n

fixed template would result in the following decoded output:
<|im_start|>user\ni am user 1<|im_end|>\n<|im_start|>assistant\ni am assistant 1<|im_end|>\n<|im_start|>user\ni am user 2<|im_end|>\n<|im_start|>assistant\ni am assistant 2<|im_end|>\n