chat_qwen.py

from transformers import AutoModelForCausalLM, AutoTokenizer

def get_response(prompt: str):
    model = AutoModelForCausalLM.from_pretrained(
        "Qwen/Qwen2.5-32B-Instruct",
        torch_dtype="auto",
        device_map="auto",
    )

    tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-32B-Instruct")

    prompt = "Give me a short introduction to large language model."
    messages = [
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": prompt},
    ]
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True,
    )
    model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

    generated_ids = model.generate(
        **model_inputs,
        max_new_tokens=512,
    )
    generated_ids = [
        output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
    ]

    response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
    return response