app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline

model_name = "Qwen/Qwen3-4B-Thinking-2507"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    trust_remote_code=True,
    device_map="cpu",
    torch_dtype="auto"
)

# 🔥 ВАЖНО: создайте pipeline
pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    device_map="cpu"
)

def respond(message, history, system_message, max_tokens):
    messages = [
        {"role": "system", "content": system_message},
        {"role": "user", "content": message}
    ]

    prompt = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )

    output = pipe(
        prompt,
        max_new_tokens=max_tokens,
        do_sample=True,
        temperature=0.7,
        top_p=0.95,
        pad_token_id=tokenizer.eos_token_id
    )

    response = output[0]["generated_text"][len(prompt):].strip()
    return response

chatbot = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Textbox(value="You are a helpful assistant.", label="System message"),
        gr.Slider(minimum=1, maximum=262144, value=64, step=1, label="Max new tokens")
    ],
    title="Qwen Chat (Local, No History)",
)

if __name__ == "__main__":
    chatbot.launch()