app.py

import huggingface_hub as hf_hub
import time
import openvino_genai as ov_genai
import numpy as np
import gradio as gr
import re

# 下載模型
model_id = "OpenVINO/Qwen3-0.6B-int4-ov"
model_path = "Qwen3-0.6B-int4-ov"

hf_hub.snapshot_download(model_id, local_dir=model_path, local_dir_use_symlinks=False)

# 建立推理管線
device = "CPU"
pipe = ov_genai.LLMPipeline(model_path, device)
tokenizer = pipe.get_tokenizer()
tokenizer.set_chat_template(tokenizer.chat_template)


def generate_response(prompt):
    full_response = ""  # 用於儲存完整的回應

    def streamer(subword):
        nonlocal full_response
        full_response += subword
        yield full_response  # 使用 yield 使 streamer 成為生成器
        return ov_genai.StreamingStatus.RUNNING # 返回 StreamingStatus.RUNNING

    try:
        # 使用流式生成
        generated = pipe.generate(prompt, streamer=streamer, max_new_tokens=100)
        tokenpersec = f'{generated.perf_metrics.get_throughput().mean:.2f}'  # 恢復原本計算 tokenpersec 的方式

        return tokenpersec, full_response
    except Exception as e:
        return "發生錯誤", "發生錯誤", f"生成回應時發生錯誤：{e}"


# 建立 Gradio 介面
demo = gr.Interface(
    fn=generate_response,
    inputs=gr.Textbox(lines=5, label="輸入提示 (Prompt)"),
    outputs=[
        gr.Textbox(label="tokens/sec"),
        gr.Textbox(label="回應"]
    ],
    title="Qwen3-0.6B-int4-ov ",
    description="基於 Qwen3-0.6B-int4-ov 推理應用，支援思考過程分離與 GUI。"
)

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