Update app.py

app.py

@@ -18,62 +18,19 @@ tokenizer.set_chat_template(tokenizer.chat_template)
 
 
 def generate_response(prompt):
-    """
-    Generates a response using the OpenVINO LLM pipeline.
-
-    Args:
-        prompt (str): The input prompt.
-
-    Returns:
-        str: The generated response.
-    """
-    start_time = time.time()
-    output = pipe.generate([prompt], max_length=1024)
-    end_time = time.time()
-
-    generated_text = output.text[0]
-
-    performance_metrics = f"Generate duration: {output.perf_metrics.get_generate_duration().mean:.2f}ms\n"
-    performance_metrics += f'Throughput: {output.perf_metrics.get_throughput().mean:.2f} tokens/s'
-
-    return generated_text, performance_metrics
-
-
-def main():
-    """
-    Creates and launches the Gradio interface.
-    """
-
-    with gr.Blocks() as demo:
-        gr.Markdown("# OpenVINO Qwen3-8B Demo")
-        prompt_input = gr.Textbox(lines=3, label="Enter your prompt:")
-        output_text = gr.Textbox(label="Generated Response")
-        performance_text = gr.Textbox(label="Performance Metrics", visible=False)
-        visibility_state = gr.State(False) # Store visibility state
-
-        def update_output(prompt):
-            response, performance = generate_response(prompt)
-            return response, performance
-
-        prompt_input.change(
-            fn=update_output,
-            inputs=prompt_input,
-            outputs=[output_text, performance_text],
-        )
-
-        # Button to show/hide performance metrics
-        show_metrics_button = gr.Button("Show/Hide Performance Metrics")
-        def toggle_visibility(visible):
-            return not visible, not visible
-        show_metrics_button.click(
-            fn=toggle_visibility,
-            inputs=[visibility_state],
-            outputs=[performance_text.visible, visibility_state],
-        )
-
-
-    demo.launch()
-
+    generated = pipe.generate([prompt], max_length=1024)   
+    return generated
+
+# 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="回應")
+    ],
+    title="Qwen3-0.6B OpenVINO + Gradio",
+    description="基於 OpenVINO 最佳化的 Qwen3-0.6B 推理應用，支援思考過程分離與 GUI。"
+)
 
 if __name__ == "__main__":
-    main()
+    demo.launch()