Update app.py

app.py

@@ -641,11 +641,12 @@ def build_demo():
     return demo
 # --- API endpoint: /chat ---
 # --- API endpoint: /chat ---
+from fastapi import FastAPI
+import gradio as gr
+import uvicorn
 
+# --- API function ---
 def http_api_infer(question: str):
-    # NOTE: No request here – Gradio will not inject it automatically
-    # So remove token checking (or add later via FastAPI if needed)
-
     dummy_state = init_state()
     dummy_state.set_system_message("You are a helpful assistant.")
     dummy_state.append_message(Conversation.USER, question)
@@ -678,8 +679,7 @@ def http_api_infer(question: str):
     except Exception as e:
         return f"❌ Error: {str(e)}"
 
-
-# ✅ Define the API Interface after the function is declared
+# Gradio API Interface
 api_interface = gr.Interface(
     fn=http_api_infer,
     inputs=[gr.Textbox(label="Question")],
@@ -687,33 +687,20 @@ api_interface = gr.Interface(
     api_name="/chat"
 )
 
+# UI block
+ui_demo = build_demo()
 
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--host", type=str, default="0.0.0.0")
-    parser.add_argument("--port", type=int, default=7860)
-    parser.add_argument("--concurrency-count", type=int, default=10)
-    parser.add_argument("--share", action="store_true")
-    parser.add_argument("--moderate", action="store_true")
-    args = parser.parse_args()
-    logger.info(f"args: {args}")
-
-    # UI block
-    ui_demo = build_demo()
-
-    # Combine UI + API endpoint
-    demo = gr.TabbedInterface(
-        interface_list=[ui_demo, api_interface],
-        tab_names=["UI", "API"]
-    )
+# Setup FastAPI app
+app = FastAPI()
 
-    # Launch with API access
-    demo.queue(api_open=True).launch(
-        server_name=args.host,
-        server_port=args.port,
-        share=args.share,
-        max_threads=args.concurrency_count,
-    )
+# Mount UI
+app = gr.mount_gradio_app(app, ui_demo, path="/")
 
+# Mount API
+app = gr.mount_gradio_app(app, api_interface, path="/api")
+
+# Run
+if __name__ == "__main__":
+    uvicorn.run(app, host="0.0.0.0", port=7860)