Update app.py

app.py

@@ -278,28 +278,29 @@ def simple_chat(message: dict, temperature: float = 0.8, max_length: int = 4096,
 
         gen_kwargs = {**input_ids, **generate_kwargs}
 
-        # Define the function to run generation
-        def generate_text():
-            with torch.no_grad():
-                model.generate(**gen_kwargs, streamer=streamer)
 
-        # Start the generation in a separate thread
-        thread = Thread(target=generate_text)
-        thread.start()
-
-        def stream_response():
+        with torch.no_grad():
+            thread = Thread(target=model.generate, kwargs=gen_kwargs)
+            thread.start()
+            print("--------------")
+            print("  ")
+            print(thread)
+            print("  ")
+            print("--------------")
             buffer = ""
             for new_text in streamer:
                 buffer += new_text
                 yield new_text
             print("--------------")
+            print(new_text)
             print("Buffer: ")
             print("  ")
             print(buffer)
             print("  ")
             print("--------------")
 
-        return StreamingResponse(stream_response(), media_type="text/plain")
+
+        return StreamingResponse(new_text), media_type="text/plain")
 
     except Exception as e:
         return PlainTextResponse(f"Error: {str(e)}")