Update app.py

app.py

@@ -134,13 +134,9 @@ class ModelManager:
                 print(f"Error loading model {model_config['name']}: {e}")
 
     def load_all_models(self):
-        with ThreadPoolExecutor() as executor:
-            futures = [
-                executor.submit(self.load_model, config)
-                for config in model_configs
-            ]
-            for future in as_completed(futures):
-                pass
+        loop = asyncio.get_event_loop()
+        tasks = [loop.run_in_executor(None, self.load_model, config) for config in model_configs]
+        loop.run_until_complete(asyncio.gather(*tasks))
         return self.models
 
 model_manager = ModelManager()
@@ -210,22 +206,16 @@ def get_best_response(responses):
 
 async def generate_model_response(model, inputs):
     try:
-        response = model(inputs)
+        response = await model(inputs)
         return remove_duplicates(response['choices'][0]['text'])
     except Exception as e:
         return ""
 
 async def process_message(message):
     inputs = normalize_input(message)
-    with ThreadPoolExecutor() as executor:
-        futures = [
-            executor.submit(generate_model_response, model, inputs)
-            for model in global_data['models'].values()
-        ]
-        responses = [
-            future.result()
-            for future in as_completed(futures)
-        ]
+    # Use asyncio.gather to handle multiple coroutines concurrently
+    tasks = [generate_model_response(model, inputs) for model in global_data['models'].values()]
+    responses = await asyncio.gather(*tasks)
     best_response = get_best_response(responses)
     return best_response