Update app.py

app.py

@@ -101,10 +101,10 @@ def loadModel():
                     device_map="auto",
                     quantization_config=quantization_config,
                     max_memory={
-                        0: "10GB",  # GPU 0 with 20GB memory
-                        1: "10GB",  # GPU 0 with 20GB memory
-                        2: "10GB",  # GPU 0 with 20GB memory
-                        3: "10GB",  # GPU 0 with 20GB memory
+                        0: "5GB",  # GPU 0 with 20GB memory
+                        1: "20GB",  # GPU 0 with 20GB memory
+                        2: "20GB",  # GPU 0 with 20GB memory
+                        3: "20GB",  # GPU 0 with 20GB memory
                         "cpu": "40GB"  # CPU with 100GB memory
                     },
                 )