Update model.py

model.py

@@ -10,14 +10,14 @@ if torch.cuda.is_available():
     model = AutoModelForCausalLM.from_pretrained(
         model_id,
         torch_dtype=torch.float16,
-        device_map='cuda',
+        device_map='auto',
         
-    ).to("cuda")
+    )
 else:
     model = AutoModelForCausalLM.from_pretrained(
         model_id,
-        torch_dtype=torch.float16,
-        load_in_4bit=True,
+        torch_dtype=torch.float32,
+        #load_in_4bit=True,
         device_map='auto',
         use_safetensors=False,
     )