Update app.py

app.py

@@ -18,8 +18,9 @@ print('Loading model...')
 pipe = transformers.pipeline(
     "text-generation", 
     model="dx2102/llama-midi",
+    revision="c303c108399aba837146e893375849b918f413b3",
     torch_dtype="bfloat16",
-    device_map="auto",
+    device="cuda",
 )
 print('Done')