Upload app.py

app.py

@@ -9,7 +9,6 @@ def main():
     # Enable CUDA if available and load in tokenizer
     device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
     tokenizer, EMPTY_TOKENS = load_tokenizer(device)
-    model.cuda()
 
     st.title("Scaling Transformers")
     st.subheader("UCLA DSU Project, Fall 2023")
@@ -72,6 +71,7 @@ def main():
     # small_model = load_casey_model(tokenizer, device)
     model = load_big_model(tokenizer, device)
     model.to('cuda')
+    model.cuda()