update max tokens

app.py

@@ -16,9 +16,10 @@ def code_gen(text):
     pipe = pipeline(
         model=checkpoint,
         # tokenizer=tokenizer,
-        max_new_tokens=64,
+        max_new_tokens=124,
         temperature=0.7,
         top_p=0.95,
+        device= "cuda" if torch.cuda.is_available() else "cpu",
         repetition_penalty=1.15
     )