Update handler.py

handler.py

@@ -67,7 +67,7 @@ class EndpointHandler:
             current_time = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
 
             # Tokenize input text
-            input_ids = self.tokenizer(user_text, return_tensors="pt").input_ids
+            input_ids = self.tokenizer(user_text, return_tensors="pt").input_ids.to("cuda")
 
             # Measure inference time
             start_time = time.time()