Update handler.py

handler.py

@@ -89,11 +89,11 @@ class EndpointHandler:
             prompt = self.processor.apply_chat_template(
                 messages, add_generation_prompt=True
             )
-            processed_inputs = self.processor(
-                text=prompt, images=images, return_tensors="pt"
-            )
 
             with torch.no_grad(), torch.autocast("cuda"):
+                processed_inputs = self.processor(
+                    text=prompt, images=images, return_tensors="pt"
+                )
                 generated_ids = self.model.generate(
                     **processed_inputs, generation_config=generation_config
                 )