update handler

handler.py

@@ -39,7 +39,7 @@ class EndpointHandler():
         """
 
         inputs = data.pop("inputs", data)
-        # with torch.cuda.amp.autocast():
-        with torch.no_grad():
+        with torch.cuda.amp.autocast():
             prediction = self.pipeline(inputs, generate_kwargs={"forced_decoder_ids": self.forced_decoder_ids}, max_new_tokens=255)
+            # prediction = self.pipeline(inputs, return_timestamps=False)
         return prediction