Upload handler.py

handler.py

@@ -30,7 +30,7 @@ class EndpointHandler:
         encodings = self.tokenizer(inputs, padding=False, truncation=False)
 
         truncated_input_ids = middle_truncate(
-            encodings["input_ids"][0].tolist(), self.MAX_LENGTH, self.tokenizer
+            encodings["input_ids"], self.MAX_LENGTH, self.tokenizer
         )
 
         attention_masks = [
@@ -42,8 +42,6 @@ class EndpointHandler:
             "attention_mask": torch.tensor([attention_masks]),
         }
 
-        truncated_encodings.set_format("torch")
-
         outputs = self.model(**truncated_encodings)
 
         # transform logits to probabilities and apply threshold