adding back truncation

handler.py

@@ -31,11 +31,11 @@ class EndpointHandler():
 
         input_ids = self.tokenizer.encode(inputs, return_tensors="pt")
         max_generation_length = 75  # Desired number of tokens to generate
-        # max_input_length = 4092 - max_generation_length  # Maximum input length to allow space for generation
+        max_input_length = 4092 - max_generation_length  # Maximum input length to allow space for generation
 
         # # Truncate input_ids to the most recent tokens that fit within the max_input_length
-        # if input_ids.shape[1] > max_input_length:
-        #     input_ids = input_ids[:, -max_input_length:]
+        if input_ids.shape[1] > max_input_length:
+            input_ids = input_ids[:, -max_input_length:]
 
         max_length = input_ids.shape[1] + max_generation_length