done:

tokenizer.py

@@ -8,13 +8,12 @@ class ModernDecoderBERTTokenizer(PreTrainedTokenizerFast):
         outputs = super()._batch_encode_plus(*args, **kwargs)
         del outputs["token_type_ids"]
         for key in ['input_ids', 'attention_mask']:
-            if isinstance(outputs[key], (list, numpy.ndarray, torch.Tensor)):
-                if isinstance(outputs[key], list):
-                    outputs[key] = [sequence[:-1] for sequence in outputs[key]]
-                elif isinstance(outputs[key], numpy.ndarray):
-                    outputs[key] = numpy.array([sequence[:-1] for sequence in outputs[key]], dtype=outputs[key].dtype)
-                elif isinstance(outputs[key], torch.Tensor):
-                    outputs[key] = torch.tensor([sequence[:-1] for sequence in outputs[key]], dtype=outputs[key].dtype, device=outputs[key].device)
+            if isinstance(outputs[key], torch.Tensor):
+                outputs[key] = outputs[key][..., :-1]
+            elif isinstance(outputs[key], numpy.ndarray): 
+                outputs[key] = outputs[key][..., :-1]
+            elif isinstance(outputs[key], list):
+                outputs[key] = [sequence[:-1] for sequence in outputs[key]]
         return outputs
 
 # Register the class