Fix cache format

Transformers now only uses the Cache classes

custom_generate/generate.py

@@ -282,10 +282,8 @@ def _contrastive_search(
                     f"{model.__class__.__name__} does not support caching and therefore **can't** be used "
                     "for contrastive search."
                 )
-            elif (
-                not isinstance(past_key_values[0], (tuple, torch.Tensor))
-                or past_key_values[0][0].shape[0] != batch_size
-            ):
+            # We now only use Cache classes, but a few models have custom cache class, so we use this check instead of an instance check
+            elif not hasattr(past_key_values, "update"):
                 raise ValueError(
                     f"{model.__class__.__name__} does not have a standard cache format and therefore **can't** be "
                     "used for contrastive search without further modifications."