TF 2.13 has a specific optimization for

dataset.shuffle(dataset.cardinality()), so use that
instead of dataset.shuffle(len(dataset))

src/datasets/utils/tf_utils.py

@@ -213,7 +213,7 @@ def scan_random_indices(state, indices):
 
         tf_dataset = tf_dataset.scan(base_seed, scan_random_indices)
     elif shuffle:
-        tf_dataset = tf_dataset.shuffle(len(tf_dataset))
+        tf_dataset = tf_dataset.shuffle(tf_dataset.cardinality())
         tf_dataset = tf_dataset.batch(batch_size, drop_remainder=drop_remainder)
     else:
         tf_dataset = tf_dataset.batch(batch_size, drop_remainder=drop_remainder)