TF: serializable hubert (#20966)

* serializable hubert

src/transformers/models/hubert/modeling_tf_hubert.py

@@ -221,13 +221,17 @@ def _compute_mask_indices(
     if mask_length < 1:
         raise ValueError("`mask_length` has to be bigger than 0.")
 
-    if mask_length > sequence_length:
-        raise ValueError(
+    tf.debugging.assert_less(
+        mask_length,
+        sequence_length,
+        message=(
             f"`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: {mask_length} and"
             f" `sequence_length`: {sequence_length}`"
+        ),
     )
+
     # compute number of masked spans in batch
-    num_masked_spans = mask_prob * sequence_length / mask_length + tf.random.uniform((1,))
+    num_masked_spans = mask_prob * tf.cast(sequence_length, tf.float32) / mask_length + tf.random.uniform((1,))
     num_masked_spans = tf.maximum(num_masked_spans, min_masks)
     num_masked_spans = tf.cast(num_masked_spans, tf.int32)
 

src/transformers/models/wav2vec2/modeling_tf_wav2vec2.py

@@ -262,13 +262,17 @@ def _compute_mask_indices(
     if mask_length < 1:
         raise ValueError("`mask_length` has to be bigger than 0.")
 
-    if mask_length > sequence_length:
-        raise ValueError(
+    tf.debugging.assert_less(
+        mask_length,
+        sequence_length,
+        message=(
             f"`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: {mask_length} and"
             f" `sequence_length`: {sequence_length}`"
+        ),
     )
+
     # compute number of masked spans in batch
-    num_masked_spans = mask_prob * sequence_length / mask_length + tf.random.uniform((1,))
+    num_masked_spans = mask_prob * tf.cast(sequence_length, tf.float32) / mask_length + tf.random.uniform((1,))
     num_masked_spans = tf.maximum(num_masked_spans, min_masks)
     num_masked_spans = tf.cast(num_masked_spans, tf.int32)