Add in a comment about legacy input masks for XLNet.

PiperOrigin-RevId: 339267061

official/nlp/xlnet/xlnet_modeling.py

@@ -795,6 +795,8 @@ class PretrainingXLNetModel(tf.keras.Model):
     masked_tokens = features["input_q"]
     seg_ids = features["seg_id"]
     if self._use_legacy_mask:
+      # Legacy input mask assumes `real` values are 0 and `padding`
+      # values are 1.
       perm_mask = 1 - features["perm_mask"]
     else:
       perm_mask = features["perm_mask"]
@@ -885,6 +887,8 @@ class ClassificationXLNetModel(tf.keras.Model):
     input_ids = features["input_ids"]
     segment_ids = features["segment_ids"]
     if self._use_legacy_mask:
+      # Legacy input mask assumes `real` values are 0 and `padding`
+      # values are 1.
       input_mask = 1 - features["input_mask"]
     else:
       input_mask = features["input_mask"]
@@ -1130,6 +1134,8 @@ class QAXLNetModel(tf.keras.Model):
     input_ids = features["input_ids"]
     segment_ids = features["segment_ids"]
     if self._use_legacy_mask:
+      # Legacy input mask assumes `real` values are 0 and `padding`
+      # values are 1.
       input_mask = 1 - features["input_mask"]
     else:
       input_mask = features["input_mask"]