Making TF MPNet model compliant with XLA (#10260)

* Fix XLA

* Rework cast

* Apply style

src/transformers/models/mpnet/modeling_tf_mpnet.py

@@ -348,15 +348,22 @@ class TFMPNetEncoder(tf.keras.layers.Layer):
         self.n_heads = config.num_attention_heads
         self.output_attentions = config.output_attentions
         self.output_hidden_states = config.output_hidden_states
+        self.relative_attention_num_buckets = config.relative_attention_num_buckets
+        self.initializer_range = config.initializer_range
 
         self.layer = [TFMPNetLayer(config, name="layer_._{}".format(i)) for i in range(config.num_hidden_layers)]
-        self.relative_attention_bias = tf.keras.layers.Embedding(
-            config.relative_attention_num_buckets,
-            self.n_heads,
-            name="relative_attention_bias",
-        )
         self.relative_attention_num_buckets = config.relative_attention_num_buckets
 
+    def build(self, input_shape):
+        with tf.name_scope("relative_attention_bias"):
+            self.relative_attention_bias = self.add_weight(
+                name="embeddings",
+                shape=[self.relative_attention_num_buckets, self.n_heads],
+                initializer=get_initializer(self.initializer_range),
+            )
+
+        return super().build(input_shape)
+
     def call(
         self,
         hidden_states,
@@ -405,18 +412,16 @@ class TFMPNetEncoder(tf.keras.layers.Layer):
         n = -relative_position
 
         num_buckets //= 2
-        ret += tf.dtypes.cast(tf.math.less(n, 0), tf.int32) * num_buckets
+        ret += tf.cast(tf.math.less(n, 0), dtype=relative_position.dtype) * num_buckets
         n = tf.math.abs(n)
 
         # now n is in the range [0, inf)
         max_exact = num_buckets // 2
         is_small = tf.math.less(n, max_exact)
 
-        val_if_large = max_exact + tf.dtypes.cast(
+        val_if_large = max_exact + tf.cast(
-            tf.math.log(tf.dtypes.cast(n, tf.float32) / max_exact)
+            tf.math.log(n / max_exact) / math.log(max_distance / max_exact) * (num_buckets - max_exact),
-            / math.log(max_distance / max_exact)
+            dtype=relative_position.dtype,
-            * (num_buckets - max_exact),
-            tf.int32,
         )
 
         val_if_large = tf.math.minimum(val_if_large, num_buckets - 1)
@@ -441,7 +446,7 @@ class TFMPNetEncoder(tf.keras.layers.Layer):
             relative_position,
             num_buckets=self.relative_attention_num_buckets,
         )
-        values = self.relative_attention_bias(rp_bucket)  # shape (qlen, klen, num_heads)
+        values = tf.gather(self.relative_attention_bias, rp_bucket)  # shape (qlen, klen, num_heads)
         values = tf.expand_dims(tf.transpose(values, [2, 0, 1]), axis=0)  # shape (1, num_heads, qlen, klen)
         return values
 
@@ -541,7 +546,9 @@ class TFMPNetMainLayer(tf.keras.layers.Layer):
         # Since we are adding it to the raw scores before the softmax, this is
         # effectively the same as removing these entirely.
         extended_attention_mask = tf.cast(extended_attention_mask, embedding_output.dtype)
-        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+        one_cst = tf.constant(1.0, dtype=embedding_output.dtype)
+        ten_thousand_cst = tf.constant(-10000.0, dtype=embedding_output.dtype)
+        extended_attention_mask = tf.multiply(tf.subtract(one_cst, extended_attention_mask), ten_thousand_cst)
 
         # Prepare head mask if needed
         # 1.0 in head_mask indicate we keep the head

tests/test_modeling_tf_mpnet.py

@@ -232,10 +232,6 @@ class TFMPNetModelTest(TFModelTesterMixin, unittest.TestCase):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_mpnet_for_token_classification(*config_and_inputs)
 
-    def test_xla_mode(self):
-        # TODO JP: Make MPNet XLA compliant
-        pass
-
     @slow
     def test_model_from_pretrained(self):
         for model_name in ["microsoft/mpnet-base"]: