Remove hard-coded uses of float32 to fix mixed precision use (#6648)

src/transformers/modeling_tf_bert.py

@@ -215,8 +215,8 @@ class TFBertEmbeddings(tf.keras.layers.Layer):
         if inputs_embeds is None:
             inputs_embeds = tf.gather(self.word_embeddings, input_ids)
 
-        position_embeddings = self.position_embeddings(position_ids)
-        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+        position_embeddings = tf.cast(self.position_embeddings(position_ids), inputs_embeds.dtype)
+        token_type_embeddings = tf.cast(self.token_type_embeddings(token_type_ids), inputs_embeds.dtype)
         embeddings = inputs_embeds + position_embeddings + token_type_embeddings
         embeddings = self.LayerNorm(embeddings)
         embeddings = self.dropout(embeddings, training=training)
@@ -281,7 +281,7 @@ class TFBertSelfAttention(tf.keras.layers.Layer):
         attention_scores = tf.matmul(
             query_layer, key_layer, transpose_b=True
         )  # (batch size, num_heads, seq_len_q, seq_len_k)
-        dk = tf.cast(shape_list(key_layer)[-1], tf.float32)  # scale attention_scores
+        dk = tf.cast(shape_list(key_layer)[-1], attention_scores.dtype)  # scale attention_scores
         attention_scores = attention_scores / tf.math.sqrt(dk)
 
         if attention_mask is not None:
@@ -613,6 +613,8 @@ class TFBertMainLayer(tf.keras.layers.Layer):
         if token_type_ids is None:
             token_type_ids = tf.fill(input_shape, 0)
 
+        embedding_output = self.embeddings(input_ids, position_ids, token_type_ids, inputs_embeds, training=training)
+
         # We create a 3D attention mask from a 2D tensor mask.
         # Sizes are [batch_size, 1, 1, to_seq_length]
         # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
@@ -626,7 +628,7 @@ class TFBertMainLayer(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, tf.float32)
+        extended_attention_mask = tf.cast(extended_attention_mask, embedding_output.dtype)
         extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
 
         # Prepare head mask if needed
@@ -640,7 +642,6 @@ class TFBertMainLayer(tf.keras.layers.Layer):
             head_mask = [None] * self.num_hidden_layers
             # head_mask = tf.constant([0] * self.num_hidden_layers)
 
-        embedding_output = self.embeddings(input_ids, position_ids, token_type_ids, inputs_embeds, training=training)
         encoder_outputs = self.encoder(
             embedding_output,
             extended_attention_mask,

src/transformers/modeling_tf_electra.py

@@ -134,8 +134,8 @@ class TFElectraEmbeddings(tf.keras.layers.Layer):
 
         if inputs_embeds is None:
             inputs_embeds = tf.gather(self.word_embeddings, input_ids)
-        position_embeddings = self.position_embeddings(position_ids)
-        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+        position_embeddings = tf.cast(self.position_embeddings(position_ids), inputs_embeds.dtype)
+        token_type_embeddings = tf.cast(self.token_type_embeddings(token_type_ids), inputs_embeds.dtype)
 
         embeddings = inputs_embeds + position_embeddings + token_type_embeddings
         embeddings = self.LayerNorm(embeddings)
@@ -194,7 +194,7 @@ class TFElectraPreTrainedModel(TFBertPreTrainedModel):
     config_class = ElectraConfig
     base_model_prefix = "electra"
 
-    def get_extended_attention_mask(self, attention_mask, input_shape):
+    def get_extended_attention_mask(self, attention_mask, input_shape, dtype):
         if attention_mask is None:
             attention_mask = tf.fill(input_shape, 1)
 
@@ -211,7 +211,7 @@ class TFElectraPreTrainedModel(TFBertPreTrainedModel):
         # 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, tf.float32)
+        extended_attention_mask = tf.cast(extended_attention_mask, dtype)
         extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
 
         return extended_attention_mask
@@ -314,11 +314,11 @@ class TFElectraMainLayer(TFElectraPreTrainedModel):
         if token_type_ids is None:
             token_type_ids = tf.fill(input_shape, 0)
 
-        extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_shape)
-        head_mask = self.get_head_mask(head_mask)
-
         hidden_states = self.embeddings(input_ids, position_ids, token_type_ids, inputs_embeds, training=training)
 
+        extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_shape, hidden_states.dtype)
+        head_mask = self.get_head_mask(head_mask)
+
         if hasattr(self, "embeddings_project"):
             hidden_states = self.embeddings_project(hidden_states, training=training)