Making TF Lxmert model compliant with AMP (huggingface#10257)

* Fix AMP

* Rework cast

* Apply style

src/transformers/models/lxmert/modeling_tf_lxmert.py

@@ -295,11 +295,12 @@ def call(self, hidden_states, context, attention_mask, output_attentions, traini
         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], dtype=attention_scores.dtype)  # scale attention_scores
         attention_scores = attention_scores / tf.math.sqrt(dk)
 
         if attention_mask is not None:
-            # Apply the attention mask is (precomputed for all layers in TFBertModel call() function)
+            # Apply the attention mask is (precomputed for all layers in TFLxmertModel call() function)
+            attention_mask = tf.cast(attention_mask, dtype=attention_scores.dtype)
             attention_scores = attention_scores + attention_mask
 
         # Normalize the attention scores to probabilities.
@@ -721,6 +722,11 @@ def call(
         if inputs["token_type_ids"] is None:
             inputs["token_type_ids"] = tf.fill(input_shape, 0)
 
+        # Positional Word Embeddings
+        embedding_output = self.embeddings(
+            inputs["input_ids"], inputs["token_type_ids"], inputs["inputs_embeds"], training=inputs["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]
@@ -734,8 +740,10 @@ def call(
         # 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 = (1.0 - extended_attention_mask) * -10000.0
+        extended_attention_mask = tf.cast(extended_attention_mask, dtype=embedding_output.dtype)
+        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)
 
         if inputs["visual_attention_mask"] is not None:
             extended_visual_attention_mask = tf.reshape(
@@ -745,16 +753,13 @@ def call(
                 tf.expand_dims(inputs["visual_attention_mask"], axis=1), axis=1
             )
 
-            extended_visual_attention_mask = tf.cast(extended_visual_attention_mask, tf.float32)
-            extended_visual_attention_mask = (1.0 - extended_visual_attention_mask) * -10000.0
+            extended_visual_attention_mask = tf.cast(extended_visual_attention_mask, dtype=embedding_output.dtype)
+            extended_visual_attention_mask = tf.multiply(
+                tf.subtract(one_cst, extended_visual_attention_mask), ten_thousand_cst
+            )
         else:
             extended_visual_attention_mask = None
 
-        # Positional Word Embeddings
-        embedding_output = self.embeddings(
-            inputs["input_ids"], inputs["token_type_ids"], inputs["inputs_embeds"], training=inputs["training"]
-        )
-
         # Run Lxmert encoder
         encoder_outputs = self.encoder(
             embedding_output,

tests/test_modeling_tf_lxmert.py

@@ -706,10 +706,6 @@ def test_saved_model_creation(self):
         # This test is too long (>30sec) and makes fail the CI
         pass
 
-    def test_mixed_precision(self):
-        # TODO JP: Make Lxmert float16 compliant
-        pass
-
     @slow
     def test_saved_model_creation_extended(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()