Making TF GPT2 compliant with XLA and AMP (huggingface#10230)

* Fix XLA and AMP

* Fix AMP and XLA

* Apply style

* Apply Patrick's comment

src/transformers/modeling_tf_utils.py

@@ -1331,119 +1331,6 @@ def call(self, x):
         return x
 
 
-class WordEmbeddings(tf.keras.layers.Layer):
-    def __init__(self, vocab_size: int, hidden_size: int, initializer_range: float, **kwargs):
-        super().__init__(**kwargs)
-
-        self.vocab_size = vocab_size
-        self.hidden_size = hidden_size
-        self.initializer_range = initializer_range
-
-    def build(self, input_shape):
-        self.word_embeddings = self.add_weight(
-            name="weight",
-            shape=[self.vocab_size, self.hidden_size],
-            initializer=get_initializer(initializer_range=self.initializer_range),
-        )
-
-        super().build(input_shape=input_shape)
-
-    def get_config(self):
-        config = {
-            "vocab_size": self.vocab_size,
-            "hidden_size": self.hidden_size,
-            "initializer_range": self.initializer_range,
-        }
-        base_config = super().get_config()
-
-        return dict(list(base_config.items()) + list(config.items()))
-
-    def call(self, input_ids):
-        flat_input_ids = tf.reshape(tensor=input_ids, shape=[-1])
-        embeddings = tf.gather(params=self.word_embeddings, indices=flat_input_ids)
-        embeddings = tf.reshape(
-            tensor=embeddings, shape=tf.concat(values=[shape_list(tensor=input_ids), [self.hidden_size]], axis=0)
-        )
-
-        embeddings.set_shape(shape=input_ids.shape.as_list() + [self.hidden_size])
-
-        return embeddings
-
-
-class TokenTypeEmbeddings(tf.keras.layers.Layer):
-    def __init__(self, type_vocab_size: int, hidden_size: int, initializer_range: float, **kwargs):
-        super().__init__(**kwargs)
-
-        self.type_vocab_size = type_vocab_size
-        self.hidden_size = hidden_size
-        self.initializer_range = initializer_range
-
-    def build(self, input_shape):
-        self.token_type_embeddings = self.add_weight(
-            name="embeddings",
-            shape=[self.type_vocab_size, self.hidden_size],
-            initializer=get_initializer(initializer_range=self.initializer_range),
-        )
-
-        super().build(input_shape=input_shape)
-
-    def get_config(self):
-        config = {
-            "type_vocab_size": self.type_vocab_size,
-            "hidden_size": self.hidden_size,
-            "initializer_range": self.initializer_range,
-        }
-        base_config = super().get_config()
-
-        return dict(list(base_config.items()) + list(config.items()))
-
-    def call(self, token_type_ids):
-        flat_token_type_ids = tf.reshape(tensor=token_type_ids, shape=[-1])
-        one_hot_data = tf.one_hot(indices=flat_token_type_ids, depth=self.type_vocab_size, dtype=self._compute_dtype)
-        embeddings = tf.matmul(a=one_hot_data, b=self.token_type_embeddings)
-        embeddings = tf.reshape(
-            tensor=embeddings, shape=tf.concat(values=[shape_list(tensor=token_type_ids), [self.hidden_size]], axis=0)
-        )
-
-        embeddings.set_shape(shape=token_type_ids.shape.as_list() + [self.hidden_size])
-
-        return embeddings
-
-
-class PositionEmbeddings(tf.keras.layers.Layer):
-    def __init__(self, max_position_embeddings: int, hidden_size: int, initializer_range: float, **kwargs):
-        super().__init__(**kwargs)
-
-        self.max_position_embeddings = max_position_embeddings
-        self.hidden_size = hidden_size
-        self.initializer_range = initializer_range
-
-    def build(self, input_shape):
-        self.position_embeddings = self.add_weight(
-            name="embeddings",
-            shape=[self.max_position_embeddings, self.hidden_size],
-            initializer=get_initializer(initializer_range=self.initializer_range),
-        )
-
-        super().build(input_shape)
-
-    def get_config(self):
-        config = {
-            "max_position_embeddings": self.max_position_embeddings,
-            "hidden_size": self.hidden_size,
-            "initializer_range": self.initializer_range,
-        }
-        base_config = super().get_config()
-
-        return dict(list(base_config.items()) + list(config.items()))
-
-    def call(self, position_ids):
-        input_shape = shape_list(tensor=position_ids)
-        position_embeddings = self.position_embeddings[: input_shape[1], :]
-
-        return tf.broadcast_to(input=position_embeddings, shape=input_shape)
-
-
 class TFSharedEmbeddings(tf.keras.layers.Layer):
     r"""
     Construct shared token embeddings.

src/transformers/models/gpt2/modeling_tf_gpt2.py

@@ -112,6 +112,7 @@ def _attn(self, q, k, v, attention_mask, head_mask, output_attentions, training=
 
         if attention_mask is not None:
             # Apply the attention mask
+            attention_mask = tf.cast(attention_mask, dtype=w.dtype)
             w = w + attention_mask
 
         w = tf.nn.softmax(w, axis=-1)
@@ -224,20 +225,26 @@ def __init__(self, config, *inputs, **kwargs):
         self.num_hidden_layers = config.n_layer
         self.vocab_size = config.vocab_size
         self.n_embd = config.n_embd
+        self.n_positions = config.n_positions
+        self.initializer_range = config.initializer_range
 
         self.wte = TFSharedEmbeddings(
             config.vocab_size, config.hidden_size, initializer_range=config.initializer_range, name="wte"
         )
-        self.wpe = tf.keras.layers.Embedding(
-            config.n_positions,
-            config.n_embd,
-            embeddings_initializer=get_initializer(config.initializer_range),
-            name="wpe",
-        )
         self.drop = tf.keras.layers.Dropout(config.embd_pdrop)
         self.h = [TFBlock(config.n_ctx, config, scale=True, name="h_._{}".format(i)) for i in range(config.n_layer)]
         self.ln_f = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_epsilon, name="ln_f")
 
+    def build(self, input_shape):
+        with tf.name_scope("wpe"):
+            self.wpe = self.add_weight(
+                name="embeddings",
+                shape=[self.n_positions, self.n_embd],
+                initializer=get_initializer(self.initializer_range),
+            )
+
+        super().build(input_shape)
+
     def get_input_embeddings(self):
         return self.wte
 
@@ -302,9 +309,7 @@ def call(
             past_length = shape_list(inputs["past"][0][0])[-2]
 
         if inputs["position_ids"] is None:
-            inputs["position_ids"] = tf.expand_dims(
-                tf.range(past_length, input_shape[-1] + past_length, dtype=tf.int32), axis=0
-            )
+            inputs["position_ids"] = tf.expand_dims(tf.range(past_length, input_shape[-1] + past_length), axis=0)
 
         if inputs["attention_mask"] is not None:
             # We create a 3D attention mask from a 2D tensor mask.
@@ -322,11 +327,11 @@ def call(
             # positions we want to attend and -10000.0 for masked positions.
             # Since we are adding it to the raw scores before the softmax, this is
             # effectively the same as removing these entirely.
-
-            inputs["attention_mask"] = tf.cast(inputs["attention_mask"], tf.float32)
-            inputs["attention_mask"] = (1.0 - inputs["attention_mask"]) * -10000.0
-        else:
-            inputs["attention_mask"] = None
+            one_cst = tf.constant(1.0)
+            inputs["attention_mask"] = tf.cast(inputs["attention_mask"], dtype=one_cst.dtype)
+            inputs["attention_mask"] = tf.multiply(
+                tf.subtract(one_cst, inputs["attention_mask"]), tf.constant(-10000.0)
+            )
 
         # Prepare head mask if needed
         # 1.0 in head_mask indicate we keep the head
@@ -344,15 +349,15 @@ def call(
         if inputs["inputs_embeds"] is None:
             inputs["inputs_embeds"] = self.wte(inputs["input_ids"], mode="embedding")
 
-        position_embeds = self.wpe(inputs["position_ids"])
+        position_embeds = tf.gather(self.wpe, inputs["position_ids"])
 
         if inputs["token_type_ids"] is not None:
             inputs["token_type_ids"] = tf.reshape(
                 inputs["token_type_ids"], [-1, shape_list(inputs["token_type_ids"])[-1]]
             )
             token_type_embeds = self.wte(inputs["token_type_ids"], mode="embedding")
         else:
-            token_type_embeds = 0
+            token_type_embeds = tf.constant(0.0)
 
         position_embeds = tf.cast(position_embeds, dtype=inputs["inputs_embeds"].dtype)
         token_type_embeds = tf.cast(token_type_embeds, dtype=inputs["inputs_embeds"].dtype)
@@ -1024,7 +1029,10 @@ def call(
             if inputs["input_ids"] is not None:
                 sequence_lengths = (
                     tf.reduce_sum(
-                        tf.cast(tf.math.not_equal(inputs["input_ids"], self.config.pad_token_id), tf.int32),
+                        tf.cast(
+                            tf.math.not_equal(inputs["input_ids"], self.config.pad_token_id),
+                            dtype=inputs["input_ids"].dtype,
+                        ),
                         -1,
                         keepdims=False,
                     )

tests/test_modeling_tf_gpt2.py

@@ -389,14 +389,6 @@ def test_gpt2_sequence_classification_model(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_gpt2_for_sequence_classification(*config_and_inputs)
 
-    def test_mixed_precision(self):
-        # TODO JP: Make GPT2 float16 compliant
-        pass
-
-    def test_xla_mode(self):
-        # TODO JP: Make GPT2 XLA compliant
-        pass
-
     @slow
     def test_model_from_pretrained(self):
         for model_name in TF_GPT2_PRETRAINED_MODEL_ARCHIVE_LIST[:1]: