Refactor RotaryEmbedding and GPTNeoXAttention (#1101)

* fix rotary emb * refactor + remove unnecessary typecast * fix formatting * refactor * formatting fix * refactoring rotary emb * added a kwarg in super().__init__()
keras-team · Jul 6, 2023 · f68c256 · f68c256
1 parent 50d18f5
commit f68c256
Show file tree

Hide file tree

Showing 2 changed files with 31 additions and 34 deletions.
diff --git a/keras_nlp/models/gpt_neo_x/gpt_neo_x_attention.py b/keras_nlp/models/gpt_neo_x/gpt_neo_x_attention.py
@@ -62,9 +62,8 @@ def __init__(
         self.dropout = dropout
         self.attn_head_size = hidden_dim // num_heads
         self.rotary_max_wavelength = rotary_max_wavelength
-        self.rotary_embedding = RotaryEmbedding(
-            self.rotary_percentage, rotary_max_wavelength
-        )
+        self.rotary_dim = int(self.attn_head_size * rotary_percentage)
+        self.rotary_embedding = RotaryEmbedding(rotary_max_wavelength)
         self.kernel_initializer = keras.initializers.get(kernel_initializer)
         self.bias_initializer = keras.initializers.get(bias_initializer)
         self.max_sequence_length = max_sequence_length
@@ -148,7 +147,19 @@ def call(
         ]
         value = query_key_value[..., 2 * self.attn_head_size :]
 
-        query, key = self.rotary_embedding(query, key)
+        query_rot, query_pass = (
+            query[..., : self.rotary_dim],
+            query[..., self.rotary_dim :],
+        )
+        key_rot, key_pass = (
+            key[..., : self.rotary_dim],
+            key[..., self.rotary_dim :],
+        )
+
+        query_rot, key_rot = self.rotary_embedding(query_rot, key_rot)
+
+        query = tf.concat((query_rot, query_pass), axis=-1)
+        key = tf.concat((key_rot, key_pass), axis=-1)
 
         attention_output = self._compute_attention(
             query=query,

diff --git a/keras_nlp/models/gpt_neo_x/rotary_embedding.py b/keras_nlp/models/gpt_neo_x/rotary_embedding.py
@@ -16,62 +16,48 @@
 
 
 class RotaryEmbedding(keras.layers.Layer):
-    def __init__(self, rotary_percentage, max_wavelength=10000):
-        super().__init__()
-        self.rotary_percentage = rotary_percentage
+    def __init__(self, max_wavelength=10000, **kwargs):
+        super().__init__(**kwargs)
         self.max_wavelength = max_wavelength
 
     def _apply_rotary_pos_emb(self, tensor, cos_emb, sin_emb):
         cos_emb = cos_emb[:, : tf.shape(tensor)[1], :, :]
         sin_emb = sin_emb[:, : tf.shape(tensor)[1], :, :]
+
         x1, x2 = tf.split(tensor, 2, axis=-1)
         half_rot_tensor = tf.concat((-x2, x1), axis=-1)
-        ret = (tensor * cos_emb) + (half_rot_tensor * sin_emb)
-        return ret
 
-    def _compute_cos_sin_embedding(self, x, rotary_ndims, seq_dim=1):
+        return (tensor * cos_emb) + (half_rot_tensor * sin_emb)
+
+    def _compute_cos_sin_embedding(self, x, rotary_dim, seq_dim=1):
         seq_len = tf.shape(x)[seq_dim]
-        rotary_ndims = tf.cast(rotary_ndims, tf.float32)
-        range = tf.range(start=0, limit=rotary_ndims, delta=2, dtype="float32")
-        inverse_freq = 1.0 / (self.max_wavelength ** (range / rotary_ndims))
+        rotary_dim = tf.cast(rotary_dim, "float32")
+
+        range = tf.range(start=0, limit=rotary_dim, delta=2, dtype="float32")
+        inverse_freq = 1.0 / (self.max_wavelength ** (range / rotary_dim))
+
         tensor = tf.range(seq_len, dtype=inverse_freq.dtype)
         freqs = tf.einsum("i, j -> ij", tensor, inverse_freq)
         embedding = tf.concat((freqs, freqs), axis=-1)[None, :, None, :]
+
         return tf.cos(embedding), tf.sin(embedding)
 
     def call(self, query, key):
-        attn_head_size = tf.shape(query)[-1]
-        rotary_ndims = tf.cast(
-            tf.cast(attn_head_size, self.compute_dtype)
-            * self.rotary_percentage,
-            tf.int32,
-        )
-
-        query_rot, query_pass = (
-            query[..., :rotary_ndims],
-            query[..., rotary_ndims:],
-        )
-        key_rot, key_pass = (
-            key[..., :rotary_ndims],
-            key[..., rotary_ndims:],
-        )
+        rotary_dim = tf.shape(query)[-1]
 
         cos_emb, sin_emb = self._compute_cos_sin_embedding(
-            key_rot, rotary_ndims, seq_dim=1
+            query, rotary_dim, seq_dim=1
         )
-        query_emb = self._apply_rotary_pos_emb(query_rot, cos_emb, sin_emb)
-        key_emb = self._apply_rotary_pos_emb(key_rot, cos_emb, sin_emb)
 
-        query = tf.concat((query_emb, query_pass), axis=-1)
-        key = tf.concat((key_emb, key_pass), axis=-1)
+        query = self._apply_rotary_pos_emb(query, cos_emb, sin_emb)
+        key = self._apply_rotary_pos_emb(key, cos_emb, sin_emb)
 
         return query, key
 
     def get_config(self):
         config = super().get_config()
         config.update(
             {
-                "rotary_percentage": self.rotary_percentage,
                 "max_wavelength": self.max_wavelength,
             }
         )