Add a mixed precision test and fix mixed precision errors for layers (#1242)

* Add a mixed precision test and fix mixed precision errors for layers

* Address comments

* fix torch cpu

Author: mattdangerw

keras_nlp/layers/modeling/cached_multi_head_attention_test.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from keras_nlp.backend import config
 from keras_nlp.backend import ops
 from keras_nlp.layers.modeling.cached_multi_head_attention import (
     CachedMultiHeadAttention,
@@ -34,6 +35,9 @@ def test_layer_behaviors(self):
             expected_output_shape=(2, 4, 6),
             expected_num_trainable_weights=8,
             expected_num_non_trainable_variables=1,
+            # tf.keras does not handle mixed precision correctly when not set
+            # globally.
+            run_mixed_precision_check=config.multi_backend(),
         )
 
     def test_cache_call_is_correct(self):

keras_nlp/layers/modeling/f_net_encoder.py

@@ -97,11 +97,13 @@ def build(self, inputs_shape):
         # Layer Norm layers.
         self._mixing_layer_norm = keras.layers.LayerNormalization(
             epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
             name="mixing_layer_norm",
         )
         self._mixing_layer_norm.build(inputs_shape)
         self._output_layer_norm = keras.layers.LayerNormalization(
             epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
             name="output_layer_norm",
         )
         self._output_layer_norm.build(inputs_shape)
@@ -112,19 +114,25 @@ def build(self, inputs_shape):
             activation=self.activation,
             kernel_initializer=clone_initializer(self.kernel_initializer),
             bias_initializer=clone_initializer(self.bias_initializer),
+            dtype=self.dtype_policy,
             name="intermediate_dense",
         )
         self._intermediate_dense.build(inputs_shape)
         self._output_dense = keras.layers.Dense(
             feature_size,
             kernel_initializer=clone_initializer(self.kernel_initializer),
             bias_initializer=clone_initializer(self.bias_initializer),
+            dtype=self.dtype_policy,
             name="output_dense",
         )
         self._output_dense.build(
             self._intermediate_dense.compute_output_shape(inputs_shape)
         )
-        self._output_dropout = keras.layers.Dropout(rate=self.dropout)
+        self._output_dropout = keras.layers.Dropout(
+            rate=self.dropout,
+            dtype=self.dtype_policy,
+            name="output_dropout",
+        )
         self.built = True
 
     def call(self, inputs):
@@ -140,9 +148,12 @@ def call(self, inputs):
 
         def fourier_transform(input):
             # Apply FFT on the input and take the real part.
+            input_dtype = input.dtype
+            # FFT transforms do not support float16.
+            input = ops.cast(input, "float32")
             real_in, imaginary_in = (input, ops.zeros_like(input))
             real_out, _ = ops.fft2((real_in, imaginary_in))
-            return real_out
+            return ops.cast(real_out, input_dtype)
 
         def add_and_norm(input1, input2, norm_layer):
             return norm_layer(input1 + input2)

keras_nlp/layers/modeling/masked_lm_head.py

@@ -141,10 +141,12 @@ def build(self, inputs_shape, mask_positions_shape=None):
             activation=self.intermediate_activation,
             kernel_initializer=self.kernel_initializer,
             bias_initializer=self.bias_initializer,
+            dtype=self.dtype_policy,
             name="intermediate_dense",
         )
         self._intermediate_layer_norm = keras.layers.LayerNormalization(
             epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
             name="intermediate_layer_norm",
         )
         # The gather length does not affect any of our built variables, so
@@ -185,6 +187,7 @@ def call(self, inputs, mask_positions):
             outputs = self.token_embedding(x, reverse=True)
         else:
             outputs = ops.matmul(x, self._kernel)
+        outputs = ops.cast(outputs, self.compute_dtype)
         outputs = outputs + self._bias
 
         # Apply a final activation.

keras_nlp/layers/modeling/reversible_embedding.py

@@ -49,6 +49,9 @@ class ReversibleEmbedding(keras.layers.Embedding):
             the `embeddings` matrix (see `keras.constraints`).
         mask_zero: Boolean, whether or not the input value 0 is a special
             "padding" value that should be masked out.
+        reverse_dtype: The dtype for the reverse projection computation.
+            For stability, it is usually best to use full precision even when
+            working with half or mixed precision training.
 
     Call args:
         inputs: The tensor inputs to the layer.
@@ -87,6 +90,7 @@ def __init__(
         embeddings_regularizer=None,
         embeddings_constraint=None,
         mask_zero=False,
+        reverse_dtype="float32",
         **kwargs,
     ):
         super().__init__(
@@ -99,6 +103,7 @@ def __init__(
             **kwargs,
         )
         self.tie_weights = tie_weights
+        self.reverse_dtype = reverse_dtype
 
     def build(self, inputs_shape=None):
         super().build(inputs_shape)
@@ -114,12 +119,12 @@ def build(self, inputs_shape=None):
     def call(self, inputs, reverse=False):
         if reverse:
             if self.tie_weights:
-                reverse_embeddings = ops.transpose(
-                    ops.convert_to_tensor(self.embeddings)
-                )
+                kernel = ops.transpose(ops.convert_to_tensor(self.embeddings))
             else:
-                reverse_embeddings = self.reverse_embeddings
-            return ops.matmul(inputs, reverse_embeddings)
+                kernel = self.reverse_embeddings
+            inputs = ops.cast(inputs, self.reverse_dtype)
+            kernel = ops.cast(kernel, self.reverse_dtype)
+            return ops.matmul(inputs, kernel)
 
         return super().call(inputs)
 
@@ -128,6 +133,7 @@ def get_config(self):
         config.update(
             {
                 "tie_weights": self.tie_weights,
+                "reverse_dtype": self.reverse_dtype,
             }
         )
         return config

keras_nlp/layers/modeling/reversible_embedding_test.py

@@ -17,6 +17,7 @@
 import numpy as np
 from absl.testing import parameterized
 
+from keras_nlp.backend import config
 from keras_nlp.backend import keras
 from keras_nlp.backend import ops
 from keras_nlp.layers.modeling.reversible_embedding import ReversibleEmbedding
@@ -73,3 +74,22 @@ def test_tied_checkpoint_untied_weights(self):
 
         input_data = ops.ones(shape=(4, 10), dtype="int32")
         self.assertAllClose(untied_model(input_data), tied_model(input_data))
+
+    def test_reverse_dtype(self):
+        embedding = ReversibleEmbedding(100, 16, reverse_dtype="float32")
+        input_data = ops.ones(shape=(4, 10, 16))
+        output_data = embedding(input_data, reverse=True)
+        self.assertEqual(output_data.shape, (4, 10, 100))
+        self.assertDTypeEqual(output_data, "float32")
+
+        if config.backend() == "torch":
+            import torch
+
+            if not torch.cuda.is_available():
+                self.skipTest("Torch CPU does not support float16")
+
+        embedding = ReversibleEmbedding(100, 16, reverse_dtype="float16")
+        input_data = ops.ones(shape=(4, 10, 16))
+        output_data = embedding(input_data, reverse=True)
+        self.assertEqual(output_data.shape, (4, 10, 100))
+        self.assertDTypeEqual(output_data, "float16")

keras_nlp/layers/modeling/token_and_position_embedding.py

@@ -93,11 +93,13 @@ def __init__(
                 self.embeddings_initializer
             ),
             mask_zero=mask_zero,
+            dtype=self.dtype_policy,
             name="token_embedding",
         )
         self.position_embedding = PositionEmbedding(
             sequence_length=sequence_length,
             initializer=clone_initializer(self.embeddings_initializer),
+            dtype=self.dtype_policy,
             name="position_embedding",
         )
         self.supports_masking = self.token_embedding.supports_masking

keras_nlp/layers/modeling/transformer_decoder.py

@@ -111,15 +111,14 @@ def __init__(
         kernel_initializer="glorot_uniform",
         bias_initializer="zeros",
         normalize_first=False,
-        name=None,
         **kwargs,
     ):
         # Work around for model saving, we need to ensure our model is built
         # immediately after restoring from config.
         decoder_sequence_shape = kwargs.pop("decoder_sequence_shape", None)
         encoder_sequence_shape = kwargs.pop("encoder_sequence_shape", None)
 
-        super().__init__(name=name, **kwargs)
+        super().__init__(**kwargs)
         self.intermediate_dim = intermediate_dim
         self.num_heads = num_heads
         self.dropout = dropout
@@ -160,6 +159,7 @@ def build(
             dropout=self.dropout,
             kernel_initializer=clone_initializer(self.kernel_initializer),
             bias_initializer=clone_initializer(self.bias_initializer),
+            dtype=self.dtype_policy,
             name="self_attention",
         )
         if hasattr(self._self_attention_layer, "_build_from_signature"):
@@ -174,11 +174,14 @@ def build(
             )
         self._self_attention_layer_norm = keras.layers.LayerNormalization(
             epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
             name="self_attention_layer_norm",
         )
         self._self_attention_layer_norm.build(decoder_sequence_shape)
         self._self_attention_dropout = keras.layers.Dropout(
             rate=self.dropout,
+            dtype=self.dtype_policy,
+            name="self_attention_dropout",
         )
 
         # Cross attention layers are optional.
@@ -191,6 +194,7 @@ def build(
                 dropout=self.dropout,
                 kernel_initializer=clone_initializer(self.kernel_initializer),
                 bias_initializer=clone_initializer(self.bias_initializer),
+                dtype=self.dtype_policy,
                 name="cross_attention",
             )
             if hasattr(self._cross_attention_layer, "_build_from_signature"):
@@ -205,11 +209,14 @@ def build(
                 )
             self._cross_attention_layer_norm = keras.layers.LayerNormalization(
                 epsilon=self.layer_norm_epsilon,
+                dtype=self.dtype_policy,
                 name="cross_attention_layer_norm",
             )
             self._cross_attention_layer_norm.build(encoder_sequence_shape)
             self._cross_attention_dropout = keras.layers.Dropout(
                 rate=self.dropout,
+                dtype=self.dtype_policy,
+                name="cross_attention_dropout",
             )
 
         # Feedforward layers.
@@ -218,25 +225,30 @@ def build(
             activation=self.activation,
             kernel_initializer=clone_initializer(self.kernel_initializer),
             bias_initializer=clone_initializer(self.bias_initializer),
-            name="intermediate_dense",
+            dtype=self.dtype_policy,
+            name="feedforward_intermediate_dense",
         )
         self._feedforward_intermediate_dense.build(decoder_sequence_shape)
         self._feedforward_output_dense = keras.layers.Dense(
             hidden_dim,
             kernel_initializer=clone_initializer(self.kernel_initializer),
             bias_initializer=clone_initializer(self.bias_initializer),
-            name="output_dense",
+            dtype=self.dtype_policy,
+            name="feedforward_output_dense",
         )
         intermediate_shape = list(decoder_sequence_shape)
         intermediate_shape[-1] = self.intermediate_dim
         self._feedforward_output_dense.build(tuple(intermediate_shape))
         self._feedforward_layer_norm = keras.layers.LayerNormalization(
             epsilon=self.layer_norm_epsilon,
-            name="output_layer_norm",
+            dtype=self.dtype_policy,
+            name="feedforward_layer_norm",
         )
         self._feedforward_layer_norm.build(decoder_sequence_shape)
         self._feedforward_dropout = keras.layers.Dropout(
             rate=self.dropout,
+            dtype=self.dtype_policy,
+            name="feedforward_dropout",
         )
         # Create layers based on input shape.
         self.built = True

keras_nlp/layers/modeling/transformer_encoder.py

@@ -92,10 +92,9 @@ def __init__(
         kernel_initializer="glorot_uniform",
         bias_initializer="zeros",
         normalize_first=False,
-        name=None,
         **kwargs,
     ):
-        super().__init__(name=name, **kwargs)
+        super().__init__(**kwargs)
         self.intermediate_dim = intermediate_dim
         self.num_heads = num_heads
         self.dropout = dropout
@@ -125,6 +124,7 @@ def build(self, inputs_shape):
             dropout=self.dropout,
             kernel_initializer=clone_initializer(self.kernel_initializer),
             bias_initializer=clone_initializer(self.bias_initializer),
+            dtype=self.dtype_policy,
             name="self_attention_layer",
         )
         if hasattr(self._self_attention_layer, "_build_from_signature"):
@@ -139,38 +139,46 @@ def build(self, inputs_shape):
             )
         self._self_attention_layer_norm = keras.layers.LayerNormalization(
             epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
             name="self_attention_layer_norm",
         )
         self._self_attention_layer_norm.build(inputs_shape)
         self._self_attention_dropout = keras.layers.Dropout(
             rate=self.dropout,
+            dtype=self.dtype_policy,
             name="self_attention_dropout",
         )
 
         # Feedforward layers.
         self._feedforward_layer_norm = keras.layers.LayerNormalization(
             epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
+            name="feedforward_layer_norm",
         )
         self._feedforward_layer_norm.build(inputs_shape)
         self._feedforward_intermediate_dense = keras.layers.Dense(
             self.intermediate_dim,
             activation=self.activation,
             kernel_initializer=clone_initializer(self.kernel_initializer),
             bias_initializer=clone_initializer(self.bias_initializer),
+            dtype=self.dtype_policy,
             name="feedforward_intermediate_dense",
         )
         self._feedforward_intermediate_dense.build(inputs_shape)
         self._feedforward_output_dense = keras.layers.Dense(
             hidden_dim,
             kernel_initializer=clone_initializer(self.kernel_initializer),
             bias_initializer=clone_initializer(self.bias_initializer),
+            dtype=self.dtype_policy,
             name="feedforward_output_dense",
         )
         intermediate_shape = list(inputs_shape)
         intermediate_shape[-1] = self.intermediate_dim
         self._feedforward_output_dense.build(tuple(intermediate_shape))
         self._feedforward_dropout = keras.layers.Dropout(
             rate=self.dropout,
+            dtype=self.dtype_policy,
+            name="feedforward_dropout",
         )
         self.built = True
 

keras_nlp/samplers/beam_sampler.py

@@ -16,7 +16,6 @@
 import tree
 
 from keras_nlp.api_export import keras_nlp_export
-from keras_nlp.backend import keras
 from keras_nlp.backend import ops
 from keras_nlp.samplers.sampler import Sampler
 from keras_nlp.samplers.sampler import call_args_docstring
@@ -161,7 +160,7 @@ def body(prompt, cache, index, log_probs):
             # Compute the softmax distribution for the next token.
             logits, _, cache = next(prompt, cache, index)
             vocab_size = ops.shape(logits)[-1]
-            probs = keras.activations.softmax(logits / self.temperature)
+            probs = self.compute_probabilities(logits)
 
             # Compute the running log-likelihood of each new candidate.
             next_log_probs = ops.log(probs) + log_probs[..., None]

keras_nlp/samplers/contrastive_sampler.py

@@ -15,7 +15,6 @@
 import tree
 
 from keras_nlp.api_export import keras_nlp_export
-from keras_nlp.backend import keras
 from keras_nlp.backend import ops
 from keras_nlp.samplers.sampler import Sampler
 from keras_nlp.samplers.sampler import call_args_docstring
@@ -131,7 +130,7 @@ def cond(prompt, cache, index, logits, hidden_states):
 
         def body(prompt, cache, index, logits, hidden_states):
             # Compute the softmax distribution for the next token.
-            probabilities = keras.activations.softmax(logits / self.temperature)
+            probabilities = self.compute_probabilities(logits)
 
             # Replicate for `self.k` times to find the best token in top-k
             # candidates.