keras-hub-nightly 0.15.0.dev20240823171555__py3-none-any.whl

keras_hub/src/models/bloom/bloom_causal_lm.py

@@ -0,0 +1,298 @@
+# Copyright 2024 The KerasHub Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from keras import ops
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.bloom.bloom_backbone import BloomBackbone
+from keras_hub.src.models.bloom.bloom_causal_lm_preprocessor import (
+    BloomCausalLMPreprocessor,
+)
+from keras_hub.src.models.causal_lm import CausalLM
+from keras_hub.src.utils.tensor_utils import any_equal
+
+
+@keras_hub_export("keras_hub.models.BloomCausalLM")
+class BloomCausalLM(CausalLM):
+    """An end-to-end BLOOM model for causal language modeling.
+
+    A causal language model (LM) predicts the next token based on previous
+    tokens. This task setup can be used to train the model unsupervised on
+    plain text input, or to autoregressively generate plain text similar to
+    the data used for training. This task can be used for pre-training or
+    fine-tuning a BLOOM model, simply by calling `fit()`.
+
+    This model has a `generate()` method, which generates text based on a
+    prompt. The generation strategy used is controlled by an additional
+    `sampler` argument on `compile()`. You can recompile the model with
+    different `keras_hub.samplers` objects to control the generation. By
+    default, `"greedy"` sampling will be used.
+
+    This model can optionally be configured with a `preprocessor` layer, in
+    which case it will automatically apply preprocessing to string inputs during
+    `fit()`, `predict()`, `evaluate()` and `generate()`. This is done by default
+    when creating the model with `from_preset()`.
+
+    Args:
+        backbone: A `keras_hub.models.BloomBackbone` instance.
+        preprocessor: A `keras_hub.models.BloomCausalLMPreprocessor` or `None`.
+            If `None`, this model will not apply preprocessing, and inputs
+            should be preprocessed before calling the model.
+
+    Examples:
+
+    Use `generate()` to do text generation.
+    ```python
+    bloom_lm = keras_hub.models.BloomCausalLM.from_preset("bloom_560m_multi")
+    bloom_lm.generate("I want to say", max_length=30)
+
+    # Generate with batched prompts.
+    bloom_lm.generate(["This is a", "Where are you"], max_length=30)
+    ```
+
+    Compile the `generate()` function with a custom sampler.
+    ```python
+    bloom_lm = keras_hub.models.BloomCausalLM.from_preset("bloom_560m_multi")
+    bloom_lm.compile(sampler="top_k")
+    bloom_lm.generate("I want to say", max_length=30)
+
+    bloom_lm.compile(sampler=keras_hub.samplers.BeamSampler(num_beams=2))
+    bloom_lm.generate("I want to say", max_length=30)
+    ```
+
+    Use `generate()` without preprocessing.
+    ```python
+    prompt = {
+        # Token ids for "<s> Keras is".
+        "token_ids": np.array([[1, 46, 15762, 632, 3, 3, 3, 3, 3]] * 2),
+        # Use `"padding_mask"` to indicate values that should not be overridden.
+        "padding_mask": np.array([[1, 1, 1, 1, 0, 0, 0, 0, 0]] * 2),
+    }
+
+    bloom_lm = keras_hub.models.BloomCausalLM.from_preset(
+        "bloom_560m_multi",
+        preprocessor=None,
+    )
+    bloom_lm.generate(prompt)
+    ```
+
+    Call `fit()` on a single batch.
+    ```python
+    features = ["The quick brown fox jumped.", "I forgot my homework."]
+    bloom_lm = keras_hub.models.BloomCausalLM.from_preset("bloom_560m_multi")
+    bloom_lm.fit(x=features, batch_size=2)
+    ```
+
+    Call `fit()` without preprocessing.
+    ```python
+    x = {
+        # Token ids for "<bos> Keras is deep learning library<eos>"
+        "token_ids": np.array([[2, 214064, 603, 5271, 6044, 9581, 1, 0]] * 2),
+        "padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 0]] * 2),
+    }
+    y = np.array([[214064, 603, 5271, 6044, 9581, 3, 0, 0]] * 2)
+    sw = np.array([[1, 1, 1, 1, 1, 1, 0, 0]] * 2)
+
+    bloom_lm = keras_hub.models.BloomCausalLM.from_preset(
+        "bloom_560m_multi",
+        preprocessor=None,
+    )
+    bloom_lm.fit(x=x, y=y, sample_weight=sw, batch_size=2)
+    ```
+
+    Custom backbone and vocabulary.
+    ```python
+    features = [
+        " airplane at airport",
+        " airplane airport",
+    ]
+    vocab = ["<unk>", "<s>", "</s>", "<pad>"]
+    vocab += ["!", "air", "Ġair", "plane", "Ġat", "port"]
+    vocab = dict([(token, i) for i, token in enumerate(vocab)])
+    merges = ["Ġ a", "Ġ t", "Ġ i", "Ġ b", "a i", "p l", "n e"]
+    merges += ["Ġa t", "p o", "r t", "Ġt h", "ai r", "pl a", "po rt"]
+    merges += ["Ġai r", "Ġa i", "pla ne"]
+    tokenizer = keras_hub.models.BloomTokenizer(vocabulary=vocab, merges=merges)
+    preprocessor = keras_hub.models.BloomCausalLMPreprocessor(
+        tokenizer=tokenizer,
+        sequence_length=128,
+    )
+    backbone = keras_hub.models.BloomBackbone(
+        vocabulary_size=tokenizer.vocabulary_size(),
+        num_layers=4,
+        num_heads=4,
+        hidden_dim=32,
+        intermediate_dim=128,
+    )
+    bloom_lm = keras_hub.models.BloomCausalLM(
+        backbone=backbone,
+        preprocessor=preprocessor,
+    )
+    bloom_lm.fit(x=features, batch_size=2)
+    ```
+    """
+
+    backbone_cls = BloomBackbone
+    preprocessor_cls = BloomCausalLMPreprocessor
+
+    def __init__(
+        self,
+        backbone,
+        preprocessor=None,
+        **kwargs,
+    ):
+        # === Layers ===
+        self.backbone = backbone
+        self.preprocessor = preprocessor
+
+        # === Functional Model ===
+        inputs = backbone.input
+        hidden_states = backbone(inputs)
+        outputs = backbone.token_embedding(hidden_states, reverse=True)
+        super().__init__(
+            inputs=inputs,
+            outputs=outputs,
+            **kwargs,
+        )
+
+    def call_with_cache(
+        self,
+        token_ids,
+        cache,
+        cache_update_index,
+    ):
+        """Forward pass of `BloomCausalLM` with cache.
+
+        `call_with_cache` adds an additional forward pass for the model for
+        autoregressive inference. Unlike calling the model directly, this method
+        allows caching previous key/value Tensors in multi-head attention layer,
+        and avoids recomputing the outputs of seen tokens.
+
+        Args:
+            token_ids: a dense int Tensor with shape `(batch_size, max_length)`.
+            cache: a dense float Tensor, the cache of key and value.
+            cache_update_index: int, or int Tensor. The index of current inputs
+                in the whole sequence.
+
+        Returns:
+            A (logits, hidden_states, cache) tuple. Where `logits` is the
+            language model logits for the input token_ids, `hidden_states` is
+            the final hidden representation of the input tokens, and `cache` is
+            the decoding cache.
+        """
+        x = self.backbone.token_embedding(token_ids)
+        x = self.backbone.embeddings_layer_norm(x)
+        # Each decoder layer has a cache; we update them separately.
+        caches = []
+        for i, transformer_layer in enumerate(self.backbone.transformer_layers):
+            current_cache = cache[:, i, ...]
+            x, next_cache = transformer_layer(
+                x,
+                cache=current_cache,
+                cache_update_index=cache_update_index,
+            )
+            caches.append(next_cache)
+        cache = ops.stack(caches, axis=1)
+        hidden_states = x = self.backbone.layer_norm(x)
+        logits = self.backbone.token_embedding(x, reverse=True)
+        return logits, hidden_states, cache
+
+    def _build_cache(self, token_ids):
+        """Build an empty cache for use with `call_with_cache()`."""
+        batch_size = ops.shape(token_ids)[0]
+        max_length = ops.shape(token_ids)[1]
+        num_layers = self.backbone.num_layers
+        num_heads = self.backbone.num_heads
+        head_dim = self.backbone.hidden_dim // num_heads
+        shape = [batch_size, num_layers, 2, max_length, num_heads, head_dim]
+        cache = ops.zeros(shape, dtype=self.compute_dtype)
+        # Seed the cache.
+        _, hidden_states, cache = self.call_with_cache(token_ids, cache, 0)
+        return hidden_states, cache
+
+    def generate_step(
+        self,
+        inputs,
+        stop_token_ids=None,
+    ):
+        """A compilable generation function for a single batch of inputs.
+
+        This function represents the inner, XLA-compilable, generation function
+        for a single batch of inputs. Inputs should have the same structure as
+        model inputs, a dictionary with keys `"token_ids"` and `"padding_mask"`.
+
+        Args:
+            inputs: A dictionary with two keys `"token_ids"` and
+                `"padding_mask"` and batched tensor values.
+            stop_token_ids: Tuple of id's of end token's to stop on. If all
+                sequences have produced a new stop token, generation
+                will stop.
+        """
+        token_ids, padding_mask = inputs["token_ids"], inputs["padding_mask"]
+        # Create and seed cache with a single forward pass.
+        hidden_states, cache = self._build_cache(token_ids)
+        # Compute the lengths of all user inputted tokens ids.
+        row_lengths = ops.sum(ops.cast(padding_mask, "int32"), axis=-1)
+        # Start at the first index that has no user inputted id.
+        index = ops.min(row_lengths)
+
+        def next(prompt, cache, index):
+            # The cache index is the index of our previous token.
+            cache_update_index = index - 1
+            batch_size = ops.shape(prompt)[0]
+            prompt = ops.slice(prompt, [0, cache_update_index], [batch_size, 1])
+            logits, hidden_states, cache = self.call_with_cache(
+                prompt,
+                cache,
+                cache_update_index,
+            )
+            return (
+                ops.squeeze(logits, axis=1),
+                ops.squeeze(hidden_states, axis=1),
+                cache,
+            )
+
+        token_ids = self.sampler(
+            next=next,
+            prompt=token_ids,
+            cache=cache,
+            index=index,
+            mask=padding_mask,
+            stop_token_ids=stop_token_ids,
+            hidden_states=hidden_states,
+            model=self,
+        )
+
+        # Compute an output padding mask with the token ids we updated.
+        if stop_token_ids is not None:
+            # Build a mask of stop token locations not in the original
+            # prompt (not in locations where `padding_mask` is True).
+            end_locations = any_equal(
+                token_ids, stop_token_ids, ops.logical_not(padding_mask)
+            )
+
+            end_locations = ops.cast(end_locations, "int32")
+            # Use cumsum to get ones in all locations after end_locations.
+            cumsum = ops.cast(ops.cumsum(end_locations, axis=-1), "int32")
+            overflow = cumsum - end_locations
+            # Our padding mask is the inverse of these overflow locations.
+            padding_mask = ops.logical_not(ops.cast(overflow, "bool"))
+        else:
+            # Without early stopping, all locations will have been updated.
+            padding_mask = ops.ones_like(token_ids, dtype="bool")
+        return {
+            "token_ids": token_ids,
+            "padding_mask": padding_mask,
+        }

keras_hub/src/models/bloom/bloom_causal_lm_preprocessor.py

@@ -0,0 +1,176 @@
+# Copyright 2024 The KerasHub Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import keras
+from absl import logging
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.bloom.bloom_preprocessor import BloomPreprocessor
+from keras_hub.src.utils.keras_utils import (
+    convert_inputs_to_list_of_tensor_segments,
+)
+from keras_hub.src.utils.tensor_utils import strip_to_ragged
+
+
+@keras_hub_export("keras_hub.models.BloomCausalLMPreprocessor")
+class BloomCausalLMPreprocessor(BloomPreprocessor):
+    """BLOOM Causal LM preprocessor.
+
+    This preprocessing layer is meant for use with
+    `keras_hub.models.BloomCausalLM`. By default, it will take in batches of
+    strings, and return outputs in a `(x, y, sample_weight)` format, where the
+    `y` label is the next token id in the `x` sequence.
+
+    For use with generation, the layer also exposes two methods
+    `generate_preprocess()` and `generate_postprocess()`. When this preprocessor
+    is attached to a `keras_hub.models.BloomCausalLM` instance, these methods
+    will be called implicitly in `generate()`. They can also be called
+    standalone (e.g. to precompute preprocessing inputs for generation in a
+    separate process).
+
+    Args:
+        tokenizer: A `keras_hub.models.BloomTokenizer` instance.
+        sequence_length: The length of the packed inputs.
+        add_start_token: If `True`, the preprocessor will prepend the tokenizer
+            start token to each input sequence.
+        add_end_token: If `True`, the preprocessor will append the tokenizer
+            end token to each input sequence.
+
+    Call arguments:
+        x: A string, `tf.Tensor` or list of python strings.
+        y: Label data. Should always be `None` as the layer generates labels.
+        sample_weight: Label weights. Should always be `None` as the layer
+            generates label weights.
+        sequence_length: Pass to override the configured `sequence_length` of
+            the layer.
+
+    Examples:
+    ```python
+    # Load the preprocessor from a preset.
+    preprocessor = keras_hub.models.BloomCausalLMPreprocessor.from_preset(
+        "bloom_560m_multi"
+    )
+
+    # Tokenize and pack a single sentence.
+    sentence = tf.constant("League of legends")
+    preprocessor(sentence)
+    # Same output.
+    preprocessor("League of legends")
+
+    # Tokenize a batch of sentences.
+    sentences = tf.constant(["Taco tuesday", "Fish taco please!"])
+    preprocessor(sentences)
+    # Same output.
+    preprocessor(["Taco tuesday", "Fish taco please!"])
+
+    # Map a dataset to preprocess a single sentence.
+    features = tf.constant(
+        [
+            "Avatar 2 is amazing!",
+            "Well, I am not sure.",
+        ]
+    )
+    labels = tf.constant([1, 0])
+    ds = tf.data.Dataset.from_tensor_slices((features, labels))
+    ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
+
+    # Map a dataset to preprocess unlabled sentences.
+    ds = tf.data.Dataset.from_tensor_slices(features)
+    ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
+    ```
+    """
+
+    def call(
+        self,
+        x,
+        y=None,
+        sample_weight=None,
+        sequence_length=None,
+    ):
+        if y is not None or sample_weight is not None:
+            logging.warning(
+                "`BloomCausalLMPreprocessor` generates `y` and `sample_weight` "
+                "based on your input data, but your data already contains `y` "
+                "or `sample_weight`. Your `y` and `sample_weight` will be "
+                "ignored."
+            )
+        sequence_length = sequence_length or self.sequence_length
+
+        x = convert_inputs_to_list_of_tensor_segments(x)[0]
+        x = self.tokenizer(x)
+        # Pad with one extra token to account for the truncation below.
+        token_ids, padding_mask = self.packer(
+            x,
+            sequence_length=sequence_length + 1,
+            add_start_value=self.add_start_token,
+            add_end_value=self.add_end_token,
+        )
+        # The last token does not have a next token, so we truncate it out.
+        x = {
+            "token_ids": token_ids[..., :-1],
+            "padding_mask": padding_mask[..., :-1],
+        }
+        # Target `y` will be the next token.
+        y, sample_weight = token_ids[..., 1:], padding_mask[..., 1:]
+        return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)
+
+    def generate_preprocess(
+        self,
+        x,
+        sequence_length=None,
+    ):
+        """Convert strings to integer token input for generation.
+
+        Similar to calling the layer for training, this method takes in strings
+        or tensor strings, tokenizes and packs the input, and computes a padding
+        mask masking all inputs not filled in with a padded value.
+
+        Unlike calling the layer for training, this method does not compute
+        labels and will never append a `tokenizer.end_token_id` to the end of
+        the sequence (as generation is expected to continue at the end of the
+        inputted prompt).
+        """
+        if not self.built:
+            self.build(None)
+
+        x = convert_inputs_to_list_of_tensor_segments(x)[0]
+        x = self.tokenizer(x)
+        token_ids, padding_mask = self.packer(
+            x, sequence_length=sequence_length, add_end_value=False
+        )
+        return {
+            "token_ids": token_ids,
+            "padding_mask": padding_mask,
+        }
+
+    def generate_postprocess(
+        self,
+        x,
+    ):
+        """Convert integer token output to strings for generation.
+
+        This method reverses `generate_preprocess()`, by first removing all
+        padding and start/end tokens, and then converting the integer sequence
+        back to a string.
+        """
+        if not self.built:
+            self.build(None)
+
+        token_ids, padding_mask = x["token_ids"], x["padding_mask"]
+        ids_to_strip = (
+            self.tokenizer.start_token_id,
+            self.tokenizer.end_token_id,
+        )
+        token_ids = strip_to_ragged(token_ids, padding_mask, ids_to_strip)
+        return self.tokenizer.detokenize(token_ids)

keras_hub/src/models/bloom/bloom_decoder.py

@@ -0,0 +1,206 @@
+# Copyright 2024 The KerasHub Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# import keras
+# from keras import ops
+import keras
+from keras import ops
+
+from keras_hub.src.layers.modeling.transformer_layer_utils import (
+    compute_causal_mask,
+)
+from keras_hub.src.layers.modeling.transformer_layer_utils import (
+    merge_padding_and_attention_mask,
+)
+from keras_hub.src.models.bloom.bloom_attention import BloomAttention
+from keras_hub.src.utils.keras_utils import clone_initializer
+
+
+class BloomDecoder(keras.layers.Layer):
+    def __init__(
+        self,
+        num_heads,
+        intermediate_dim,
+        dropout=0.0,
+        layer_norm_epsilon=1e-5,
+        kernel_initializer="glorot_uniform",
+        bias_initializer="zeros",
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.num_heads = num_heads
+        self.intermediate_dim = intermediate_dim
+        self.dropout = dropout
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.kernel_initializer = keras.initializers.get(kernel_initializer)
+        self.bias_initializer = keras.initializers.get(bias_initializer)
+
+    def build(self, decoder_sequence_shape):
+        hidden_dim = decoder_sequence_shape[-1]
+        head_dim = int(hidden_dim // self.num_heads)
+
+        if head_dim * self.num_heads != hidden_dim:
+            raise ValueError(
+                f"`hidden_dim` must be divisible by num_heads (got `hidden_dim`"
+                f": {hidden_dim} and `num_heads`: {self.num_heads})."
+            )
+
+        self._pre_attention_layernorm = keras.layers.LayerNormalization(
+            epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
+            name="pre_attention_layernorm",
+        )
+        self._pre_attention_layernorm.build(decoder_sequence_shape)
+
+        self._self_attention_layer = BloomAttention(
+            num_heads=self.num_heads,
+            dropout=self.dropout,
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            bias_initializer=clone_initializer(self.bias_initializer),
+            dtype=self.dtype_policy,
+            name="self_attention",
+        )
+        self._self_attention_layer.build(decoder_sequence_shape)
+
+        self._post_attention_layernorm = keras.layers.LayerNormalization(
+            epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
+            name="post_attention_layernorm",
+        )
+        self._post_attention_layernorm.build(decoder_sequence_shape)
+
+        self._mlp_intermediate_dense = keras.layers.Dense(
+            self.intermediate_dim,
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            bias_initializer=clone_initializer(self.bias_initializer),
+            dtype=self.dtype_policy,
+            name="mlp_intermediate_dense",
+        )
+        self._mlp_intermediate_dense.build(decoder_sequence_shape)
+
+        self._mlp_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="mlp_output_dense",
+        )
+        intermediate_shape = list(decoder_sequence_shape)
+        intermediate_shape[-1] = self.intermediate_dim
+        self._mlp_output_dense.build(tuple(intermediate_shape))
+
+        self._dropout_layer = keras.layers.Dropout(
+            rate=self.dropout, dtype=self.dtype_policy, name="dropout"
+        )
+
+        self.built = True
+
+    def call(
+        self,
+        decoder_sequence,
+        decoder_padding_mask=None,
+        decoder_attention_mask=None,
+        cache=None,
+        cache_update_index=None,
+        use_causal_mask=True,
+    ):
+        self_attention_mask = self._compute_attention_mask(
+            decoder_sequence=decoder_sequence,
+            decoder_padding_mask=decoder_padding_mask,
+            decoder_attention_mask=decoder_attention_mask,
+            use_causal_mask=use_causal_mask,
+            cache=cache,
+            cache_update_index=cache_update_index,
+        )
+
+        residual = decoder_sequence
+        x = self._pre_attention_layernorm(decoder_sequence)
+
+        attention_output = self._self_attention_layer(
+            hidden_states=x,
+            attention_mask=self_attention_mask,
+            cache=cache,
+            cache_update_index=cache_update_index,
+        )
+
+        if cache is None:
+            x = attention_output
+        else:
+            x, cache = attention_output
+
+        x = x + residual
+        residual = x
+        x = self._post_attention_layernorm(x)
+        x = self._mlp_intermediate_dense(x)
+        x = keras.activations.gelu(x, approximate=True)
+        x = self._mlp_output_dense(x)
+        x = self._dropout_layer(x)
+        x = x + residual
+
+        if cache is not None:
+            return x, cache
+        else:
+            return x
+
+    def _compute_attention_mask(
+        self,
+        decoder_sequence,
+        decoder_padding_mask,
+        decoder_attention_mask,
+        use_causal_mask,
+        cache,
+        cache_update_index,
+    ):
+        decoder_mask = merge_padding_and_attention_mask(
+            decoder_sequence, decoder_padding_mask, decoder_attention_mask
+        )
+        if use_causal_mask:
+            batch_size = ops.shape(decoder_sequence)[0]
+            input_length = output_length = ops.shape(decoder_sequence)[1]
+            if cache is not None:
+                input_length = ops.shape(cache)[2]
+
+            causal_mask = compute_causal_mask(
+                batch_size,
+                input_length,
+                output_length,
+                (0 if cache_update_index is None else cache_update_index),
+            )
+            return (
+                ops.minimum(decoder_mask, causal_mask)
+                if decoder_mask is not None
+                else causal_mask
+            )
+        return decoder_mask
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "num_heads": self.num_heads,
+                "intermediate_dim": self.intermediate_dim,
+                "dropout": self.dropout,
+                "layer_norm_epsilon": self.layer_norm_epsilon,
+                "kernel_initializer": keras.initializers.serialize(
+                    self.kernel_initializer
+                ),
+                "bias_initializer": keras.initializers.serialize(
+                    self.bias_initializer
+                ),
+            }
+        )
+        return config
+
+    def compute_output_shape(self, decoder_sequence_shape):
+        return decoder_sequence_shape