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

keras_hub/src/models/phi3/phi3_causal_lm.py

@@ -0,0 +1,218 @@
+# 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.causal_lm import CausalLM
+from keras_hub.src.models.phi3.phi3_backbone import Phi3Backbone
+from keras_hub.src.models.phi3.phi3_causal_lm_preprocessor import (
+    Phi3CausalLMPreprocessor,
+)
+from keras_hub.src.utils.python_utils import classproperty
+from keras_hub.src.utils.tensor_utils import any_equal
+
+
+@keras_hub_export("keras_hub.models.Phi3CausalLM")
+class Phi3CausalLM(CausalLM):
+    """An end-to-end Phi3 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 Phi-3 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, `"top_k"` sampling will be used.
+
+    Args:
+        backbone: A `keras_hub.models.Phi3Backbone` instance.
+        preprocessor: A `keras_hub.models.Phi3CausalLMPreprocessor` or `None`.
+            If `None`, this model will not apply preprocessing, and inputs
+            should be preprocessed before calling the model.
+    """
+
+    def __init__(self, backbone, preprocessor=None, **kwargs):
+        # === Layers ===
+        self.backbone = backbone
+        self.preprocessor = preprocessor
+
+        # === Functional Model ===
+        inputs = backbone.inputs
+        hidden_states = backbone(inputs)
+        outputs = backbone.token_embedding(hidden_states, reverse=True)
+        super().__init__(
+            inputs=inputs,
+            outputs=outputs,
+            **kwargs,
+        )
+
+    @classproperty
+    def backbone_cls(cls):
+        return Phi3Backbone
+
+    @classproperty
+    def preprocessor_cls(cls):
+        return Phi3CausalLMPreprocessor
+
+    def call_with_cache(
+        self,
+        token_ids,
+        cache,
+        cache_update_index,
+    ):
+        """Forward pass of `Phi3CausalLM` 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)
+        # Each decoder layer has a cache; we update them separately.
+        updated_cache = []
+        for i in range(self.backbone.num_layers):
+            current_cache = cache[:, i, ...]
+            x, next_cache = self.backbone.transformer_layers[i](
+                x,
+                attention_cache=current_cache,
+                attention_cache_update_index=cache_update_index,
+            )
+            updated_cache.append(next_cache)
+        cache = ops.stack(updated_cache, 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_key_value_heads = self.backbone.num_key_value_heads
+        head_dim = self.backbone.hidden_dim // self.backbone.num_query_heads
+        shape = [
+            batch_size,
+            num_layers,
+            2,
+            max_length,
+            num_key_value_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 the end token 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,
+        }
+
+    def generate(self, inputs, max_length=None, stop_token_ids="auto"):
+        if self.preprocessor and stop_token_ids == "auto":
+            # Stop at:
+            # `<|endoftext|>` (end of sequence token).
+            # `<|end|>` (end of turn token).
+            stop_token_ids = [self.preprocessor.tokenizer.end_token_id]
+            end_of_turn_id = self.preprocessor.tokenizer.token_to_id("<|end|>")
+            if end_of_turn_id != 0:
+                # If `<|end|>` exists in the vocabulary.
+                stop_token_ids.append(end_of_turn_id)
+
+        return super().generate(inputs, max_length, stop_token_ids)

keras_hub/src/models/phi3/phi3_causal_lm_preprocessor.py

@@ -0,0 +1,173 @@
+# 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.phi3.phi3_preprocessor import Phi3Preprocessor
+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.Phi3CausalLMPreprocessor")
+class Phi3CausalLMPreprocessor(Phi3Preprocessor):
+    """Phi3 Causal LM preprocessor.
+
+    This preprocessing layer is meant for use with
+    `keras_hub.models.Phi3CausalLM`. 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.Phi3CausalLM` 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.Phi3Tokenizer` 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. Default is `True`.
+        add_end_token: If `True`, the preprocessor will append the tokenizer
+            end token to each input sequence. Default is `False`.
+
+    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.Phi3CausalLMPreprocessor.from_preset(
+        "phi3_mini_4k_instruct_en"
+    )
+
+    # 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(
+                "`Phi3CausalLMPreprocessor` 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.
+        """
+        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/phi3/phi3_decoder.py

@@ -0,0 +1,260 @@
+# 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
+
+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.phi3.phi3_attention import Phi3Attention
+from keras_hub.src.models.phi3.phi3_layernorm import Phi3LayerNorm
+from keras_hub.src.utils.keras_utils import clone_initializer
+
+
+class Phi3Decoder(keras.layers.Layer):
+    """A Transformer decoder layer for the Phi-3 backbone."""
+
+    def __init__(
+        self,
+        hidden_dim,
+        intermediate_dim,
+        num_query_heads,
+        num_key_value_heads,
+        activation="silu",
+        layer_norm_epsilon=1e-5,
+        kernel_initializer="glorot_uniform",
+        dropout=0,
+        max_sequence_length=4096,
+        pretraining_sequence_length=4096,
+        rope_max_wavelength=10000,
+        rope_scaling_type=None,
+        rope_scaling_short_factor=None,
+        rope_scaling_long_factor=None,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.hidden_dim = hidden_dim
+        self.intermediate_dim = intermediate_dim
+        self.num_query_heads = num_query_heads
+        self.num_key_value_heads = num_key_value_heads
+
+        self.max_sequence_length = max_sequence_length
+        self.pretraining_sequence_length = pretraining_sequence_length
+        self.rope_max_wavelength = rope_max_wavelength
+        self.rope_scaling_type = rope_scaling_type
+        self.rope_scaling_short_factor = rope_scaling_short_factor
+        self.rope_scaling_long_factor = rope_scaling_long_factor
+
+        self.dropout = dropout
+
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.activation = keras.activations.get(activation)
+        self.kernel_initializer = keras.initializers.get(kernel_initializer)
+
+    def build(self, decoder_sequence_shape):
+
+        # Pre-attention layernorm.
+        self.pre_attention_layernorm = Phi3LayerNorm(
+            epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
+            name="pre_attention_layernorm",
+        )
+        self.pre_attention_layernorm.build(decoder_sequence_shape)
+
+        # Self attention layer.
+        self.attention = Phi3Attention(
+            num_query_heads=self.num_query_heads,
+            num_key_value_heads=self.num_key_value_heads,
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            dropout=self.dropout,
+            max_sequence_length=self.max_sequence_length,
+            pretraining_sequence_length=self.pretraining_sequence_length,
+            rope_max_wavelength=self.rope_max_wavelength,
+            rope_scaling_type=self.rope_scaling_type,
+            rope_scaling_short_factor=self.rope_scaling_short_factor,
+            rope_scaling_long_factor=self.rope_scaling_long_factor,
+            dtype=self.dtype_policy,
+            name="attention",
+        )
+        self.attention.build(decoder_sequence_shape)
+
+        # Post-attention layernorm.
+        self.post_attention_layernorm = Phi3LayerNorm(
+            epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
+            name="post_attention_layernorm",
+        )
+        self.post_attention_layernorm.build(decoder_sequence_shape)
+
+        # feedforward layers.
+        self.feedforward_intermediate_dense = keras.layers.Dense(
+            self.intermediate_dim,
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            use_bias=False,
+            dtype=self.dtype_policy,
+            name="feedforward_intermediate_dense",
+        )
+        self.feedforward_intermediate_dense.build(decoder_sequence_shape)
+
+        self.feedforward_gate_dense = keras.layers.Dense(
+            self.intermediate_dim,
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            use_bias=False,
+            dtype=self.dtype_policy,
+            name="feedforward_gate_dense",
+        )
+        self.feedforward_gate_dense.build(decoder_sequence_shape)
+
+        self.feedforward_output_dense = keras.layers.Dense(
+            self.hidden_dim,
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            use_bias=False,
+            dtype=self.dtype_policy,
+            name="feedforward_output_dense",
+        )
+
+        self.feedforward_output_dense.build(
+            self.feedforward_gate_dense.compute_output_shape(
+                decoder_sequence_shape
+            )
+        )
+
+        # Dropout
+        self.attention_dropout = keras.layers.Dropout(
+            rate=self.dropout,
+            dtype=self.dtype_policy,
+            name="attention_dropout",
+        )
+        self.feedforward_dropout = keras.layers.Dropout(
+            rate=self.dropout,
+            dtype=self.dtype_policy,
+            name="feedforward_dropout",
+        )
+
+        self.built = True
+
+    def call(
+        self,
+        decoder_sequence,
+        decoder_padding_mask=None,
+        decoder_attention_mask=None,
+        attention_cache=None,
+        attention_cache_update_index=None,
+    ):
+        self_attention_mask = self._compute_self_attention_mask(
+            decoder_sequence=decoder_sequence,
+            decoder_padding_mask=decoder_padding_mask,
+            decoder_attention_mask=decoder_attention_mask,
+            attention_cache=attention_cache,
+            attention_cache_update_index=attention_cache_update_index,
+        )
+        residual = decoder_sequence
+        x = self.pre_attention_layernorm(decoder_sequence)
+        x = self.attention(
+            hidden_states=x,
+            attention_mask=self_attention_mask,
+            cache=attention_cache,
+            cache_update_index=attention_cache_update_index,
+        )
+        if attention_cache is not None:
+            x, attention_cache = x
+        x = self.attention_dropout(x)
+        x = x + residual
+
+        residual = x
+        x = self.post_attention_layernorm(x)
+        # Note that we run the activation function in full 32-bit
+        # precision since this is what `torch.nn.functional.silu`
+        # does. Internally, `torch.nn.functional.silu` converts the
+        # inputs to float32, computes SiLU, and converts the outputs
+        # back to compute dtype.
+        # CPU Kernel: https://github.com/pytorch/pytorch/blob/35c493f2cf9b623bfdc7e6b34dc1cb39690a7919/aten/src/ATen/native/cpu/Activation.cpp#L1221-L1235  # noqa: E501
+        # CUDA Kernel: https://github.com/pytorch/pytorch/blob/35c493f2cf9b623bfdc7e6b34dc1cb39690a7919/aten/src/ATen/native/cuda/ActivationSiluKernel.cu  # noqa: E501
+        gate_output = self.feedforward_gate_dense(x)
+        gate_output = ops.cast(gate_output, "float32")
+        gate_output = self.activation(gate_output)
+        gate_output = ops.cast(gate_output, self.compute_dtype)
+        x = self.feedforward_intermediate_dense(x)
+        x = self.feedforward_output_dense(ops.multiply(x, gate_output))
+        x = self.feedforward_dropout(x)
+        decoder_output = x + residual
+
+        if attention_cache is not None:
+            return decoder_output, attention_cache
+        return decoder_output
+
+    def _compute_self_attention_mask(
+        self,
+        decoder_sequence,
+        decoder_padding_mask,
+        decoder_attention_mask,
+        attention_cache,
+        attention_cache_update_index,
+    ):
+        decoder_mask = merge_padding_and_attention_mask(
+            decoder_sequence, decoder_padding_mask, decoder_attention_mask
+        )
+        batch_size = ops.shape(decoder_sequence)[0]
+        input_length = output_length = ops.shape(decoder_sequence)[1]
+        # We need to handle a rectangular causal mask when doing cached
+        # decoding. For generative inference, `decoder_sequence` will
+        # generally be length 1, and `cache` will be the full generation length.
+        if attention_cache is not None:
+            input_length = ops.shape(attention_cache)[2]
+
+        cache_update_index = (
+            0
+            if attention_cache_update_index is None
+            else attention_cache_update_index
+        )
+
+        causal_mask = compute_causal_mask(
+            batch_size, input_length, output_length, cache_update_index
+        )
+
+        return (
+            ops.minimum(decoder_mask, causal_mask)
+            if decoder_mask is not None
+            else causal_mask
+        )
+
+    def compute_output_shape(self, decoder_sequence_shape):
+        return decoder_sequence_shape
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "hidden_dim": self.hidden_dim,
+                "intermediate_dim": self.intermediate_dim,
+                "num_query_heads": self.num_query_heads,
+                "num_key_value_heads": self.num_key_value_heads,
+                "activation": keras.activations.serialize(self.activation),
+                "layer_norm_epsilon": self.layer_norm_epsilon,
+                "kernel_initializer": keras.initializers.serialize(
+                    self.kernel_initializer
+                ),
+                "dropout": self.dropout,
+                "max_sequence_length": self.max_sequence_length,
+                "pretraining_sequence_length": self.pretraining_sequence_length,
+                "rope_max_wavelength": self.rope_max_wavelength,
+                "rope_scaling_type": self.rope_scaling_type,
+                "rope_scaling_short_factor": self.rope_scaling_short_factor,
+                "rope_scaling_long_factor": self.rope_scaling_long_factor,
+            }
+        )
+        return config

keras_hub/src/models/phi3/phi3_layernorm.py

@@ -0,0 +1,48 @@
+# 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
+
+
+# TODO: Deprecate this in favor of
+# `keras.layers.LayerNormalization(rms_scaling=True)` once Keras 2 support is
+# removed.
+class Phi3LayerNorm(keras.layers.Layer):
+    """A normalization layer for Phi-3 that implements RMS normalization."""
+
+    def __init__(self, epsilon=1e-6, **kwargs):
+        super().__init__(**kwargs)
+        self.epsilon = epsilon
+
+    def build(self, input_shape):
+        dim = input_shape[-1]
+        self.scale = self.add_weight(
+            name="scale",
+            trainable=True,
+            shape=(dim,),
+            initializer="ones",
+            dtype=self.variable_dtype,
+        )
+        self.built = True
+
+    def call(self, x):
+        x = ops.cast(x, "float32")
+        var = ops.mean(ops.power(x, 2), axis=-1, keepdims=True)
+        x = x * ops.rsqrt(var + self.epsilon)
+        return ops.cast(x * self.scale, self.compute_dtype)
+
+    def get_config(self):
+        config = super().get_config()
+        config.update({"epsilon": self.epsilon})
+        return config