keras-hub-nightly 0.21.0.dev202505050407__py3-none-any.whl → 0.21.0.dev202505070407__py3-none-any.whl

keras_hub/src/models/mixtral/mixtral_backbone.py

@@ -0,0 +1,207 @@
+import keras
+from keras import ops
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.layers.modeling.reversible_embedding import (
+    ReversibleEmbedding,
+)
+from keras_hub.src.models.backbone import Backbone
+from keras_hub.src.models.mixtral.mixtral_decoder import (
+    MixtralTransformerDecoder,
+)
+from keras_hub.src.models.mixtral.mixtral_layer_norm import (
+    MixtralLayerNormalization,
+)
+
+
+def _mixtral_kernel_initializer(stddev=0.02):
+    return keras.initializers.RandomNormal(stddev=stddev)
+
+
+@keras_hub_export("keras_hub.models.MixtralBackbone")
+class MixtralBackbone(Backbone):
+    """The Mixtral Transformer core architecture with hyperparameters.
+
+    This network implements a mixture of Experts based decoder network,
+    Mixtral, as described in
+    ["Mixtral of Experts"](https://arxiv.org/pdf/2401.04088).
+    It includes the embedding lookups and transformer layers.
+
+    The default constructor gives a fully customizable, randomly initialized
+    Mixtral model with any number of layers, heads, and embedding
+    dimensions. To load preset architectures and weights, use the `from_preset`
+    constructor.
+
+    Args:
+        vocabulary_size (int): The size of the token vocabulary.
+        num_layers (int): The number of transformer layers.
+        num_query_heads (int): The number of query attention heads for
+            each transformer.
+        hidden_dim (int): The size of the transformer encoding and pooling
+            layers.
+        intermediate_dim (int): The output dimension of the first Dense layer
+            in a three-layer feedforward network for each transformer.
+        num_key_value_heads (int): The number of key and value attention heads
+            for each transformer.
+        rope_max_wavelength (int, optional): The maximum angular wavelength of
+            the sine/cosine curves, for rotary embeddings. Defaults to `10000`.
+        rope_scaling_factor (float, optional): The scaling factor for
+            calculation of roatary embedding. Defaults to `1.0`.
+        layer_norm_epsilon (float, optional): Epsilon for the layer
+            normalization layers in the transformer decoder. Defaults to `1e-6`.
+        sliding_window (int, optional): The sliding window for the mixtral
+            attention layers. This controls the maximum cache size for the
+            attention layers in each transformer decoder. Only `sliding_window`
+            number of tokens are saved in the cache and used to generate the
+            next token. Defaults to `512`.
+        dtype: string or `keras.mixed_precision.DTypePolicy`. The dtype to use
+            for model computations and weights. Note that some computations,
+            such as softmax and layer normalization, will always be done at
+            float32 precision regardless of dtype.
+
+    Examples:
+
+    ```python
+    input_data = {
+        "token_ids": np.ones(shape=(1, 12), dtype="int32"),
+        "padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]),
+    }
+
+    # Pretrained Mixtral decoder.
+    model = keras_hub.models.MixtralBackbone.from_preset("mixtral7b_base_en")
+    model(input_data)
+
+    # Randomly initialized Mixtral decoder with custom config.
+    model = keras_hub.models.MixtralBackbone(
+        vocabulary_size=10,
+        hidden_dim=512,
+        num_layers=2,
+        num_query_heads=32,
+        num_key_value_heads=8,
+        intermediate_dim=1024,
+        sliding_window=512,
+        layer_norm_epsilon=1e-6,
+        dtype="float32"
+    )
+    model(input_data)
+    ```
+    """
+
+    def __init__(
+        self,
+        vocabulary_size,
+        num_layers,
+        num_query_heads,
+        hidden_dim,
+        intermediate_dim,
+        num_key_value_heads,
+        num_experts,
+        top_k=2,
+        router_jitter_noise=0.0,
+        rope_max_wavelength=10000,
+        rope_scaling_factor=1.0,
+        layer_norm_epsilon=1e-6,
+        router_aux_loss_coef=0.02,
+        sliding_window=512,
+        dropout=0,
+        dtype=None,
+        output_router_logits=False,
+        **kwargs,
+    ):
+        # === Layers ===
+        self.token_embedding = ReversibleEmbedding(
+            input_dim=vocabulary_size,
+            output_dim=hidden_dim,
+            tie_weights=False,
+            embeddings_initializer=_mixtral_kernel_initializer(stddev=0.01),
+            dtype=dtype,
+            name="token_embedding",
+        )
+        self.transformer_layers = []
+        for i in range(num_layers):
+            layer = MixtralTransformerDecoder(
+                intermediate_dim=intermediate_dim,
+                num_query_heads=num_query_heads,
+                num_key_value_heads=num_key_value_heads,
+                num_experts=num_experts,
+                top_k=top_k,
+                router_jitter_noise=router_jitter_noise,
+                output_router_logits=output_router_logits,
+                rope_max_wavelength=rope_max_wavelength,
+                rope_scaling_factor=rope_scaling_factor,
+                layer_norm_epsilon=layer_norm_epsilon,
+                activation=ops.silu,
+                router_aux_loss_coef=router_aux_loss_coef,
+                kernel_initializer=_mixtral_kernel_initializer(stddev=0.02),
+                sliding_window=sliding_window,
+                dropout=dropout,
+                dtype=dtype,
+                name=f"transformer_layer_{i}",
+            )
+            self.transformer_layers.append(layer)
+        self.layer_norm = MixtralLayerNormalization(
+            epsilon=layer_norm_epsilon,
+            dtype=dtype,
+            name="sequence_output_layernorm",
+        )
+
+        # === Functional Model ===
+        token_id_input = keras.Input(
+            shape=(None,), dtype="int32", name="token_ids"
+        )
+        padding_mask_input = keras.Input(
+            shape=(None,), dtype="int32", name="padding_mask"
+        )
+        x = self.token_embedding(token_id_input)
+        for transformer_layer in self.transformer_layers:
+            x = transformer_layer(x, decoder_padding_mask=padding_mask_input)
+        sequence_output = self.layer_norm(x)
+        super().__init__(
+            inputs={
+                "token_ids": token_id_input,
+                "padding_mask": padding_mask_input,
+            },
+            outputs=sequence_output,
+            dtype=dtype,
+            **kwargs,
+        )
+
+        # === Config ===
+        self.vocabulary_size = vocabulary_size
+        self.num_layers = num_layers
+        self.num_query_heads = num_query_heads
+        self.hidden_dim = hidden_dim
+        self.intermediate_dim = intermediate_dim
+        self.num_key_value_heads = num_key_value_heads
+        self.num_experts = num_experts
+        self.top_k = top_k
+        self.router_jitter_noise = router_jitter_noise
+        self.rope_max_wavelength = rope_max_wavelength
+        self.router_aux_loss_coef = router_aux_loss_coef
+        self.rope_scaling_factor = rope_scaling_factor
+        self.sliding_window = sliding_window
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.dropout = dropout
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "vocabulary_size": self.vocabulary_size,
+                "num_layers": self.num_layers,
+                "num_query_heads": self.num_query_heads,
+                "hidden_dim": self.hidden_dim,
+                "intermediate_dim": self.intermediate_dim,
+                "num_experts": self.num_experts,
+                "top_k": self.top_k,
+                "router_jitter_noise": self.router_jitter_noise,
+                "rope_max_wavelength": self.rope_max_wavelength,
+                "rope_scaling_factor": self.rope_scaling_factor,
+                "num_key_value_heads": self.num_key_value_heads,
+                "router_aux_loss_coef": self.router_aux_loss_coef,
+                "sliding_window": self.sliding_window,
+                "layer_norm_epsilon": self.layer_norm_epsilon,
+                "dropout": self.dropout,
+            }
+        )
+        return config

keras_hub/src/models/mixtral/mixtral_causal_lm.py

@@ -0,0 +1,281 @@
+import keras
+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.mixtral.mixtral_backbone import MixtralBackbone
+from keras_hub.src.models.mixtral.mixtral_causal_lm_preprocessor import (
+    MixtralCausalLMPreprocessor,
+)
+from keras_hub.src.utils.tensor_utils import any_equal
+
+
+@keras_hub_export("keras_hub.models.MixtralCausalLM")
+class MixtralCausalLM(CausalLM):
+    """An end-to-end Mixtral 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 GPT-NeoX 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.MixtralBackbone` instance.
+        preprocessor: A `keras_hub.models.MixtralCausalLMPreprocessor` or
+            `None`. If `None`, this model will not apply preprocessing, and
+            inputs should be preprocessed before calling the model.
+    """
+
+    backbone_cls = MixtralBackbone
+    preprocessor_cls = MixtralCausalLMPreprocessor
+
+    def __init__(self, backbone, preprocessor=None, **kwargs):
+        # === Layers ===
+        self.backbone = backbone
+        self.preprocessor = preprocessor
+
+        # === Functional Model ===
+        # This must be "backbone.input" i.e. the full input structure,
+        # rather than "backbone.inputs" which is the flattened list of inputs.
+        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 `MixtralCausalLM` 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,
+                self_attention_cache=current_cache,
+                self_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: List 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_tokens 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 score(
+        self,
+        token_ids,
+        padding_mask=None,
+        scoring_mode="logits",
+        layer_intercept_fn=None,
+        target_ids=None,
+    ):
+        """Score a generation represented by the provided token ids.
+
+        Args:
+            token_ids: A <int>[batch_size, num_tokens] tensor containing tokens
+                to score. Typically, this tensor captures the output from a call
+                to `MixtralCausalLM.generate()`, i.e., tokens for both the input
+                text and the model-generated text.
+            padding_mask: A <bool>[batch_size, num_tokens] tensor indicating the
+                tokens that should be preserved during generation. This is an
+                artifact required by the MixtralBackbone and isn't influential
+                on the computation of this function. If omitted, this function
+                uses `keras.ops.ones()` to create a tensor of the appropriate
+                shape.
+            scoring_mode: The type of scores to return, either "logits" or
+                "loss", both will be per input token.
+            layer_intercept_fn: An optional function for augmenting activations
+                with additional computation, for example, as part of
+                interpretability research. This function will be passed the
+                activations as its first parameter and a numeric index
+                associated with that backbone layer. _This index _is not_ an
+                index into `self.backbone.layers`. The index -1 accompanies the
+                embeddings returned by calling `self.backbone.token_embedding()`
+                on `token_ids` in the forward direction. All subsequent indexes
+                will be 0-based indices for the activations returned by each of
+                the Transformers layers in the backbone. This function must
+                return a <float>[batch_size, num_tokens, hidden_dims] tensor
+                that can be passed as an input to the next layer in the model.
+            target_ids: An <bool>[batch_size, num_tokens] tensor containing the
+                predicted tokens against which the loss should be computed. If a
+                span of tokens is provided (sequential truthy values along
+                axis=1 in the tensor), the loss will be computed as the
+                aggregate across those tokens.
+
+        Raises:
+            ValueError: If an unsupported scoring_mode is provided, or if the
+                target_ids are not provided when using ScoringMode.LOSS.
+
+        Returns:
+            The per-token scores as a tensor of size
+            <float>[batch_size, num_tokens, vocab_size] in "logits" mode, or
+            <float>[batch_size, num_tokens] in "loss" mode.
+        ```
+        """
+        if scoring_mode not in ("logits", "loss"):
+            raise ValueError(
+                "Unsupported scoring_mode. Must be one of 'logits' or 'loss'."
+            )
+
+        if scoring_mode == "loss" and target_ids is None:
+            raise ValueError(
+                "Cannot compute loss without targets. Please provide target "
+                "token ids via the target_ids parameter."
+            )
+
+        batch_shape = ops.shape(token_ids)[:2]
+        assert len(batch_shape) == 2
+
+        if layer_intercept_fn is None:
+
+            def default_layer_intercept_fn(x, unused_i):
+                return x
+
+            layer_intercept_fn = default_layer_intercept_fn
+
+        token_embeddings = self.backbone.token_embedding(token_ids)
+        x = layer_intercept_fn(token_embeddings, -1)
+
+        for i, transformer_layer in enumerate(self.backbone.transformer_layers):
+            x = transformer_layer(x, decoder_padding_mask=padding_mask)
+            x = layer_intercept_fn(x, i)
+
+        x = self.backbone.layer_norm(x)
+        logits = self.backbone.token_embedding(x, reverse=True)
+
+        if scoring_mode == "logits":
+            return logits
+
+        per_token_loss_fn = keras.losses.SparseCategoricalCrossentropy(
+            from_logits=True, reduction="none"
+        )
+        per_token_loss = per_token_loss_fn(target_ids, logits)
+        return per_token_loss

keras_hub/src/models/mixtral/mixtral_causal_lm_preprocessor.py

@@ -0,0 +1,76 @@
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.causal_lm_preprocessor import CausalLMPreprocessor
+from keras_hub.src.models.mixtral.mixtral_backbone import MixtralBackbone
+from keras_hub.src.models.mixtral.mixtral_tokenizer import MixtralTokenizer
+
+
+@keras_hub_export("keras_hub.models.MixtralCausalLMPreprocessor")
+class MixtralCausalLMPreprocessor(CausalLMPreprocessor):
+    """Mixtral Causal LM preprocessor.
+
+    This preprocessing layer is meant for use with
+    `keras_hub.models.MixtralCausalLM`. 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.MixtralCausalLM` 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.MixtralTokenizer` 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.MixtralCausalLMPreprocessor.from_preset(
+        "mixtral_base_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)
+    ```
+    """
+
+    backbone_cls = MixtralBackbone
+    tokenizer_cls = MixtralTokenizer