keras-hub 0.20.0.dev1__py3-none-any.whl → 0.21.0.dev1__py3-none-any.whl

keras_hub/src/models/mixtral/mixtral_attention.py

@@ -0,0 +1,252 @@
+import inspect
+import math
+
+import keras
+from keras import ops
+
+from keras_hub.src.layers.modeling.rotary_embedding import RotaryEmbedding
+from keras_hub.src.utils.keras_utils import clone_initializer
+from keras_hub.src.utils.keras_utils import fused_attention_op_available
+from keras_hub.src.utils.keras_utils import gpu_supports_fused_attention_op
+from keras_hub.src.utils.keras_utils import running_on_gpu
+from keras_hub.src.utils.keras_utils import running_on_tpu
+
+
+class CachedMixtralAttention(keras.layers.Layer):
+    """A cached grounded query attention layer with sliding window."""
+
+    def __init__(
+        self,
+        num_query_heads,
+        num_key_value_heads,
+        rope_max_wavelength=10000,
+        rope_scaling_factor=1.0,
+        kernel_initializer="glorot_uniform",
+        sliding_window=512,
+        dropout=0,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.num_query_heads = num_query_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.sliding_window = sliding_window
+        self.dropout = dropout
+
+        self.num_key_value_groups = num_query_heads // num_key_value_heads
+        self.rope_max_wavelength = rope_max_wavelength
+
+        self._kernel_initializer = keras.initializers.get(
+            clone_initializer(kernel_initializer)
+        )
+
+        self.rope_scaling_factor = rope_scaling_factor
+
+    def build(self, inputs_shape):
+        # Einsum variables:
+        # b = batch size
+        # q = query length
+        # k = key/value length
+        # m = model dim
+        # u = num query heads
+        # v = num key/value heads
+        # h = head dim
+        self._hidden_dim = inputs_shape[-1]
+        self._head_dim = self._hidden_dim // self.num_query_heads
+        self._inv_norm_factor = 1.0 / math.sqrt(self._head_dim)
+
+        self.query_dense = keras.layers.EinsumDense(
+            equation="bqm,muh->bquh",
+            output_shape=(None, self.num_query_heads, self._head_dim),
+            kernel_initializer=self._kernel_initializer,
+            dtype=self.dtype_policy,
+            name="query",
+        )
+        self.query_dense.build(inputs_shape)
+
+        self.key_dense = keras.layers.EinsumDense(
+            equation="bkm,mvh->bkvh",
+            output_shape=(
+                None,
+                self.num_key_value_heads,
+                self._head_dim,
+            ),
+            kernel_initializer=self._kernel_initializer,
+            dtype=self.dtype_policy,
+            name="key",
+        )
+        self.key_dense.build(inputs_shape)
+
+        self.value_dense = keras.layers.EinsumDense(
+            equation="bkm,mvh->bkvh",
+            output_shape=(
+                None,
+                self.num_key_value_heads,
+                self._head_dim,
+            ),
+            kernel_initializer=self._kernel_initializer,
+            dtype=self.dtype_policy,
+            name="value",
+        )
+        self.value_dense.build(inputs_shape)
+
+        self.softmax = keras.layers.Softmax(
+            axis=-1,
+            dtype="float32",
+            name="attention_softmax",
+        )
+
+        self.dropout_layer = keras.layers.Dropout(
+            rate=self.dropout,
+            dtype=self.dtype_policy,
+        )
+
+        self.output_dense = keras.layers.EinsumDense(
+            equation="bquh,uhm->bqm",
+            output_shape=(None, self._hidden_dim),
+            kernel_initializer=self._kernel_initializer,
+            dtype=self.dtype_policy,
+            name="attention_output",
+        )
+        self.output_dense.build(
+            (None, None, self.num_query_heads, self._head_dim)
+        )
+
+        self.rotary_embedding_layer = RotaryEmbedding(
+            max_wavelength=self.rope_max_wavelength,
+            scaling_factor=self.rope_scaling_factor,
+            dtype=self.dtype_policy,
+        )
+
+        self._dot_product_equation = "bquh,bkuh->buqk"
+        self._combine_equation = "buqk,bkuh->bquh"
+
+        self.built = True
+
+    def call(
+        self,
+        hidden_states,
+        attention_mask=None,
+        cache=None,
+        cache_update_index=None,
+        training=None,
+    ):
+        start_index = (
+            cache_update_index if cache_update_index is not None else 0
+        )
+
+        query = self.query_dense(hidden_states)
+
+        # Compute RoPE for queries
+        query = self.rotary_embedding_layer(query, start_index=start_index)
+
+        def _compute_key_value(x):
+            key, value = self.key_dense(x), self.value_dense(x)
+            # Compute RoPE for keys
+            key = self.rotary_embedding_layer(key, start_index=start_index)
+            return key, value
+
+        if cache is not None:
+            key_cache = cache[:, 0, ...]
+            value_cache = cache[:, 1, ...]
+            if cache_update_index is None:
+                key = key_cache
+                value = value_cache
+            else:
+                key_update, value_update = _compute_key_value(hidden_states)
+                start = [0, cache_update_index, 0, 0]
+                key = ops.slice_update(key_cache, start, key_update)
+                value = ops.slice_update(value_cache, start, value_update)
+                cache = ops.stack((key, value), axis=1)
+        else:
+            if cache_update_index is not None:
+                raise ValueError(
+                    "`cache_update_index` should not be set if `cache` is "
+                    f"`None`. Received: cache={cache}, "
+                    f"cache_update_index={cache_update_index}"
+                )
+            key, value = _compute_key_value(hidden_states)
+
+        # [batch_shape, seq_len, num_key_value_heads, head_dim]
+        # -> [batch_shape, seq_len, num_heads, head_dim]
+        key = ops.repeat(key, repeats=self.num_key_value_groups, axis=2)
+        value = ops.repeat(value, repeats=self.num_key_value_groups, axis=2)
+
+        attention_output = self._compute_attention(
+            query, key, value, attention_mask
+        )
+
+        attention_output = self.dropout_layer(
+            attention_output, training=training
+        )
+
+        attention_output = self.output_dense(attention_output)
+
+        if cache is not None:
+            return attention_output, cache
+        return attention_output
+
+    def _masked_softmax(self, attention_scores, attention_mask=None):
+        if attention_mask is not None:
+            return self.softmax(attention_scores, attention_mask[:, None, :, :])
+        return self.softmax(attention_scores)
+
+    def _use_fused_attention_op(self):
+        if not fused_attention_op_available():
+            return False
+        if self.dropout > 0.0:
+            return False
+        if running_on_gpu():
+            return gpu_supports_fused_attention_op()
+        elif running_on_tpu():
+            # TPU supports softcap with on keras >= 3.10.
+            sig = inspect.signature(ops.dot_product_attention)
+            return "attn_logits_soft_cap" in sig.parameters
+        else:
+            return False
+
+    def _compute_attention(self, query, key, value, attention_mask=None):
+        if self._use_fused_attention_op():
+            if attention_mask is not None:
+                attention_mask = ops.expand_dims(attention_mask, axis=1)
+                attention_mask = ops.cast(attention_mask, dtype="bool")
+
+            attention_output = ops.dot_product_attention(
+                query,
+                key,
+                value,
+                mask=attention_mask,
+                scale=self._inv_norm_factor,
+            )
+            return attention_output
+
+        attention_scores = ops.einsum(self._dot_product_equation, query, key)
+        attention_scores = ops.multiply(
+            attention_scores,
+            ops.cast(self._inv_norm_factor, self.compute_dtype),
+        )
+        attention_scores = self._masked_softmax(
+            attention_scores, attention_mask
+        )
+        attention_scores = ops.cast(attention_scores, self.compute_dtype)
+        attention_output = ops.einsum(
+            self._combine_equation, attention_scores, value
+        )
+
+        return attention_output
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "num_query_heads": self.num_query_heads,
+                "num_key_value_heads": self.num_key_value_heads,
+                "rope_max_wavelength": self.rope_max_wavelength,
+                "rope_scaling_factor": self.rope_scaling_factor,
+                "kernel_initializer": keras.initializers.serialize(
+                    self._kernel_initializer
+                ),
+                "sliding_window": self.sliding_window,
+                "dropout": self.dropout,
+            }
+        )
+        return config

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