keras-hub-nightly 0.21.0.dev202505280410__py3-none-any.whl → 0.22.0.dev202505300409__py3-none-any.whl

keras_hub/models/__init__.py

@@ -444,6 +444,15 @@ from keras_hub.src.models.qwen.qwen_tokenizer import (
 from keras_hub.src.models.qwen.qwen_tokenizer import (
     QwenTokenizer as QwenTokenizer,
 )
+from keras_hub.src.models.qwen3.qwen3_backbone import (
+    Qwen3Backbone as Qwen3Backbone,
+)
+from keras_hub.src.models.qwen3.qwen3_causal_lm_preprocessor import (
+    Qwen3CausalLMPreprocessor as Qwen3CausalLMPreprocessor,
+)
+from keras_hub.src.models.qwen3.qwen3_tokenizer import (
+    Qwen3Tokenizer as Qwen3Tokenizer,
+)
 from keras_hub.src.models.qwen_moe.qwen_moe_backbone import (
     QwenMoeBackbone as QwenMoeBackbone,
 )

keras_hub/src/models/mixtral/mixtral_presets.py

@@ -4,8 +4,8 @@ backbone_presets = {
     "mixtral_8_7b_en": {
         "metadata": {
             "description": (
-                "32-layer Mixtral MoE model with 7 billion",
-                "active parameters and 8 experts per MoE layer.",
+                "32-layer Mixtral MoE model with 7 billion"
+                "active parameters and 8 experts per MoE layer."
             ),
             "params": 46702792704,
             "path": "mixtral",
@@ -15,8 +15,8 @@ backbone_presets = {
     "mixtral_8_instruct_7b_en": {
         "metadata": {
             "description": (
-                "Instruction fine-tuned 32-layer Mixtral MoE model",
-                "with 7 billion active parameters and 8 experts per MoE layer.",
+                "Instruction fine-tuned 32-layer Mixtral MoE model"
+                "with 7 billion active parameters and 8 experts per MoE layer."
             ),
             "params": 46702792704,
             "path": "mixtral",

keras_hub/src/models/qwen/qwen_presets.py

@@ -28,8 +28,8 @@ backbone_presets = {
     "qwen2.5_instruct_0.5b_en": {
         "metadata": {
             "description": (
-                "Instruction fine-tuned 24-layer Qwen model with 0.5 ",
-                "billion parameters.",
+                "Instruction fine-tuned 24-layer Qwen model with 0.5 "
+                "billion parameters."
             ),
             "params": 494032768,
             "path": "qwen",
@@ -39,8 +39,8 @@ backbone_presets = {
     "qwen2.5_instruct_32b_en": {
         "metadata": {
             "description": (
-                "Instruction fine-tuned 64-layer Qwen model with 32 ",
-                "billion parameters.",
+                "Instruction fine-tuned 64-layer Qwen model with 32 "
+                "billion parameters."
             ),
             "params": 32763876352,
             "path": "qwen",
@@ -50,8 +50,8 @@ backbone_presets = {
     "qwen2.5_instruct_72b_en": {
         "metadata": {
             "description": (
-                "Instruction fine-tuned 80-layer Qwen model with 72 ",
-                "billion parameters.",
+                "Instruction fine-tuned 80-layer Qwen model with 72 "
+                "billion parameters."
             ),
             "params": 72706203648,
             "path": "qwen",

keras_hub/src/models/qwen3/qwen3_attention.py

@@ -0,0 +1,369 @@
+import math
+
+import keras
+from keras import ops
+
+from keras_hub.src.layers.modeling.rotary_embedding import RotaryEmbedding
+from keras_hub.src.models.qwen3.qwen3_layernorm import Qwen3LayerNorm
+from keras_hub.src.utils.keras_utils import clone_initializer
+from keras_hub.src.utils.keras_utils import fused_attention_op_available
+
+
+class Qwen3Attention(keras.layers.Layer):
+    """A multi-head attention layer for Qwen3 models
+
+    This attention implementation supports grouped-query attention (GQA) where
+    the number of key-value heads can be less than the number of query heads.
+
+    Args:
+        num_query_heads: Number of query heads.
+        num_key_value_heads: Number of key/value heads (for GQA).
+        rope_max_wavelength: Maximum wavelength for RoPE (Rotary Position
+            Embedding).
+        rope_scaling_factor: Scaling factor for RoPE, used for extending
+            context length.
+        kernel_initializer: Initializer for the kernel weights.
+        dropout: Dropout rate for attention weights.
+        sliding_window_size: Size of the sliding window for attention.
+        **kwargs: Additional keyword arguments to pass to the Layer.
+    """
+
+    def __init__(
+        self,
+        num_query_heads,
+        num_key_value_heads,
+        head_dim,
+        rope_max_wavelength=10000,
+        rope_scaling_factor=1,
+        kernel_initializer="glorot_uniform",
+        dropout=0.0,
+        layer_norm_epsilon=1e-5,
+        sliding_window_size=None,
+        **kwargs,
+    ):
+        super().__init__(
+            **kwargs,
+        )
+        self.num_query_heads = num_query_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.head_dim = head_dim
+        self.dropout = dropout
+
+        self.layer_norm_epsilon = layer_norm_epsilon
+
+        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
+        self.sliding_window_size = sliding_window_size
+
+    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
+        hidden_dim = inputs_shape[-1]
+        if not self.head_dim:
+            self.head_dim = 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._query_dense_layer_norm = Qwen3LayerNorm(
+            epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
+            head_dim=self.head_dim,
+            name="query_dense_layernorm",
+        )
+        self._query_dense_layer_norm.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._key_dense_layer_norm = Qwen3LayerNorm(
+            epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
+            head_dim=self.head_dim,
+            name="key_dense_layernorm",
+        )
+        self._key_dense_layer_norm.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, 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,
+    ):
+        """Applies attention mechanism to the input hidden states.
+
+        Args:
+            hidden_states: Input tensor of shape [batch_size, seq_length,
+                hidden_size].
+            attention_mask: Mask tensor of shape [batch_size, seq_length,
+                seq_length].
+            cache: Optional cached key and value tensors.
+            cache_update_index: Index at which to update the cache.
+            training: Boolean indicating whether in training mode.
+
+        Returns:
+            attention_output: Output tensor after applying attention.
+            cache: Updated cache tensors (if cache is provided).
+        """
+        start_index = (
+            cache_update_index if cache_update_index is not None else 0
+        )
+
+        query = self._query_dense(hidden_states)
+        query = self._query_dense_layer_norm(query)
+
+        # Compute RoPE for queries
+        query = self.rotary_embedding_layer(query, start_index=start_index)
+
+        def _compute_key_value(x):
+            key = self._key_dense(x)
+            key = self._key_dense_layer_norm(key)
+            key = self.rotary_embedding_layer(key, start_index=start_index)
+
+            value = self._value_dense(x)
+
+            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,
+            cache_update_index=cache_update_index,
+        )
+
+        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):
+        """Applies softmax with optional masking.
+
+        Args:
+            attention_scores: Attention score tensor.
+            attention_mask: Optional mask tensor.
+
+        Returns:
+            Masked softmax attention weights.
+        """
+        if attention_mask is not None:
+            return self._softmax(
+                attention_scores, attention_mask[:, None, :, :]
+            )
+        return self._softmax(attention_scores)
+
+    def _compute_attention(
+        self, query, key, value, attention_mask=None, cache_update_index=None
+    ):
+        """Computes attention using query, key, and value tensors.
+
+        Uses Flash Attention when available for better performance.
+
+        Args:
+            query: Query tensor.
+            key: Key tensor.
+            value: Value tensor.
+            attention_mask: Optional mask tensor.
+            cache_update_index: Index for sliding window computation.
+
+        Returns:
+            attention_output: Output tensor after applying attention.
+        """
+        if fused_attention_op_available():
+            # Use `dot_product_attention` with Flash Attention support if
+            # available.
+            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),
+        )
+        if not self.sliding_window_size:
+            attention_mask = self._mask_sliding_window(
+                attention_mask,
+                cache_update_index=cache_update_index
+                if cache_update_index
+                else 0,
+            )
+        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 _mask_sliding_window(
+        self,
+        attention_mask,
+        cache_update_index=0,
+    ):
+        """Creates and combines a sliding window mask with the attention mask.
+
+        Args:
+            attention_mask: Original attention mask.
+            cache_update_index: Starting index for the sliding window.
+
+        Returns:
+            Combined attention mask with sliding window constraints.
+        """
+        _, query_len, key_len = ops.shape(attention_mask)
+        # Compute the sliding window for square attention.
+        all_ones = ops.ones((key_len, key_len), "bool")
+        if keras.config.backend() == "tensorflow":
+            # TODO: trui/tril has issues with dynamic shape on the tensorflow
+            # backend. We should fix, but use `band_part` for now.
+            import tensorflow as tf
+
+            band_size = ops.minimum(key_len, self.sliding_window_size - 1)
+            band_size = ops.cast(band_size, "int32")
+            sliding_mask = tf.linalg.band_part(all_ones, band_size, band_size)
+        else:
+            sliding_mask = ops.triu(
+                all_ones, -1 * self.sliding_window_size + 1
+            ) * ops.tril(all_ones, self.sliding_window_size - 1)
+        # Slice the window for short queries during generation.
+        start = (cache_update_index, 0)
+        sliding_mask = ops.slice(sliding_mask, start, (query_len, key_len))
+        sliding_mask = ops.expand_dims(sliding_mask, 0)
+        return ops.logical_and(attention_mask, ops.cast(sliding_mask, "bool"))
+
+    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
+                ),
+                "dropout": self.dropout,
+                "sliding_window_size": self.sliding_window_size,
+            }
+        )
+        return config

keras_hub/src/models/qwen3/qwen3_backbone.py

@@ -0,0 +1,191 @@
+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.qwen3.qwen3_decoder import Qwen3TransformerDecoder
+from keras_hub.src.models.qwen3.qwen3_layernorm import Qwen3LayerNorm
+
+
+def _qwen3_kernel_initializer(stddev=0.02):
+    return keras.initializers.RandomNormal(stddev=stddev)
+
+
+@keras_hub_export("keras_hub.models.Qwen3Backbone")
+class Qwen3Backbone(Backbone):
+    """The Qwen3 Transformer core architecture with hyperparameters.
+
+    This network implements a Transformer-based decoder network,
+    Qwen3, as described in the Qwen3 model architecture.
+    It includes the embedding lookups and transformer layers.
+
+    The default constructor gives a fully customizable, randomly initialized
+    Qwen3 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 rotary embedding. Defaults to `1.0`.
+        layer_norm_epsilon (float, optional): Epsilon for the layer
+            normalization layers in the transformer decoder. Defaults to `1e-6`.
+        dropout (float, optional): Dropout rate for attention and hidden layers.
+            Defaults to `0`.
+        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.
+        tie_word_embeddings (bool, optional): Whether to tie input and output
+            embeddings. Defaults to `True`.
+        sliding_window_size (int, optional): Size of the sliding window for
+            attention when enabled. Defaults to `32768`.
+
+    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 Qwen3 decoder.
+    model = keras_hub.models.Qwen3Backbone.from_preset("qwen32.5_0.5b_en")
+    model(input_data)
+
+    # Randomly initialized Qwen3 decoder with custom config.
+    model = keras_hub.models.Qwen3Backbone(
+        vocabulary_size=10,
+        hidden_dim=512,
+        num_layers=2,
+        num_query_heads=32,
+        num_key_value_heads=8,
+        intermediate_dim=1024,
+        layer_norm_epsilon=1e-6,
+        dtype="float32"
+    )
+    model(input_data)
+    ```
+    """
+
+    def __init__(
+        self,
+        vocabulary_size,
+        num_layers,
+        num_query_heads,
+        num_key_value_heads,
+        head_dim,
+        hidden_dim,
+        intermediate_dim,
+        rope_max_wavelength=10000,
+        rope_scaling_factor=1.0,
+        layer_norm_epsilon=1e-6,
+        dropout=0.0,
+        tie_word_embeddings=True,
+        sliding_window_size=32768,
+        dtype=None,
+        **kwargs,
+    ):
+        # === Layers ===
+        self.token_embedding = ReversibleEmbedding(
+            input_dim=vocabulary_size,
+            output_dim=hidden_dim,
+            tie_weights=tie_word_embeddings,
+            embeddings_initializer=_qwen3_kernel_initializer(stddev=0.01),
+            dtype=dtype,
+            name="token_embedding",
+        )
+        self.transformer_layers = []
+        for i in range(num_layers):
+            layer = Qwen3TransformerDecoder(
+                intermediate_dim=intermediate_dim,
+                head_dim=head_dim,
+                num_query_heads=num_query_heads,
+                num_key_value_heads=num_key_value_heads,
+                rope_max_wavelength=rope_max_wavelength,
+                rope_scaling_factor=rope_scaling_factor,
+                layer_norm_epsilon=layer_norm_epsilon,
+                activation=ops.silu,
+                kernel_initializer=_qwen3_kernel_initializer(stddev=0.02),
+                dropout=dropout,
+                sliding_window_size=sliding_window_size,
+                dtype=dtype,
+                name=f"transformer_layer_{i}",
+            )
+            self.transformer_layers.append(layer)
+        self.layer_norm = Qwen3LayerNorm(
+            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.head_dim = head_dim
+        self.intermediate_dim = intermediate_dim
+        self.rope_max_wavelength = rope_max_wavelength
+        self.num_key_value_heads = num_key_value_heads
+        self.rope_scaling_factor = rope_scaling_factor
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.dropout = dropout
+        self.tie_word_embeddings = tie_word_embeddings
+        self.sliding_window_size = sliding_window_size
+
+    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,
+                "head_dim": self.head_dim,
+                "intermediate_dim": self.intermediate_dim,
+                "rope_max_wavelength": self.rope_max_wavelength,
+                "rope_scaling_factor": self.rope_scaling_factor,
+                "num_key_value_heads": self.num_key_value_heads,
+                "layer_norm_epsilon": self.layer_norm_epsilon,
+                "dropout": self.dropout,
+                "tie_word_embeddings": self.tie_word_embeddings,
+                "sliding_window_size": self.sliding_window_size,
+            }
+        )
+        return config

keras_hub/src/models/qwen3/qwen3_causal_lm_preprocessor.py

@@ -0,0 +1,10 @@
+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.qwen3.qwen3_backbone import Qwen3Backbone
+from keras_hub.src.models.qwen3.qwen3_tokenizer import Qwen3Tokenizer
+
+
+@keras_hub_export("keras_hub.models.Qwen3CausalLMPreprocessor")
+class Qwen3CausalLMPreprocessor(CausalLMPreprocessor):
+    backbone_cls = Qwen3Backbone
+    tokenizer_cls = Qwen3Tokenizer