keras-hub-nightly 0.22.0.dev202505290412__py3-none-any.whl → 0.22.0.dev202505310408__py3-none-any.whl

keras_hub/src/models/qwen3/qwen3_decoder.py

@@ -0,0 +1,309 @@
+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.qwen3.qwen3_attention import Qwen3Attention
+from keras_hub.src.models.qwen3.qwen3_layernorm import Qwen3LayerNorm
+from keras_hub.src.utils.keras_utils import clone_initializer
+
+
+class Qwen3TransformerDecoder(keras.layers.Layer):
+    """A Transformer decoder layer for the Qwen3 backbone.
+
+    This layer implements a Transformer decoder block that includes
+    self-attention with optional sliding window attention and a feed-forward
+    network.
+
+    Args:
+    intermediate_dim: Output dimension of the first dense layer in the
+        feed-forward network.
+    num_query_heads: Number of query attention heads.
+    num_key_value_heads: Number of key/value attention 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.
+    activation: Activation function to use in the feed-forward network.
+    layer_norm_epsilon: Small float added to variance to avoid dividing
+        by zero in layer norm.
+    kernel_initializer: Initializer for the kernel weights.
+    dropout: Dropout rate for attention and hidden layers.
+    sliding_window_size: Size of the sliding window for attention when
+        enabled.
+    **kwargs: Additional keyword arguments to pass to the Layer.
+    """
+
+    def __init__(
+        self,
+        intermediate_dim,
+        num_query_heads,
+        num_key_value_heads,
+        head_dim,
+        rope_max_wavelength=10000,
+        rope_scaling_factor=1.0,
+        activation="silu",
+        layer_norm_epsilon=1e-5,
+        kernel_initializer="glorot_uniform",
+        dropout=0.0,
+        sliding_window_size=None,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.intermediate_dim = intermediate_dim
+        self.num_query_heads = num_query_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.head_dim = head_dim
+
+        self.rope_max_wavelength = rope_max_wavelength
+        self.rope_scaling_factor = rope_scaling_factor
+
+        self.dropout = dropout
+
+        self.sliding_window_size = sliding_window_size
+
+        self.activation = keras.activations.get(activation)
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.kernel_initializer = keras.initializers.get(kernel_initializer)
+
+        self.supports_masking = True
+
+    def build(self, decoder_sequence_shape):
+        self._decoder_sequence_shape = decoder_sequence_shape
+        self.hidden_dim = decoder_sequence_shape[-1]
+
+        # Self attention layer.
+        self._self_attention_layer = Qwen3Attention(
+            num_query_heads=self.num_query_heads,
+            num_key_value_heads=self.num_key_value_heads,
+            rope_max_wavelength=self.rope_max_wavelength,
+            head_dim=self.head_dim,
+            rope_scaling_factor=self.rope_scaling_factor,
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            dropout=self.dropout,
+            sliding_window_size=self.sliding_window_size,
+            dtype=self.dtype_policy,
+            name="self_attention",
+        )
+        self._self_attention_layer.build(decoder_sequence_shape)
+
+        self._self_attention_layernorm = Qwen3LayerNorm(
+            epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
+            name="self_attention_layernorm",
+        )
+
+        self._self_attention_layernorm.build(decoder_sequence_shape)
+        self._self_attention_dropout = keras.layers.Dropout(
+            rate=self.dropout,
+            dtype=self.dtype_policy,
+            name="self_attention_dropout",
+        )
+
+        # 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
+            )
+        )
+
+        self._feedforward_layernorm = Qwen3LayerNorm(
+            epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
+            name="feedforward_layernorm",
+        )
+        self._feedforward_layernorm.build(decoder_sequence_shape)
+
+        self.built = True
+
+    def call(
+        self,
+        decoder_sequence,
+        decoder_padding_mask=None,
+        decoder_attention_mask=None,
+        self_attention_cache=None,
+        self_attention_cache_update_index=None,
+        training=None,
+    ):
+        """Forward pass for the decoder layer.
+
+        Args:
+            decoder_sequence: Input tensor of shape [batch_size, seq_length,
+                hidden_size].
+            decoder_padding_mask: Mask tensor for padding tokens.
+            decoder_attention_mask: Additional attention mask.
+            self_attention_cache: Optional cached key and value tensors for
+                self-attention.
+            self_attention_cache_update_index: Index at which to update the
+                cache.
+            training: Boolean indicating whether in training mode.
+
+        Returns:
+            decoder_output: Output tensor after applying transformer decoder
+                block.
+            self_attention_cache: Updated cache tensors (if cache is provided).
+        """
+        self_attention_mask = self._compute_self_attention_mask(
+            decoder_sequence=decoder_sequence,
+            decoder_padding_mask=decoder_padding_mask,
+            decoder_attention_mask=decoder_attention_mask,
+            self_attention_cache=self_attention_cache,
+            self_attention_cache_update_index=self_attention_cache_update_index,
+        )
+        residual = decoder_sequence
+
+        x = self._self_attention_layernorm(decoder_sequence)
+
+        # Self attention block.
+        x = self._self_attention_layer(
+            hidden_states=x,
+            attention_mask=self_attention_mask,
+            cache=self_attention_cache,
+            cache_update_index=self_attention_cache_update_index,
+        )
+
+        if self_attention_cache is not None:
+            x, self_attention_cache = x
+
+        x = self._self_attention_dropout(x, training=training)
+
+        x = x + residual
+        residual = x
+
+        x = self._feedforward_layernorm(x)
+        gate_output = self._feedforward_gate_dense(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 = 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))
+
+        decoder_output = x + residual
+
+        if self_attention_cache is not None:
+            return decoder_output, self_attention_cache
+        return decoder_output
+
+    def _compute_self_attention_mask(
+        self,
+        decoder_sequence,
+        decoder_padding_mask,
+        decoder_attention_mask,
+        self_attention_cache,
+        self_attention_cache_update_index,
+    ):
+        """Computes the self-attention mask combining causal, padding and
+        attention masks.
+
+        Args:
+            decoder_sequence: Input tensor.
+            decoder_padding_mask: Mask tensor for padding tokens.
+            decoder_attention_mask: Additional attention mask.
+            self_attention_cache: Optional cached key and value tensors.
+            self_attention_cache_update_index: Index at which to update the
+                cache.
+
+        Returns:
+            Combined attention mask tensor.
+        """
+        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 self_attention_cache is not None:
+            input_length = ops.shape(self_attention_cache)[2]
+
+        cache_update_index = (
+            0
+            if self_attention_cache_update_index is None
+            else self_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):
+        """Computes the output shape of the layer.
+
+        Args:
+            decoder_sequence_shape: Shape of the decoder sequence input.
+
+        Returns:
+            Output shape, which is the same as the input shape.
+        """
+        return decoder_sequence_shape
+
+    def get_config(self):
+        """Returns the config of the layer.
+
+        Returns:
+            Dictionary containing the parameters used to initialize this layer.
+        """
+        config = super().get_config()
+        config.update(
+            {
+                "intermediate_dim": self.intermediate_dim,
+                "num_query_heads": self.num_query_heads,
+                "rope_max_wavelength": self.rope_max_wavelength,
+                "rope_scaling_factor": self.rope_scaling_factor,
+                "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,
+            }
+        )
+        return config

keras_hub/src/models/qwen3/qwen3_layernorm.py

@@ -0,0 +1,38 @@
+import keras
+from keras import ops
+
+
+class Qwen3LayerNorm(keras.layers.Layer):
+    """A normalization layer for Qwen that implements RMS normalization."""
+
+    def __init__(self, head_dim=None, epsilon=1e-6, **kwargs):
+        super().__init__(**kwargs)
+        self.head_dim = head_dim
+        self.epsilon = epsilon
+
+    def build(self, input_shape):
+        if self.head_dim:
+            dim = self.head_dim
+        else:
+            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):
+        input_dtype = x.dtype
+        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, input_dtype)
+
+    def get_config(self):
+        config = super().get_config()
+        config.update({"epsilon": self.epsilon})
+        return config

keras_hub/src/models/qwen3/qwen3_tokenizer.py

@@ -0,0 +1,48 @@
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.qwen3.qwen3_backbone import Qwen3Backbone
+from keras_hub.src.tokenizers.byte_pair_tokenizer import BytePairTokenizer
+
+
+@keras_hub_export(
+    "keras_hub.models.Qwen3Tokenizer",
+)
+class Qwen3Tokenizer(BytePairTokenizer):
+    """Tokenizer for Qwen3 models.
+
+    This tokenizer implements byte-pair encoding (BPE) for Qwen3 models,
+    handling special tokens like BOS (beginning of sequence) and EOS (end of
+    sequence).
+
+    Args:
+        vocabulary: Dictionary mapping tokens to token IDs, or path to
+            vocabulary file.
+        merges: List of BPE merges, or path to merges file.
+        bos_token: Beginning of sequence token. Defaults to None.
+        eos_token: End of sequence token. Defaults to "<|endoftext|>".
+        misc_special_tokens: Set of additional special tokens. Defaults to
+            empty set.
+    """
+
+    backbone_cls = Qwen3Backbone
+
+    def __init__(
+        self,
+        vocabulary=None,
+        merges=None,
+        **kwargs,
+    ):
+        # Add EOS token
+        eos_token = "<|im_end|>"
+        self._add_special_token(eos_token, "end_token")
+
+        pad_token = "<|endoftext|>"
+        self._add_special_token(pad_token, "pad_token")
+
+        self.start_token_id = None
+        self.start_token = None
+
+        super().__init__(
+            vocabulary=vocabulary,
+            merges=merges,
+            **kwargs,
+        )

keras_hub/src/models/qwen_moe/qwen_moe_presets.py

@@ -4,8 +4,8 @@ backbone_presets = {
     "qwen1.5_moe_2.7b_en": {
         "metadata": {
             "description": (
-                "24-layer Qwen MoE model with 2.7 billion active parameters ",
-                "and 8 experts per MoE layer.",
+                "24-layer Qwen MoE model with 2.7 billion active parameters "
+                "and 8 experts per MoE layer."
             ),
             "params": 14315784192,
             "path": "qwen-1.5-moe",

keras_hub/src/utils/transformers/convert_deit.py

@@ -0,0 +1,155 @@
+import numpy as np
+
+from keras_hub.src.models.deit.deit_backbone import DeiTBackbone
+
+backbone_cls = DeiTBackbone
+
+
+def convert_backbone_config(transformers_config):
+    image_size = transformers_config["image_size"]
+    return {
+        "image_shape": (image_size, image_size, 3),
+        "patch_size": transformers_config["patch_size"],
+        "num_layers": transformers_config["num_hidden_layers"],
+        "num_heads": transformers_config["num_attention_heads"],
+        "hidden_dim": transformers_config["hidden_size"],
+        "intermediate_dim": transformers_config["intermediate_size"],
+        "dropout_rate": transformers_config["hidden_dropout_prob"],
+        "attention_dropout": transformers_config[
+            "attention_probs_dropout_prob"
+        ],
+        "layer_norm_epsilon": transformers_config["layer_norm_eps"],
+    }
+
+
+def convert_weights(backbone, loader, transformers_config):
+    def port_ln(keras_variable, weight_key):
+        loader.port_weight(keras_variable.gamma, f"{weight_key}.weight")
+        loader.port_weight(keras_variable.beta, f"{weight_key}.bias")
+
+    def port_dense(keras_variable, weight_key):
+        loader.port_weight(
+            keras_variable.kernel,
+            f"{weight_key}.weight",
+            hook_fn=lambda x, _: x.T,
+        )
+        if keras_variable.bias is not None:
+            loader.port_weight(keras_variable.bias, f"{weight_key}.bias")
+
+    def port_mha(keras_variable, weight_key, num_heads, hidden_dim):
+        # query
+        loader.port_weight(
+            keras_variable.query_dense.kernel,
+            f"{weight_key}.attention.query.weight",
+            hook_fn=lambda x, _: np.reshape(
+                x.T, (hidden_dim, num_heads, hidden_dim // num_heads)
+            ),
+        )
+        loader.port_weight(
+            keras_variable.query_dense.bias,
+            f"{weight_key}.attention.query.bias",
+            hook_fn=lambda x, _: np.reshape(
+                x, (num_heads, hidden_dim // num_heads)
+            ),
+        )
+        # key
+        loader.port_weight(
+            keras_variable.key_dense.kernel,
+            f"{weight_key}.attention.key.weight",
+            hook_fn=lambda x, _: np.reshape(
+                x.T, (hidden_dim, num_heads, hidden_dim // num_heads)
+            ),
+        )
+        loader.port_weight(
+            keras_variable.key_dense.bias,
+            f"{weight_key}.attention.key.bias",
+            hook_fn=lambda x, _: np.reshape(
+                x, (num_heads, hidden_dim // num_heads)
+            ),
+        )
+        # value
+        loader.port_weight(
+            keras_variable.value_dense.kernel,
+            f"{weight_key}.attention.value.weight",
+            hook_fn=lambda x, _: np.reshape(
+                x.T, (hidden_dim, num_heads, hidden_dim // num_heads)
+            ),
+        )
+        loader.port_weight(
+            keras_variable.value_dense.bias,
+            f"{weight_key}.attention.value.bias",
+            hook_fn=lambda x, _: np.reshape(
+                x, (num_heads, hidden_dim // num_heads)
+            ),
+        )
+        # output
+        loader.port_weight(
+            keras_variable.output_dense.kernel,
+            f"{weight_key}.output.dense.weight",
+            hook_fn=lambda x, _: np.reshape(
+                x.T, (num_heads, hidden_dim // num_heads, hidden_dim)
+            ),
+        )
+        loader.port_weight(
+            keras_variable.output_dense.bias, f"{weight_key}.output.dense.bias"
+        )
+
+    loader.port_weight(
+        keras_variable=backbone.layers[1].patch_embedding.kernel,
+        hf_weight_key="deit.embeddings.patch_embeddings.projection.weight",
+        hook_fn=lambda x, _: np.transpose(x, (2, 3, 1, 0)),
+    )
+
+    loader.port_weight(
+        backbone.layers[1].patch_embedding.bias,
+        "deit.embeddings.patch_embeddings.projection.bias",
+    )
+
+    loader.port_weight(
+        backbone.layers[1].class_token,
+        "deit.embeddings.cls_token",
+    )
+
+    loader.port_weight(
+        backbone.layers[1].distillation_token,
+        "deit.embeddings.distillation_token",
+    )
+
+    loader.port_weight(
+        backbone.layers[1].position_embedding,
+        "deit.embeddings.position_embeddings",
+    )
+
+    encoder_layers = backbone.layers[2].encoder_layers
+    for i, encoder_block in enumerate(encoder_layers):
+        prefix = "deit.encoder.layer"
+        num_heads = encoder_block.num_heads
+        hidden_dim = encoder_block.hidden_dim
+
+        port_mha(
+            encoder_block.mha,
+            f"{prefix}.{i}.attention",
+            num_heads,
+            hidden_dim,
+        )
+        port_ln(encoder_block.layer_norm_1, f"{prefix}.{i}.layernorm_before")
+        port_ln(encoder_block.layer_norm_2, f"{prefix}.{i}.layernorm_after")
+
+        port_dense(encoder_block.mlp.dense, f"{prefix}.{i}.intermediate.dense")
+        port_dense(
+            encoder_block.output_layer.dense, f"{prefix}.{i}.output.dense"
+        )
+    port_ln(backbone.layers[2].layer_norm, "deit.layernorm")
+
+
+def convert_head(task, loader, transformers_config):
+    prefix = "cls_classifier."
+    loader.port_weight(
+        task.output_dense.kernel,
+        hf_weight_key=prefix + "weight",
+        hook_fn=lambda x, _: x.T,
+    )
+    loader.port_weight(
+        task.output_dense.bias,
+        hf_weight_key=prefix + "bias",
+    )