keras-hub-nightly 0.21.0.dev202505280410__py3-none-any.whl → 0.22.0.dev202505300409__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/models/vit/vit_backbone.py

@@ -18,10 +18,10 @@ class ViTBackbone(Backbone):
 
     Args:
         image_shape: A tuple or list of 3 integers representing the shape of the
-            input image `(height, width, channels)`, `height` and `width` must
-            be equal.
-        patch_size: int. The size of each image patch, the input image will be
-            divided into patches of shape `(patch_size, patch_size)`.
+            input image `(height, width, channels)`.
+        patch_size: int or (int, int). The size of each image patch, the input
+            image will be divided into patches of shape
+            `(patch_size_h, patch_size_w)`.
         num_layers: int. The number of transformer encoder layers.
         num_heads: int. specifying the number of attention heads in each
             Transformer encoder layer.
@@ -37,6 +37,10 @@ class ViTBackbone(Backbone):
         use_mha_bias: bool. Whether to use bias in the multi-head
             attention layers.
         use_mlp_bias: bool. Whether to use bias in the MLP layers.
+        use_class_token: bool. Whether to use class token to be part of
+            patch embedding. Defaults to `True`.
+        use_patch_bias: bool. Whether to use bias in Conv2d of patch embedding
+            layer. Defaults to `True`.
         data_format: str.  `"channels_last"` or `"channels_first"`, specifying
             the data format for the input image. If `None`, defaults to
             `"channels_last"`.
@@ -58,6 +62,8 @@ class ViTBackbone(Backbone):
         layer_norm_epsilon=1e-6,
         use_mha_bias=True,
         use_mlp_bias=True,
+        use_class_token=True,
+        use_patch_bias=True,
         data_format=None,
         dtype=None,
         **kwargs,
@@ -74,24 +80,34 @@ class ViTBackbone(Backbone):
                 f"at index {h_axis} (height) or {w_axis} (width). "
                 f"Image shape: {image_shape}"
             )
-        if image_shape[h_axis] != image_shape[w_axis]:
+
+        if isinstance(patch_size, int):
+            patch_size = (patch_size, patch_size)
+
+        if image_shape[h_axis] % patch_size[0] != 0:
+            raise ValueError(
+                f"Input height {image_shape[h_axis]} should be divisible by "
+                f"patch size {patch_size[0]}."
+            )
+
+        if image_shape[w_axis] % patch_size[1] != 0:
             raise ValueError(
-                f"Image height and width must be equal. Found height: "
-                f"{image_shape[h_axis]}, width: {image_shape[w_axis]} at "
-                f"indices {h_axis} and {w_axis} respectively. Image shape: "
-                f"{image_shape}"
+                f"Input width {image_shape[h_axis]} should be divisible by "
+                f"patch size {patch_size[1]}."
             )
 
         num_channels = image_shape[channels_axis]
 
         # === Functional Model ===
-        inputs = keras.layers.Input(shape=image_shape)
+        inputs = keras.layers.Input(shape=image_shape, name="images")
 
         x = ViTPatchingAndEmbedding(
-            image_size=image_shape[h_axis],
+            image_size=(image_shape[h_axis], image_shape[w_axis]),
             patch_size=patch_size,
             hidden_dim=hidden_dim,
             num_channels=num_channels,
+            use_class_token=use_class_token,
+            use_patch_bias=use_patch_bias,
             data_format=data_format,
             dtype=dtype,
             name="vit_patching_and_embedding",
@@ -130,6 +146,8 @@ class ViTBackbone(Backbone):
         self.layer_norm_epsilon = layer_norm_epsilon
         self.use_mha_bias = use_mha_bias
         self.use_mlp_bias = use_mlp_bias
+        self.use_class_token = use_class_token
+        self.use_patch_bias = use_patch_bias
         self.data_format = data_format
 
     def get_config(self):
@@ -147,6 +165,8 @@ class ViTBackbone(Backbone):
                 "layer_norm_epsilon": self.layer_norm_epsilon,
                 "use_mha_bias": self.use_mha_bias,
                 "use_mlp_bias": self.use_mlp_bias,
+                "use_class_token": self.use_class_token,
+                "use_patch_bias": self.use_patch_bias,
             }
         )
         return config

keras_hub/src/models/vit/vit_image_converter.py

@@ -1,78 +1,8 @@
 from keras_hub.src.api_export import keras_hub_export
 from keras_hub.src.layers.preprocessing.image_converter import ImageConverter
 from keras_hub.src.models.vit.vit_backbone import ViTBackbone
-from keras_hub.src.utils.tensor_utils import preprocessing_function
 
 
 @keras_hub_export("keras_hub.layers.ViTImageConverter")
 class ViTImageConverter(ImageConverter):
-    """Converts images to the format expected by a ViT model.
-
-    This layer performs image normalization using mean and standard deviation
-    values. By default, it uses the same normalization as the
-    "google/vit-large-patch16-224" model on Hugging Face:
-    `norm_mean=[0.5, 0.5, 0.5]` and `norm_std=[0.5, 0.5, 0.5]`
-    ([reference](https://huggingface.co/google/vit-large-patch16-224/blob/main/preprocessor_config.json)).
-    These defaults are suitable for models pretrained using this normalization.
-
-    Args:
-        norm_mean: list or tuple of floats. Mean values for image normalization.
-            Defaults to `[0.5, 0.5, 0.5]`.
-        norm_std: list or tuple of floats. Standard deviation values for
-            image normalization. Defaults to `[0.5, 0.5, 0.5]`.
-        **kwargs: Additional keyword arguments passed to
-            `keras_hub.layers.preprocessing.ImageConverter`.
-
-    Examples:
-    ```python
-    import keras
-    import numpy as np
-    from keras_hub.src.layers import ViTImageConverter
-
-    # Example image (replace with your actual image data)
-    image = np.random.rand(1, 224, 224, 3)  # Example: (B, H, W, C)
-
-    # Create a ViTImageConverter instance
-    converter = ViTImageConverter(
-        image_size=(28,28),
-        scale=1/255.
-    )
-    # Preprocess the image
-    preprocessed_image = converter(image)
-    ```
-    """
-
     backbone_cls = ViTBackbone
-
-    def __init__(
-        self, norm_mean=[0.5, 0.5, 0.5], norm_std=[0.5, 0.5, 0.5], **kwargs
-    ):
-        super().__init__(**kwargs)
-        self.norm_mean = norm_mean
-        self.norm_std = norm_std
-
-    @preprocessing_function
-    def call(self, inputs):
-        # TODO: Remove this whole function. Why can just use scale and offset
-        # in the base class.
-        x = super().call(inputs)
-        if self.norm_mean:
-            norm_mean = self._expand_non_channel_dims(self.norm_mean, x)
-            x, norm_mean = self._convert_types(x, norm_mean, self.compute_dtype)
-            x = x - norm_mean
-        if self.norm_std:
-            norm_std = self._expand_non_channel_dims(self.norm_std, x)
-            x, norm_std = self._convert_types(x, norm_std, x.dtype)
-            x = x / norm_std
-
-        return x
-
-    def get_config(self):
-        config = super().get_config()
-        config.update(
-            {
-                "norm_mean": self.norm_mean,
-                "norm_std": self.norm_std,
-            }
-        )
-        return config

keras_hub/src/models/vit/vit_layers.py

@@ -75,12 +75,13 @@ class ViTPatchingAndEmbedding(keras.layers.Layer):
     """Patches the image and embeds the patches.
 
     Args:
-        image_size: int. Size of the input image (height or width).
-            Assumed to be square.
-        patch_size: int. Size of each image patch.
+        image_size: (int, int). Size of the input image.
+        patch_size: (int, int). Size of each image patch.
         hidden_dim: int. Dimensionality of the patch embeddings.
         num_channels: int. Number of channels in the input image. Defaults to
             `3`.
+        use_class_token: bool. Whether to use class token to be part of
+            patch embedding. Defaults to `True`.
         data_format: str. `"channels_last"` or `"channels_first"`. Defaults to
             `None` (which uses `"channels_last"`).
         **kwargs: Additional keyword arguments passed to `keras.layers.Layer`
@@ -92,12 +93,15 @@ class ViTPatchingAndEmbedding(keras.layers.Layer):
         patch_size,
         hidden_dim,
         num_channels=3,
+        use_class_token=True,
+        use_patch_bias=True,
         data_format=None,
         **kwargs,
     ):
         super().__init__(**kwargs)
-        num_patches = (image_size // patch_size) ** 2
-        num_positions = num_patches + 1
+        grid_size = tuple([s // p for s, p in zip(image_size, patch_size)])
+        num_patches = grid_size[0] * grid_size[1]
+        num_positions = num_patches + 1 if use_class_token else num_patches
 
         # === Config ===
         self.image_size = image_size
@@ -106,19 +110,22 @@ class ViTPatchingAndEmbedding(keras.layers.Layer):
         self.num_channels = num_channels
         self.num_patches = num_patches
         self.num_positions = num_positions
+        self.use_class_token = use_class_token
+        self.use_patch_bias = use_patch_bias
         self.data_format = standardize_data_format(data_format)
 
     def build(self, input_shape):
-        self.class_token = self.add_weight(
-            shape=(
-                1,
-                1,
-                self.hidden_dim,
-            ),
-            initializer="random_normal",
-            dtype=self.variable_dtype,
-            name="class_token",
-        )
+        if self.use_class_token:
+            self.class_token = self.add_weight(
+                shape=(
+                    1,
+                    1,
+                    self.hidden_dim,
+                ),
+                initializer="random_normal",
+                dtype=self.variable_dtype,
+                name="class_token",
+            )
         self.patch_embedding = keras.layers.Conv2D(
             filters=self.hidden_dim,
             kernel_size=self.patch_size,
@@ -127,6 +134,7 @@ class ViTPatchingAndEmbedding(keras.layers.Layer):
             activation=None,
             dtype=self.dtype_policy,
             data_format=self.data_format,
+            use_bias=self.use_patch_bias,
             name="patch_embedding",
         )
         self.patch_embedding.build(input_shape)
@@ -153,10 +161,16 @@ class ViTPatchingAndEmbedding(keras.layers.Layer):
         patch_embeddings = ops.reshape(
             patch_embeddings, [embeddings_shape[0], -1, embeddings_shape[-1]]
         )
-        class_token = ops.tile(self.class_token, (embeddings_shape[0], 1, 1))
         position_embeddings = self.position_embedding(self.position_ids)
-        embeddings = ops.concatenate([class_token, patch_embeddings], axis=1)
-        return ops.add(embeddings, position_embeddings)
+
+        if self.use_class_token:
+            class_token = ops.tile(
+                self.class_token, (embeddings_shape[0], 1, 1)
+            )
+            patch_embeddings = ops.concatenate(
+                [class_token, patch_embeddings], axis=1
+            )
+        return ops.add(patch_embeddings, position_embeddings)
 
     def compute_output_shape(self, input_shape):
         return (
@@ -175,6 +189,7 @@ class ViTPatchingAndEmbedding(keras.layers.Layer):
                 "num_channels": self.num_channels,
                 "num_patches": self.num_patches,
                 "num_positions": self.num_positions,
+                "use_class_token": self.use_class_token,
             }
         )
         return config