keras-hub 0.22.1__tar.gz → 0.23.0.dev0__tar.gz

{keras_hub-0.22.1 → keras_hub-0.23.0.dev0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: keras-hub
-Version: 0.22.1
+Version: 0.23.0.dev0
 Summary: Pretrained models for Keras.
 Author-email: Keras team <keras-users@googlegroups.com>
 License-Expression: Apache-2.0

{keras_hub-0.22.1 → keras_hub-0.23.0.dev0}/keras_hub/api/layers/__init__.py

@@ -75,6 +75,9 @@ from keras_hub.src.models.clip.clip_image_converter import (
 from keras_hub.src.models.cspnet.cspnet_image_converter import (
     CSPNetImageConverter as CSPNetImageConverter,
 )
+from keras_hub.src.models.d_fine.d_fine_image_converter import (
+    DFineImageConverter as DFineImageConverter,
+)
 from keras_hub.src.models.deeplab_v3.deeplab_v3_image_converter import (
     DeepLabV3ImageConverter as DeepLabV3ImageConverter,
 )
@@ -84,6 +87,9 @@ from keras_hub.src.models.deit.deit_image_converter import (
 from keras_hub.src.models.densenet.densenet_image_converter import (
     DenseNetImageConverter as DenseNetImageConverter,
 )
+from keras_hub.src.models.depth_anything.depth_anything_image_converter import (
+    DepthAnythingImageConverter as DepthAnythingImageConverter,
+)
 from keras_hub.src.models.dinov2.dinov2_image_converter import (
     DINOV2ImageConverter as DINOV2ImageConverter,
 )
@@ -102,12 +108,18 @@ from keras_hub.src.models.mit.mit_image_converter import (
 from keras_hub.src.models.mobilenet.mobilenet_image_converter import (
     MobileNetImageConverter as MobileNetImageConverter,
 )
+from keras_hub.src.models.mobilenetv5.mobilenetv5_image_converter import (
+    MobileNetV5ImageConverter as MobileNetV5ImageConverter,
+)
 from keras_hub.src.models.moonshine.moonshine_audio_converter import (
     MoonshineAudioConverter as MoonshineAudioConverter,
 )
 from keras_hub.src.models.pali_gemma.pali_gemma_image_converter import (
     PaliGemmaImageConverter as PaliGemmaImageConverter,
 )
+from keras_hub.src.models.parseq.parseq_image_converter import (
+    PARSeqImageConverter as PARSeqImageConverter,
+)
 from keras_hub.src.models.resnet.resnet_image_converter import (
     ResNetImageConverter as ResNetImageConverter,
 )

{keras_hub-0.22.1 → keras_hub-0.23.0.dev0}/keras_hub/api/models/__init__.py

@@ -108,6 +108,15 @@ from keras_hub.src.models.cspnet.cspnet_image_classifier import (
 from keras_hub.src.models.cspnet.cspnet_image_classifier_preprocessor import (
     CSPNetImageClassifierPreprocessor as CSPNetImageClassifierPreprocessor,
 )
+from keras_hub.src.models.d_fine.d_fine_backbone import (
+    DFineBackbone as DFineBackbone,
+)
+from keras_hub.src.models.d_fine.d_fine_object_detector import (
+    DFineObjectDetector as DFineObjectDetector,
+)
+from keras_hub.src.models.d_fine.d_fine_object_detector_preprocessor import (
+    DFineObjectDetectorPreprocessor as DFineObjectDetectorPreprocessor,
+)
 from keras_hub.src.models.deberta_v3.deberta_v3_backbone import (
     DebertaV3Backbone as DebertaV3Backbone,
 )
@@ -157,6 +166,21 @@ from keras_hub.src.models.densenet.densenet_image_classifier import (
 from keras_hub.src.models.densenet.densenet_image_classifier_preprocessor import (
     DenseNetImageClassifierPreprocessor as DenseNetImageClassifierPreprocessor,
 )
+from keras_hub.src.models.depth_anything.depth_anything_backbone import (
+    DepthAnythingBackbone as DepthAnythingBackbone,
+)
+from keras_hub.src.models.depth_anything.depth_anything_depth_estimator import (
+    DepthAnythingDepthEstimator as DepthAnythingDepthEstimator,
+)
+from keras_hub.src.models.depth_anything.depth_anything_depth_estimator_preprocessor import (
+    DepthAnythingDepthEstimatorPreprocessor as DepthAnythingDepthEstimatorPreprocessor,
+)
+from keras_hub.src.models.depth_estimator import (
+    DepthEstimator as DepthEstimator,
+)
+from keras_hub.src.models.depth_estimator_preprocessor import (
+    DepthEstimatorPreprocessor as DepthEstimatorPreprocessor,
+)
 from keras_hub.src.models.dinov2.dinov2_backbone import (
     DINOV2Backbone as DINOV2Backbone,
 )
@@ -404,6 +428,15 @@ from keras_hub.src.models.mobilenet.mobilenet_image_classifier import (
 from keras_hub.src.models.mobilenet.mobilenet_image_classifier_preprocessor import (
     MobileNetImageClassifierPreprocessor as MobileNetImageClassifierPreprocessor,
 )
+from keras_hub.src.models.mobilenetv5.mobilenetv5_backbone import (
+    MobileNetV5Backbone as MobileNetV5Backbone,
+)
+from keras_hub.src.models.mobilenetv5.mobilenetv5_image_classifier import (
+    MobileNetV5ImageClassifier as MobileNetV5ImageClassifier,
+)
+from keras_hub.src.models.mobilenetv5.mobilenetv5_image_classifier_preprocessor import (
+    MobileNetV5ImageClassifierPreprocessor as MobileNetV5ImageClassifierPreprocessor,
+)
 from keras_hub.src.models.moonshine.moonshine_audio_to_text import (
     MoonshineAudioToText as MoonshineAudioToText,
 )
@@ -446,6 +479,18 @@ from keras_hub.src.models.pali_gemma.pali_gemma_causal_lm_preprocessor import (
 from keras_hub.src.models.pali_gemma.pali_gemma_tokenizer import (
     PaliGemmaTokenizer as PaliGemmaTokenizer,
 )
+from keras_hub.src.models.parseq.parseq_backbone import (
+    PARSeqBackbone as PARSeqBackbone,
+)
+from keras_hub.src.models.parseq.parseq_causal_lm import (
+    PARSeqCausalLM as PARSeqCausalLM,
+)
+from keras_hub.src.models.parseq.parseq_causal_lm_preprocessor import (
+    PARSeqCausalLMPreprocessor as PARSeqCausalLMPreprocessor,
+)
+from keras_hub.src.models.parseq.parseq_tokenizer import (
+    PARSeqTokenizer as PARSeqTokenizer,
+)
 from keras_hub.src.models.phi3.phi3_backbone import Phi3Backbone as Phi3Backbone
 from keras_hub.src.models.phi3.phi3_causal_lm import (
     Phi3CausalLM as Phi3CausalLM,
@@ -491,6 +536,15 @@ from keras_hub.src.models.qwen3.qwen3_causal_lm_preprocessor import (
 from keras_hub.src.models.qwen3.qwen3_tokenizer import (
     Qwen3Tokenizer as Qwen3Tokenizer,
 )
+from keras_hub.src.models.qwen3_moe.qwen3_moe_backbone import (
+    Qwen3MoeBackbone as Qwen3MoeBackbone,
+)
+from keras_hub.src.models.qwen3_moe.qwen3_moe_causal_lm import (
+    Qwen3MoeCausalLM as Qwen3MoeCausalLM,
+)
+from keras_hub.src.models.qwen3_moe.qwen3_moe_causal_lm_preprocessor import (
+    Qwen3MoeCausalLMPreprocessor as Qwen3MoeCausalLMPreprocessor,
+)
 from keras_hub.src.models.qwen_moe.qwen_moe_backbone import (
     QwenMoeBackbone as QwenMoeBackbone,
 )
@@ -595,6 +649,30 @@ from keras_hub.src.models.siglip.siglip_tokenizer import (
 from keras_hub.src.models.siglip.siglip_vision_encoder import (
     SigLIPVisionEncoder as SigLIPVisionEncoder,
 )
+from keras_hub.src.models.smollm3.smollm3_backbone import (
+    SmolLM3Backbone as SmolLM3Backbone,
+)
+from keras_hub.src.models.smollm3.smollm3_backbone import (
+    SmolLM3Backbone as SmolLMBackbone,
+)
+from keras_hub.src.models.smollm3.smollm3_causal_lm import (
+    SmolLM3CausalLM as SmolLM3CausalLM,
+)
+from keras_hub.src.models.smollm3.smollm3_causal_lm import (
+    SmolLM3CausalLM as SmolLMCausalLM,
+)
+from keras_hub.src.models.smollm3.smollm3_causal_lm_preprocessor import (
+    SmolLM3CausalLMPreprocessor as SmolLM3CausalLMPreprocessor,
+)
+from keras_hub.src.models.smollm3.smollm3_causal_lm_preprocessor import (
+    SmolLM3CausalLMPreprocessor as SmolLMCausalLMPreprocessor,
+)
+from keras_hub.src.models.smollm3.smollm3_tokenizer import (
+    SmolLM3Tokenizer as SmolLM3Tokenizer,
+)
+from keras_hub.src.models.smollm3.smollm3_tokenizer import (
+    SmolLM3Tokenizer as SmolLMTokenizer,
+)
 from keras_hub.src.models.stable_diffusion_3.stable_diffusion_3_backbone import (
     StableDiffusion3Backbone as StableDiffusion3Backbone,
 )
@@ -615,6 +693,18 @@ from keras_hub.src.models.t5.t5_preprocessor import (
     T5Preprocessor as T5Preprocessor,
 )
 from keras_hub.src.models.t5.t5_tokenizer import T5Tokenizer as T5Tokenizer
+from keras_hub.src.models.t5gemma.t5gemma_backbone import (
+    T5GemmaBackbone as T5GemmaBackbone,
+)
+from keras_hub.src.models.t5gemma.t5gemma_seq_2_seq_lm import (
+    T5GemmaSeq2SeqLM as T5GemmaSeq2SeqLM,
+)
+from keras_hub.src.models.t5gemma.t5gemma_seq_2_seq_lm_preprocessor import (
+    T5GemmaSeq2SeqLMPreprocessor as T5GemmaSeq2SeqLMPreprocessor,
+)
+from keras_hub.src.models.t5gemma.t5gemma_tokenizer import (
+    T5GemmaTokenizer as T5GemmaTokenizer,
+)
 from keras_hub.src.models.task import Task as Task
 from keras_hub.src.models.text_classifier import TextClassifier as Classifier
 from keras_hub.src.models.text_classifier import (

{keras_hub-0.22.1 → keras_hub-0.23.0.dev0}/keras_hub/api/tokenizers/__init__.py

@@ -66,6 +66,9 @@ from keras_hub.src.models.opt.opt_tokenizer import OPTTokenizer as OPTTokenizer
 from keras_hub.src.models.pali_gemma.pali_gemma_tokenizer import (
     PaliGemmaTokenizer as PaliGemmaTokenizer,
 )
+from keras_hub.src.models.parseq.parseq_tokenizer import (
+    PARSeqTokenizer as PARSeqTokenizer,
+)
 from keras_hub.src.models.phi3.phi3_tokenizer import (
     Phi3Tokenizer as Phi3Tokenizer,
 )
@@ -75,6 +78,9 @@ 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_moe.qwen3_moe_tokenizer import (
+    Qwen3MoeTokenizer as Qwen3MoeTokenizer,
+)
 from keras_hub.src.models.qwen_moe.qwen_moe_tokenizer import (
     QwenMoeTokenizer as QwenMoeTokenizer,
 )
@@ -87,7 +93,16 @@ from keras_hub.src.models.roformer_v2.roformer_v2_tokenizer import (
 from keras_hub.src.models.siglip.siglip_tokenizer import (
     SigLIPTokenizer as SigLIPTokenizer,
 )
+from keras_hub.src.models.smollm3.smollm3_tokenizer import (
+    SmolLM3Tokenizer as SmolLM3Tokenizer,
+)
+from keras_hub.src.models.smollm3.smollm3_tokenizer import (
+    SmolLM3Tokenizer as SmolLMTokenizer,
+)
 from keras_hub.src.models.t5.t5_tokenizer import T5Tokenizer as T5Tokenizer
+from keras_hub.src.models.t5gemma.t5gemma_tokenizer import (
+    T5GemmaTokenizer as T5GemmaTokenizer,
+)
 from keras_hub.src.models.whisper.whisper_tokenizer import (
     WhisperTokenizer as WhisperTokenizer,
 )

{keras_hub-0.22.1 → keras_hub-0.23.0.dev0}/keras_hub/src/layers/modeling/position_embedding.py

@@ -31,6 +31,11 @@ class PositionEmbedding(keras.layers.Layer):
         start_index: An integer or integer tensor. The starting position to
             compute the position embedding from. This is useful during cached
             decoding, where each position is predicted separately in a loop.
+        positions: Tensor of shape `(sequence_length,)` or
+            `(batch_size, sequence_length)`. Custom positions for the input
+            sequence. If specified, this tensor will be used to
+            compute the position embedding, and the `start_index` argument will
+            be ignored. This is useful for cases with non-standard positions.
 
     Example:
 
@@ -91,18 +96,28 @@ class PositionEmbedding(keras.layers.Layer):
         )
         self.built = True
 
-    def call(self, inputs, start_index=0):
+    def call(self, inputs, start_index=0, positions=None):
         shape = ops.shape(inputs)
         feature_length = shape[-1]
         sequence_length = shape[-2]
         # trim to match the length of the input sequence, which might be less
         # than the sequence_length of the layer.
         position_embeddings = ops.convert_to_tensor(self.position_embeddings)
-        position_embeddings = ops.slice(
-            position_embeddings,
-            (start_index, 0),
-            (sequence_length, feature_length),
-        )
+        if positions is None:
+            position_embeddings = ops.slice(
+                position_embeddings,
+                (start_index, 0),
+                (sequence_length, feature_length),
+            )
+        else:
+            # Take care of unbatched `positions`.
+            if len(ops.shape(positions)) == 1:
+                positions = ops.expand_dims(positions, axis=0)
+
+            position_embeddings = ops.take(
+                position_embeddings, positions, axis=0
+            )
+
         return ops.broadcast_to(position_embeddings, shape)
 
     def compute_output_shape(self, input_shape):

{keras_hub-0.22.1 → keras_hub-0.23.0.dev0}/keras_hub/src/layers/modeling/reversible_embedding.py

@@ -235,7 +235,8 @@ class ReversibleEmbedding(keras.layers.Embedding):
 
         return super()._int8_call(inputs)
 
-    def quantize(self, mode, type_check=True):
+    def quantize(self, mode, type_check=True, config=None):
+        del config
         if type_check and type(self) is not ReversibleEmbedding:
             raise self._not_implemented_error(self.quantize)
 
@@ -244,6 +245,12 @@ class ReversibleEmbedding(keras.layers.Embedding):
                 inputs, axis=axis, to_numpy=True
             )
 
+        if mode != "int8":
+            raise NotImplementedError(
+                "Invalid quantization mode. Expected 'int8'. "
+                f"Received: quantization_mode={mode}"
+            )
+
         embeddings_shape = (self.input_dim, self.output_dim)
         if mode == "int8":
             embeddings, embeddings_scale = abs_max_quantize(

{keras_hub-0.22.1 → keras_hub-0.23.0.dev0}/keras_hub/src/layers/modeling/rotary_embedding.py

@@ -37,6 +37,11 @@ class RotaryEmbedding(keras.layers.Layer):
         start_index: An integer or integer tensor. The starting position to
             compute the rotary embedding from. This is useful during cached
             decoding, where each position is predicted separately in a loop.
+        positions: Tensor of shape `(sequence_length,)` or
+            `(batch_size, sequence_length)`. Custom positions for the input
+            sequence. If specified, this tensor will be used to
+            compute the rotary embedding, and the `start_index` argument will
+            be ignored. This is useful for cases with non-standard positions.
 
     Examples:
 
@@ -76,6 +81,11 @@ class RotaryEmbedding(keras.layers.Layer):
         self.built = True
 
     def call(self, inputs, start_index=0, positions=None):
+        # Take care of unbatched `positions`.
+        if positions is not None:
+            if len(ops.shape(positions)) == 1:
+                positions = ops.expand_dims(positions, axis=0)
+
         inputs = ops.moveaxis(
             inputs, (self.feature_axis, self.sequence_axis), (-1, 1)
         )
@@ -103,6 +113,7 @@ class RotaryEmbedding(keras.layers.Layer):
         return positions + ops.cast(start_index, dtype="float32")
 
     def _compute_cos_sin_embedding(self, inputs, start_index=0, positions=None):
+        batch_axis = 0
         feature_axis = len(inputs.shape) - 1
         sequence_axis = 1
 
@@ -111,21 +122,20 @@ class RotaryEmbedding(keras.layers.Layer):
 
         if positions is None:
             positions = self._compute_positions(inputs, start_index)
+            positions = ops.expand_dims(positions, axis=batch_axis)
         else:
             positions = ops.cast(positions, "float32")
-
         positions = positions / ops.cast(self.scaling_factor, "float32")
-        freq = ops.einsum("i,j->ij", positions, inverse_freq)
+
+        freq = ops.einsum("bi,j->bij", positions, inverse_freq)
+
         embedding = ops.stack((freq, freq), axis=-2)
         embedding = ops.reshape(
             embedding, (*ops.shape(freq)[:-1], ops.shape(freq)[-1] * 2)
         )
 
-        # Reshape the embedding to be broadcastable with input shape.
-        if feature_axis < sequence_axis:
-            embedding = ops.transpose(embedding)
         for axis in range(len(inputs.shape)):
-            if axis != sequence_axis and axis != feature_axis:
+            if axis not in (batch_axis, sequence_axis, feature_axis):
                 embedding = ops.expand_dims(embedding, axis)
 
         cos_emb = ops.cast(ops.cos(embedding), self.compute_dtype)

{keras_hub-0.22.1 → keras_hub-0.23.0.dev0}/keras_hub/src/layers/modeling/sine_position_encoding.py

@@ -30,6 +30,11 @@ class SinePositionEncoding(keras.layers.Layer):
         start_index: An integer or integer tensor. The starting position to
             compute the encoding from. This is useful during cached decoding,
             where each position is predicted separately in a loop.
+        positions: Tensor of shape `(sequence_length,)` or
+            `(batch_size, sequence_length)`. Custom positions for the input
+            sequence. If specified, this tensor will be used to
+            compute the position embedding, and the `start_index` argument will
+            be ignored. This is useful for cases with non-standard positions.
 
     Example:
     ```python
@@ -58,27 +63,35 @@ class SinePositionEncoding(keras.layers.Layer):
         self.max_wavelength = max_wavelength
         self.built = True
 
-    def call(self, inputs, start_index=0):
+    def call(self, inputs, start_index=0, positions=None):
         shape = ops.shape(inputs)
         seq_length = shape[-2]
         hidden_size = shape[-1]
-        positions = ops.arange(seq_length)
-        positions = ops.cast(positions + start_index, self.compute_dtype)
+
+        if positions is None:
+            positions = ops.arange(seq_length)
+            positions = ops.cast(positions + start_index, self.compute_dtype)
+
+        # Take care of unbatched `positions`.
+        if len(ops.shape(positions)) == 1:
+            positions = ops.expand_dims(positions, axis=0)
+
         min_freq = ops.cast(1 / self.max_wavelength, dtype=self.compute_dtype)
         timescales = ops.power(
             min_freq,
             ops.cast(2 * (ops.arange(hidden_size) // 2), self.compute_dtype)
             / ops.cast(hidden_size, self.compute_dtype),
         )
-        angles = ops.expand_dims(positions, 1) * ops.expand_dims(timescales, 0)
+        angles = ops.einsum("bi,j->bij", positions, timescales)
+
         # even indices are sine, odd are cosine
         cos_mask = ops.cast(ops.arange(hidden_size) % 2, self.compute_dtype)
         sin_mask = 1 - cos_mask
-        # embedding shape is [seq_length, hidden_size]
-        positional_encodings = (
-            ops.sin(angles) * sin_mask + ops.cos(angles) * cos_mask
-        )
 
+        # embedding shape is `[bsz (or 1), seq_length, hidden_size]`.
+        positional_encodings = ops.einsum(
+            "bij,j->bij", ops.sin(angles), sin_mask
+        ) + ops.einsum("bij,j->bij", ops.cos(angles), cos_mask)
         return ops.broadcast_to(positional_encodings, shape)
 
     def get_config(self):

{keras_hub-0.22.1 → keras_hub-0.23.0.dev0}/keras_hub/src/layers/modeling/token_and_position_embedding.py

@@ -120,11 +120,12 @@ class TokenAndPositionEmbedding(keras.layers.Layer):
         )
         return config
 
-    def call(self, inputs, start_index=0):
+    def call(self, inputs, start_index=0, positions=None):
         embedded_tokens = self.token_embedding(inputs)
         embedded_positions = self.position_embedding(
             embedded_tokens,
             start_index=start_index,
+            positions=positions,
         )
         outputs = embedded_tokens + embedded_positions
         return outputs

{keras_hub-0.22.1 → keras_hub-0.23.0.dev0}/keras_hub/src/models/backbone.py

@@ -91,21 +91,16 @@ class Backbone(keras.Model):
         }
 
         # Add quantization support by utilizing `DTypePolicyMap`
-        try:
-            if isinstance(
-                self.dtype_policy, keras.dtype_policies.DTypePolicyMap
-            ):
-                config.update({"dtype": self.dtype_policy})
-            else:
-                policy_map = keras.dtype_policies.DTypePolicyMap()
-                for layer in self._flatten_layers():
-                    if layer.quantization_mode is not None:
-                        policy_map[layer.path] = layer.dtype_policy
-                if len(policy_map) > 0:
-                    config.update({"dtype": policy_map})
-        # Before Keras 3.2, there is no `keras.dtype_policies.get`.
-        except AttributeError:
-            pass
+        dtype = self.dtype_policy
+        if not isinstance(dtype, keras.dtype_policies.DTypePolicyMap):
+            policy_map = keras.dtype_policies.DTypePolicyMap()
+            for layer in self._flatten_layers():
+                if layer.quantization_mode is not None:
+                    policy_map[layer.path] = layer.dtype_policy
+            if len(policy_map) > 0:
+                dtype = policy_map
+
+        config.update({"dtype": keras.dtype_policies.serialize(dtype)})
         return config
 
     @classmethod
@@ -135,7 +130,8 @@ class Backbone(keras.Model):
         1. a built-in preset identifier like `'bert_base_en'`
         2. a Kaggle Models handle like `'kaggle://user/bert/keras/bert_base_en'`
         3. a Hugging Face handle like `'hf://user/bert_base_en'`
-        4. a path to a local preset directory like `'./bert_base_en'`
+        4. a ModelScope handle like `'modelscope://user/bert_base_en'`
+        5. a path to a local preset directory like `'./bert_base_en'`
 
         This constructor can be called in one of two ways. Either from the base
         class like `keras_hub.models.Backbone.from_preset()`, or from
@@ -277,3 +273,19 @@ class Backbone(keras.Model):
             layer.lora_kernel_a.assign(lora_kernel_a)
             layer.lora_kernel_b.assign(lora_kernel_b)
         store.close()
+
+    def export_to_transformers(self, path):
+        """Export the backbone model to HuggingFace Transformers format.
+
+        This saves the backbone's configuration and weights in a format
+        compatible with HuggingFace Transformers. For unsupported model
+        architectures, a ValueError is raised.
+
+        Args:
+            path: str. Path to save the exported model.
+        """
+        from keras_hub.src.utils.transformers.export.hf_exporter import (
+            export_backbone,
+        )
+
+        export_backbone(self, path)

{keras_hub-0.22.1 → keras_hub-0.23.0.dev0}/keras_hub/src/models/causal_lm.py

@@ -132,6 +132,17 @@ class CausalLM(Task):
             return self.generate_function
 
         self.generate_function = self.generate_step
+        if keras.config.backend() == "openvino":
+            from keras_hub.src.utils.openvino_utils import ov_infer
+
+            def wrapped_generate_function(inputs, stop_token_ids=None):
+                # Convert to numpy for OpenVINO backend
+                inputs = tree.map_structure(ops.array, inputs)
+                return ov_infer(
+                    self, inputs, stop_token_ids, self.generate_step
+                )
+
+            self.generate_function = wrapped_generate_function
         if keras.config.backend() == "torch":
             import torch
 
@@ -392,3 +403,29 @@ class CausalLM(Task):
             outputs = [postprocess(x) for x in outputs]
 
         return self._normalize_generate_outputs(outputs, input_is_scalar)
+
+    def export_to_transformers(self, path):
+        """Export the full CausalLM model to HuggingFace Transformers format.
+
+        This exports the trainable model, tokenizer, and configurations in a
+        format compatible with HuggingFace Transformers. For unsupported model
+        architectures, a ValueError is raised.
+
+        If the preprocessor is attached (default), both the trainable model and
+        tokenizer are exported. To export only the trainable model, set
+        `self.preprocessor = None` before calling this method, then export the
+        preprocessor separately via `preprocessor.export_to_transformers(path)`.
+
+        Args:
+            path: str. Path to save the exported model.
+        """
+        from keras_hub.src.utils.transformers.export.hf_exporter import (
+            export_to_safetensors,
+        )
+
+        export_to_safetensors(self, path)
+
+    def _post_quantize(self, mode, **kwargs):
+        super()._post_quantize(mode, **kwargs)
+        # Reset the compiled generate function.
+        self.generate_function = None

{keras_hub-0.22.1 → keras_hub-0.23.0.dev0}/keras_hub/src/models/causal_lm_preprocessor.py

@@ -180,3 +180,17 @@ class CausalLMPreprocessor(Preprocessor):
         self._sequence_length = value
         if self.packer is not None:
             self.packer.sequence_length = value
+
+    def export_to_transformers(self, path):
+        """Export the preprocessor to HuggingFace Transformers format.
+
+        Args:
+            path: str. Path to save the exported preprocessor/tokenizer.
+        """
+        if self.tokenizer is None:
+            raise ValueError("Preprocessor must have a tokenizer for export.")
+        from keras_hub.src.utils.transformers.export.hf_exporter import (
+            export_tokenizer,
+        )
+
+        export_tokenizer(self.tokenizer, path)

{keras_hub-0.22.1 → keras_hub-0.23.0.dev0}/keras_hub/src/models/clip/clip_presets.py

@@ -11,7 +11,7 @@ backbone_presets = {
             "params": 149620934,
             "path": "clip",
         },
-        "kaggle_handle": "kaggle://keras/clip/keras/clip_vit_base_patch16/2",
+        "kaggle_handle": "kaggle://keras/clip/keras/clip_vit_base_patch16/3",
     },
     "clip_vit_base_patch32": {
         "metadata": {
@@ -22,7 +22,7 @@ backbone_presets = {
             "params": 151277363,
             "path": "clip",
         },
-        "kaggle_handle": "kaggle://keras/clip/keras/clip_vit_base_patch32/2",
+        "kaggle_handle": "kaggle://keras/clip/keras/clip_vit_base_patch32/3",
     },
     "clip_vit_large_patch14": {
         "metadata": {
@@ -33,7 +33,7 @@ backbone_presets = {
             "params": 427616770,
             "path": "clip",
         },
-        "kaggle_handle": "kaggle://keras/clip/keras/clip_vit_large_patch14/2",
+        "kaggle_handle": "kaggle://keras/clip/keras/clip_vit_large_patch14/3",
     },
     "clip_vit_large_patch14_336": {
         "metadata": {
@@ -44,7 +44,7 @@ backbone_presets = {
             "params": 427944770,
             "path": "clip",
         },
-        "kaggle_handle": "kaggle://keras/clip/keras/clip_vit_large_patch14_336/2",
+        "kaggle_handle": "kaggle://keras/clip/keras/clip_vit_large_patch14_336/3",
     },
     "clip_vit_b_32_laion2b_s34b_b79k": {
         "metadata": {
@@ -55,7 +55,7 @@ backbone_presets = {
             "params": 151277363,
             "path": "clip",
         },
-        "kaggle_handle": "kaggle://keras/clip/keras/clip_vit_b_32_laion2b_s34b_b79k/2",
+        "kaggle_handle": "kaggle://keras/clip/keras/clip_vit_b_32_laion2b_s34b_b79k/3",
     },
     "clip_vit_h_14_laion2b_s32b_b79k": {
         "metadata": {
@@ -66,7 +66,7 @@ backbone_presets = {
             "params": 986109698,
             "path": "clip",
         },
-        "kaggle_handle": "kaggle://keras/clip/keras/clip_vit_h_14_laion2b_s32b_b79k/2",
+        "kaggle_handle": "kaggle://keras/clip/keras/clip_vit_h_14_laion2b_s32b_b79k/3",
     },
     "clip_vit_g_14_laion2b_s12b_b42k": {
         "metadata": {
@@ -77,7 +77,7 @@ backbone_presets = {
             "params": 1366678530,
             "path": "clip",
         },
-        "kaggle_handle": "kaggle://keras/clip/keras/clip_vit_g_14_laion2b_s12b_b42k/2",
+        "kaggle_handle": "kaggle://keras/clip/keras/clip_vit_g_14_laion2b_s12b_b42k/3",
     },
     "clip_vit_bigg_14_laion2b_39b_b160k": {
         "metadata": {
@@ -88,6 +88,6 @@ backbone_presets = {
             "params": 2539567362,
             "path": "clip",
         },
-        "kaggle_handle": "kaggle://keras/clip/keras/clip_vit_bigg_14_laion2b_39b_b160k/2",
+        "kaggle_handle": "kaggle://keras/clip/keras/clip_vit_bigg_14_laion2b_39b_b160k/3",
     },
 }

keras_hub-0.23.0.dev0/keras_hub/src/models/d_fine/__init__.py

@@ -0,0 +1,5 @@
+from keras_hub.src.models.d_fine.d_fine_backbone import DFineBackbone
+from keras_hub.src.models.d_fine.d_fine_presets import backbone_presets
+from keras_hub.src.utils.preset_utils import register_presets
+
+register_presets(backbone_presets, DFineBackbone)