keras-hub-nightly 0.24.0.dev202511160427__tar.gz → 0.26.0.dev202601130438__tar.gz

{keras_hub_nightly-0.24.0.dev202511160427 → keras_hub_nightly-0.26.0.dev202601130438}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: keras-hub-nightly
-Version: 0.24.0.dev202511160427
+Version: 0.26.0.dev202601130438
 Summary: Pretrained models for Keras.
 Author-email: Keras team <keras-users@googlegroups.com>
 License-Expression: Apache-2.0
@@ -8,7 +8,6 @@ Project-URL: Home, https://keras.io/keras_hub/
 Project-URL: Repository, https://github.com/keras-team/keras/keras_hub
 Classifier: Development Status :: 3 - Alpha
 Classifier: Programming Language :: Python :: 3
-Classifier: Programming Language :: Python :: 3.9
 Classifier: Programming Language :: Python :: 3.10
 Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3 :: Only
@@ -18,9 +17,9 @@ Classifier: Operating System :: MacOS
 Classifier: Intended Audience :: Science/Research
 Classifier: Topic :: Scientific/Engineering
 Classifier: Topic :: Software Development
-Requires-Python: >=3.10
+Requires-Python: >=3.11
 Description-Content-Type: text/markdown
-Requires-Dist: keras>=3.8
+Requires-Dist: keras>=3.13
 Requires-Dist: absl-py
 Requires-Dist: numpy
 Requires-Dist: packaging
@@ -31,7 +30,7 @@ Requires-Dist: tensorflow-text; platform_system != "Windows"
 
 # KerasHub: Multi-framework Pretrained Models
 [![](https://github.com/keras-team/keras-hub/workflows/Tests/badge.svg?branch=master)](https://github.com/keras-team/keras-hub/actions?query=workflow%3ATests+branch%3Amaster)
-![Python](https://img.shields.io/badge/python-v3.10.0+-success.svg)
+![Python](https://img.shields.io/badge/python-v3.11.0+-success.svg)
 [![contributions welcome](https://img.shields.io/badge/contributions-welcome-brightgreen.svg?style=flat)](https://github.com/keras-team/keras-hub/issues)
 
 > [!IMPORTANT]

{keras_hub_nightly-0.24.0.dev202511160427 → keras_hub_nightly-0.26.0.dev202601130438}/README.md

@@ -1,6 +1,6 @@
 # KerasHub: Multi-framework Pretrained Models
 [![](https://github.com/keras-team/keras-hub/workflows/Tests/badge.svg?branch=master)](https://github.com/keras-team/keras-hub/actions?query=workflow%3ATests+branch%3Amaster)
-![Python](https://img.shields.io/badge/python-v3.10.0+-success.svg)
+![Python](https://img.shields.io/badge/python-v3.11.0+-success.svg)
 [![contributions welcome](https://img.shields.io/badge/contributions-welcome-brightgreen.svg?style=flat)](https://github.com/keras-team/keras-hub/issues)
 
 > [!IMPORTANT]
@@ -186,4 +186,4 @@ Thank you to all of our wonderful contributors!
 
 <a href="https://github.com/keras-team/keras-hub/graphs/contributors">
   <img src="https://contrib.rocks/image?repo=keras-team/keras-hub" />
-</a>
+</a>

{keras_hub_nightly-0.24.0.dev202511160427 → keras_hub_nightly-0.26.0.dev202601130438}/keras_hub/api/models/__init__.py

@@ -340,6 +340,18 @@ from keras_hub.src.models.gpt_neo_x.gpt_neo_x_causal_lm_preprocessor import (
 from keras_hub.src.models.gpt_neo_x.gpt_neo_x_tokenizer import (
     GPTNeoXTokenizer as GPTNeoXTokenizer,
 )
+from keras_hub.src.models.gpt_oss.gpt_oss_backbone import (
+    GptOssBackbone as GptOssBackbone,
+)
+from keras_hub.src.models.gpt_oss.gpt_oss_causal_lm import (
+    GptOssCausalLM as GptOssCausalLM,
+)
+from keras_hub.src.models.gpt_oss.gpt_oss_causal_lm_preprocessor import (
+    GptOssCausalLMPreprocessor as GptOssCausalLMPreprocessor,
+)
+from keras_hub.src.models.gpt_oss.gpt_oss_tokenizer import (
+    GptOssTokenizer as GptOssTokenizer,
+)
 from keras_hub.src.models.hgnetv2.hgnetv2_backbone import (
     HGNetV2Backbone as HGNetV2Backbone,
 )
@@ -617,6 +629,15 @@ from keras_hub.src.models.roformer_v2.roformer_v2_text_classifier_preprocessor i
 from keras_hub.src.models.roformer_v2.roformer_v2_tokenizer import (
     RoformerV2Tokenizer as RoformerV2Tokenizer,
 )
+from keras_hub.src.models.rwkv7.rwkv7_backbone import (
+    RWKV7Backbone as RWKV7Backbone,
+)
+from keras_hub.src.models.rwkv7.rwkv7_causal_lm import (
+    RWKV7CausalLM as RWKV7CausalLM,
+)
+from keras_hub.src.models.rwkv7.rwkv7_causal_lm_preprocessor import (
+    RWKV7CausalLMPreprocessor as RWKV7CausalLMPreprocessor,
+)
 from keras_hub.src.models.sam.sam_backbone import SAMBackbone as SAMBackbone
 from keras_hub.src.models.sam.sam_image_segmenter import (
     SAMImageSegmenter as SAMImageSegmenter,

{keras_hub_nightly-0.24.0.dev202511160427 → keras_hub_nightly-0.26.0.dev202601130438}/keras_hub/api/tokenizers/__init__.py

@@ -47,6 +47,9 @@ from keras_hub.src.models.gpt2.gpt2_tokenizer import (
 from keras_hub.src.models.gpt_neo_x.gpt_neo_x_tokenizer import (
     GPTNeoXTokenizer as GPTNeoXTokenizer,
 )
+from keras_hub.src.models.gpt_oss.gpt_oss_tokenizer import (
+    GptOssTokenizer as GptOssTokenizer,
+)
 from keras_hub.src.models.llama.llama_tokenizer import (
     LlamaTokenizer as LlamaTokenizer,
 )
@@ -90,6 +93,9 @@ from keras_hub.src.models.roberta.roberta_tokenizer import (
 from keras_hub.src.models.roformer_v2.roformer_v2_tokenizer import (
     RoformerV2Tokenizer as RoformerV2Tokenizer,
 )
+from keras_hub.src.models.rwkv7.rwkv7_tokenizer import (
+    RWKVTokenizer as RWKVTokenizer,
+)
 from keras_hub.src.models.siglip.siglip_tokenizer import (
     SigLIPTokenizer as SigLIPTokenizer,
 )

{keras_hub_nightly-0.24.0.dev202511160427 → keras_hub_nightly-0.26.0.dev202601130438}/keras_hub/src/layers/modeling/non_max_supression.py

@@ -290,16 +290,19 @@ def non_max_suppression(
         "int32",
     )
     idx = ops.minimum(idx, num_boxes - 1)
+    idx = ops.cast(idx, "int32")
 
     index_offsets = ops.cast(ops.arange(batch_size) * num_boxes, "int32")
     take_along_axis_idx = ops.reshape(
         idx + ops.expand_dims(index_offsets, 1), [-1]
     )
+    take_along_axis_idx = ops.cast(take_along_axis_idx, "int32")
 
     if keras.backend.backend() != "tensorflow":
-        idx = ops.take_along_axis(
-            ops.reshape(sorted_indices, [-1]), take_along_axis_idx
+        sorted_indices_int = ops.cast(
+            ops.reshape(sorted_indices, [-1]), "int32"
         )
+        idx = ops.take_along_axis(sorted_indices_int, take_along_axis_idx)
     else:
         import tensorflow as tf
 

keras_hub_nightly-0.26.0.dev202601130438/keras_hub/src/layers/modeling/reversible_embedding.py

@@ -0,0 +1,8 @@
+import keras
+
+from keras_hub.src.api_export import keras_hub_export
+
+
+@keras_hub_export("keras_hub.layers.ReversibleEmbedding")
+class ReversibleEmbedding(keras.layers.ReversibleEmbedding):
+    pass

keras_hub_nightly-0.26.0.dev202601130438/keras_hub/src/layers/modeling/rotary_embedding.py

@@ -0,0 +1,340 @@
+import keras
+import numpy as np
+from keras import ops
+
+from keras_hub.src.api_export import keras_hub_export
+
+
+@keras_hub_export("keras_hub.layers.RotaryEmbedding")
+class RotaryEmbedding(keras.layers.Layer):
+    """Rotary positional encoding layer.
+
+    This layer encodes absolute positional information with a rotation
+    matrix. It calculates the rotary encoding with a mix of sine and
+    cosine functions with geometrically increasing wavelengths.
+    Defined and formulated in
+    [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/abs/2104.09864v4).
+    The input must be a tensor with shape a sequence dimension and a feature
+    dimension. Typically, this will either an input with shape
+    `(batch_size, sequence_length, feature_length)` or
+    `(batch_size, sequence_length, num_heads, feature_length)`.
+    This layer will return a new tensor with the rotary embedding applied to
+    the input tensor.
+
+    Args:
+        max_wavelength: int. The maximum angular wavelength of the sine/cosine
+            curves.
+        scaling_factor: float. The scaling factor used to scale positions of
+            the tokens.
+        rope_type: str. The type of RoPE scaling to apply. Supported types:
+            "linear", "dynamic", "yarn". Defaults to "linear".
+        beta_fast: float. Beta fast parameter for YaRN scaling. Only used
+            when rope_type="yarn". Defaults to 32.0.
+        beta_slow: float. Beta slow parameter for YaRN scaling. Only used
+            when rope_type="yarn". Defaults to 1.0.
+        original_max_position_embeddings: int. Original maximum position
+            embeddings for YaRN scaling. Only used when rope_type="yarn".
+            Defaults to 4096.
+        truncate: bool. Whether to apply truncation for YaRN scaling. Only used
+            when rope_type="yarn". Defaults to False.
+        sequence_axis: int. Sequence axis in the input tensor.
+        feature_axis: int. Feature axis in the input tensor.
+        **kwargs: other keyword arguments passed to `keras.layers.Layer`,
+            including `name`, `trainable`, `dtype` etc.
+
+    Call arguments:
+        inputs: The tensor inputs to apply the embedding to. This can have
+            any shape, but must contain both a sequence and feature axis. The
+            rotary embedding will be applied to `inputs` and returned.
+        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:
+
+    ```python
+    batch_size = 16
+    feature_length = 18
+    sequence_length = 256
+    num_heads = 8
+
+    # No multi-head dimension.
+    tensor = np.ones((batch_size, sequence_length, feature_length))
+    rot_emb_layer = RotaryEmbedding()
+    tensor_rot = rot_emb_layer(tensor)
+
+    # With multi-head dimension.
+    tensor = np.ones((batch_size, sequence_length, num_heads, feature_length))
+    tensor_rot = rot_emb_layer(tensor)
+    ```
+
+    References:
+     - [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/abs/2104.09864v4)
+    """
+
+    def __init__(
+        self,
+        max_wavelength=10000,
+        scaling_factor=1.0,
+        rope_type="linear",
+        beta_fast=32.0,
+        beta_slow=1.0,
+        original_max_position_embeddings=4096,
+        truncate=False,
+        sequence_axis=1,
+        feature_axis=-1,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.max_wavelength = max_wavelength
+        self.sequence_axis = sequence_axis
+        self.feature_axis = feature_axis
+        self.scaling_factor = scaling_factor
+        self.rope_type = rope_type
+
+        # YaRN-specific parameters (only used when rope_type="yarn")
+        self.beta_fast = beta_fast
+        self.beta_slow = beta_slow
+        self.original_max_position_embeddings = original_max_position_embeddings
+        self.truncate = truncate
+        self.built = True
+
+    def _normalize_axes(self, input_shape):
+        """Normalize and validate axis indices for the given input shape."""
+        rank = len(input_shape)
+
+        # Normalize negative indices
+        sequence_axis = self.sequence_axis
+        feature_axis = self.feature_axis
+
+        if sequence_axis < 0:
+            sequence_axis += rank
+        if feature_axis < 0:
+            feature_axis += rank
+
+        if sequence_axis < 0 or sequence_axis >= rank:
+            raise ValueError(
+                f"sequence_axis {self.sequence_axis} "
+                f"is out of range for input with rank {rank}"
+            )
+        if feature_axis < 0 or feature_axis >= rank:
+            raise ValueError(
+                f"feature_axis {self.feature_axis} "
+                f"is out of range for input with rank {rank}"
+            )
+        if sequence_axis == feature_axis:
+            raise ValueError("sequence_axis and feature_axis must be different")
+
+        return sequence_axis, feature_axis
+
+    def _validate_rotary_dimension(self, rotary_dim):
+        if rotary_dim % 2 != 0:
+            raise ValueError(
+                f"Rotary dimension must be even, got {rotary_dim}."
+                "The rotary embedding splits the feature dimension "
+                "into two halves. Consider using a different feature "
+                "dimension or padding."
+            )
+
+    def call(self, inputs, start_index=0, positions=None):
+        input_shape = ops.shape(inputs)
+        sequence_axis, feature_axis = self._normalize_axes(input_shape)
+
+        rotary_dim = input_shape[feature_axis]
+        self._validate_rotary_dimension(rotary_dim)
+
+        # 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, (feature_axis, sequence_axis), (-1, 1))
+        cos_emb, sin_emb = self._compute_cos_sin_embedding(
+            inputs, start_index, positions
+        )
+        output = self._apply_rotary_pos_emb(inputs, cos_emb, sin_emb)
+        return ops.moveaxis(output, (-1, 1), (feature_axis, sequence_axis))
+
+    def _apply_rotary_pos_emb(self, tensor, cos_emb, sin_emb):
+        x1, x2 = ops.split(tensor, 2, axis=-1)
+        # Avoid `ops.concatenate` for now, to avoid a obscure bug with XLA
+        # compilation on jax. We should be able to remove this once the
+        # following PR is in all jax releases we care about:
+        # https://github.com/openxla/xla/pull/7875
+        half_rot_tensor = ops.stack((-x2, x1), axis=-2)
+        half_rot_tensor = ops.reshape(half_rot_tensor, ops.shape(tensor))
+        return (tensor * cos_emb) + (half_rot_tensor * sin_emb)
+
+    def _compute_positions(self, inputs, start_index=0):
+        seq_len = ops.shape(inputs)[1]
+        positions = ops.arange(seq_len, dtype="float32")
+        return positions + ops.cast(start_index, dtype="float32")
+
+    def _compute_cos_sin_embedding(self, inputs, start_index=0, positions=None):
+        """Compute cos & sin RoPE embeddings with optional YaRN scaling.
+        Uses tensor ops only to remain JIT/backends friendly.
+        """
+        batch_axis = 0
+        sequence_axis = 1
+        feature_axis = len(inputs.shape) - 1
+
+        rotary_dim = ops.shape(inputs)[feature_axis]
+        inverse_freq = self._get_inverse_freq(rotary_dim)
+
+        if positions is None:
+            positions = self._compute_positions(inputs, start_index)
+            positions = ops.expand_dims(
+                positions, axis=batch_axis
+            )  # shape (1, seq_len)
+        else:
+            positions = ops.cast(positions, "float32")
+            if len(ops.shape(positions)) == 1:
+                positions = ops.expand_dims(positions, axis=batch_axis)
+
+        if (
+            self.rope_type == "yarn"
+            and self.truncate
+            and self.original_max_position_embeddings is not None
+        ):
+            positions = ops.minimum(
+                positions,
+                ops.cast(self.original_max_position_embeddings, "float32"),
+            )
+
+        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)
+        )
+
+        for axis in range(len(inputs.shape)):
+            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)
+        sin_emb = ops.cast(ops.sin(embedding), self.compute_dtype)
+
+        if self.rope_type == "yarn":
+            # YaRN temperature scaling
+            factor = ops.add(
+                ops.multiply(
+                    ops.cast(0.1, self.compute_dtype),
+                    ops.log(ops.cast(self.scaling_factor, self.compute_dtype)),
+                ),
+                ops.cast(1.0, self.compute_dtype),
+            )
+            cos_emb = cos_emb * factor
+            sin_emb = sin_emb * factor
+        return cos_emb, sin_emb
+
+    def _get_inverse_freq(self, rotary_dim):
+        """Return inverse frequencies."""
+        idx = ops.arange(0, rotary_dim, 2, dtype="float32")
+        denom = ops.cast(rotary_dim, "float32")
+        freq_range = idx / denom
+        inv = ops.power(ops.cast(self.max_wavelength, "float32"), -freq_range)
+
+        if self.rope_type == "linear":
+            return inv / ops.cast(self.scaling_factor, "float32")
+        elif self.rope_type == "dynamic":
+            exponent = ops.cast(rotary_dim, "float32") / ops.cast(
+                max(1, rotary_dim - 2), "float32"
+            )
+            return inv / ops.power(
+                ops.cast(self.scaling_factor, "float32"), exponent
+            )
+        elif self.rope_type == "yarn":
+            return self._get_yarn_inverse_freq(rotary_dim)
+        else:
+            return inv
+
+    def _get_yarn_inverse_freq(self, rotary_dim):
+        # Get the base (rope_theta equivalent) from max_wavelength
+        base = ops.cast(self.max_wavelength, "float32")
+
+        # Compute base frequencies: base ** (idx / dim)
+        idx = ops.arange(0, rotary_dim, 2, dtype="float32")
+        pos_freqs = ops.power(base, idx / ops.cast(rotary_dim, "float32"))
+
+        # Compute interpolation and extrapolation frequencies
+        inv_freq_extrapolation = 1.0 / pos_freqs
+        inv_freq_interpolation = 1.0 / (
+            ops.cast(self.scaling_factor, "float32") * pos_freqs
+        )
+
+        # Find correction range
+        beta_fast = ops.cast(self.beta_fast, "float32")
+        beta_slow = ops.cast(self.beta_slow, "float32")
+
+        # Find correction dimensions for beta_fast and beta_slow
+        def find_correction_dim_tensor(num_rotations, dim):
+            max_pos = ops.cast(self.original_max_position_embeddings, "float32")
+            return (dim * ops.log(max_pos / (num_rotations * 2 * np.pi))) / (
+                2 * ops.log(base)
+            )
+
+        low = find_correction_dim_tensor(
+            beta_fast, ops.cast(rotary_dim, "float32")
+        )
+        high = find_correction_dim_tensor(
+            beta_slow, ops.cast(rotary_dim, "float32")
+        )
+
+        # Apply truncation if specified
+        if self.truncate:
+            low = ops.floor(low)
+            high = ops.ceil(high)
+
+        # Clamp to valid range
+        low = ops.maximum(low, ops.cast(0, "float32"))
+        high = ops.minimum(high, ops.cast(rotary_dim // 2 - 1, "float32"))
+
+        # Linear ramp function
+        dim_half = rotary_dim // 2
+        idx_half = ops.arange(0, dim_half, dtype="float32")
+
+        # Prevent singularity
+        diff = high - low
+        diff = ops.maximum(diff, ops.cast(0.001, "float32"))
+
+        linear_func = (idx_half - low) / diff
+        ramp_func = ops.clip(linear_func, 0, 1)
+
+        # Apply the ramp to get extrapolation factor
+        inv_freq_extrapolation_factor = 1 - ramp_func
+
+        # Combine interpolation and extrapolation
+        scaled_inverse_freq = (
+            inv_freq_interpolation * (1 - inv_freq_extrapolation_factor)
+            + inv_freq_extrapolation * inv_freq_extrapolation_factor
+        )
+
+        return scaled_inverse_freq
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "max_wavelength": self.max_wavelength,
+                "scaling_factor": self.scaling_factor,
+                "rope_type": self.rope_type,
+                "beta_fast": self.beta_fast,
+                "beta_slow": self.beta_slow,
+                "original_max_position_embeddings": (
+                    self.original_max_position_embeddings
+                ),
+                "truncate": self.truncate,
+                "sequence_axis": self.sequence_axis,
+                "feature_axis": self.feature_axis,
+            }
+        )
+        return config
+
+    def compute_output_shape(self, input_shape):
+        return input_shape

{keras_hub_nightly-0.24.0.dev202511160427 → keras_hub_nightly-0.26.0.dev202601130438}/keras_hub/src/layers/modeling/token_and_position_embedding.py

@@ -1,10 +1,10 @@
 import keras
+from keras.layers import ReversibleEmbedding
+from keras.src.backend import get_keras_mask
+from keras.src.backend import set_keras_mask
 
 from keras_hub.src.api_export import keras_hub_export
 from keras_hub.src.layers.modeling.position_embedding import PositionEmbedding
-from keras_hub.src.layers.modeling.reversible_embedding import (
-    ReversibleEmbedding,
-)
 from keras_hub.src.utils.keras_utils import clone_initializer
 
 
@@ -128,10 +128,10 @@ class TokenAndPositionEmbedding(keras.layers.Layer):
             positions=positions,
         )
         outputs = embedded_tokens + embedded_positions
+        mask = get_keras_mask(embedded_tokens)
+        if mask is not None:
+            set_keras_mask(outputs, mask)
         return outputs
 
-    def compute_mask(self, inputs, mask=None):
-        return self.token_embedding.compute_mask(inputs, mask=mask)
-
     def compute_output_shape(self, input_shape):
         return tuple(input_shape) + (self.embedding_dim,)

{keras_hub_nightly-0.24.0.dev202511160427 → keras_hub_nightly-0.26.0.dev202601130438}/keras_hub/src/layers/modeling/transformer_layer_utils.py

@@ -1,11 +1,12 @@
 from absl import logging
 from keras import ops
+from keras.src.backend import get_keras_mask
 
 
 def _check_masks_shapes(inputs, padding_mask, attention_mask):
     mask = padding_mask
-    if hasattr(inputs, "_keras_mask") and mask is None:
-        mask = inputs._keras_mask
+    if mask is None:
+        mask = get_keras_mask(inputs)
     if mask is not None:
         if len(mask.shape) != 2:
             raise ValueError(
@@ -68,17 +69,16 @@ def merge_padding_and_attention_mask(
         returned mask is padding_mask with one additional axis.
     """
     _check_masks_shapes(inputs, padding_mask, attention_mask)
-    mask = padding_mask
-    if hasattr(inputs, "_keras_mask"):
-        if mask is None:
-            # If no padding mask is explicitly provided, we look for padding
-            # mask from the input data.
-            mask = inputs._keras_mask
-        else:
+    # We look for a padding mask from the input data.
+    mask = get_keras_mask(inputs)
+    # But if padding mask is explicitly provided, we use it.
+    if padding_mask is not None:
+        if mask is not None:
             logging.warning(
                 "You are explicitly setting `padding_mask` while the `inputs` "
                 "have built-in mask, so the built-in mask is ignored."
             )
+        mask = padding_mask
     if mask is not None:
         # Add an axis for broadcasting, the attention mask should be 2D
         # (not including the batch axis).

{keras_hub_nightly-0.24.0.dev202511160427 → keras_hub_nightly-0.26.0.dev202601130438}/keras_hub/src/models/albert/albert_backbone.py

@@ -1,10 +1,8 @@
 import keras
+from keras.layers import ReversibleEmbedding
 
 from keras_hub.src.api_export import keras_hub_export
 from keras_hub.src.layers.modeling.position_embedding import PositionEmbedding
-from keras_hub.src.layers.modeling.reversible_embedding import (
-    ReversibleEmbedding,
-)
 from keras_hub.src.layers.modeling.transformer_encoder import TransformerEncoder
 from keras_hub.src.models.backbone import Backbone
 from keras_hub.src.utils.keras_utils import gelu_approximate

{keras_hub_nightly-0.24.0.dev202511160427 → keras_hub_nightly-0.26.0.dev202601130438}/keras_hub/src/models/backbone.py

@@ -107,6 +107,9 @@ class Backbone(keras.Model):
     def from_config(cls, config):
         # The default `from_config()` for functional models will return a
         # vanilla `keras.Model`. We override it to get a subclass instance back.
+        config = config.copy()
+        if "dtype" in config and isinstance(config["dtype"], dict):
+            config["dtype"] = keras.dtype_policies.get(config["dtype"])
         return cls(**config)
 
     @classproperty

{keras_hub_nightly-0.24.0.dev202511160427 → keras_hub_nightly-0.26.0.dev202601130438}/keras_hub/src/models/bart/bart_backbone.py

@@ -1,10 +1,8 @@
 import keras
+from keras.layers import ReversibleEmbedding
 
 from keras_hub.src.api_export import keras_hub_export
 from keras_hub.src.layers.modeling.position_embedding import PositionEmbedding
-from keras_hub.src.layers.modeling.reversible_embedding import (
-    ReversibleEmbedding,
-)
 from keras_hub.src.layers.modeling.transformer_decoder import TransformerDecoder
 from keras_hub.src.layers.modeling.transformer_encoder import TransformerEncoder
 from keras_hub.src.models.backbone import Backbone

{keras_hub_nightly-0.24.0.dev202511160427 → keras_hub_nightly-0.26.0.dev202601130438}/keras_hub/src/models/bert/bert_backbone.py

@@ -1,10 +1,8 @@
 import keras
+from keras.layers import ReversibleEmbedding
 
 from keras_hub.src.api_export import keras_hub_export
 from keras_hub.src.layers.modeling.position_embedding import PositionEmbedding
-from keras_hub.src.layers.modeling.reversible_embedding import (
-    ReversibleEmbedding,
-)
 from keras_hub.src.layers.modeling.transformer_encoder import TransformerEncoder
 from keras_hub.src.models.backbone import Backbone
 from keras_hub.src.utils.keras_utils import gelu_approximate

{keras_hub_nightly-0.24.0.dev202511160427 → keras_hub_nightly-0.26.0.dev202601130438}/keras_hub/src/models/bloom/bloom_backbone.py

@@ -1,9 +1,7 @@
 import keras
+from keras.layers import ReversibleEmbedding
 
 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.bloom.bloom_decoder import BloomDecoder
 

{keras_hub_nightly-0.24.0.dev202511160427 → keras_hub_nightly-0.26.0.dev202601130438}/keras_hub/src/models/causal_lm.py

@@ -196,7 +196,7 @@ class CausalLM(Task):
 
                 # Create an explicit tuple of all variable state.
                 state = (
-                    self.sampler.variables,
+                    [v.value for v in self.sampler.variables],
                     # Use the explicit variable.value to preserve the
                     # sharding spec of distribution.
                     [v.value for v in self.trainable_variables],
@@ -429,3 +429,25 @@ class CausalLM(Task):
         super()._post_quantize(mode, **kwargs)
         # Reset the compiled generate function.
         self.generate_function = None
+
+    def get_quantization_layer_structure(self, mode):
+        if mode not in ["gptq", "awq"]:
+            return None
+
+        backbone = self.backbone
+        # Check for standard backbone structure.
+        if not hasattr(backbone, "transformer_layers"):
+            return None
+
+        # Check for embedding.
+        embedding = getattr(backbone, "token_embedding", None)
+        if embedding is None:
+            embedding = getattr(backbone, "embedding", None)
+
+        if embedding is None:
+            return None
+
+        return {
+            "pre_block_layers": [embedding],
+            "sequential_blocks": backbone.transformer_layers,
+        }

{keras_hub_nightly-0.24.0.dev202511160427 → keras_hub_nightly-0.26.0.dev202601130438}/keras_hub/src/models/deberta_v3/deberta_v3_backbone.py

@@ -1,9 +1,7 @@
 import keras
+from keras.layers import ReversibleEmbedding
 
 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.deberta_v3.disentangled_attention_encoder import (
     DisentangledAttentionEncoder,