keras-hub 0.25.0.dev0__py3-none-any.whl → 0.26.0.dev0__py3-none-any.whl

keras_hub/src/models/gemma3/{gemma3_interleave_embeddings.py → gemma3_layers.py}

@@ -2,6 +2,107 @@ import keras
 from keras import ops
 
 
+class RMSNormalization(keras.layers.Layer):
+    def __init__(self, epsilon=1e-6, **kwargs):
+        super().__init__(**kwargs)
+        self.epsilon = epsilon
+
+    def build(self, input_shape):
+        self.scale = self.add_weight(
+            name="scale",
+            trainable=True,
+            shape=(input_shape[-1],),
+            initializer="zeros",
+        )
+        self.built = True
+
+    def call(self, x):
+        # Always compute normalization in float32.
+        x = ops.cast(x, "float32")
+        scale = ops.cast(self.scale, "float32")
+        var = ops.mean(ops.square(x), axis=-1, keepdims=True)
+        normed_inputs = x * ops.reciprocal(ops.sqrt(var + self.epsilon))
+        normed_inputs = normed_inputs * (1 + scale)
+        return ops.cast(normed_inputs, self.compute_dtype)
+
+
+class Gemma3MeanPooling(keras.layers.Layer):
+    """Mean pooling layer that computes the average of token embeddings.
+
+    This layer correctly handles variable-length sequences by ignoring
+    padded tokens in the mean calculation, using a `padding_mask`.
+
+    Example:
+    ```python
+    import numpy as np
+
+    sequence_output = np.random.rand(2, 4, 8).astype("float32")
+    padding_mask = np.array([[1, 1, 1, 0], [1, 1, 0, 0]], dtype="int32")
+    mean_pool_layer = Gemma3MeanPooling()
+    pooled = mean_pool_layer([sequence_output, padding_mask])
+    # pooled.shape -> (2, 8)
+    ```
+    """
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self.supports_masking = True
+
+    def call(self, inputs, padding_mask=None):
+        """Performs masked mean pooling on the token embeddings.
+
+        Args:
+            inputs: The sequence of embeddings to pool, with a shape of
+                `(batch_size, seq_len, hidden_dim)`.
+            padding_mask: The mask indicating valid tokens, with a shape of
+                `(batch_size, seq_len)`.
+
+        Returns:
+            A tensor representing the pooled embeddings, with a shape of
+            `(batch_size, hidden_dim)`.
+        """
+        if padding_mask is None:
+            inputs, padding_mask = inputs
+
+        sequence_output = inputs
+        mask = ops.expand_dims(
+            ops.cast(padding_mask, sequence_output.dtype), axis=-1
+        )
+
+        masked_output = sequence_output * mask
+
+        sum_embeddings = ops.sum(masked_output, axis=1)
+
+        num_tokens = ops.sum(
+            ops.cast(padding_mask, sequence_output.dtype), axis=1
+        )
+        num_tokens = ops.expand_dims(num_tokens, axis=1)
+        # Avoid division by zero
+        num_tokens = ops.maximum(num_tokens, 1e-9)
+
+        mean_embeddings = sum_embeddings / num_tokens
+        return ops.cast(mean_embeddings, self.compute_dtype)
+
+    def compute_output_shape(self, input_shape):
+        """Computes the output shape of the layer.
+
+        Args:
+            input_shape: A tuple or list of tuples representing input shapes.
+
+        Returns:
+            A tuple representing the output shape.
+        """
+        if isinstance(input_shape, list):
+            sequence_output_shape = input_shape[0]
+        else:
+            sequence_output_shape = input_shape
+        return sequence_output_shape[:-2] + (sequence_output_shape[-1],)
+
+    def get_config(self):
+        """Returns the config of the layer."""
+        return super().get_config()
+
+
 class Gemma3InterleaveEmbeddings(keras.layers.Layer):
     """Places image embeddings in the correct position in an embedding sequence.
 

keras_hub/src/models/gemma3/gemma3_presets.py

@@ -181,12 +181,25 @@ backbone_presets = {
         },
         "kaggle_handle": "kaggle://keras/gemma3/keras/gemma3_instruct_270m/4",
     },
+    "medgemma_4b": {
+        "metadata": {
+            "description": (
+                "A 4 billion parameter model based on Gemma 3. "
+                "This model is pre-trained for performance on medical text "
+                "and image comprehension and is optimized for medical "
+                "applications that involve a text generation component."
+            ),
+            "params": 4300079472,
+            "path": "gemma3",
+        },
+        "kaggle_handle": "kaggle://keras/medgemma/keras/medgemma_4b/1",
+    },
     "medgemma_instruct_4b": {
         "metadata": {
             "description": (
                 "A 4 billion parameter model based on Gemma 3. "
-                "This model is trained for performance on medical text"
-                "and image comprehension and is optimized for medical"
+                "This model is instruction-tuned for performance on medical "
+                "text and image comprehension and is optimized for medical "
                 "applications that involve a text generation component."
             ),
             "params": 4300079472,
@@ -198,8 +211,8 @@ backbone_presets = {
         "metadata": {
             "description": (
                 "A 27 billion parameter model based on Gemma 3. "
-                "This model trained for performance on medical text "
-                "and image comprehension and is optimized for medical "
+                "This model is instruction-tuned for performance on medical "
+                " text and image comprehension and is optimized for medical "
                 "applications that involve a text generation component."
             ),
             "params": 27432406640,
@@ -211,13 +224,72 @@ backbone_presets = {
         "metadata": {
             "description": (
                 "A 27 billion parameter text-only model based on Gemma 3. "
-                "This model is trained for performance on medical text "
-                "comprehension and is optimized for medical applications "
-                "that involve a text generation component."
+                "This model is instruction-tuned (No images) for performance "
+                "on medical text comprehension and is optimized for medical "
+                "applications that involve a text generation component."
             ),
             "params": 27009002240,
             "path": "gemma3",
         },
         "kaggle_handle": "kaggle://keras/medgemma/keras/medgemma_instruct_27b_text/1",
     },
+    "medgemma_1.5_instruct_4b": {
+        "metadata": {
+            "description": (
+                "A 4 billion parameter,Instruct-tuned MedGemma 1.5 4B is an "
+                "updated version of the Instruction-tuned MedGemma 4B model."
+            ),
+            "params": 4300079472,
+            "path": "gemma3",
+        },
+        "kaggle_handle": "kaggle://keras/medgemma/keras/medgemma_1.5_instruct_4b/1",
+    },
+    "function_gemma_instruct_270m": {
+        "metadata": {
+            "description": (
+                "A 270M Million parameter text-only model based on Gemma 3. "
+                "This model is trained specifically for function calling "
+                "improvements."
+            ),
+            "params": 268098176,
+            "path": "gemma3",
+        },
+        "kaggle_handle": "kaggle://keras/function-gemma/keras/function_gemma_instruct_270m/1",
+    },
+    "translategemma_4b_it": {
+        "metadata": {
+            "description": (
+                "4 billion parameter, 34-layer, multimodal instruction-tuned "
+                "translation model based on Gemma 3. Supports text and image "
+                "input for translation across 55 languages."
+            ),
+            "params": 4299915632,
+            "path": "gemma3",
+        },
+        "kaggle_handle": "kaggle://keras/translategemma/keras/translategemma_4b_it/1",
+    },
+    "translategemma_12b_it": {
+        "metadata": {
+            "description": (
+                "12 billion parameter, 48-layer, multimodal instruction-tuned "
+                "translation model based on Gemma 3. Supports text and image "
+                "input for translation across 55 languages."
+            ),
+            "params": 12187079280,
+            "path": "gemma3",
+        },
+        "kaggle_handle": "kaggle://keras/translategemma/keras/translategemma_12b_it/1",
+    },
+    "translategemma_27b_it": {
+        "metadata": {
+            "description": (
+                "27 billion parameter, 62-layer, multimodal instruction-tuned "
+                "translation model based on Gemma 3. Supports text and image "
+                "input for translation across 55 languages."
+            ),
+            "params": 27432062576,
+            "path": "gemma3",
+        },
+        "kaggle_handle": "kaggle://keras/translategemma/keras/translategemma_27b_it/1",
+    },
 }

keras_hub/src/models/gemma3/gemma3_vision_encoder.py

@@ -2,7 +2,7 @@ import keras
 from keras import ops
 
 from keras_hub.src.api_export import keras_hub_export
-from keras_hub.src.models.gemma.rms_normalization import RMSNormalization
+from keras_hub.src.models.gemma3.gemma3_layers import RMSNormalization
 from keras_hub.src.utils.keras_utils import clone_initializer
 
 

keras_hub/src/models/gpt2/gpt2_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.models.backbone import Backbone
 from keras_hub.src.utils.keras_utils import gelu_approximate

keras_hub/src/models/gpt2/gpt2_causal_lm.py

@@ -422,7 +422,7 @@ class GPT2CausalLM(CausalLM):
         return per_token_loss
 
     def get_quantization_layer_structure(self, mode):
-        if mode != "gptq":
+        if mode not in ["gptq", "awq"]:
             return None
 
         backbone = self.backbone

keras_hub/src/models/gpt_neo_x/gpt_neo_x_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.gpt_neo_x.gpt_neo_x_decoder import GPTNeoXDecoder
 from keras_hub.src.utils.keras_utils import gelu_approximate

keras_hub/src/models/gpt_oss/gpt_oss_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.gpt_oss.gpt_oss_decoder import (
     GptOssTransformerDecoder,

keras_hub/src/models/llama/llama_backbone.py

@@ -1,10 +1,8 @@
 import keras
 from keras import ops
+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.llama.llama_decoder import LlamaTransformerDecoder
 from keras_hub.src.models.llama.llama_layernorm import LlamaLayerNorm

keras_hub/src/models/masked_lm.py

@@ -86,7 +86,7 @@ class MaskedLM(Task):
         )
 
     def get_quantization_layer_structure(self, mode):
-        if mode != "gptq":
+        if mode not in ["gptq", "awq"]:
             return None
 
         backbone = self.backbone

keras_hub/src/models/mistral/mistral_backbone.py

@@ -1,10 +1,8 @@
 import keras
 from keras import ops
+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.mistral.mistral_layer_norm import (
     MistralLayerNormalization,

keras_hub/src/models/mixtral/mixtral_backbone.py

@@ -1,10 +1,8 @@
 import keras
 from keras import ops
+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.mixtral.mixtral_decoder import (
     MixtralTransformerDecoder,

keras_hub/src/models/moonshine/moonshine_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.moonshine.moonshine_decoder import (
     MoonshineDecoderBlock,

keras_hub/src/models/pali_gemma/pali_gemma_backbone.py

@@ -1,10 +1,8 @@
 import keras
 from keras import ops
+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.gemma.rms_normalization import RMSNormalization
 from keras_hub.src.models.pali_gemma.pali_gemma_decoder_block import (

keras_hub/src/models/parseq/parseq_tokenizer.py

@@ -13,9 +13,11 @@ from keras_hub.src.utils.tensor_utils import preprocessing_function
 
 try:
     import tensorflow as tf
-    import tensorflow_text as tf_text
 except ImportError:
     tf = None
+try:
+    import tensorflow_text as tf_text
+except ImportError:
     tf_text = None
 
 PARSEQ_VOCAB = list(

keras_hub/src/models/phi3/phi3_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.phi3.phi3_decoder import Phi3Decoder
 from keras_hub.src.models.phi3.phi3_layernorm import Phi3LayerNorm

keras_hub/src/models/qwen/qwen_backbone.py

@@ -1,10 +1,8 @@
 import keras
 from keras import ops
+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.qwen.qwen_decoder import QwenTransformerDecoder
 from keras_hub.src.models.qwen.qwen_layernorm import QwenLayerNorm

keras_hub/src/models/qwen/qwen_presets.py

@@ -1,6 +1,7 @@
 """Qwen preset configurations."""
 
 backbone_presets = {
+    # Qwen 2.5 Models
     "qwen2.5_0.5b_en": {
         "metadata": {
             "description": ("24-layer Qwen model with 0.5 billion parameters."),
@@ -58,4 +59,212 @@ backbone_presets = {
         },
         "kaggle_handle": "kaggle://keras/qwen/keras/qwen2.5_instruct_72b_en/2",
     },
+    # Qwen 2.5 Coder Models
+    "qwen2.5_coder_0.5b": {
+        "metadata": {
+            "description": (
+                "Code-focused fine-tuned Qwen-2.5 model with 0.5 "
+                "billion parameters."
+            ),
+            "params": 494032768,
+            "path": "qwen",
+        },
+        "kaggle_handle": (
+            "kaggle://keras/qwen2-5-coder/keras/qwen2.5_coder_0.5b/1"
+        ),
+    },
+    "qwen2.5_coder_1.5b": {
+        "metadata": {
+            "description": (
+                "Code-focused fine-tuned 28-layer Qwen-2.5 model with 1.5 "
+                "billion parameters."
+            ),
+            "params": 1543434240,
+            "path": "qwen",
+        },
+        "kaggle_handle": (
+            "kaggle://keras/qwen2-5-coder/keras/qwen2.5_coder_1.5b/1"
+        ),
+    },
+    "qwen2.5_coder_3b": {
+        "metadata": {
+            "description": (
+                "Code-focused fine-tuned Qwen-2.5 model with 3 "
+                "billion parameters."
+            ),
+            "params": 3085938688,
+            "path": "qwen",
+        },
+        "kaggle_handle": (
+            "kaggle://keras/qwen2-5-coder/keras/qwen2.5_coder_3b/1"
+        ),
+    },
+    "qwen2.5_coder_7b": {
+        "metadata": {
+            "description": (
+                "Code-focused fine-tuned Qwen-2.5 model with 7 "
+                "billion parameters."
+            ),
+            "params": 6993420288,
+            "path": "qwen",
+        },
+        "kaggle_handle": (
+            "kaggle://keras/qwen2-5-coder/keras/qwen2.5_coder_7b/1"
+        ),
+    },
+    "qwen2.5_coder_14b": {
+        "metadata": {
+            "description": (
+                "Code-focused fine-tuned Qwen-2.5 model with 14 "
+                "billion parameters."
+            ),
+            "params": 14000000000,
+            "path": "qwen",
+        },
+        "kaggle_handle": (
+            "kaggle://keras/qwen2-5-coder/keras/qwen2.5_coder_14b/1"
+        ),
+    },
+    "qwen2.5_coder_32b": {
+        "metadata": {
+            "description": (
+                "Code-focused fine-tuned Qwen-2.5 model with 32 "
+                "billion parameters."
+            ),
+            "params": 32763876352,
+            "path": "qwen",
+        },
+        "kaggle_handle": (
+            "kaggle://keras/qwen2-5-coder/keras/qwen2.5_coder_32b/1"
+        ),
+    },
+    "qwen2.5_coder_instruct_0.5b": {
+        "metadata": {
+            "description": (
+                "Instruction-tuned code-focused Qwen-2.5 model with "
+                "0.5 billion parameters."
+            ),
+            "params": 494032768,
+            "path": "qwen",
+        },
+        "kaggle_handle": (
+            "kaggle://keras/qwen2-5-coder/keras/qwen2.5_coder_instruct_0.5b/1"
+        ),
+    },
+    "qwen2.5_coder_instruct_1.5b": {
+        "metadata": {
+            "description": (
+                "Instruction-tuned code-focused Qwen-2.5 model with "
+                "1.5 billion parameters."
+            ),
+            "params": 1543434240,
+            "path": "qwen",
+        },
+        "kaggle_handle": (
+            "kaggle://keras/qwen2-5-coder/keras/qwen2.5_coder_instruct_1.5b/1"
+        ),
+    },
+    "qwen2.5_coder_instruct_3b": {
+        "metadata": {
+            "description": (
+                "Instruction-tuned code-focused Qwen-2.5 model with "
+                "3 billion parameters."
+            ),
+            "params": 3085938688,
+            "path": "qwen",
+        },
+        "kaggle_handle": (
+            "kaggle://keras/qwen2-5-coder/keras/qwen2.5_coder_instruct_3b/1"
+        ),
+    },
+    "qwen2.5_coder_instruct_7b": {
+        "metadata": {
+            "description": (
+                "Instruction-tuned code-focused Qwen-2.5 model with "
+                "7 billion parameters."
+            ),
+            "params": 6993420288,
+            "path": "qwen",
+        },
+        "kaggle_handle": (
+            "kaggle://keras/qwen2-5-coder/keras/qwen2.5_coder_instruct_7b/1"
+        ),
+    },
+    "qwen2.5_coder_instruct_14b": {
+        "metadata": {
+            "description": (
+                "Instruction-tuned code-focused Qwen-2.5 model with "
+                "14 billion parameters."
+            ),
+            "params": 14000000000,
+            "path": "qwen",
+        },
+        "kaggle_handle": (
+            "kaggle://keras/qwen2-5-coder/keras/qwen2.5_coder_instruct_14b/1"
+        ),
+    },
+    "qwen2.5_coder_instruct_32b": {
+        "metadata": {
+            "description": (
+                "Instruction-tuned code-focused Qwen-2.5 model with "
+                "32 billion parameters."
+            ),
+            "params": 32763876352,
+            "path": "qwen",
+        },
+        "kaggle_handle": (
+            "kaggle://keras/qwen2-5-coder/keras/qwen2.5_coder_instruct_32b/1"
+        ),
+    },
+    # Qwen 2.5 Math Models
+    "qwen2.5_math_1.5b_en": {
+        "metadata": {
+            "description": (
+                "Math-focused Qwen-2.5 model with 1.5 billion parameters."
+            ),
+            "params": 1543714304,
+            "path": "qwen",
+        },
+        "kaggle_handle": (
+            "kaggle://keras/qwen2-5-math/keras/qwen2.5_math_1.5b_en/1"
+        ),
+    },
+    "qwen2.5_math_instruct_1.5b_en": {
+        "metadata": {
+            "description": (
+                "Instruction-tuned math-focused Qwen-2.5 model with "
+                "1.5 billion parameters."
+            ),
+            "params": 1543714304,
+            "path": "qwen",
+        },
+        "kaggle_handle": (
+            "kaggle://keras/qwen2-5-math/keras/qwen2.5_math_instruct_1.5b_en/1"
+        ),
+    },
+    "qwen2.5_math_7b_en": {
+        "metadata": {
+            "description": (
+                "Math-focused Qwen-2.5 model with 7 billion parameters."
+            ),
+            "params": 7615616512,
+            "path": "qwen",
+        },
+        "kaggle_handle": (
+            "kaggle://keras/qwen2-5-math/keras/qwen2.5_math_7b_en/1"
+        ),
+    },
+    "qwen2.5_math_instruct_7b_en": {
+        "metadata": {
+            "description": (
+                "Instruction-tuned math-focused Qwen-2.5 model with "
+                "7 billion parameters."
+            ),
+            "params": 7615616512,
+            "path": "qwen",
+        },
+        "kaggle_handle": (
+            "kaggle://keras/qwen2-5-math/keras/qwen2.5_math_instruct_7b_en/1"
+        ),
+    },
 }

keras_hub/src/models/qwen3/qwen3_backbone.py

@@ -1,10 +1,8 @@
 import keras
 from keras import ops
+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.qwen3.qwen3_decoder import Qwen3TransformerDecoder
 from keras_hub.src.models.qwen3.qwen3_layernorm import Qwen3LayerNorm

keras_hub/src/models/qwen3_moe/qwen3_moe_backbone.py

@@ -1,10 +1,8 @@
 import keras
 from keras import ops
+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.qwen3_moe.qwen3_moe_decoder import (
     Qwen3MoeTransformerDecoder,

keras_hub/src/models/qwen3_moe/qwen3_moe_presets.py

@@ -1,6 +1,7 @@
 """Qwen3 MoE model preset configurations."""
 
 backbone_presets = {
+    #  Qwen-3 MoE Models
     "qwen3_moe_30b_a3b_en": {
         "metadata": {
             "description": (
@@ -27,4 +28,18 @@ backbone_presets = {
         },
         "kaggle_handle": "kaggle://keras/qwen-3-moe/keras/qwen3_moe_235b_a22b_en/1",
     },
+    # Qwen-3 Coder MoE Models
+    "qwen3_coder_instruct_30b_a3b_en": {
+        "metadata": {
+            "description": (
+                "A Code-Specific Model,Mixture-of-Experts (MoE) model "
+                "has 30.5 billion total parameters with 3.3 billion "
+                "activated, built on 48 layers and utilizes 32 query "
+                "and 4 key/value attention heads with 128 experts (8 active)."
+            ),
+            "params": 30532122624,
+            "path": "qwen3_moe",
+        },
+        "kaggle_handle": "kaggle://keras/qwen3-coder/keras/qwen3_coder_instruct_30b_a3b_en/1",
+    },
 }

keras_hub/src/models/qwen_moe/qwen_moe_backbone.py

@@ -1,10 +1,8 @@
 import keras
 from keras import ops
+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.qwen.qwen_layernorm import QwenLayerNorm
 from keras_hub.src.models.qwen_moe.qwen_moe_decoder import (