keras-hub-nightly 0.23.0.dev202509190415__py3-none-any.whl → 0.23.0.dev202509280419__py3-none-any.whl

keras_hub/src/models/qwen3_moe/qwen3_moe_layernorm.py

@@ -0,0 +1,45 @@
+import keras
+from keras import ops
+
+
+class Qwen3MoeLayerNorm(keras.layers.Layer):
+    """A normalization layer for Qwen that implements RMS normalization.
+
+    Args:
+        head_dim: int. The dimension of each attention head, used for per-head
+            normalization. Defaults to `None`.
+        epsilon: float. A small float added to variance to avoid dividing by
+            zero. Defaults to `1e-6`.
+    """
+
+    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, "head_dim": self.head_dim})
+        return config

keras_hub/src/models/qwen3_moe/qwen3_moe_tokenizer.py

@@ -0,0 +1,48 @@
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.qwen3_moe.qwen3_moe_backbone import Qwen3MoeBackbone
+from keras_hub.src.tokenizers.byte_pair_tokenizer import BytePairTokenizer
+
+
+@keras_hub_export(
+    "keras_hub.tokenizers.Qwen3MoeTokenizer",
+)
+class Qwen3MoeTokenizer(BytePairTokenizer):
+    """Tokenizer for Qwen Moe model.
+
+    This tokenizer implements byte-pair encoding (BPE) for Qwen 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 = Qwen3MoeBackbone
+
+    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/tests/test_case.py

@@ -538,10 +538,11 @@ class TestCase(tf.test.TestCase, parameterized.TestCase):
 
             self.assertIsInstance(output_data, dict)
             self.assertEqual(
-                list(output_data.keys()), list(backbone.pyramid_outputs.keys())
+                sorted(output_data.keys()),
+                sorted(backbone.pyramid_outputs.keys()),
             )
             self.assertEqual(
-                list(output_data.keys()), expected_pyramid_output_keys
+                sorted(output_data.keys()), sorted(expected_pyramid_output_keys)
             )
             # check height and width of each level.
             for i, (k, v) in enumerate(output_data.items()):

keras_hub/src/utils/transformers/convert_dinov2.py

@@ -29,6 +29,7 @@ def convert_backbone_config(transformers_config):
         "image_shape": (image_size, image_size, 3),
         "position_embedding_shape": (image_size, image_size),
         "antialias_in_interpolation": antialias_in_interpolation,
+        "apply_layernorm": transformers_config.get("apply_layernorm", False),
     }
 
 

keras_hub/src/utils/transformers/convert_qwen3_moe.py

@@ -0,0 +1,216 @@
+import numpy as np
+
+from keras_hub.src.models.qwen3_moe.qwen3_moe_backbone import Qwen3MoeBackbone
+from keras_hub.src.utils.preset_utils import load_json
+
+backbone_cls = Qwen3MoeBackbone
+
+
+def convert_backbone_config(transformers_config):
+    return {
+        "vocabulary_size": transformers_config["vocab_size"],
+        "hidden_dim": transformers_config["hidden_size"],
+        "head_dim": transformers_config["head_dim"],
+        "num_layers": transformers_config["num_hidden_layers"],
+        "num_query_heads": transformers_config["num_attention_heads"],
+        "num_key_value_heads": transformers_config["num_key_value_heads"],
+        "intermediate_dim": transformers_config["intermediate_size"],
+        "moe_intermediate_dim": transformers_config["moe_intermediate_size"],
+        "num_experts": transformers_config["num_experts"],
+        "top_k": transformers_config["num_experts_per_tok"],
+        "norm_top_k_prob": transformers_config["norm_topk_prob"],
+        "decoder_sparse_step": transformers_config["decoder_sparse_step"],
+        "layer_norm_epsilon": transformers_config["rms_norm_eps"],
+        "rope_max_wavelength": transformers_config["rope_theta"],
+        "sliding_window_size": transformers_config["sliding_window"],
+        "router_aux_loss_coefficient": transformers_config[
+            "router_aux_loss_coef"
+        ],
+        "tie_word_embeddings": transformers_config.get(
+            "tie_word_embeddings", False
+        ),
+    }
+
+
+def convert_weights(backbone, loader, transformers_config):
+    loader.port_weight(
+        keras_variable=backbone.get_layer("token_embedding").embeddings,
+        hf_weight_key="model.embed_tokens.weight",
+    )
+    if not backbone.tie_word_embeddings:
+        loader.port_weight(
+            keras_variable=backbone.get_layer(
+                "token_embedding"
+            ).reverse_embeddings,
+            hf_weight_key="lm_head.weight",
+            # rearrange_pattern="b a -> a b",
+            hook_fn=lambda hf_tensor, _: np.transpose(hf_tensor, axes=(1, 0)),
+        )
+
+    def transpose_and_reshape(x, shape):
+        return np.reshape(np.transpose(x), shape)
+
+    for i in range(backbone.num_layers):
+        decoder_layer = backbone.get_layer(f"transformer_layer_{i}")
+
+        # Input layernorm
+        loader.port_weight(
+            keras_variable=decoder_layer._self_attention_layernorm.scale,
+            hf_weight_key=f"model.layers.{i}.input_layernorm.weight",
+        )
+
+        # Attention layers
+
+        ## Query
+        loader.port_weight(
+            keras_variable=decoder_layer._self_attention_layer._query_dense.kernel,
+            hf_weight_key=f"model.layers.{i}.self_attn.q_proj.weight",
+            hook_fn=transpose_and_reshape,
+        )
+        loader.port_weight(
+            keras_variable=decoder_layer._self_attention_layer._query_dense_layer_norm.scale,
+            hf_weight_key=f"model.layers.{i}.self_attn.q_norm.weight",
+        )
+        ## Key
+        loader.port_weight(
+            keras_variable=decoder_layer._self_attention_layer._key_dense.kernel,
+            hf_weight_key=f"model.layers.{i}.self_attn.k_proj.weight",
+            hook_fn=transpose_and_reshape,
+        )
+        loader.port_weight(
+            keras_variable=decoder_layer._self_attention_layer._key_dense_layer_norm.scale,
+            hf_weight_key=f"model.layers.{i}.self_attn.k_norm.weight",
+        )
+        ## Value
+        loader.port_weight(
+            keras_variable=decoder_layer._self_attention_layer._value_dense.kernel,
+            hf_weight_key=f"model.layers.{i}.self_attn.v_proj.weight",
+            hook_fn=transpose_and_reshape,
+        )
+        ## Output
+        loader.port_weight(
+            keras_variable=decoder_layer._self_attention_layer._output_dense.kernel,
+            hf_weight_key=f"model.layers.{i}.self_attn.o_proj.weight",
+            # rearrange_patterns="c (a b) -> a b c",
+            # rearrange_dims={"a": backbone.num_query_heads},
+            hook_fn=transpose_and_reshape,
+        )
+
+        # MLP layers
+        if (
+            (i not in backbone.mlp_only_layers)
+            and backbone.num_experts > 0
+            and ((i + 1) % backbone.decoder_sparse_step == 0)
+        ):
+            # MoE layers
+            loader.port_weight(
+                keras_variable=decoder_layer.mlp._sparse_feedforward_gate_dense.kernel,
+                hf_weight_key=f"model.layers.{i}.mlp.gate.weight",
+                # rearrange_patterns="b a -> a b",
+                hook_fn=lambda hf_tensor, _: np.transpose(
+                    hf_tensor, axes=(1, 0)
+                ),
+            )
+            # Batched experts: gate_up_proj and down_proj
+            gate_up_proj_list = []
+            down_proj_list = []
+            for expert_idx in range(backbone.num_experts):
+                # Load gate_proj and up_proj for each expert
+                gate_proj = loader.get_tensor(
+                    f"model.layers.{i}.mlp.experts.{expert_idx}.gate_proj.weight"
+                )
+                up_proj = loader.get_tensor(
+                    f"model.layers.{i}.mlp.experts.{expert_idx}.up_proj.weight"
+                )
+                # Transpose to (hidden_dim, intermediate_dim)
+                gate_proj = np.transpose(gate_proj, axes=(1, 0))
+                up_proj = np.transpose(up_proj, axes=(1, 0))
+                # Concatenate gate_proj and up_proj along the last dimension
+                gate_up_proj = np.concatenate([gate_proj, up_proj], axis=-1)
+                gate_up_proj_list.append(gate_up_proj)
+
+                # Load down_proj for each expert
+                down_proj = loader.get_tensor(
+                    f"model.layers.{i}.mlp.experts.{expert_idx}.down_proj.weight"
+                )
+                down_proj = np.transpose(
+                    down_proj, axes=(1, 0)
+                )  # (intermediate_dim, hidden_dim)
+                down_proj_list.append(down_proj)
+
+            # Stack the lists to create batched weights
+            gate_up_proj_batched = np.stack(
+                gate_up_proj_list, axis=0
+            )  # (num_experts, hidden_dim, 2 * intermediate_dim)
+            down_proj_batched = np.stack(
+                down_proj_list, axis=0
+            )  # (num_experts, intermediate_dim, hidden_dim)
+
+            # Assign batched weights to expert_bank
+            decoder_layer.mlp.expert_bank._expert_feedforward_gate_dense.assign(
+                gate_up_proj_batched
+            )
+            decoder_layer.mlp.expert_bank._expert_feedforward_output_dense.assign(
+                down_proj_batched
+            )
+        else:
+            loader.port_weight(
+                keras_variable=decoder_layer._feedforward_intermediate_dense.kernel,
+                hf_weight_key=f"model.layers.{i}.mlp.up_proj.weight",
+                # rearrange_patterns="b a -> a b",
+                hook_fn=lambda hf_tensor, _: np.transpose(
+                    hf_tensor, axes=(1, 0)
+                ),
+            )
+            loader.port_weight(
+                keras_variable=decoder_layer._feedforward_output_dense.kernel,
+                hf_weight_key=f"model.layers.{i}.mlp.down_proj.weight",
+                # rearrange_patterns="b a -> a b",
+                hook_fn=lambda hf_tensor, _: np.transpose(
+                    hf_tensor, axes=(1, 0)
+                ),
+            )
+            loader.port_weight(
+                keras_variable=decoder_layer._feedforward_gate_dense.kernel,
+                hf_weight_key=f"model.layers.{i}.mlp.gate_proj.weight",
+                # rearrange_patterns="b a -> a b",
+                hook_fn=lambda hf_tensor, _: np.transpose(
+                    hf_tensor, axes=(1, 0)
+                ),
+            )
+
+        # Feedforward layernorm
+        loader.port_weight(
+            keras_variable=decoder_layer._feedforward_layernorm.scale,
+            hf_weight_key=f"model.layers.{i}.post_attention_layernorm.weight",
+        )
+
+    # Final normalization layer
+    loader.port_weight(
+        keras_variable=backbone.get_layer("sequence_output_layernorm").scale,
+        hf_weight_key="model.norm.weight",
+    )
+
+    return backbone
+
+
+def convert_tokenizer(cls, preset, **kwargs):
+    tokenizer_config = load_json(preset, "tokenizer.json")
+    vocab = tokenizer_config["model"]["vocab"]
+    merges = tokenizer_config["model"]["merges"]
+    merges = [" ".join(item) for item in merges]
+
+    # Load all special tokens with the exception of "reserved" ones.
+    special_tokens = set()
+    for token in tokenizer_config["added_tokens"]:
+        if not token["content"].startswith("<|reserved_special_token_"):
+            vocab[token["content"]] = token["id"]
+            special_tokens.add(token["content"])
+
+    kwargs.update(
+        {
+            "unsplittable_tokens": list(special_tokens),
+        }
+    )
+
+    return cls(vocabulary=vocab, merges=merges, **kwargs)

keras_hub/src/utils/transformers/preset_loader.py

@@ -18,6 +18,7 @@ from keras_hub.src.utils.transformers import convert_mixtral
 from keras_hub.src.utils.transformers import convert_pali_gemma
 from keras_hub.src.utils.transformers import convert_qwen
 from keras_hub.src.utils.transformers import convert_qwen3
+from keras_hub.src.utils.transformers import convert_qwen3_moe
 from keras_hub.src.utils.transformers import convert_qwen_moe
 from keras_hub.src.utils.transformers import convert_t5gemma
 from keras_hub.src.utils.transformers import convert_vit
@@ -61,6 +62,8 @@ class TransformersPresetLoader(PresetLoader):
             self.converter = convert_mixtral
         elif model_type == "qwen2_moe":
             self.converter = convert_qwen_moe
+        elif model_type == "qwen3_moe":
+            self.converter = convert_qwen3_moe
         elif model_type == "qwen3":
             self.converter = convert_qwen3
         elif model_type == "t5gemma":

keras_hub/src/version.py

@@ -1,7 +1,7 @@
 from keras_hub.src.api_export import keras_hub_export
 
 # Unique source of truth for the version number.
-__version__ = "0.23.0.dev202509190415"
+__version__ = "0.23.0.dev202509280419"
 
 
 @keras_hub_export("keras_hub.version")

keras_hub/tokenizers/__init__.py

@@ -78,6 +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,
 )

{keras_hub_nightly-0.23.0.dev202509190415.dist-info → keras_hub_nightly-0.23.0.dev202509280419.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: keras-hub-nightly
-Version: 0.23.0.dev202509190415
+Version: 0.23.0.dev202509280419
 Summary: Pretrained models for Keras.
 Author-email: Keras team <keras-users@googlegroups.com>
 License-Expression: Apache-2.0

{keras_hub_nightly-0.23.0.dev202509190415.dist-info → keras_hub_nightly-0.23.0.dev202509280419.dist-info}/RECORD

@@ -1,11 +1,11 @@
 keras_hub/__init__.py,sha256=bJbUZkqwhZvTb1Tqx1fbkq6mzBYiEyq-Hin3oQIkhdE,558
-keras_hub/layers/__init__.py,sha256=GUDgi0KdORQnv-rH_IRQQ1cCwb-wGQFHy6Vdb7H6FA8,5660
+keras_hub/layers/__init__.py,sha256=8FTy8HwjgFdBvbl_QKTxXmOc13TXjUUBgLYrSTtkc0M,5807
 keras_hub/metrics/__init__.py,sha256=KYalsMPBnfwim9BdGHFfJ5WxUKFXOQ1QoKIMT_0lwlM,439
-keras_hub/models/__init__.py,sha256=Est6LugIjoAFkpTgqZWfISk-1NVMH_k-4soHCHaMmyM,29696
+keras_hub/models/__init__.py,sha256=wy75CGuTVxRIEXSCqmXgMyf23vUbuRbByWrlAaPWXB0,30737
 keras_hub/samplers/__init__.py,sha256=aFQIkiqbZpi8vjrPp2MVII4QUfE-eQjra5fMeHsoy7k,886
 keras_hub/src/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 keras_hub/src/api_export.py,sha256=9pQZK27JObxWZ96QPLBp1OBsjWigh1iuV6RglPGMRk0,1499
-keras_hub/src/version.py,sha256=qBNsfvVj06y9TaC0rRyBoOFXlhAoTAMHAYMCMEyXsDY,222
+keras_hub/src/version.py,sha256=0tAlNHf1D3kDAgKZGBPUBzVIyetqgBNYLSG1UnyUu7I,222
 keras_hub/src/layers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 keras_hub/src/layers/modeling/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 keras_hub/src/layers/modeling/alibi_bias.py,sha256=1XBTHI52L_iJDhN_w5ydu_iMhCuTgQAxEPwcLA6BPuk,4411
@@ -46,6 +46,8 @@ keras_hub/src/models/audio_to_text_preprocessor.py,sha256=GS-WWyJ6aSsPRxi_0bxvxA
 keras_hub/src/models/backbone.py,sha256=BdqPsne7lIITIxn6jY6AN4vZ-Rc9VnpqTxvVNR3CS7M,12210
 keras_hub/src/models/causal_lm.py,sha256=FHGpbyFrDrnPSv5eRLfDgcpvjS6jDhSokMRl-kuumyg,18164
 keras_hub/src/models/causal_lm_preprocessor.py,sha256=nxl-sfmCfkfl6JmVRASa878QbaZUgWSA6Jdu48x4-dY,7155
+keras_hub/src/models/depth_estimator.py,sha256=JR7wtunOPrfEoDkLspoZnL2ItWhZFDeAxxw2vue5QLs,8992
+keras_hub/src/models/depth_estimator_preprocessor.py,sha256=2iE8NAUyiD2AvjZwNoXKUaOUogcE1fRzTNXLQ75GZpQ,2822
 keras_hub/src/models/feature_pyramid_backbone.py,sha256=clEW-TTQSVJ_5qFNdDF0iABkin1p_xlBUFjJrC7T0IA,2247
 keras_hub/src/models/image_classifier.py,sha256=yt6cjhPfqs8A_eWXBsXdXFzn-aRgH2rVHUq7Zu7CyK8,7804
 keras_hub/src/models/image_classifier_preprocessor.py,sha256=Bf7jSqHB1hX2ZWoWQS4GcXNOY_EjeoJi-_vtzCAqw4o,2690
@@ -160,10 +162,19 @@ keras_hub/src/models/densenet/densenet_image_classifier.py,sha256=ye-Ix3oU42pfsD
 keras_hub/src/models/densenet/densenet_image_classifier_preprocessor.py,sha256=xDZbTw_h6pjLDzf8QmbDyMnMsFzgh-dPX1ldg9kddhg,563
 keras_hub/src/models/densenet/densenet_image_converter.py,sha256=DoxYlJVZ9uaabFhVjWOmzvhONoc8KNcQj2vQ6Z1AUpU,354
 keras_hub/src/models/densenet/densenet_presets.py,sha256=d2GEB9cWYrzP8Qj1w8CWiRW976MibQBuk_YQYvgCzr4,1222
+keras_hub/src/models/depth_anything/__init__.py,sha256=NAU7sgJ68ddqKO0X8h5C1V0QC8PQ2lT8QdmM-oZe91M,321
+keras_hub/src/models/depth_anything/depth_anything_backbone.py,sha256=bZzfCJI5Altrs85t8f_QYfQ-C-uhgaWXz_TWQBdaJsU,9308
+keras_hub/src/models/depth_anything/depth_anything_depth_estimator.py,sha256=sqAgwA5x2U8EhoRQhLNgHynFzFY5dKUZs8xFTMrxrMM,2904
+keras_hub/src/models/depth_anything/depth_anything_depth_estimator_preprocessor.py,sha256=ucmXNfY5RoRRYlSvRBZzZm6CuBGbHVlbiehXiM_yOg8,621
+keras_hub/src/models/depth_anything/depth_anything_image_converter.py,sha256=Xutwc8IyklFilDcc4psNBwPGRDcFlzalWXsHvEz7rUc,395
+keras_hub/src/models/depth_anything/depth_anything_layers.py,sha256=_43iEE7F8P7BL4xssjpPeFyhiDk4gGLu-wPxuqQ-nT8,27739
+keras_hub/src/models/depth_anything/depth_anything_loss.py,sha256=GJqzvLkCZrWsMDO6T2Gt_9-TYJqorfNnyOXSNgLUjQg,3389
+keras_hub/src/models/depth_anything/depth_anything_presets.py,sha256=mJ1Kg6J5lP1GkrEDH_qS1XJ__98Tjt5noRZVZaHCn9k,121
+keras_hub/src/models/depth_anything/interpolate.py,sha256=qwrPGP6wA4jZ-XcSeulhkyxPDiMRxHlC92EqSd0H5Tk,2041
 keras_hub/src/models/dinov2/__init__.py,sha256=qacZi82EfAloVND4gDLZjqgR5_yVdz_dc4mMKyCsjOA,257
-keras_hub/src/models/dinov2/dinov2_backbone.py,sha256=kwzd5eqftMS0m5v1HB_4y7JiHxp13ECgG9dNsDoknWo,9491
+keras_hub/src/models/dinov2/dinov2_backbone.py,sha256=QH3lzE1EnxTcOSii9KS1Qx3lq0XcZMsvElB7AL_ejZY,10672
 keras_hub/src/models/dinov2/dinov2_image_converter.py,sha256=gfFROdYV5rOzo3kJFlRvRHYjek8z9YirKfrFwlVJO3g,342
-keras_hub/src/models/dinov2/dinov2_layers.py,sha256=-G3elRWDy09_VPJDJa0qYS5P8vkBGgxPooMZhy2ifu0,33140
+keras_hub/src/models/dinov2/dinov2_layers.py,sha256=UCcia2kWA1O37SMmUbyhUcSXmUpLfNjk1E6mPTPDrF0,33647
 keras_hub/src/models/dinov2/dinov2_presets.py,sha256=ho493GPH98K4LH1E54UV2qZZ4h7Un9ylbBmMQjNoKh4,2937
 keras_hub/src/models/distil_bert/__init__.py,sha256=3Z0w-Mt3aOR0u9RGzjHQ7B3J3qBF2pGjupDGQ9yyzoc,303
 keras_hub/src/models/distil_bert/distil_bert_backbone.py,sha256=rnAf_GokB3wAeJwVZtgUKQO_bKJIa8RavhL_ykTJpNw,6440
@@ -371,6 +382,13 @@ keras_hub/src/models/qwen3/qwen3_decoder.py,sha256=68s9jQj53zFmXE4-SGXKYHu546fXO
 keras_hub/src/models/qwen3/qwen3_layernorm.py,sha256=EJxjf7Pr6ufPQnNeuYQxkExzPjPk4PQxqMsoBeSEkDo,1073
 keras_hub/src/models/qwen3/qwen3_presets.py,sha256=eAqRbjLyRTSXcN-jnGHqoCHejKm2gmt8_zL4EPoE-JA,2518
 keras_hub/src/models/qwen3/qwen3_tokenizer.py,sha256=LmPtg0vprMchDvYfTj8m5PraXI2QS3-YgdIIpIm5iAs,1448
+keras_hub/src/models/qwen3_moe/qwen3_moe_attention.py,sha256=rZnzWA-cAhuWSuHSJfrNqf5_Cu0PNEe7PKbPNbhJdeM,13355
+keras_hub/src/models/qwen3_moe/qwen3_moe_backbone.py,sha256=gguc_M5akemEaQCklTDFiABSRa4nwa4IuDzlfzRRpKM,14618
+keras_hub/src/models/qwen3_moe/qwen3_moe_causal_lm.py,sha256=g3IlpuNzKbcAt_VBYnm895GBLQIPDuMP9eVbL44tf-A,13286
+keras_hub/src/models/qwen3_moe/qwen3_moe_causal_lm_preprocessor.py,sha256=CU5sH0bljNCPuN7sKNnP1FV-jexc12WK0HFU7RWsAvU,499
+keras_hub/src/models/qwen3_moe/qwen3_moe_decoder.py,sha256=lUmDkxrikv4s40tcT9a8muCbEbfUBN97nTFWQEelIJw,25926
+keras_hub/src/models/qwen3_moe/qwen3_moe_layernorm.py,sha256=T6BjJm93F37_0XrrqkWPPXXg4DFOt3f6Al0LDF8N15Y,1360
+keras_hub/src/models/qwen3_moe/qwen3_moe_tokenizer.py,sha256=tDx1WSxmpiWn39NhzkQO-YUbdy713RYHKc_F-EUa6Tw,1473
 keras_hub/src/models/qwen_moe/__init__.py,sha256=5D8GUmVDsJs0J4sVZHcXOLkZf12U96l-WtwyVee4lu8,267
 keras_hub/src/models/qwen_moe/qwen_moe_attention.py,sha256=o0mcVTDMtElMYq3NSYRCfuYVdF-W8YDSU5ogensrVJg,13277
 keras_hub/src/models/qwen_moe/qwen_moe_backbone.py,sha256=nrfELvIvRLmrgKrUNXci2CrecmeI6bWzJj7HH-RcWJA,15341
@@ -523,7 +541,7 @@ keras_hub/src/samplers/serialization.py,sha256=K6FC4AY1sfOLLIk2k4G783XWnQ_Rk3z1Q
 keras_hub/src/samplers/top_k_sampler.py,sha256=WSyrhmOCan55X2JYAnNWE88rkx66sXqdoerl87nOrDQ,2250
 keras_hub/src/samplers/top_p_sampler.py,sha256=9r29WdqBlrW_2TBma6QqkRps2Uit4a6iZPmq1Gsiuko,3400
 keras_hub/src/tests/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-keras_hub/src/tests/test_case.py,sha256=MgvZrz9bz_ubOEAt0D4q5ZiX_UUarHaV8taXoXFuz4U,29260
+keras_hub/src/tests/test_case.py,sha256=Ie0L8NbFJjSO0NWcjU0BRQRdbyCLxT1ru2M0FdU_uHo,29291
 keras_hub/src/tests/mocks/mock_gemma3_tokenizer.py,sha256=a4mSer84-xh9dVJUVpFUPzglCh-7NcFqHRKPDR35c8c,4888
 keras_hub/src/tokenizers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 keras_hub/src/tokenizers/byte_pair_tokenizer.py,sha256=WeUlHMAf5y_MUjFIfVhEcFoOZu-z4kkSj-Dq-pegM9w,24052
@@ -558,7 +576,7 @@ keras_hub/src/utils/transformers/convert_albert.py,sha256=VdKclZpCxtDWq3UbUUQZf4
 keras_hub/src/utils/transformers/convert_bart.py,sha256=Tk4h9Md9rwN5wjQbGIVrC7qzDpF8kI8qm-FKL8HlUok,14411
 keras_hub/src/utils/transformers/convert_bert.py,sha256=4gQqXCJzC9QWdLPDUAq741K8t_kjPIET050YjUnLeDA,5977
 keras_hub/src/utils/transformers/convert_deit.py,sha256=ubcqYzMlhWTCE2S_TsXICCMmqjN9RsQPaw_70vArnjo,5306
-keras_hub/src/utils/transformers/convert_dinov2.py,sha256=FvmB3ggEgowVFPSO5WOzC2hKkG2JvjSb-DeVffu78iU,6908
+keras_hub/src/utils/transformers/convert_dinov2.py,sha256=Zmxz33hKJCcykQOcW8XhG_Yy1l8XqIYam1cjzM69-Mk,6986
 keras_hub/src/utils/transformers/convert_distilbert.py,sha256=SlfIRhSRk5c1ir2HGiDPiXa5XdOId_DbcnZO9lbwyZ8,6498
 keras_hub/src/utils/transformers/convert_esm.py,sha256=rOgGnNY37ZbYnoVC3L-Y-yGGAxTRmYtQV0nJoandH2Y,6214
 keras_hub/src/utils/transformers/convert_gemma.py,sha256=ElCgwBpSN5Q7rV5PJawTsoytPzs5ZjuwoY60YAe8y_A,6533
@@ -569,16 +587,17 @@ keras_hub/src/utils/transformers/convert_mixtral.py,sha256=PxeCY8Xe7U_caICugwOCE
 keras_hub/src/utils/transformers/convert_pali_gemma.py,sha256=B1leeDw96Yvu81hYumf66hIid07k5NLqoeWAJgPnaLs,10649
 keras_hub/src/utils/transformers/convert_qwen.py,sha256=WUxMAEFVqRs7TRw7QU5TH3_ev4yf02R1xFVliMvTQqg,5886
 keras_hub/src/utils/transformers/convert_qwen3.py,sha256=LIormvCMWPq6X9Wo2eNbADjtFZ0nI7tFGZFBxmo4GKw,5700
+keras_hub/src/utils/transformers/convert_qwen3_moe.py,sha256=4Cr2haS20VNHF1iqKRpKeZ47baV0TAXg87zkCfl-JTg,8876
 keras_hub/src/utils/transformers/convert_qwen_moe.py,sha256=a7R28aln-PdAcNuKAXdrtzvslho2Co6GypChxLMKPpc,10618
 keras_hub/src/utils/transformers/convert_t5gemma.py,sha256=DPOwd61UhjspKuCsk3_EaNvSADGP_f8KLcZARHYVk5Y,9490
 keras_hub/src/utils/transformers/convert_vit.py,sha256=YAmXh519ecSgEO5B4g-aEQg1Bb_6ifFafLMqDTfLn_c,5259
-keras_hub/src/utils/transformers/preset_loader.py,sha256=JZn5mfKnVTN5aAvdZ6GWbS_CK3wP42iDkEJsmA58BVw,4925
+keras_hub/src/utils/transformers/preset_loader.py,sha256=_NX7T6ecEQ8I01FkaZApJf-u8JUymNplixR8vcriRlQ,5075
 keras_hub/src/utils/transformers/safetensor_utils.py,sha256=CYUHyA4y-B61r7NDnCsFb4t_UmSwZ1k9L-8gzEd6KRg,3339
 keras_hub/src/utils/transformers/export/gemma.py,sha256=xX_vfQwvFZ_-lQX4kgMNOGKL7fL_1yk7QyGYV2Qyly4,4699
 keras_hub/src/utils/transformers/export/hf_exporter.py,sha256=Qk52c6LIA2eMHUNY9Vy4STJSpnhLMdJ_t-3ljqhSr4k,5081
-keras_hub/tokenizers/__init__.py,sha256=YEr_cwyX6MACxQOgyRwETilOFYBXpQLNXH22ZdSSv3o,4450
+keras_hub/tokenizers/__init__.py,sha256=kyFWYm4mb--U4xYU-2Gb1COM8xEFWNK6LcKxr8h9Ivc,4561
 keras_hub/utils/__init__.py,sha256=jXPqVGBpJr_PpYmqD8aDG-fRMlxH-ulqCR2SZMn288Y,646
-keras_hub_nightly-0.23.0.dev202509190415.dist-info/METADATA,sha256=r_hK_l4RrMqkzmqufjPnhoSiOp7v7B6TP0pduHAo7Rs,7395
-keras_hub_nightly-0.23.0.dev202509190415.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-keras_hub_nightly-0.23.0.dev202509190415.dist-info/top_level.txt,sha256=N4J6piIWBKa38A4uV-CnIopnOEf8mHAbkNXafXm_CuA,10
-keras_hub_nightly-0.23.0.dev202509190415.dist-info/RECORD,,
+keras_hub_nightly-0.23.0.dev202509280419.dist-info/METADATA,sha256=hNCnpYrO1qf6MHV8MDBOXPwgAB4D7yqDsVvuWGuJ8Ww,7395
+keras_hub_nightly-0.23.0.dev202509280419.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+keras_hub_nightly-0.23.0.dev202509280419.dist-info/top_level.txt,sha256=N4J6piIWBKa38A4uV-CnIopnOEf8mHAbkNXafXm_CuA,10
+keras_hub_nightly-0.23.0.dev202509280419.dist-info/RECORD,,