keras-hub-nightly 0.22.0.dev202507150421__py3-none-any.whl → 0.22.0.dev202507170424__py3-none-any.whl

keras_hub/src/utils/transformers/export/hf_exporter.py

@@ -0,0 +1,98 @@
+import json
+import os
+import shutil
+import warnings
+
+import keras
+
+from keras_hub.src.utils.transformers.export.gemma import get_gemma_config
+from keras_hub.src.utils.transformers.export.gemma import get_gemma_weights_map
+
+MODEL_CONFIGS = {
+    "GemmaBackbone": get_gemma_config,
+    # Add future models here, e.g., "LlamaBackbone": get_llama_config,
+}
+
+MODEL_EXPORTERS = {
+    "GemmaBackbone": get_gemma_weights_map,
+    # Add future models here, e.g., "LlamaBackbone": get_llama_weights_map,
+}
+
+
+def export_to_safetensors(keras_model, path):
+    """Converts a Keras model to Hugging Face safetensor format.
+
+    It does the following:
+    - Extracts and maps weights from the Keras backbone to safetensors.
+    - Saves the configuration as 'config.json'.
+    - Saves weights in 'model.safetensors'.
+    - Saves tokenizer assets.
+
+    Args:
+        keras_model: The Keras model to convert.
+        path: str. Path of the directory to which the safetensors file,
+          config and tokenizer will be saved.
+    """
+    backend = keras.config.backend()
+    backbone = keras_model.backbone
+    model_type = backbone.__class__.__name__
+
+    if model_type not in MODEL_CONFIGS:
+        raise ValueError(f"Config not implemented for {model_type}")
+
+    if model_type not in MODEL_EXPORTERS:
+        raise ValueError(f"Exporter not implemented for {model_type}")
+
+    get_config_fn = MODEL_CONFIGS[model_type]
+    hf_config = get_config_fn(backbone)
+
+    get_weights_fn = MODEL_EXPORTERS[model_type]
+    weights_dict = get_weights_fn(backbone)
+
+    if not weights_dict:
+        raise ValueError("No weights to save.")
+
+    # Save config
+    os.makedirs(path, exist_ok=True)
+    config_path = os.path.join(path, "config.json")
+    with open(config_path, "w") as f:
+        json.dump(hf_config, f)
+
+    # Save weights based on backend
+    weights_path = os.path.join(path, "model.safetensors")
+    if backend == "torch":
+        from safetensors.torch import save_file
+
+        weights_dict_contiguous = {
+            k: v.value.contiguous() if hasattr(v, "value") else v.contiguous()
+            for k, v in weights_dict.items()
+        }
+        save_file(
+            weights_dict_contiguous, weights_path, metadata={"format": "pt"}
+        )
+    elif backend == "tensorflow":
+        from safetensors.tensorflow import save_file
+
+        save_file(weights_dict, weights_path, metadata={"format": "pt"})
+    elif backend == "jax":
+        from safetensors.flax import save_file
+
+        save_file(weights_dict, weights_path, metadata={"format": "pt"})
+    else:
+        raise ValueError(f"Unsupported backend: {backend}")
+
+    # Save tokenizer assets
+    keras_model.preprocessor.tokenizer.save_assets(path)
+
+    # Rename vocabulary file
+    vocab_spm_path = os.path.join(path, "vocabulary.spm")
+    tokenizer_model_path = os.path.join(path, "tokenizer.model")
+    if os.path.exists(vocab_spm_path):
+        shutil.move(vocab_spm_path, tokenizer_model_path)
+    else:
+        warnings.warn(
+            f"{vocab_spm_path} not found. Tokenizer may not load "
+            "correctly. Ensure that the tokenizer configuration "
+            "is correct and that the vocabulary file is present "
+            "in the original model."
+        )

keras_hub/src/utils/transformers/preset_loader.py

@@ -7,6 +7,7 @@ from keras_hub.src.utils.transformers import convert_albert
 from keras_hub.src.utils.transformers import convert_bart
 from keras_hub.src.utils.transformers import convert_bert
 from keras_hub.src.utils.transformers import convert_deit
+from keras_hub.src.utils.transformers import convert_dinov2
 from keras_hub.src.utils.transformers import convert_distilbert
 from keras_hub.src.utils.transformers import convert_gemma
 from keras_hub.src.utils.transformers import convert_gpt2
@@ -35,7 +36,9 @@ class TransformersPresetLoader(PresetLoader):
             self.converter = convert_deit
         elif model_type == "distilbert":
             self.converter = convert_distilbert
-        elif model_type == "gemma" or model_type == "gemma2":
+        elif model_type in ("dinov2", "dinov2_with_registers"):
+            self.converter = convert_dinov2
+        elif model_type in ("gemma", "gemma2"):
             self.converter = convert_gemma
         elif model_type == "gpt2":
             self.converter = convert_gpt2

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__ = "0.22.0.dev202507150421"
+__version__ = "0.22.0.dev202507170424"
 
 
 @keras_hub_export("keras_hub.version")

{keras_hub_nightly-0.22.0.dev202507150421.dist-info → keras_hub_nightly-0.22.0.dev202507170424.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: keras-hub-nightly
-Version: 0.22.0.dev202507150421
+Version: 0.22.0.dev202507170424
 Summary: Pretrained models for Keras.
 Author-email: Keras team <keras-users@googlegroups.com>
 License-Expression: Apache-2.0

{keras_hub_nightly-0.22.0.dev202507150421.dist-info → keras_hub_nightly-0.22.0.dev202507170424.dist-info}/RECORD

@@ -1,11 +1,11 @@
 keras_hub/__init__.py,sha256=bJbUZkqwhZvTb1Tqx1fbkq6mzBYiEyq-Hin3oQIkhdE,558
-keras_hub/layers/__init__.py,sha256=T1XBtpT0UH9-r0Jc-ljCxtZD_ccapf86ByvUgzdRbvg,5311
+keras_hub/layers/__init__.py,sha256=SMkchjCbNydCBULOFC1pzZRaD-KWZ2CaH6CEVf1MRWE,5428
 keras_hub/metrics/__init__.py,sha256=KYalsMPBnfwim9BdGHFfJ5WxUKFXOQ1QoKIMT_0lwlM,439
-keras_hub/models/__init__.py,sha256=52UNIL7my_9g6ubPtOMDnGYeuGD4SOldfnGTVRMKTeE,27558
+keras_hub/models/__init__.py,sha256=-Rz2qSdQezPihXYnZlaG3S18HLDrCe37pZjzNIPU7D8,27656
 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=xBxcEBC7BLhOh8DULfCuLeohknkMZJpgPhVlQNz_tNc,222
+keras_hub/src/version.py,sha256=-Bevtrx2qCQT5lthAimfMBv46nlIAEIb7EDasoOPGcE,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
@@ -102,15 +102,14 @@ keras_hub/src/models/bloom/bloom_decoder.py,sha256=fda8iX4wzx2M8AoLX7fDHkyoir89K
 keras_hub/src/models/bloom/bloom_presets.py,sha256=7RptuZi__oJyiX6X4xE5ToANcEwsmLDqhuEKwFyKIPU,3215
 keras_hub/src/models/bloom/bloom_tokenizer.py,sha256=6Konh7B_L9BqgjkA0z8-APFpr9sQmQPuAJFZSsCIClU,2574
 keras_hub/src/models/clip/__init__.py,sha256=NcjBkTNWxLY4Ss9wV-NW9iS8k6AwMiS2ARMcxr6KEps,245
-keras_hub/src/models/clip/clip_backbone.py,sha256=AyVhLwFg5nLFSaoaL8mLuNkK9uBPJ9y5FMQu4psTGvo,9877
-keras_hub/src/models/clip/clip_encoder_block.py,sha256=4Jxqb0Pq3Joh-lHDq-Y2c8v-gcMm1sDjPID4eRGK0DE,3823
+keras_hub/src/models/clip/clip_backbone.py,sha256=DRAXEJFVPcgf1-AeVDDmuoxplwTCl4Xt7-D4whM4w04,6619
 keras_hub/src/models/clip/clip_image_converter.py,sha256=XyHEDB4RbYiveMN1hLQxHgGADb_goyWyE0TceAd2owM,330
-keras_hub/src/models/clip/clip_preprocessor.py,sha256=nUYu8Bgf3TU7jrR10kr0BIe7ph3aABvGtIqnjqrIb9k,4752
+keras_hub/src/models/clip/clip_layers.py,sha256=ns3Zzm5UzMpm-ynyU3aJu2d4i3HmzNiZKdAea624ako,10184
+keras_hub/src/models/clip/clip_preprocessor.py,sha256=xj-FzK7gLIUyvTo2iM1zHh9f2Ff25tZCYFxsPE3dwFU,4771
 keras_hub/src/models/clip/clip_presets.py,sha256=b9Azial1dUtuNV96Q0Ahz-bcBRmlIjnZPUzMvAMb8OY,3348
-keras_hub/src/models/clip/clip_text_encoder.py,sha256=BCIE24eKZJ3yc4T0sjD6-Msjr1FQRKpdTP7vpGEn_7M,5456
+keras_hub/src/models/clip/clip_text_encoder.py,sha256=lZa9Ditvn4DH9As3NEML_Wl6g2qeYer_LzRHGu1hqCM,5449
 keras_hub/src/models/clip/clip_tokenizer.py,sha256=6gIm_LWRbCeBQUI9M2gA8-OXb4tXGygixkbcL6joV1c,7444
-keras_hub/src/models/clip/clip_vision_embedding.py,sha256=6_qC7T1dqKd-39EreGmHZj-YfjOLEDDKjWnEKcKIyuY,3667
-keras_hub/src/models/clip/clip_vision_encoder.py,sha256=q62MXySZN38uCsjqq8cttfBxD7P5abaKQV2i8_u4N6E,6385
+keras_hub/src/models/clip/clip_vision_encoder.py,sha256=C5grKgIgFF8ls-kkGdYorpw5tbfgbmBQe6VJg_3yWII,6368
 keras_hub/src/models/cspnet/__init__.py,sha256=TOpvk2cfOVv1bPA1BOGZj0mhmhc6E98zZmW9e0PIvhk,257
 keras_hub/src/models/cspnet/cspnet_backbone.py,sha256=meHzxubG_9vHQHSelDfrROaQERkDiWkjTtk_gKaWsDc,42457
 keras_hub/src/models/cspnet/cspnet_image_classifier.py,sha256=JqfBHIBTFxaLOyAWx6TdXs0aAOMbcCx1oo47RoQnytc,510
@@ -148,6 +147,11 @@ 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/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_image_converter.py,sha256=gfFROdYV5rOzo3kJFlRvRHYjek8z9YirKfrFwlVJO3g,342
+keras_hub/src/models/dinov2/dinov2_layers.py,sha256=-G3elRWDy09_VPJDJa0qYS5P8vkBGgxPooMZhy2ifu0,33140
+keras_hub/src/models/dinov2/dinov2_presets.py,sha256=maGHlMM_9Fa9UXm7xRS0h9w7nQwMkS53TTmtY2juF0M,114
 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
 keras_hub/src/models/distil_bert/distil_bert_masked_lm.py,sha256=axeZd5UcxFr3_Q8H4yG10CINh93wbcyjlPLauqe5N9E,4289
@@ -191,7 +195,7 @@ keras_hub/src/models/flux/flux_maths.py,sha256=2pnHW8HW7V2JZ8HIrUwE-UU4klpFQaOko
 keras_hub/src/models/flux/flux_model.py,sha256=K92PyeFHIp8SwXuxhv__XCEaQ2wqSW1jOb97I4S24Rw,8991
 keras_hub/src/models/flux/flux_presets.py,sha256=z7C_FbI1_F5YETXuWpc7Yh_0w-5N0eBQy6Oks_X9W88,54
 keras_hub/src/models/flux/flux_text_to_image.py,sha256=Rf5dD2EhG0bE8Gyg9sqaA8YEexS1kdraofIkxiZDjvc,4166
-keras_hub/src/models/flux/flux_text_to_image_preprocessor.py,sha256=Fs9jr97QtmRUbRRz1kITpkuhDM2GoV3n0XSFC-qQA14,2252
+keras_hub/src/models/flux/flux_text_to_image_preprocessor.py,sha256=2kI2vSZvTia5ISb4BVPgC_e1l5rkirLSjhm13P-UR_k,2362
 keras_hub/src/models/gemma/__init__.py,sha256=rVzOJMJ39bgVlT8UdC0t8PlN2c237GKTBmfHIsbPuOQ,251
 keras_hub/src/models/gemma/gemma_attention.py,sha256=wmU5FgQu1Ajg-KHKVXTLHWH7pXqN4_zVJTCp_FXMcAs,10095
 keras_hub/src/models/gemma/gemma_backbone.py,sha256=GzAUSArw_pN9dtWQzTVhWDbW-XyWt4GyMcFLn9hwmh0,13391
@@ -227,14 +231,14 @@ keras_hub/src/models/gpt_neo_x/gpt_neo_x_causal_lm.py,sha256=HriMXNVjGlFTjCIgfLR
 keras_hub/src/models/gpt_neo_x/gpt_neo_x_causal_lm_preprocessor.py,sha256=YiVz9qBHjQlwKgtUVrgBTFitHcX5pbmhhfHwaulyRxY,1957
 keras_hub/src/models/gpt_neo_x/gpt_neo_x_decoder.py,sha256=hmB81V0SuI6bEsxEuFkYgq58wbcrv1YLvmXGin5T3E0,9732
 keras_hub/src/models/gpt_neo_x/gpt_neo_x_tokenizer.py,sha256=aKso-8yGrynn3tZ5xm2egcXIBQo3__sWZDBtjmS3ZgU,1991
-keras_hub/src/models/hgnetv2/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+keras_hub/src/models/hgnetv2/__init__.py,sha256=hGilfTnRPpVFS3YpRhJWEyK8CaPIzkRh6zUC1_5imaY,263
 keras_hub/src/models/hgnetv2/hgnetv2_backbone.py,sha256=eqVrbU2EyB2ToxK1g2QRW90zd5GyvJ8I7PKVBgqRpfY,7966
 keras_hub/src/models/hgnetv2/hgnetv2_encoder.py,sha256=VL6XCqyXieUPkqXS7fhsAT-EV6jzyN_i31EjsAizgVU,6464
 keras_hub/src/models/hgnetv2/hgnetv2_image_classifier.py,sha256=62Xual9pRBkU6G_RUdCblx68Z827SCA_5q9utCXxwa0,7897
 keras_hub/src/models/hgnetv2/hgnetv2_image_classifier_preprocessor.py,sha256=df7OKvJmz2UqOXrqECvI9QdVMVkVMWhK0go9sltajnI,553
 keras_hub/src/models/hgnetv2/hgnetv2_image_converter.py,sha256=qaGRtDeQwmC0PR69KWC7GzYNdWZ5cHu_exhNzdYyYzM,348
 keras_hub/src/models/hgnetv2/hgnetv2_layers.py,sha256=OMUKW5VWL0xkEQl7RJYGAbTTB7qeqH3FHtMMuiQ0QmI,36418
-keras_hub/src/models/hgnetv2/hgnetv2_presets.py,sha256=azRtaBynFqI1ccmP8_LMG16tjNSSeMvgo_ZFneG-bg8,1767
+keras_hub/src/models/hgnetv2/hgnetv2_presets.py,sha256=kbwxp8Nh4jdDN6egSmSxxwpY7CP5AklINXlWI0K3ZYA,2078
 keras_hub/src/models/llama/__init__.py,sha256=svVZjGi71R3lVbq0AdbqlXj909mr3Rp9EPXdiO0w0G0,251
 keras_hub/src/models/llama/llama_attention.py,sha256=UFHOWr69vTkOxLdgSUckGaSuUUyqlJ_xYoswWHVnTOU,8977
 keras_hub/src/models/llama/llama_backbone.py,sha256=AT8kUPHEn6DT-aGY838_sZkBhByIdh82DWW8y-Sp3mE,13614
@@ -406,14 +410,14 @@ keras_hub/src/models/siglip/siglip_text_encoder.py,sha256=xOVvzyQHLX9ne30y4ussar
 keras_hub/src/models/siglip/siglip_tokenizer.py,sha256=j_67JbIHJDRk-CbiemG2dgAO6lp3_0_JdnfroZ90G18,2579
 keras_hub/src/models/siglip/siglip_vision_encoder.py,sha256=CaNaFq5thBC3TUXXOf2qknk5vWsauM20ZoaDPYRnXcs,5927
 keras_hub/src/models/stable_diffusion_3/__init__.py,sha256=ZKYQuaRObyhKq8GVAHmoRvlXp6FpU8ChvutVCHyXKuc,343
-keras_hub/src/models/stable_diffusion_3/flow_match_euler_discrete_scheduler.py,sha256=vtVhieAv277mAiZj7Kvvqg_Ba7klfQxZVk4PPxNNQ0s,3062
-keras_hub/src/models/stable_diffusion_3/mmdit.py,sha256=jkO7uP3fNrbuFLiOJV-7_S8hz-DqkasZNkoJIdsg58Q,40859
-keras_hub/src/models/stable_diffusion_3/stable_diffusion_3_backbone.py,sha256=M1e8G83CzwOgFvtDVYbm8HKSODa9keN2Jan18EWpWU0,24370
-keras_hub/src/models/stable_diffusion_3/stable_diffusion_3_image_to_image.py,sha256=oQcVCWOwrdUTrr_JNekoMqdSlKYMGz5tG6v8uD25lTc,5479
-keras_hub/src/models/stable_diffusion_3/stable_diffusion_3_inpaint.py,sha256=t4uw920Jn1k80air3WRGimKf71aMVu6q73oWFH348vk,6384
+keras_hub/src/models/stable_diffusion_3/flow_match_euler_discrete_scheduler.py,sha256=1K_B3d3fNn50eY84OgxVHyIHHZhmlJY03b71pMSmE9s,3246
+keras_hub/src/models/stable_diffusion_3/mmdit.py,sha256=emyDmtpJiFU_9crSDBC5CaXoZnM1Eti8uAQtwv2v8B0,42794
+keras_hub/src/models/stable_diffusion_3/stable_diffusion_3_backbone.py,sha256=odg1nA02dmINSjAfwKnzPt_HC6b7nQfP000swHxIfaI,24055
+keras_hub/src/models/stable_diffusion_3/stable_diffusion_3_image_to_image.py,sha256=uNsNSQ4EFceGfIMzgjYWFMuL0XdfM58rubTcrCVPrts,5532
+keras_hub/src/models/stable_diffusion_3/stable_diffusion_3_inpaint.py,sha256=2UIRz11DRbHJ7IVbkjpBjtbkZGC3-eYhMtVUWTmWMH8,6437
 keras_hub/src/models/stable_diffusion_3/stable_diffusion_3_presets.py,sha256=x7Ez4L955MJE4ABtBy-63YpU9XpR0Ro8QWPzYYJs1yE,2167
-keras_hub/src/models/stable_diffusion_3/stable_diffusion_3_text_to_image.py,sha256=Yt-UIatVKANjjKFCFEj1rIHhOrt8hqefKKQJIAWcTLc,4567
-keras_hub/src/models/stable_diffusion_3/stable_diffusion_3_text_to_image_preprocessor.py,sha256=sQvrhuWvC4hM99cxFiI5532QLeiW4YXpH0zAvnnCXQA,2801
+keras_hub/src/models/stable_diffusion_3/stable_diffusion_3_text_to_image.py,sha256=-xmmCaoPc1ixJvyIBwVTW1yKBA-rP4nWReovcs7OLKE,4620
+keras_hub/src/models/stable_diffusion_3/stable_diffusion_3_text_to_image_preprocessor.py,sha256=crUT82moaPx8RVKrLtUHx1zry602f8DWItek9aFkojg,2903
 keras_hub/src/models/stable_diffusion_3/t5_encoder.py,sha256=oV7P1uwCKdGiD93zXq7kmqX0elMZQU4UvBa8wg6P1hs,5113
 keras_hub/src/models/t5/__init__.py,sha256=OWyoUeDY3v4DnO8Ry02DWV1bNSVGcC89PF9oCftyi1s,233
 keras_hub/src/models/t5/t5_backbone.py,sha256=MUmabugPx5_BkAHkuJXr2-8z_yZfKD19SO0KJtlcHhA,10331
@@ -493,7 +497,7 @@ keras_hub/src/tokenizers/word_piece_tokenizer_trainer.py,sha256=cylrs02ZrYQ1TuZr
 keras_hub/src/utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 keras_hub/src/utils/keras_utils.py,sha256=IWsbg-p-XVLuOkba8PAYNf9zDo4G2RkINLr58p12MhA,5291
 keras_hub/src/utils/pipeline_model.py,sha256=jgzB6NQPSl0KOu08N-TazfOnXnUJbZjH2EXXhx25Ftg,9084
-keras_hub/src/utils/preset_utils.py,sha256=GKYFKK9YcdIrMm0_hC_KTIXgpiMYD6SauMnSRpNsDQo,34975
+keras_hub/src/utils/preset_utils.py,sha256=dEOAGjkjnu69nhWuS1wnHVyMmkYnlzUQAUPzbLexLhY,35142
 keras_hub/src/utils/python_utils.py,sha256=N8nWeO3san4YnGkffRXG3Ix7VEIMTKSN21FX5TuL7G8,202
 keras_hub/src/utils/tensor_utils.py,sha256=WrohV6-hvxtLE6rRRhtN4hy8GkHikV-NrRnVEYUwJQo,16133
 keras_hub/src/utils/coco/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -513,6 +517,7 @@ 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_distilbert.py,sha256=SlfIRhSRk5c1ir2HGiDPiXa5XdOId_DbcnZO9lbwyZ8,6498
 keras_hub/src/utils/transformers/convert_gemma.py,sha256=ElCgwBpSN5Q7rV5PJawTsoytPzs5ZjuwoY60YAe8y_A,6533
 keras_hub/src/utils/transformers/convert_gpt2.py,sha256=HCeHN_-GiQJRxLCM9OCJJ1watPVpIBF8ujS8pGbBOWc,5703
@@ -524,11 +529,13 @@ keras_hub/src/utils/transformers/convert_qwen.py,sha256=WUxMAEFVqRs7TRw7QU5TH3_e
 keras_hub/src/utils/transformers/convert_qwen3.py,sha256=LIormvCMWPq6X9Wo2eNbADjtFZ0nI7tFGZFBxmo4GKw,5700
 keras_hub/src/utils/transformers/convert_qwen_moe.py,sha256=a7R28aln-PdAcNuKAXdrtzvslho2Co6GypChxLMKPpc,10618
 keras_hub/src/utils/transformers/convert_vit.py,sha256=9SUZ9utNJhW_5cj3acMn9cRy47u2eIcDsrhmzj77o9k,5187
-keras_hub/src/utils/transformers/preset_loader.py,sha256=K5FzDAtCuXS9rmZc0Zj7UCwbz5J9_pf7ozWov1qRAfg,4495
+keras_hub/src/utils/transformers/preset_loader.py,sha256=uEmDzEa6v-W59WssBXi8RnvW_dR9km9_3K6bJmBGpcY,4649
 keras_hub/src/utils/transformers/safetensor_utils.py,sha256=CYUHyA4y-B61r7NDnCsFb4t_UmSwZ1k9L-8gzEd6KRg,3339
+keras_hub/src/utils/transformers/export/gemma.py,sha256=NpTSgRUSWp3WXQil1CjYUVFVyyVhpO-4-3q2en2Wxwg,3264
+keras_hub/src/utils/transformers/export/hf_exporter.py,sha256=oTdRS8SalPCbi_cZPC55aZUBc-1_pdviUIp0XysA4cI,3234
 keras_hub/tokenizers/__init__.py,sha256=uMjjm0mzUkRb0e4Ac_JK8aJ9cKGUi5UqmzWoWAFJprE,4164
 keras_hub/utils/__init__.py,sha256=jXPqVGBpJr_PpYmqD8aDG-fRMlxH-ulqCR2SZMn288Y,646
-keras_hub_nightly-0.22.0.dev202507150421.dist-info/METADATA,sha256=0qyFbIeEjvLqHG2Z0mz3yw_ThuOgKpAg1sQqp-x5k7g,7393
-keras_hub_nightly-0.22.0.dev202507150421.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-keras_hub_nightly-0.22.0.dev202507150421.dist-info/top_level.txt,sha256=N4J6piIWBKa38A4uV-CnIopnOEf8mHAbkNXafXm_CuA,10
-keras_hub_nightly-0.22.0.dev202507150421.dist-info/RECORD,,
+keras_hub_nightly-0.22.0.dev202507170424.dist-info/METADATA,sha256=S2Qlx04gz55z7gcbYbvnkHBfqT-oR3LCrZ164IdLORM,7393
+keras_hub_nightly-0.22.0.dev202507170424.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+keras_hub_nightly-0.22.0.dev202507170424.dist-info/top_level.txt,sha256=N4J6piIWBKa38A4uV-CnIopnOEf8mHAbkNXafXm_CuA,10
+keras_hub_nightly-0.22.0.dev202507170424.dist-info/RECORD,,

keras_hub/src/models/clip/clip_encoder_block.py

@@ -1,111 +0,0 @@
-from keras import dtype_policies
-from keras import layers
-from keras import ops
-
-
-def quick_gelu(x):
-    return x * ops.sigmoid(1.702 * x)
-
-
-# TODO: Deprecate this in favor of `keras.layers.MultiHeadAttention` once the
-# dtype compatibility issue is resolved.
-class CLIPMultiHeadAttention(layers.MultiHeadAttention):
-    def _masked_softmax(self, attention_scores, attention_mask=None):
-        attention_scores = super()._masked_softmax(
-            attention_scores, attention_mask
-        )
-        return ops.cast(attention_scores, self._value_dense.compute_dtype)
-
-
-class CLIPEncoderBlock(layers.Layer):
-    def __init__(
-        self,
-        hidden_dim,
-        num_heads,
-        intermediate_dim,
-        intermediate_activation="quick_gelu",
-        use_causal_mask=True,
-        **kwargs,
-    ):
-        super().__init__(**kwargs)
-        if hidden_dim % num_heads != 0:
-            raise ValueError(
-                "`hidden_dim` must be divisible by `num_heads`. "
-                f"Received: hidden_dim={hidden_dim}, num_heads={num_heads}"
-            )
-        self.hidden_dim = hidden_dim
-        self.num_heads = num_heads
-        self.intermediate_dim = intermediate_dim
-        self.intermediate_activation = intermediate_activation
-        self.use_causal_mask = use_causal_mask
-
-        if intermediate_activation == "quick_gelu":
-            intermediate_activation = quick_gelu
-
-        self.layer_norm_1 = layers.LayerNormalization(
-            epsilon=1e-5, dtype=self.dtype_policy, name="layer_norm_1"
-        )
-        self.attention = CLIPMultiHeadAttention(
-            num_heads,
-            hidden_dim // num_heads,
-            dtype=self.dtype_policy,
-            name="attention",
-        )
-        self.layer_norm_2 = layers.LayerNormalization(
-            epsilon=1e-5, dtype=self.dtype_policy, name="layer_norm_2"
-        )
-        self.dense_1 = layers.Dense(
-            self.intermediate_dim, dtype=self.dtype_policy, name="dense_1"
-        )
-        self.activation = layers.Activation(
-            intermediate_activation, dtype=self.dtype_policy, name="activation"
-        )
-        self.dense_2 = layers.Dense(
-            self.hidden_dim, dtype=self.dtype_policy, name="dense_2"
-        )
-
-    def build(self, input_shape):
-        self.layer_norm_1.build(input_shape)
-        self.attention.build(input_shape, input_shape, input_shape)
-        # Before Keras 3.2, there was no setter for `dtype_policy`. Directly
-        # assign a `DTypePolicy` instead.
-        self.attention._softmax.dtype_policy = dtype_policies.DTypePolicy(
-            "float32"
-        )
-        self.layer_norm_2.build(input_shape)
-        self.dense_1.build(input_shape)
-        input_shape = self.dense_1.compute_output_shape(input_shape)
-        self.dense_2.build(input_shape)
-
-    def compute_output_shape(self, inputs_shape):
-        outputs_shape = list(inputs_shape)
-        outputs_shape[-1] = self.hidden_dim
-        return outputs_shape
-
-    def call(self, x, training=None):
-        residual = x
-        x = self.layer_norm_1(x)
-        x = self.attention(
-            x, x, x, training=training, use_causal_mask=self.use_causal_mask
-        )
-        x = ops.add(residual, x)
-
-        residual = x
-        x = self.dense_1(self.layer_norm_2(residual))
-        x = self.activation(x)
-        x = self.dense_2(x)
-        x = ops.add(residual, x)
-        return x
-
-    def get_config(self):
-        config = super().get_config()
-        config.update(
-            {
-                "hidden_dim": self.hidden_dim,
-                "num_heads": self.num_heads,
-                "intermediate_dim": self.intermediate_dim,
-                "intermediate_activation": self.intermediate_activation,
-                "use_causal_mask": self.use_causal_mask,
-            }
-        )
-        return config

keras_hub/src/models/clip/clip_vision_embedding.py

@@ -1,101 +0,0 @@
-from keras import layers
-from keras import ops
-
-from keras_hub.src.utils.keras_utils import standardize_data_format
-
-
-class CLIPVisionEmbedding(layers.Layer):
-    def __init__(
-        self,
-        hidden_dim,
-        patch_size,
-        image_size,
-        data_format=None,
-        dtype=None,
-        **kwargs,
-    ):
-        super().__init__(dtype=dtype, **kwargs)
-        self.hidden_dim = int(hidden_dim)
-        self.patch_size = int(patch_size)
-        self.image_size = int(image_size)
-        data_format = standardize_data_format(data_format)
-        self.data_format = data_format
-        num_patches = (image_size // patch_size) ** 2
-        self.num_positions = num_patches + 1
-
-        self.patch_embedding = layers.Conv2D(
-            hidden_dim,
-            kernel_size=patch_size,
-            strides=patch_size,
-            data_format=data_format,
-            use_bias=False,
-            dtype=dtype,
-            name="patch_embedding",
-        )
-        self.position_embedding = layers.Embedding(
-            num_patches + 1, hidden_dim, dtype=dtype, name="position_embedding"
-        )
-
-    def build(self, input_shape):
-        self.class_embedding = self.add_weight(
-            shape=(self.hidden_dim,),
-            initializer="random_normal",
-            dtype=self.variable_dtype,
-            name="class_embedding",
-        )
-        self.position_ids = self.add_weight(
-            shape=(1, self.num_positions),
-            initializer="zeros",
-            # Let the backend determine the int dtype. For example, tf
-            # requires int64 for correct device placement, whereas jax and torch
-            # don't.
-            dtype=int,
-            trainable=False,
-            name="position_ids",
-        )
-        self.patch_embedding.build(input_shape)
-        self.position_embedding.build(self.position_ids.shape)
-
-    def call(self, inputs, training=None):
-        x = inputs
-        batch_size = ops.shape(x)[0]
-        patch_embeddings = self.patch_embedding(x, training=training)
-        if self.data_format == "channels_last":
-            patch_embeddings = ops.reshape(
-                patch_embeddings, (batch_size, -1, self.hidden_dim)
-            )
-        else:
-            patch_embeddings = ops.reshape(
-                patch_embeddings, (batch_size, self.hidden_dim, -1)
-            )
-            patch_embeddings = ops.transpose(patch_embeddings, (0, 2, 1))
-        class_embeddings = ops.expand_dims(self.class_embedding, axis=(0, 1))
-        class_embeddings = ops.tile(class_embeddings, (batch_size, 1, 1))
-        position_embeddings = self.position_embedding(self.position_ids)
-        embeddings = ops.concatenate(
-            [class_embeddings, patch_embeddings], axis=1
-        )
-        return ops.add(embeddings, position_embeddings)
-
-    def get_config(self):
-        config = super().get_config()
-        config.update(
-            {
-                "hidden_dim": self.hidden_dim,
-                "patch_size": self.patch_size,
-                "image_size": self.image_size,
-            }
-        )
-        return config
-
-    def compute_output_shape(self, input_shape):
-        output_shape = [input_shape[0], None, self.hidden_dim]
-        if self.data_format == "channels_last":
-            if input_shape[1] is not None and input_shape[2] is not None:
-                patch_num = input_shape[1] // self.patch_size
-                output_shape[1] = patch_num**2 + 1
-        else:
-            if input_shape[2] is not None and input_shape[3] is not None:
-                patch_num = input_shape[2] // self.patch_size
-                output_shape[1] = patch_num**2 + 1
-        return output_shape