keras-hub-nightly 0.19.0.dev202501270344__py3-none-any.whl → 0.19.0.dev202501290343__py3-none-any.whl

keras_hub/src/layers/preprocessing/image_converter.py

@@ -98,6 +98,7 @@ class ImageConverter(PreprocessingLayer):
         scale=None,
         offset=None,
         crop_to_aspect_ratio=True,
+        pad_to_aspect_ratio=False,
         interpolation="bilinear",
         data_format=None,
         **kwargs,
@@ -112,12 +113,19 @@ class ImageConverter(PreprocessingLayer):
 
         super().__init__(**kwargs)
 
+        if crop_to_aspect_ratio and pad_to_aspect_ratio:
+            raise ValueError(
+                "Only one of 'crop_to_aspect_ratio' or 'pad_to_aspect_ratio' "
+                "can be True."
+            )
+
         # Create the `Resizing` layer here even if it's not being used. That
         # allows us to make `image_size` a settable property.
         self.resizing = keras.layers.Resizing(
             height=image_size[0] if image_size else None,
             width=image_size[1] if image_size else None,
             crop_to_aspect_ratio=crop_to_aspect_ratio,
+            pad_to_aspect_ratio=pad_to_aspect_ratio,
             interpolation=interpolation,
             data_format=data_format,
             dtype=self.dtype_policy,
@@ -126,6 +134,7 @@ class ImageConverter(PreprocessingLayer):
         self.scale = scale
         self.offset = offset
         self.crop_to_aspect_ratio = crop_to_aspect_ratio
+        self.pad_to_aspect_ratio = pad_to_aspect_ratio
         self.interpolation = interpolation
         self.data_format = standardize_data_format(data_format)
 
@@ -182,6 +191,7 @@ class ImageConverter(PreprocessingLayer):
                 "offset": self.offset,
                 "interpolation": self.interpolation,
                 "crop_to_aspect_ratio": self.crop_to_aspect_ratio,
+                "pad_to_aspect_ratio": self.pad_to_aspect_ratio,
             }
         )
         return config

keras_hub/src/models/falcon/falcon_attention.py

@@ -110,9 +110,11 @@ class FalconAttention(keras.layers.Layer):
 
         attention_scores = ops.einsum("bqnh,bknh->bnqk", query, key)
         attention_scores = ops.add(attention_scores, alibi)
-        attention_scores = (
-            attention_scores * self.inv_norm_factor
-        )  # [batch_size, num_heads, query_length, kv_length]
+        # [batch_size, num_heads, query_length, kv_length]
+        attention_scores = ops.multiply(
+            attention_scores,
+            ops.cast(self.inv_norm_factor, self.compute_dtype),
+        )
         attention_scores = self.softmax(
             attention_scores, ops.expand_dims(attention_mask, 1)
         )
@@ -120,6 +122,7 @@ class FalconAttention(keras.layers.Layer):
         attention_output = ops.einsum(
             "bnqk,bknh->bqnh", attention_scores, value
         )
+
         attention_output = ops.reshape(
             attention_output,
             [batch_size, seq_length, self.num_heads * self.head_dim],

keras_hub/src/models/gpt_neo_x/gpt_neo_x_attention.py

@@ -1,8 +1,11 @@
+import math
+
 import keras
 from keras import ops
 
 from keras_hub.src.layers.modeling.rotary_embedding import RotaryEmbedding
 from keras_hub.src.utils.keras_utils import clone_initializer
+from keras_hub.src.utils.keras_utils import has_flash_attention_support
 
 
 class GPTNeoXAttention(keras.layers.Layer):
@@ -58,6 +61,8 @@ class GPTNeoXAttention(keras.layers.Layer):
         self.bias_initializer = keras.initializers.get(bias_initializer)
         self.max_sequence_length = max_sequence_length
 
+        self._inv_norm_factor = 1.0 / math.sqrt(self.attn_head_size)
+
     def build(self, input_shape):
         self._qkv_dense = keras.layers.EinsumDense(
             equation="abc,cde->abde",
@@ -120,14 +125,26 @@ class GPTNeoXAttention(keras.layers.Layer):
     def _compute_attention(
         self, query, key, value, attention_mask=None, training=None
     ):
-        attention_scores = ops.einsum("aecd,abcd->acbe", key, query)
+        if has_flash_attention_support() and self.dropout == 0:
+            # Use `dot_product_attention` with Flash Attention support if
+            # available.
+            if attention_mask is not None:
+                attention_mask = ops.expand_dims(attention_mask, axis=1)
+                attention_mask = ops.cast(attention_mask, dtype="bool")
+            attention_output = ops.dot_product_attention(
+                query,
+                key,
+                value,
+                mask=attention_mask,
+                scale=self._inv_norm_factor,
+            )
+            return attention_output
 
-        norm_factor = ops.sqrt(
-            ops.convert_to_tensor(self.attn_head_size, self.compute_dtype)
+        attention_scores = ops.einsum("aecd,abcd->acbe", key, query)
+        attention_scores = ops.multiply(
+            attention_scores,
+            ops.cast(self._inv_norm_factor, self.compute_dtype),
         )
-
-        attention_scores /= norm_factor
-
         attention_scores = self._masked_softmax(
             attention_scores, attention_mask
         )

keras_hub/src/models/llama/llama_attention.py

@@ -1,8 +1,11 @@
+import math
+
 import keras
 from keras import ops
 
 from keras_hub.src.layers.modeling.rotary_embedding import RotaryEmbedding
 from keras_hub.src.utils.keras_utils import clone_initializer
+from keras_hub.src.utils.keras_utils import has_flash_attention_support
 
 
 class LlamaAttention(keras.layers.Layer):
@@ -43,7 +46,7 @@ class LlamaAttention(keras.layers.Layer):
         # h = head dim
         hidden_dim = inputs_shape[-1]
         head_dim = hidden_dim // self.num_query_heads
-        self._norm_factor = ops.sqrt(ops.cast(head_dim, self.compute_dtype))
+        self._inv_norm_factor = 1.0 / math.sqrt(head_dim)
 
         self._query_dense = keras.layers.EinsumDense(
             equation="bqm,muh->bquh",
@@ -182,9 +185,27 @@ class LlamaAttention(keras.layers.Layer):
         return self._softmax(attention_scores)
 
     def _compute_attention(self, query, key, value, attention_mask=None):
+        if has_flash_attention_support():
+            # Use `dot_product_attention` with Flash Attention support if
+            # available.
+            if attention_mask is not None:
+                attention_mask = ops.expand_dims(attention_mask, axis=1)
+                attention_mask = ops.cast(attention_mask, dtype="bool")
+            attention_output = ops.dot_product_attention(
+                query,
+                key,
+                value,
+                mask=attention_mask,
+                scale=self._inv_norm_factor,
+            )
+            return attention_output
+
         attention_scores = ops.einsum(self._dot_product_equation, query, key)
 
-        attention_scores = attention_scores / self._norm_factor
+        attention_scores = ops.multiply(
+            attention_scores,
+            ops.cast(self._inv_norm_factor, self.compute_dtype),
+        )
         attention_scores = self._masked_softmax(
             attention_scores, attention_mask
         )

keras_hub/src/models/mistral/mistral_attention.py

@@ -1,8 +1,11 @@
+import math
+
 import keras
 from keras import ops
 
 from keras_hub.src.layers.modeling.rotary_embedding import RotaryEmbedding
 from keras_hub.src.utils.keras_utils import clone_initializer
+from keras_hub.src.utils.keras_utils import has_flash_attention_support
 
 
 # This is just a self-attention layer in Mistral. But it can be generalized
@@ -52,6 +55,7 @@ class CachedMistralAttention(keras.layers.Layer):
         # h = head dim
         self._hidden_dim = inputs_shape[-1]
         self._head_dim = self._hidden_dim // self._num_query_heads
+        self._inv_norm_factor = 1.0 / math.sqrt(self._head_dim)
 
         self._query_dense = keras.layers.EinsumDense(
             equation="bqm,muh->bquh",
@@ -192,11 +196,26 @@ class CachedMistralAttention(keras.layers.Layer):
         return self._softmax(attention_scores)
 
     def _compute_attention(self, query, key, value, attention_mask=None):
-        attention_scores = ops.einsum(self._dot_product_equation, query, key)
-
-        norm_factor = ops.sqrt(ops.cast(self._head_dim, self.compute_dtype))
+        if has_flash_attention_support():
+            # Use `dot_product_attention` with Flash Attention support if
+            # available.
+            if attention_mask is not None:
+                attention_mask = ops.expand_dims(attention_mask, axis=1)
+                attention_mask = ops.cast(attention_mask, dtype="bool")
+            attention_output = ops.dot_product_attention(
+                query,
+                key,
+                value,
+                mask=attention_mask,
+                scale=self._inv_norm_factor,
+            )
+            return attention_output
 
-        attention_scores = attention_scores / norm_factor
+        attention_scores = ops.einsum(self._dot_product_equation, query, key)
+        attention_scores = ops.multiply(
+            attention_scores,
+            ops.cast(self._inv_norm_factor, self.compute_dtype),
+        )
         attention_scores = self._masked_softmax(
             attention_scores, attention_mask
         )

keras_hub/src/models/phi3/phi3_attention.py

@@ -1,3 +1,5 @@
+import math
+
 import keras
 from keras import ops
 
@@ -6,6 +8,7 @@ from keras_hub.src.models.phi3.phi3_rotary_embedding import (
     Phi3SuScaledRotaryEmbedding,
 )
 from keras_hub.src.utils.keras_utils import clone_initializer
+from keras_hub.src.utils.keras_utils import has_flash_attention_support
 
 
 class Phi3Attention(keras.layers.Layer):
@@ -53,7 +56,7 @@ class Phi3Attention(keras.layers.Layer):
         # h = head dim
         hidden_dim = inputs_shape[-1]
         head_dim = hidden_dim // self.num_query_heads
-        self._norm_factor = ops.sqrt(ops.cast(head_dim, self.compute_dtype))
+        self._inv_norm_factor = 1.0 / math.sqrt(head_dim)
 
         self.query_dense = keras.layers.EinsumDense(
             equation="bqm,muh->bquh",
@@ -214,8 +217,26 @@ class Phi3Attention(keras.layers.Layer):
         return self.softmax(attention_scores)
 
     def _compute_attention(self, query, key, value, attention_mask=None):
+        if has_flash_attention_support():
+            # Use `dot_product_attention` with Flash Attention support if
+            # available.
+            if attention_mask is not None:
+                attention_mask = ops.expand_dims(attention_mask, axis=1)
+                attention_mask = ops.cast(attention_mask, dtype="bool")
+            attention_output = ops.dot_product_attention(
+                query,
+                key,
+                value,
+                mask=attention_mask,
+                scale=self._inv_norm_factor,
+            )
+            return attention_output
+
         attention_scores = ops.einsum("bquh,bkuh->buqk", query, key)
-        attention_scores = attention_scores / self._norm_factor
+        attention_scores = ops.multiply(
+            attention_scores,
+            ops.cast(self._inv_norm_factor, self.compute_dtype),
+        )
         attention_scores = self._masked_softmax(
             attention_scores, attention_mask
         )

keras_hub/src/models/stable_diffusion_3/mmdit.py

@@ -7,6 +7,7 @@ from keras import ops
 from keras_hub.src.layers.modeling.position_embedding import PositionEmbedding
 from keras_hub.src.models.backbone import Backbone
 from keras_hub.src.utils.keras_utils import gelu_approximate
+from keras_hub.src.utils.keras_utils import has_flash_attention_support
 from keras_hub.src.utils.keras_utils import standardize_data_format
 
 
@@ -770,17 +771,14 @@ class MMDiTBlock(layers.Layer):
     def _compute_attention(self, query, key, value):
         batch_size = ops.shape(query)[0]
 
-        # Use the fast path when `ops.dot_product_attention` and flash attention
-        # are available.
-        if hasattr(ops, "dot_product_attention") and hasattr(
-            keras.config, "is_flash_attention_enabled"
-        ):
+        if has_flash_attention_support():
+            # Use `dot_product_attention` with Flash Attention support if
+            # available.
             encoded = ops.dot_product_attention(
                 query,
                 key,
                 value,
                 scale=self._inverse_sqrt_key_dim,
-                flash_attention=keras.config.is_flash_attention_enabled(),
             )
             return ops.reshape(
                 encoded, (batch_size, -1, self.num_heads * self.head_dim)
@@ -793,10 +791,9 @@ class MMDiTBlock(layers.Layer):
         probs = self.softmax(logits)
         probs = ops.cast(probs, self.compute_dtype)
         encoded = ops.einsum("BNTS,BSNH->BTNH", probs, value)
-        encoded = ops.reshape(
+        return ops.reshape(
             encoded, (batch_size, -1, self.num_heads * self.head_dim)
         )
-        return encoded
 
     def call(self, inputs, context, timestep_embedding, training=None):
         # Compute pre-attention.

keras_hub/src/utils/keras_utils.py

@@ -53,3 +53,10 @@ def standardize_data_format(data_format):
             f"Received: data_format={data_format}"
         )
     return data_format
+
+
+def has_flash_attention_support():
+    if hasattr(keras.config, "is_flash_attention_enabled"):
+        return True
+    else:
+        return False

keras_hub/src/utils/preset_utils.py

@@ -240,7 +240,7 @@ def tf_copy_gfile_to_cache(preset, path):
         try:
             import tensorflow as tf
 
-            os.make_dirs(os.path.dirname(local_path), exist_ok=True)
+            os.makedirs(os.path.dirname(local_path), exist_ok=True)
             tf.io.gfile.copy(url, local_path)
         except Exception as e:
             # gfile.copy will leave an empty file after an error.

keras_hub/src/version_utils.py

@@ -1,7 +1,7 @@
 from keras_hub.src.api_export import keras_hub_export
 
 # Unique source of truth for the version number.
-__version__ = "0.19.0.dev202501270344"
+__version__ = "0.19.0.dev202501290343"
 
 
 @keras_hub_export("keras_hub.version")

{keras_hub_nightly-0.19.0.dev202501270344.dist-info → keras_hub_nightly-0.19.0.dev202501290343.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: keras-hub-nightly
-Version: 0.19.0.dev202501270344
+Version: 0.19.0.dev202501290343
 Summary: Industry-strength Natural Language Processing extensions for Keras.
 Home-page: https://github.com/keras-team/keras-hub
 Author: Keras team

{keras_hub_nightly-0.19.0.dev202501270344.dist-info → keras_hub_nightly-0.19.0.dev202501290343.dist-info}/RECORD

@@ -9,7 +9,7 @@ keras_hub/api/tokenizers/__init__.py,sha256=mtJgQy1spfQnPAkeLoeinsT_W9iCWHlJXwzc
 keras_hub/api/utils/__init__.py,sha256=Gp1E6gG-RtKQS3PBEQEOz9PQvXkXaJ0ySGMqZ7myN7A,215
 keras_hub/src/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 keras_hub/src/api_export.py,sha256=9pQZK27JObxWZ96QPLBp1OBsjWigh1iuV6RglPGMRk0,1499
-keras_hub/src/version_utils.py,sha256=zEJjYSSegzR2k-8rtXOP5PNiX00pt5mYEbJZMwvWIos,222
+keras_hub/src/version_utils.py,sha256=I_PEW1lDeDRalBs5L1P1JM6o85YjT0_X-wZPbsOS_iI,222
 keras_hub/src/bounding_box/__init__.py,sha256=7i6KnGupN4AVivR_dFjQyuuTbI0GkHy8d-aMXeqZdU8,95
 keras_hub/src/bounding_box/converters.py,sha256=UUp1hwegpDZyIo8sh9TLNy1v6JjwmvwzL6wmHFMAtbk,21916
 keras_hub/src/bounding_box/formats.py,sha256=YmskOz2BOSat7NaE__J9VfpSNGPJJR0znSzA4lp8MMI,3868
@@ -35,7 +35,7 @@ keras_hub/src/layers/modeling/transformer_encoder.py,sha256=Qe19_aR6w4PTFbzvBmSP
 keras_hub/src/layers/modeling/transformer_layer_utils.py,sha256=FuznrW33iG50B-VDN8R1RjuA5JG72yNMJ1TBgWLxR0E,3487
 keras_hub/src/layers/preprocessing/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 keras_hub/src/layers/preprocessing/audio_converter.py,sha256=YGh_kQw65a1Z6S5zzSNVP-ChyLYHq3-eOYpOS53xIN8,4156
-keras_hub/src/layers/preprocessing/image_converter.py,sha256=QZr1XGsIR67-wuTspHgBt9a44mjuwIw9b5frzSF5Ia8,10542
+keras_hub/src/layers/preprocessing/image_converter.py,sha256=XwqgHYWj0Z14UMGQw5E4pOm3MmgbuvQpBcKl36e-nvo,10962
 keras_hub/src/layers/preprocessing/masked_lm_mask_generator.py,sha256=itxWq3FHYlR0I7jKarQlSKbSmRLl9ut_UTSP3ZDwP0A,8162
 keras_hub/src/layers/preprocessing/multi_segment_packer.py,sha256=ZNqnUFnc9Af122Q7T6YyUoXgIdU9AgIJfsvR1UrCjFU,12068
 keras_hub/src/layers/preprocessing/preprocessing_layer.py,sha256=WyX41b9Ev_YJ5uVQVOAqD0PQasMOPDoyDjl_PkzkAkE,687
@@ -174,7 +174,7 @@ keras_hub/src/models/f_net/f_net_text_classifier.py,sha256=YoWq08mcn-oOsdiajxLy2
 keras_hub/src/models/f_net/f_net_text_classifier_preprocessor.py,sha256=UUa7RKylLt41Z0wRxGzhSgWTaJjNAcgqkVeC-ZzJbfo,4822
 keras_hub/src/models/f_net/f_net_tokenizer.py,sha256=ZRTaSfgZnYLTVXgM51303LpryRsSL5GaC2Cl_D7g27A,2285
 keras_hub/src/models/falcon/__init__.py,sha256=IVwPgPbw0l8XJRPQETmeQNvpdn_SneXhe_3oRMOvdx8,257
-keras_hub/src/models/falcon/falcon_attention.py,sha256=nBpvh3KGElNG062NfqznNJmTqKGN-0k_VZ7j7DryjMI,4497
+keras_hub/src/models/falcon/falcon_attention.py,sha256=fRHuK_y_w64hrqq0XYfcsycs3KD1_3RmeKP7j8LEjGU,4559
 keras_hub/src/models/falcon/falcon_backbone.py,sha256=nGJcHnbqncZRTPERRi4ZuYGcODpkH2Mu0-Db59vH5io,5451
 keras_hub/src/models/falcon/falcon_causal_lm.py,sha256=2UEIeju5Tg-FstVuusejJ-MbHZ6vsNfsSJzzBM89fnU,10908
 keras_hub/src/models/falcon/falcon_causal_lm_preprocessor.py,sha256=nI9E8N9enx5DppDHpLwGslb65rqGorL2sEz1jzet4gA,3033
@@ -205,14 +205,14 @@ keras_hub/src/models/gpt2/gpt2_preprocessor.py,sha256=eYMIXw8Oebsr14GhqBh1CEhbLb
 keras_hub/src/models/gpt2/gpt2_presets.py,sha256=1mflR1dVuEwFfNe3Fkra6vt7DrjmkAckjyP-LclNLFc,1897
 keras_hub/src/models/gpt2/gpt2_tokenizer.py,sha256=-wA7ZTiRZDVdEKAskyQiXB5GLYuj9TkuADtyyYpoBuA,2615
 keras_hub/src/models/gpt_neo_x/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-keras_hub/src/models/gpt_neo_x/gpt_neo_x_attention.py,sha256=iCwNoeaHMLfnL-MLOeLir7G75XRJilvpmdKJeBAqLTY,8535
+keras_hub/src/models/gpt_neo_x/gpt_neo_x_attention.py,sha256=IwfmmAlndr8W1VfXXMK9lkncBxo-AHYqzLo_3hS-k_s,9285
 keras_hub/src/models/gpt_neo_x/gpt_neo_x_backbone.py,sha256=yov92B8j9lXz-9ZOpLa_PLT7WgcRKWG8fwB824Z_1hw,6508
 keras_hub/src/models/gpt_neo_x/gpt_neo_x_causal_lm.py,sha256=HriMXNVjGlFTjCIgfLRR3AvsOPbjGToDM_XY-bdger0,7690
 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/llama/__init__.py,sha256=svVZjGi71R3lVbq0AdbqlXj909mr3Rp9EPXdiO0w0G0,251
-keras_hub/src/models/llama/llama_attention.py,sha256=HzTWtvTjfN_j0vA9-ComstHpI81tzUrJU3RSSvSCaI4,7194
+keras_hub/src/models/llama/llama_attention.py,sha256=i2OHkBAuC7iKDZmZF9eRUilxTKjbmLYXXyO2vbSLT8A,7929
 keras_hub/src/models/llama/llama_backbone.py,sha256=tjNEIKIL9ncoEL5KNFE5i0oTUkysjmJmh3mHmCz4RCw,11861
 keras_hub/src/models/llama/llama_causal_lm.py,sha256=9bP4-XDCMgsZuH1ILIMzmwq2Fyy6vkk1Vsht-lMGCNo,13258
 keras_hub/src/models/llama/llama_causal_lm_preprocessor.py,sha256=VTboOMiRBoxHrwP343upLUTsv3AG65r2H8h_PNPVphE,3047
@@ -227,7 +227,7 @@ keras_hub/src/models/llama3/llama3_causal_lm_preprocessor.py,sha256=twbXel9hsQgG
 keras_hub/src/models/llama3/llama3_presets.py,sha256=PWEW_hLMCD9SIYm3QLhRVIcwjrPuqv-KDebXACXRNbM,1579
 keras_hub/src/models/llama3/llama3_tokenizer.py,sha256=J-KxRc08vGs4olFw_4mtJs0W_dTeUyj_XxMycazBmxI,1934
 keras_hub/src/models/mistral/__init__.py,sha256=vjBlzcrIsFSwJKnfwfTNMKstIEKGFTE3kVcdAdfwlnE,263
-keras_hub/src/models/mistral/mistral_attention.py,sha256=HCkUIc2DVIlYC5hhwomENlqLOsKTvbCKF0lx0_OBAyA,7862
+keras_hub/src/models/mistral/mistral_attention.py,sha256=00_bwlkBkxZoCuqhy4aMV6o-8kc4Ek06_m-ZeyPlsE8,8607
 keras_hub/src/models/mistral/mistral_backbone.py,sha256=oatoqSX0z-xjKfXeSveL4P0Dcr84kd0o0pK_kuTbJtY,7257
 keras_hub/src/models/mistral/mistral_causal_lm.py,sha256=ujCKfsbuYzr8VusqPYcnTH6rTb0MRfzsinEraVhQksc,13234
 keras_hub/src/models/mistral/mistral_causal_lm_preprocessor.py,sha256=_4qq-uKktfIg_i081ZWjZGEIYZpedBwtBGpchQQ-qEk,3079
@@ -261,7 +261,7 @@ keras_hub/src/models/pali_gemma/pali_gemma_presets.py,sha256=Ka1ChUUSKw-yY2th3Qt
 keras_hub/src/models/pali_gemma/pali_gemma_tokenizer.py,sha256=ljTiADHo0Ok88q-jVzwJIle2C8xcxnudLTsBLzIySaM,2415
 keras_hub/src/models/pali_gemma/pali_gemma_vit.py,sha256=ViPKfGksbxBGJ3iS3M_KWxRc8Ie4LF7rWWUKDiqECJE,18285
 keras_hub/src/models/phi3/__init__.py,sha256=zIbf1MU-ks91mEkjTRJAsk51N3BBnXDF2JM1vO-13PQ,245
-keras_hub/src/models/phi3/phi3_attention.py,sha256=dN8QwwTP9TxPBDv0MCvObLF3nHm1H6xbYr3T1K0nmg8,9243
+keras_hub/src/models/phi3/phi3_attention.py,sha256=7eeEP3za6arfqzyLtrJV6desV2_9b6bR5LkupeURrHs,9978
 keras_hub/src/models/phi3/phi3_backbone.py,sha256=fY-OY2ZrqxDHglYjTM0OCacBdEQHwj-XNmU0MnXL7iU,8885
 keras_hub/src/models/phi3/phi3_causal_lm.py,sha256=kMMq7fQ8hlb_mLO_nU1lGVqILayulVvzzZgl2EvY9_k,8389
 keras_hub/src/models/phi3/phi3_causal_lm_preprocessor.py,sha256=gNx1k-n7d0XDwpNbcZiO9yLkwdXYCvwGyA3b0QCnPAE,3043
@@ -314,7 +314,7 @@ keras_hub/src/models/segformer/segformer_image_segmenter_preprocessor.py,sha256=
 keras_hub/src/models/segformer/segformer_presets.py,sha256=ET39ospixkTaCsjoMLdJrr3wlGvTAQu5prleVC5lMZI,4793
 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=0gq2tcIqcbiGKKDDj3vrRsF67U3qE9g706XPs2BfCOY,40979
+keras_hub/src/models/stable_diffusion_3/mmdit.py,sha256=ls0DCsyAP2VMdaKSZ4xm_LaWkvxokCMFWgQCKVWJRLQ,40857
 keras_hub/src/models/stable_diffusion_3/stable_diffusion_3_backbone.py,sha256=w8lsMampk34M9xQi96mEnXmkaKQqFQtoFTW8zP7ilEA,24078
 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
@@ -391,9 +391,9 @@ keras_hub/src/tokenizers/unicode_codepoint_tokenizer.py,sha256=hRv_XxoPIPDpHfO0Z
 keras_hub/src/tokenizers/word_piece_tokenizer.py,sha256=vP6AZgbzsRiuPCt3W_n94nsF7XiERnagWcH_rqJHtVU,19943
 keras_hub/src/tokenizers/word_piece_tokenizer_trainer.py,sha256=cylrs02ZrYQ1TuZr9oyS3NrVbDwGctA3VXbIh1pFJMQ,6743
 keras_hub/src/utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-keras_hub/src/utils/keras_utils.py,sha256=0yKIfFuO_IqAH8vHbG3ncRmCVKg__xRGfQtLYWZ8YuA,1695
+keras_hub/src/utils/keras_utils.py,sha256=ZULqIQylAQen-_pNC96htvLaxSJbfAenNoCo3ZSvY5g,1843
 keras_hub/src/utils/pipeline_model.py,sha256=jgzB6NQPSl0KOu08N-TazfOnXnUJbZjH2EXXhx25Ftg,9084
-keras_hub/src/utils/preset_utils.py,sha256=gy0zjPZ3WYvB5LHekw60NU8bHdrV6qUMG84DuN5mT6M,30505
+keras_hub/src/utils/preset_utils.py,sha256=cRsviMUs-Xskg5KefJ-bQCL9y30yJFyVg3RtvmVCo8o,30504
 keras_hub/src/utils/python_utils.py,sha256=N8nWeO3san4YnGkffRXG3Ix7VEIMTKSN21FX5TuL7G8,202
 keras_hub/src/utils/tensor_utils.py,sha256=YVJesN91bk-OzJXY1mOKBppuY8noBU7zhPQNXPxZVGc,14646
 keras_hub/src/utils/imagenet/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -417,7 +417,7 @@ keras_hub/src/utils/transformers/convert_pali_gemma.py,sha256=B1leeDw96Yvu81hYum
 keras_hub/src/utils/transformers/convert_vit.py,sha256=9SUZ9utNJhW_5cj3acMn9cRy47u2eIcDsrhmzj77o9k,5187
 keras_hub/src/utils/transformers/preset_loader.py,sha256=DgGJXbTSB9Na8FIR-YWWVqQPOFxHwWrGm41EwcS_EFs,3797
 keras_hub/src/utils/transformers/safetensor_utils.py,sha256=CYUHyA4y-B61r7NDnCsFb4t_UmSwZ1k9L-8gzEd6KRg,3339
-keras_hub_nightly-0.19.0.dev202501270344.dist-info/METADATA,sha256=6tqDd98GQK3QD2ep9PaI3cMcNM3k2TXL-LOO6d6R9iw,7498
-keras_hub_nightly-0.19.0.dev202501270344.dist-info/WHEEL,sha256=In9FTNxeP60KnTkGw7wk6mJPYd_dQSjEZmXdBdMCI-8,91
-keras_hub_nightly-0.19.0.dev202501270344.dist-info/top_level.txt,sha256=N4J6piIWBKa38A4uV-CnIopnOEf8mHAbkNXafXm_CuA,10
-keras_hub_nightly-0.19.0.dev202501270344.dist-info/RECORD,,
+keras_hub_nightly-0.19.0.dev202501290343.dist-info/METADATA,sha256=oJJlEUSh0UxqwuSrTzCcAswHdqKxHDKg6NYVe0HoSGk,7498
+keras_hub_nightly-0.19.0.dev202501290343.dist-info/WHEEL,sha256=In9FTNxeP60KnTkGw7wk6mJPYd_dQSjEZmXdBdMCI-8,91
+keras_hub_nightly-0.19.0.dev202501290343.dist-info/top_level.txt,sha256=N4J6piIWBKa38A4uV-CnIopnOEf8mHAbkNXafXm_CuA,10
+keras_hub_nightly-0.19.0.dev202501290343.dist-info/RECORD,,