keras-hub-nightly 0.19.0.dev202501080345__py3-none-any.whl → 0.19.0.dev202501150344__py3-none-any.whl

keras_hub/src/metrics/bleu.py

@@ -329,8 +329,9 @@ class Bleu(keras.metrics.Metric):
                 return tf.squeeze(inputs, axis=-1)
             else:
                 raise ValueError(
-                    f"{tensor_name} must be of rank {base_rank}, {base_rank+1} "
-                    f"or {base_rank+2}. Found rank: {inputs.shape.rank}"
+                    f"{tensor_name} must be of rank {base_rank}, "
+                    f"{base_rank + 1}, or {base_rank + 2}. "
+                    f"Found rank: {inputs.shape.rank}"
                 )
 
         y_true = validate_and_fix_rank(y_true, "y_true", 1)

keras_hub/src/models/basnet/basnet_backbone.py

@@ -219,7 +219,7 @@ def get_resnet_block(_resnet, block_num):
     else:
         x = _resnet.pyramid_outputs[extractor_levels[block_num - 1]]
     y = _resnet.get_layer(
-        f"stack{block_num}_block{num_blocks[block_num]-1}_add"
+        f"stack{block_num}_block{num_blocks[block_num] - 1}_add"
     ).output
     return keras.models.Model(
         inputs=x,

keras_hub/src/models/deeplab_v3/deeplab_v3_layers.py

@@ -88,13 +88,13 @@ class SpatialPyramidPooling(keras.layers.Layer):
                         dilation_rate=dilation_rate,
                         use_bias=False,
                         data_format=self.data_format,
-                        name=f"aspp_conv_{i+2}",
+                        name=f"aspp_conv_{i + 2}",
                     ),
                     keras.layers.BatchNormalization(
-                        axis=self.channel_axis, name=f"aspp_bn_{i+2}"
+                        axis=self.channel_axis, name=f"aspp_bn_{i + 2}"
                     ),
                     keras.layers.Activation(
-                        self.activation, name=f"aspp_activation_{i+2}"
+                        self.activation, name=f"aspp_activation_{i + 2}"
                     ),
                 ]
             )

keras_hub/src/models/densenet/densenet_backbone.py

@@ -81,14 +81,14 @@ class DenseNetBackbone(FeaturePyramidBackbone):
                 channel_axis,
                 stackwise_num_repeats[stack_index],
                 growth_rate,
-                name=f"stack{stack_index+1}",
+                name=f"stack{stack_index + 1}",
             )
             pyramid_outputs[f"P{index}"] = x
             x = apply_transition_block(
                 x,
                 channel_axis,
                 compression_ratio,
-                name=f"transition{stack_index+1}",
+                name=f"transition{stack_index + 1}",
             )
 
         x = apply_dense_block(
@@ -140,7 +140,7 @@ def apply_dense_block(x, channel_axis, num_repeats, growth_rate, name=None):
 
     for i in range(num_repeats):
         x = apply_conv_block(
-            x, channel_axis, growth_rate, name=f"{name}_block{i+1}"
+            x, channel_axis, growth_rate, name=f"{name}_block{i + 1}"
         )
     return x
 

keras_hub/src/models/flux/flux_text_to_image.py

@@ -81,7 +81,7 @@ class FluxTextToImage(TextToImage):
 
     def fit(self, *args, **kwargs):
         raise NotImplementedError(
-            "Currently, `fit` is not supported for " "`FluxTextToImage`."
+            "Currently, `fit` is not supported for `FluxTextToImage`."
         )
 
     def generate_step(

keras_hub/src/models/pali_gemma/pali_gemma_presets.py

@@ -5,7 +5,7 @@ backbone_presets = {
     "pali_gemma_3b_mix_224": {
         "metadata": {
             "description": (
-                "image size 224, mix fine tuned, text sequence " "length is 256"
+                "image size 224, mix fine tuned, text sequence length is 256"
             ),
             "params": 2923335408,
             "path": "pali_gemma",
@@ -45,7 +45,7 @@ backbone_presets = {
     "pali_gemma_3b_896": {
         "metadata": {
             "description": (
-                "image size 896, pre trained, text sequence length " "is 512"
+                "image size 896, pre trained, text sequence length is 512"
             ),
             "params": 2927759088,
             "path": "pali_gemma",

keras_hub/src/models/resnet/resnet_backbone.py

@@ -177,7 +177,7 @@ class ResNetBackbone(FeaturePyramidBackbone):
                 use_bias=False,
                 padding="same",
                 dtype=dtype,
-                name=f"conv{conv_index+1}_conv",
+                name=f"conv{conv_index + 1}_conv",
             )(x)
 
         if not use_pre_activation:

keras_hub/src/models/retinanet/feature_pyramid.py

@@ -209,9 +209,9 @@ class FeaturePyramid(keras.layers.Layer):
             )
             if i == backbone_max_level + 1 and self.use_p5:
                 self.output_conv_layers[level].build(
-                    (None, None, None, input_shapes[f"P{i-1}"][-1])
+                    (None, None, None, input_shapes[f"P{i - 1}"][-1])
                     if self.data_format == "channels_last"
-                    else (None, input_shapes[f"P{i-1}"][1], None, None)
+                    else (None, input_shapes[f"P{i - 1}"][1], None, None)
                 )
             else:
                 self.output_conv_layers[level].build(
@@ -277,7 +277,7 @@ class FeaturePyramid(keras.layers.Layer):
             if i < backbone_max_level:
                 # for the top most output, it doesn't need to merge with any
                 # upper stream outputs
-                upstream_output = self.top_down_op(output_features[f"P{i+1}"])
+                upstream_output = self.top_down_op(output_features[f"P{i + 1}"])
                 output = self.merge_op([output, upstream_output])
             output_features[level] = (
                 self.lateral_batch_norm_layers[level](output)
@@ -296,9 +296,9 @@ class FeaturePyramid(keras.layers.Layer):
         for i in range(backbone_max_level + 1, self.max_level + 1):
             level = f"P{i}"
             feats_in = (
-                inputs[f"P{i-1}"]
+                inputs[f"P{i - 1}"]
                 if i == backbone_max_level + 1 and self.use_p5
-                else output_features[f"P{i-1}"]
+                else output_features[f"P{i - 1}"]
             )
             if i > backbone_max_level + 1:
                 feats_in = self.activation(feats_in)

keras_hub/src/models/stable_diffusion_3/mmdit.py

@@ -15,9 +15,8 @@ class AdaptiveLayerNormalization(layers.Layer):
 
     Args:
         embedding_dim: int. The size of each embedding vector.
-        residual_modulation: bool. Whether to output the modulation parameters
-            of the residual connection within the block of the diffusion
-            transformers. Defaults to `False`.
+        num_modulations: int. The number of the modulation parameters. The
+            available values are `2`, `6` and `9`. Defaults to `2`.
         **kwargs: other keyword arguments passed to `keras.layers.Layer`,
             including `name`, `dtype` etc.
 
@@ -28,11 +27,17 @@ class AdaptiveLayerNormalization(layers.Layer):
     https://arxiv.org/abs/2212.09748).
     """
 
-    def __init__(self, hidden_dim, residual_modulation=False, **kwargs):
+    def __init__(self, hidden_dim, num_modulations=2, **kwargs):
         super().__init__(**kwargs)
-        self.hidden_dim = int(hidden_dim)
-        self.residual_modulation = bool(residual_modulation)
-        num_modulations = 6 if self.residual_modulation else 2
+        hidden_dim = int(hidden_dim)
+        num_modulations = int(num_modulations)
+        if num_modulations not in (2, 6, 9):
+            raise ValueError(
+                "`num_modulations` must be `2`, `6` or `9`. "
+                f"Received: num_modulations={num_modulations}"
+            )
+        self.hidden_dim = hidden_dim
+        self.num_modulations = num_modulations
 
         self.silu = layers.Activation("silu", dtype=self.dtype_policy)
         self.dense = layers.Dense(
@@ -52,40 +57,84 @@ class AdaptiveLayerNormalization(layers.Layer):
         self.norm.build(inputs_shape)
 
     def call(self, inputs, embeddings, training=None):
-        x = inputs
+        hidden_states = inputs
         emb = self.dense(self.silu(embeddings), training=training)
-        if self.residual_modulation:
-            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
-                ops.split(emb, 6, axis=1)
-            )
+        if self.num_modulations == 9:
+            (
+                shift_msa,
+                scale_msa,
+                gate_msa,
+                shift_mlp,
+                scale_mlp,
+                gate_mlp,
+                shift_msa2,
+                scale_msa2,
+                gate_msa2,
+            ) = ops.split(emb, self.num_modulations, axis=1)
+        elif self.num_modulations == 6:
+            (
+                shift_msa,
+                scale_msa,
+                gate_msa,
+                shift_mlp,
+                scale_mlp,
+                gate_mlp,
+            ) = ops.split(emb, self.num_modulations, axis=1)
         else:
-            shift_msa, scale_msa = ops.split(emb, 2, axis=1)
+            shift_msa, scale_msa = ops.split(emb, self.num_modulations, axis=1)
+
         scale_msa = ops.expand_dims(scale_msa, axis=1)
         shift_msa = ops.expand_dims(shift_msa, axis=1)
-        x = ops.add(
-            ops.multiply(
-                self.norm(x, training=training),
-                ops.add(1.0, scale_msa),
-            ),
-            shift_msa,
+        norm_hidden_states = ops.cast(
+            self.norm(hidden_states, training=training), scale_msa.dtype
+        )
+        hidden_states = ops.add(
+            ops.multiply(norm_hidden_states, ops.add(1.0, scale_msa)), shift_msa
         )
-        if self.residual_modulation:
-            return x, gate_msa, shift_mlp, scale_mlp, gate_mlp
+
+        if self.num_modulations == 9:
+            scale_msa2 = ops.expand_dims(scale_msa2, axis=1)
+            shift_msa2 = ops.expand_dims(shift_msa2, axis=1)
+            hidden_states2 = ops.add(
+                ops.multiply(norm_hidden_states, ops.add(1.0, scale_msa2)),
+                shift_msa2,
+            )
+            return (
+                hidden_states,
+                gate_msa,
+                shift_mlp,
+                scale_mlp,
+                gate_mlp,
+                hidden_states2,
+                gate_msa2,
+            )
+        elif self.num_modulations == 6:
+            return hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp
         else:
-            return x
+            return hidden_states
 
     def get_config(self):
         config = super().get_config()
         config.update(
             {
                 "hidden_dim": self.hidden_dim,
-                "residual_modulation": self.residual_modulation,
+                "num_modulations": self.num_modulations,
             }
         )
         return config
 
     def compute_output_shape(self, inputs_shape, embeddings_shape):
-        if self.residual_modulation:
+        if self.num_modulations == 9:
+            return (
+                inputs_shape,
+                embeddings_shape,
+                embeddings_shape,
+                embeddings_shape,
+                embeddings_shape,
+                inputs_shape,
+                embeddings_shape,
+            )
+        elif self.num_modulations == 6:
             return (
                 inputs_shape,
                 embeddings_shape,
@@ -345,6 +394,27 @@ class TimestepEmbedding(layers.Layer):
         return output_shape
 
 
+def get_qk_norm(qk_norm=None, q_norm_name="q_norm", k_norm_name="k_norm"):
+    """Helper function to instantiate `LayerNormalization` layers."""
+    q_norm = None
+    k_norm = None
+    if qk_norm is None:
+        pass
+    elif qk_norm == "rms_norm":
+        q_norm = layers.LayerNormalization(
+            epsilon=1e-6, rms_scaling=True, dtype="float32", name=q_norm_name
+        )
+        k_norm = layers.LayerNormalization(
+            epsilon=1e-6, rms_scaling=True, dtype="float32", name=k_norm_name
+        )
+    else:
+        raise NotImplementedError(
+            "Supported `qk_norm` are `'rms_norm'` and `None`. "
+            f"Received: qk_norm={qk_norm}."
+        )
+    return q_norm, k_norm
+
+
 class DismantledBlock(layers.Layer):
     """A dismantled block used to compute pre- and post-attention.
 
@@ -356,6 +426,8 @@ class DismantledBlock(layers.Layer):
             the end of the block.
         qk_norm: Optional str. Whether to normalize the query and key tensors.
             Available options are `None` and `"rms_norm"`. Defaults to `None`.
+        use_dual_attention: bool. Whether to use a dual attention in the
+            block. Defaults to `False`.
         **kwargs: other keyword arguments passed to `keras.layers.Layer`,
             including `name`, `dtype` etc.
     """
@@ -367,6 +439,7 @@ class DismantledBlock(layers.Layer):
         mlp_ratio=4.0,
         use_projection=True,
         qk_norm=None,
+        use_dual_attention=False,
         **kwargs,
     ):
         super().__init__(**kwargs)
@@ -375,6 +448,7 @@ class DismantledBlock(layers.Layer):
         self.mlp_ratio = mlp_ratio
         self.use_projection = use_projection
         self.qk_norm = qk_norm
+        self.use_dual_attention = use_dual_attention
 
         head_dim = hidden_dim // num_heads
         self.head_dim = head_dim
@@ -384,7 +458,7 @@ class DismantledBlock(layers.Layer):
         if use_projection:
             self.ada_layer_norm = AdaptiveLayerNormalization(
                 hidden_dim,
-                residual_modulation=True,
+                num_modulations=9 if use_dual_attention else 6,
                 dtype=self.dtype_policy,
                 name="ada_layer_norm",
             )
@@ -395,18 +469,10 @@ class DismantledBlock(layers.Layer):
         self.attention_qkv = layers.Dense(
             hidden_dim * 3, dtype=self.dtype_policy, name="attention_qkv"
         )
-        if qk_norm is not None and qk_norm == "rms_norm":
-            self.q_norm = layers.LayerNormalization(
-                epsilon=1e-6, rms_scaling=True, dtype="float32", name="q_norm"
-            )
-            self.k_norm = layers.LayerNormalization(
-                epsilon=1e-6, rms_scaling=True, dtype="float32", name="q_norm"
-            )
-        elif qk_norm is not None:
-            raise NotImplementedError(
-                "Supported `qk_norm` are `'rms_norm'` and `None`. "
-                f"Received: qk_norm={qk_norm}."
-            )
+        q_norm, k_norm = get_qk_norm(qk_norm)
+        if q_norm is not None:
+            self.q_norm = q_norm
+            self.k_norm = k_norm
         if use_projection:
             self.attention_proj = layers.Dense(
                 hidden_dim, dtype=self.dtype_policy, name="attention_proj"
@@ -426,6 +492,19 @@ class DismantledBlock(layers.Layer):
                 name="mlp",
             )
 
+        if use_dual_attention:
+            self.attention_qkv2 = layers.Dense(
+                hidden_dim * 3, dtype=self.dtype_policy, name="attention_qkv2"
+            )
+            q_norm2, k_norm2 = get_qk_norm(qk_norm, "q_norm2", "k_norm2")
+            if q_norm is not None:
+                self.q_norm2 = q_norm2
+                self.k_norm2 = k_norm2
+            if use_projection:
+                self.attention_proj2 = layers.Dense(
+                    hidden_dim, dtype=self.dtype_policy, name="attention_proj2"
+                )
+
     def build(self, inputs_shape, timestep_embedding):
         self.ada_layer_norm.build(inputs_shape, timestep_embedding)
         self.attention_qkv.build(inputs_shape)
@@ -437,6 +516,13 @@ class DismantledBlock(layers.Layer):
             self.attention_proj.build(inputs_shape)
             self.norm2.build(inputs_shape)
             self.mlp.build(inputs_shape)
+        if self.use_dual_attention:
+            self.attention_qkv2.build(inputs_shape)
+            if self.qk_norm is not None:
+                self.q_norm2.build([None, None, self.num_heads, self.head_dim])
+                self.k_norm2.build([None, None, self.num_heads, self.head_dim])
+            if self.use_projection:
+                self.attention_proj2.build(inputs_shape)
 
     def _modulate(self, inputs, shift, scale):
         inputs = ops.cast(inputs, self.compute_dtype)
@@ -456,8 +542,12 @@ class DismantledBlock(layers.Layer):
             )
             q, k, v = ops.unstack(qkv, 3, axis=2)
             if self.qk_norm is not None:
-                q = self.q_norm(q, training=training)
-                k = self.k_norm(k, training=training)
+                q = ops.cast(
+                    self.q_norm(q, training=training), self.compute_dtype
+                )
+                k = ops.cast(
+                    self.k_norm(k, training=training), self.compute_dtype
+                )
             return (q, k, v), (inputs, gate_msa, shift_mlp, scale_mlp, gate_mlp)
         else:
             x = self.ada_layer_norm(
@@ -469,8 +559,12 @@ class DismantledBlock(layers.Layer):
             )
             q, k, v = ops.unstack(qkv, 3, axis=2)
             if self.qk_norm is not None:
-                q = self.q_norm(q, training=training)
-                k = self.k_norm(k, training=training)
+                q = ops.cast(
+                    self.q_norm(q, training=training), self.compute_dtype
+                )
+                k = ops.cast(
+                    self.k_norm(k, training=training), self.compute_dtype
+                )
             return (q, k, v)
 
     def _compute_post_attention(
@@ -495,22 +589,95 @@ class DismantledBlock(layers.Layer):
         )
         return x
 
+    def _compute_pre_attention_with_dual_attention(
+        self, inputs, timestep_embedding, training=None
+    ):
+        batch_size = ops.shape(inputs)[0]
+        x, gate_msa, shift_mlp, scale_mlp, gate_mlp, x2, gate_msa2 = (
+            self.ada_layer_norm(inputs, timestep_embedding, training=training)
+        )
+        # Compute the main attention
+        qkv = self.attention_qkv(x, training=training)
+        qkv = ops.reshape(
+            qkv, (batch_size, -1, 3, self.num_heads, self.head_dim)
+        )
+        q, k, v = ops.unstack(qkv, 3, axis=2)
+        if self.qk_norm is not None:
+            q = ops.cast(self.q_norm(q, training=training), self.compute_dtype)
+            k = ops.cast(self.k_norm(k, training=training), self.compute_dtype)
+        # Compute the dual attention
+        qkv2 = self.attention_qkv2(x2, training=training)
+        qkv2 = ops.reshape(
+            qkv2, (batch_size, -1, 3, self.num_heads, self.head_dim)
+        )
+        q2, k2, v2 = ops.unstack(qkv2, 3, axis=2)
+        if self.qk_norm is not None:
+            q2 = ops.cast(
+                self.q_norm2(q2, training=training), self.compute_dtype
+            )
+            k2 = ops.cast(
+                self.k_norm2(k2, training=training), self.compute_dtype
+            )
+        return (
+            (q, k, v),
+            (q2, k2, v2),
+            (inputs, gate_msa, shift_mlp, scale_mlp, gate_mlp, gate_msa2),
+        )
+
+    def _compute_post_attention_with_dual_attention(
+        self, inputs, inputs2, inputs_intermediates, training=None
+    ):
+        x, gate_msa, shift_mlp, scale_mlp, gate_mlp, gate_msa2 = (
+            inputs_intermediates
+        )
+        gate_msa = ops.expand_dims(gate_msa, axis=1)
+        shift_mlp = ops.expand_dims(shift_mlp, axis=1)
+        scale_mlp = ops.expand_dims(scale_mlp, axis=1)
+        gate_mlp = ops.expand_dims(gate_mlp, axis=1)
+        gate_msa2 = ops.expand_dims(gate_msa2, axis=1)
+        attn = self.attention_proj(inputs, training=training)
+        x = ops.add(x, ops.multiply(gate_msa, attn))
+        attn2 = self.attention_proj2(inputs2, training=training)
+        x = ops.add(x, ops.multiply(gate_msa2, attn2))
+        x = ops.add(
+            x,
+            ops.multiply(
+                gate_mlp,
+                self.mlp(
+                    self._modulate(self.norm2(x), shift_mlp, scale_mlp),
+                    training=training,
+                ),
+            ),
+        )
+        return x
+
     def call(
         self,
         inputs,
         timestep_embedding=None,
         inputs_intermediates=None,
+        inputs2=None,  # For the dual attention.
         pre_attention=True,
         training=None,
     ):
         if pre_attention:
-            return self._compute_pre_attention(
-                inputs, timestep_embedding, training=training
-            )
+            if self.use_dual_attention:
+                return self._compute_pre_attention_with_dual_attention(
+                    inputs, timestep_embedding, training=training
+                )
+            else:
+                return self._compute_pre_attention(
+                    inputs, timestep_embedding, training=training
+                )
         else:
-            return self._compute_post_attention(
-                inputs, inputs_intermediates, training=training
-            )
+            if self.use_dual_attention:
+                return self._compute_post_attention_with_dual_attention(
+                    inputs, inputs2, inputs_intermediates, training=training
+                )
+            else:
+                return self._compute_post_attention(
+                    inputs, inputs_intermediates, training=training
+                )
 
     def get_config(self):
         config = super().get_config()
@@ -521,6 +688,7 @@ class DismantledBlock(layers.Layer):
                 "mlp_ratio": self.mlp_ratio,
                 "use_projection": self.use_projection,
                 "qk_norm": self.qk_norm,
+                "use_dual_attention": self.use_dual_attention,
             }
         )
         return config
@@ -542,6 +710,8 @@ class MMDiTBlock(layers.Layer):
             layer at the end of the context block.
         qk_norm: Optional str. Whether to normalize the query and key tensors.
             Available options are `None` and `"rms_norm"`. Defaults to `None`.
+        use_dual_attention: bool. Whether to use a dual attention in the
+            block. Defaults to `False`.
         **kwargs: other keyword arguments passed to `keras.layers.Layer`,
             including `name`, `dtype` etc.
 
@@ -557,6 +727,7 @@ class MMDiTBlock(layers.Layer):
         mlp_ratio=4.0,
         use_context_projection=True,
         qk_norm=None,
+        use_dual_attention=False,
         **kwargs,
     ):
         super().__init__(**kwargs)
@@ -565,6 +736,7 @@ class MMDiTBlock(layers.Layer):
         self.mlp_ratio = mlp_ratio
         self.use_context_projection = use_context_projection
         self.qk_norm = qk_norm
+        self.use_dual_attention = use_dual_attention
 
         head_dim = hidden_dim // num_heads
         self.head_dim = head_dim
@@ -576,6 +748,7 @@ class MMDiTBlock(layers.Layer):
             mlp_ratio=mlp_ratio,
             use_projection=True,
             qk_norm=qk_norm,
+            use_dual_attention=use_dual_attention,
             dtype=self.dtype_policy,
             name="x_block",
         )
@@ -602,8 +775,6 @@ class MMDiTBlock(layers.Layer):
         if hasattr(ops, "dot_product_attention") and hasattr(
             keras.config, "is_flash_attention_enabled"
         ):
-            # `ops.dot_product_attention` is slower than the vanilla
-            # implementation in the tensorflow backend.
             encoded = ops.dot_product_attention(
                 query,
                 key,
@@ -643,9 +814,14 @@ class MMDiTBlock(layers.Layer):
                 training=training,
             )
         context_len = ops.shape(context_qkv[0])[1]
-        x_qkv, x_intermediates = self.x_block(
-            x, timestep_embedding=timestep_embedding, training=training
-        )
+        if self.x_block.use_dual_attention:
+            x_qkv, x_qkv2, x_intermediates = self.x_block(
+                x, timestep_embedding=timestep_embedding, training=training
+            )
+        else:
+            x_qkv, x_intermediates = self.x_block(
+                x, timestep_embedding=timestep_embedding, training=training
+            )
         q = ops.concatenate([context_qkv[0], x_qkv[0]], axis=1)
         k = ops.concatenate([context_qkv[1], x_qkv[1]], axis=1)
         v = ops.concatenate([context_qkv[2], x_qkv[2]], axis=1)
@@ -656,12 +832,23 @@ class MMDiTBlock(layers.Layer):
         x_attention = attention[:, context_len:]
 
         # Compute post-attention.
-        x = self.x_block(
-            x_attention,
-            inputs_intermediates=x_intermediates,
-            pre_attention=False,
-            training=training,
-        )
+        if self.x_block.use_dual_attention:
+            q2, k2, v2 = x_qkv2
+            x_attention2 = self._compute_attention(q2, k2, v2)
+            x = self.x_block(
+                x_attention,
+                inputs_intermediates=x_intermediates,
+                inputs2=x_attention2,
+                pre_attention=False,
+                training=training,
+            )
+        else:
+            x = self.x_block(
+                x_attention,
+                inputs_intermediates=x_intermediates,
+                pre_attention=False,
+                training=training,
+            )
         if self.use_context_projection:
             context = self.context_block(
                 context_attention,
@@ -682,6 +869,7 @@ class MMDiTBlock(layers.Layer):
                 "mlp_ratio": self.mlp_ratio,
                 "use_context_projection": self.use_context_projection,
                 "qk_norm": self.qk_norm,
+                "use_dual_attention": self.use_dual_attention,
             }
         )
         return config
@@ -761,6 +949,9 @@ class MMDiT(Backbone):
         qk_norm: Optional str. Whether to normalize the query and key tensors in
             the intermediate blocks. Available options are `None` and
             `"rms_norm"`. Defaults to `None`.
+        dual_attention_indices: Optional tuple. Specifies the indices of
+            the blocks that serve as dual attention blocks. Typically, this is
+            for 3.5 version. Defaults to `None`.
         data_format: `None` or str. If specified, either `"channels_last"` or
             `"channels_first"`. The ordering of the dimensions in the
             inputs. `"channels_last"` corresponds to inputs with shape
@@ -786,6 +977,7 @@ class MMDiT(Backbone):
         context_shape=(None, 4096),
         pooled_projection_shape=(2048,),
         qk_norm=None,
+        dual_attention_indices=None,
         data_format=None,
         dtype=None,
         **kwargs,
@@ -799,6 +991,7 @@ class MMDiT(Backbone):
         image_width = latent_shape[1] // patch_size
         output_dim = latent_shape[-1]
         output_dim_in_final = patch_size**2 * output_dim
+        dual_attention_indices = dual_attention_indices or ()
         data_format = standardize_data_format(data_format)
         if data_format != "channels_last":
             raise NotImplementedError(
@@ -840,6 +1033,7 @@ class MMDiT(Backbone):
                 mlp_ratio,
                 use_context_projection=not (i == num_layers - 1),
                 qk_norm=qk_norm,
+                use_dual_attention=i in dual_attention_indices,
                 dtype=dtype,
                 name=f"joint_block_{i}",
             )
@@ -910,6 +1104,7 @@ class MMDiT(Backbone):
         self.context_shape = context_shape
         self.pooled_projection_shape = pooled_projection_shape
         self.qk_norm = qk_norm
+        self.dual_attention_indices = dual_attention_indices
 
     def get_config(self):
         config = super().get_config()
@@ -925,6 +1120,7 @@ class MMDiT(Backbone):
                 "context_shape": self.context_shape,
                 "pooled_projection_shape": self.pooled_projection_shape,
                 "qk_norm": self.qk_norm,
+                "dual_attention_indices": self.dual_attention_indices,
             }
         )
         return config

keras_hub/src/models/stable_diffusion_3/stable_diffusion_3_backbone.py

@@ -205,7 +205,10 @@ class StableDiffusion3Backbone(Backbone):
         mmdit_qk_norm: Optional str. Whether to normalize the query and key
             tensors for each transformer in MMDiT. Available options are `None`
             and `"rms_norm"`. Typically, this is set to `None` for 3.0 version
-            and to `"rms_norm" for 3.5 version.
+            and to `"rms_norm"` for 3.5 version.
+        mmdit_dual_attention_indices: Optional tuple. Specifies the indices of
+            the blocks that serve as dual attention blocks. Typically, this is
+            for 3.5 version. Defaults to `None`.
         vae: The VAE used for transformations between pixel space and latent
             space.
         clip_l: The CLIP text encoder for encoding the inputs.
@@ -253,6 +256,7 @@ class StableDiffusion3Backbone(Backbone):
         mmdit_depth=4,
         mmdit_position_size=192,
         mmdit_qk_norm=None,
+        mmdit_dual_attention_indices=None,
         vae=vae,
         clip_l=clip_l,
         clip_g=clip_g,
@@ -268,6 +272,7 @@ class StableDiffusion3Backbone(Backbone):
         mmdit_num_heads,
         mmdit_position_size,
         mmdit_qk_norm,
+        mmdit_dual_attention_indices,
         vae,
         clip_l,
         clip_g,
@@ -319,6 +324,7 @@ class StableDiffusion3Backbone(Backbone):
             context_shape=context_shape,
             pooled_projection_shape=pooled_projection_shape,
             qk_norm=mmdit_qk_norm,
+            dual_attention_indices=mmdit_dual_attention_indices,
             data_format=data_format,
             dtype=dtype,
             name="diffuser",
@@ -454,6 +460,7 @@ class StableDiffusion3Backbone(Backbone):
         self.mmdit_num_heads = mmdit_num_heads
         self.mmdit_position_size = mmdit_position_size
         self.mmdit_qk_norm = mmdit_qk_norm
+        self.mmdit_dual_attention_indices = mmdit_dual_attention_indices
         self.latent_channels = latent_channels
         self.output_channels = output_channels
         self.num_train_timesteps = num_train_timesteps
@@ -590,6 +597,9 @@ class StableDiffusion3Backbone(Backbone):
                 "mmdit_num_heads": self.mmdit_num_heads,
                 "mmdit_position_size": self.mmdit_position_size,
                 "mmdit_qk_norm": self.mmdit_qk_norm,
+                "mmdit_dual_attention_indices": (
+                    self.mmdit_dual_attention_indices
+                ),
                 "vae": layers.serialize(self.vae),
                 "clip_l": layers.serialize(self.clip_l),
                 "clip_g": layers.serialize(self.clip_g),
@@ -638,7 +648,10 @@ class StableDiffusion3Backbone(Backbone):
             )
 
         # To maintain backward compatibility, we need to ensure that
-        # `mmdit_qk_norm` is included in the config.
+        # `mmdit_qk_norm` and `mmdit_dual_attention_indices` is included in the
+        # config.
         if "mmdit_qk_norm" not in config:
             config["mmdit_qk_norm"] = None
+        if "mmdit_dual_attention_indices" not in config:
+            config["mmdit_dual_attention_indices"] = None
         return cls(**config)

keras_hub/src/models/stable_diffusion_3/stable_diffusion_3_inpaint.py

@@ -82,8 +82,7 @@ class StableDiffusion3Inpaint(Inpaint):
 
     def fit(self, *args, **kwargs):
         raise NotImplementedError(
-            "Currently, `fit` is not supported for "
-            "`StableDiffusion3Inpaint`."
+            "Currently, `fit` is not supported for `StableDiffusion3Inpaint`."
         )
 
     def generate_step(

keras_hub/src/models/stable_diffusion_3/stable_diffusion_3_presets.py

@@ -13,6 +13,18 @@ backbone_presets = {
         },
         "kaggle_handle": "kaggle://keras/stablediffusion3/keras/stable_diffusion_3_medium/4",
     },
+    "stable_diffusion_3.5_medium": {
+        "metadata": {
+            "description": (
+                "3 billion parameter, including CLIP L and CLIP G text "
+                "encoders, MMDiT-X generative model, and VAE autoencoder. "
+                "Developed by Stability AI."
+            ),
+            "params": 3371793763,
+            "path": "stable_diffusion_3",
+        },
+        "kaggle_handle": "kaggle://keras/stablediffusion3/keras/stable_diffusion_3.5_medium/1",
+    },
     "stable_diffusion_3.5_large": {
         "metadata": {
             "description": (

keras_hub/src/models/vit/vit_layers.py

@@ -351,7 +351,7 @@ class ViTEncoder(keras.layers.Layer):
                 attention_dropout=self.attention_dropout,
                 layer_norm_epsilon=self.layer_norm_epsilon,
                 dtype=self.dtype_policy,
-                name=f"tranformer_block_{i+1}",
+                name=f"tranformer_block_{i + 1}",
             )
             encoder_block.build((None, None, self.hidden_dim))
             self.encoder_layers.append(encoder_block)

keras_hub/src/tokenizers/byte_tokenizer.py

@@ -150,8 +150,7 @@ class ByteTokenizer(tokenizer.Tokenizer):
     ):
         if not is_int_dtype(dtype):
             raise ValueError(
-                "Output dtype must be an integer type. "
-                f"Received: dtype={dtype}"
+                f"Output dtype must be an integer type. Received: dtype={dtype}"
             )
 
         # Check normalization_form.

keras_hub/src/tokenizers/sentence_piece_tokenizer_trainer.py

@@ -1,5 +1,7 @@
 import io
 
+from keras_hub.src.utils.tensor_utils import assert_tf_libs_installed
+
 try:
     import sentencepiece as spm
     import tensorflow as tf
@@ -77,6 +79,7 @@ def compute_sentence_piece_proto(
     tf.Tensor([ 4  8 12  5  9 14  5  6 13  4  7 10 11  6 13],
     shape=(15,), dtype=int32)
     """
+    assert_tf_libs_installed("compute_sentence_piece_proto")
 
     if spm is None:
         raise ImportError(

keras_hub/src/tokenizers/unicode_codepoint_tokenizer.py

@@ -203,8 +203,7 @@ class UnicodeCodepointTokenizer(tokenizer.Tokenizer):
     ) -> None:
         if not is_int_dtype(dtype):
             raise ValueError(
-                "Output dtype must be an integer type. "
-                f"Received: dtype={dtype}"
+                f"Output dtype must be an integer type. Received: dtype={dtype}"
             )
 
         # Check normalization_form.

keras_hub/src/tokenizers/word_piece_tokenizer_trainer.py

@@ -1,5 +1,6 @@
 from keras_hub.src.api_export import keras_hub_export
 from keras_hub.src.tokenizers.word_piece_tokenizer import pretokenize
+from keras_hub.src.utils.tensor_utils import assert_tf_libs_installed
 
 try:
     import tensorflow as tf
@@ -117,6 +118,8 @@ def compute_word_piece_vocabulary(
     inputs.map(tokenizer.tokenize)
     ```
     """  # noqa: E501
+    assert_tf_libs_installed("compute_word_piece_vocabulary")
+
     # Read data files.
     if not isinstance(data, (list, tf.data.Dataset)):
         raise ValueError(

keras_hub/src/utils/timm/convert_densenet.py

@@ -59,9 +59,11 @@ def convert_weights(backbone, loader, timm_config):
     num_stacks = len(backbone.stackwise_num_repeats)
     for stack_index in range(num_stacks):
         for block_idx in range(backbone.stackwise_num_repeats[stack_index]):
-            keras_name = f"stack{stack_index+1}_block{block_idx+1}"
+            keras_name = f"stack{stack_index + 1}_block{block_idx + 1}"
             hf_name = (
-                f"features.denseblock{stack_index+1}.denselayer{block_idx+1}"
+                "features."
+                f"denseblock{stack_index + 1}"
+                f".denselayer{block_idx + 1}"
             )
             port_batch_normalization(f"{keras_name}_1_bn", f"{hf_name}.norm1")
             port_conv2d(f"{keras_name}_1_conv", f"{hf_name}.conv1")
@@ -69,8 +71,8 @@ def convert_weights(backbone, loader, timm_config):
             port_conv2d(f"{keras_name}_2_conv", f"{hf_name}.conv2")
 
     for stack_index in range(num_stacks - 1):
-        keras_transition_name = f"transition{stack_index+1}"
-        hf_transition_name = f"features.transition{stack_index+1}"
+        keras_transition_name = f"transition{stack_index + 1}"
+        hf_transition_name = f"features.transition{stack_index + 1}"
         port_batch_normalization(
             f"{keras_transition_name}_bn", f"{hf_transition_name}.norm"
         )

keras_hub/src/utils/timm/convert_efficientnet.py

@@ -268,7 +268,7 @@ def convert_weights(backbone, loader, timm_config):
             # 97 is the start of the lowercase alphabet.
             letter_identifier = chr(block_idx + 97)
 
-            keras_block_prefix = f"block{stack_index+1}{letter_identifier}_"
+            keras_block_prefix = f"block{stack_index + 1}{letter_identifier}_"
             hf_block_prefix = f"blocks.{stack_index}.{block_idx}."
 
             if block_type == "v1":

keras_hub/src/utils/timm/convert_resnet.py

@@ -89,7 +89,7 @@ def convert_weights(backbone, loader, timm_config):
         for block_idx in range(backbone.stackwise_num_blocks[stack_index]):
             if version == "v1":
                 keras_name = f"stack{stack_index}_block{block_idx}"
-                hf_name = f"layer{stack_index+1}.{block_idx}"
+                hf_name = f"layer{stack_index + 1}.{block_idx}"
             else:
                 keras_name = f"stack{stack_index}_block{block_idx}"
                 hf_name = f"stages.{stack_index}.blocks.{block_idx}"

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.dev202501080345"
+__version__ = "0.19.0.dev202501150344"
 
 
 @keras_hub_export("keras_hub.version")

{keras_hub_nightly-0.19.0.dev202501080345.dist-info → keras_hub_nightly-0.19.0.dev202501150344.dist-info}/METADATA

@@ -1,6 +1,6 @@
-Metadata-Version: 2.1
+Metadata-Version: 2.2
 Name: keras-hub-nightly
-Version: 0.19.0.dev202501080345
+Version: 0.19.0.dev202501150344
 Summary: Industry-strength Natural Language Processing extensions for Keras.
 Home-page: https://github.com/keras-team/keras-hub
 Author: Keras team
@@ -31,6 +31,17 @@ Requires-Dist: tensorflow-text
 Provides-Extra: extras
 Requires-Dist: rouge-score; extra == "extras"
 Requires-Dist: sentencepiece; extra == "extras"
+Dynamic: author
+Dynamic: author-email
+Dynamic: classifier
+Dynamic: description
+Dynamic: description-content-type
+Dynamic: home-page
+Dynamic: license
+Dynamic: provides-extra
+Dynamic: requires-dist
+Dynamic: requires-python
+Dynamic: summary
 
 # KerasHub: Multi-framework Pretrained Models
 [![](https://github.com/keras-team/keras-hub/workflows/Tests/badge.svg?branch=master)](https://github.com/keras-team/keras-hub/actions?query=workflow%3ATests+branch%3Amaster)

{keras_hub_nightly-0.19.0.dev202501080345.dist-info → keras_hub_nightly-0.19.0.dev202501150344.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=PULLssHhM5UCqNcvlGVVyAQo9rDMkhLThc6DLU9Kz2g,222
+keras_hub/src/version_utils.py,sha256=XXUJ1oMuODMzez6Sqr-8PGIem4zG0YD-78PlkMxNEXI,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
@@ -43,7 +43,7 @@ keras_hub/src/layers/preprocessing/random_deletion.py,sha256=x23nRo0ir2J4Ps42i9X
 keras_hub/src/layers/preprocessing/random_swap.py,sha256=w2z7yNQsII5g4sEFi4GXfgxIc1S6UUt3a8YWZew_f4U,9504
 keras_hub/src/layers/preprocessing/start_end_packer.py,sha256=lY2K937z6JucxNe7VknynhhjrcUfFigU6mqIdv2gS-Y,7973
 keras_hub/src/metrics/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-keras_hub/src/metrics/bleu.py,sha256=N4hnlCIFk558nZAHxGlzLYBx6gwpWS3Wvw1iFM69xiA,13665
+keras_hub/src/metrics/bleu.py,sha256=pnid5azpAxO6vKEfUtAby3nH29OGbwYKgVGOGeoaA3I,13694
 keras_hub/src/metrics/edit_distance.py,sha256=kjhe8uNjvv8aN49RyrKAbNi7a8_OlB8fMza0J_CfNQg,6353
 keras_hub/src/metrics/perplexity.py,sha256=dDUQcfE5JbAruG3spEkgue6IjHcynqgmGpJAqWg22Tw,6139
 keras_hub/src/metrics/rouge_base.py,sha256=Pt2DUznhTTeR-fX1nQ_wSbPtmuTgxQTvrGpu8LRVapE,6264
@@ -87,7 +87,7 @@ keras_hub/src/models/bart/bart_seq_2_seq_lm_preprocessor.py,sha256=3_e-ULIcm_3DK
 keras_hub/src/models/bart/bart_tokenizer.py,sha256=Q7IXmIwXzhPSN427oQRyF9ufoExQGS184Yo_4boaOZo,2811
 keras_hub/src/models/basnet/__init__.py,sha256=4N6XvIUYYJl5xtoaL3_9fawUX_qP3WmTYNEEU7tn8Gw,253
 keras_hub/src/models/basnet/basnet.py,sha256=JA58Q9lmygdSOm5MUaPAlaL6B8XnmqCcRaGrk9c8P3Q,4287
-keras_hub/src/models/basnet/basnet_backbone.py,sha256=t_52WW6jetONS7AnPf9YsiMLDqOjVwjNuayQEv6ZAk4,13503
+keras_hub/src/models/basnet/basnet_backbone.py,sha256=P-jogkYIu9j7_28fl2RFQRMl87BXz1wcY_LtIrxBy1E,13505
 keras_hub/src/models/basnet/basnet_image_converter.py,sha256=DwzAwtZeggYw_qyRQ-Abnnm885Wobv3wClxRzOTscI0,342
 keras_hub/src/models/basnet/basnet_preprocessor.py,sha256=uM504utaXODSqR5zpKnopRuaV_l84zCg06RkNoNSKIs,510
 keras_hub/src/models/basnet/basnet_presets.py,sha256=z6tR2q_EvYnUmGfsWIWYfmR_8gvWYPH3QmtpAu_T8f8,63
@@ -135,11 +135,11 @@ keras_hub/src/models/deeplab_v3/__init__.py,sha256=FHAUPM4a1DJj4EsNTbYEd1riNq__u
 keras_hub/src/models/deeplab_v3/deeplab_v3_backbone.py,sha256=dH7HHu_NAnE-HP6ivOL7fFLQZHt_MWmehlMccLljhPc,7764
 keras_hub/src/models/deeplab_v3/deeplab_v3_image_converter.py,sha256=mRkH3HdhpV0fCcQcVXEvIX7SNk-bAMb3SAHzgK-FD5c,371
 keras_hub/src/models/deeplab_v3/deeplab_v3_image_segmeter_preprocessor.py,sha256=hR9S6lNYamY0EBDBo3e1qTCiwtftmLXrN-UYuzfw5Io,581
-keras_hub/src/models/deeplab_v3/deeplab_v3_layers.py,sha256=qmEiolOOriLAojXB67xXW9IOo717kaCGeDVZJLaGY98,7834
+keras_hub/src/models/deeplab_v3/deeplab_v3_layers.py,sha256=mz9nG55gdXSTDE96AXgeTCwUFB95DIpTuqrvWIt5Lco,7840
 keras_hub/src/models/deeplab_v3/deeplab_v3_presets.py,sha256=ZKYY8A7mV2QvwXwjDUd9xAbVHo58-Hgj_IqNUbuyCIU,625
 keras_hub/src/models/deeplab_v3/deeplab_v3_segmenter.py,sha256=pubi30sPJKLOpz9fRQff2FZt_53KBvwf2uyaJ5YL7J8,3726
 keras_hub/src/models/densenet/__init__.py,sha256=r7StyamnWeeZxOk9r4ZYNbS_YVhu9YGPyXhNxljvdPg,269
-keras_hub/src/models/densenet/densenet_backbone.py,sha256=5QawyB4EhyaXpmm8l_QUYveU7kEet3jRD3s94XAz8Tw,6738
+keras_hub/src/models/densenet/densenet_backbone.py,sha256=f2nfsXyXQert2aYHq-L-JZtp8inq1fs1K47rzZQ9nTI,6744
 keras_hub/src/models/densenet/densenet_image_classifier.py,sha256=ye-Ix3oU42pfsDoh-h1PG4di1kzldO0ZO7Nj304p_X4,544
 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
@@ -186,7 +186,7 @@ keras_hub/src/models/flux/flux_layers.py,sha256=wevcAEbayBD8bVm-21FBi2LQ13pZzB99
 keras_hub/src/models/flux/flux_maths.py,sha256=2pnHW8HW7V2JZ8HIrUwE-UU4klpFQaOkoAvG5nWVfyY,7502
 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=mI_QxOzjXl3b5s7Q1LZemceCdeboqPD5ilEPEEyer40,4169
+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/gemma/__init__.py,sha256=rVzOJMJ39bgVlT8UdC0t8PlN2c237GKTBmfHIsbPuOQ,251
 keras_hub/src/models/gemma/gemma_attention.py,sha256=1CVN5z9GKoU8TuNMih2_MweDkpd98xSqdic9F8xIBE8,8317
@@ -257,7 +257,7 @@ keras_hub/src/models/pali_gemma/pali_gemma_causal_lm.py,sha256=AViEs6YltUqWnIVo7
 keras_hub/src/models/pali_gemma/pali_gemma_causal_lm_preprocessor.py,sha256=F57y0fZ0wYYxfGIjfrJc1W9uQpViYFx5bvFjj5CqUbI,4814
 keras_hub/src/models/pali_gemma/pali_gemma_decoder_block.py,sha256=24ABQ1vGlppV-KfWh0YqJjzM_Lu2GIwvyJ4X2XXie_A,5616
 keras_hub/src/models/pali_gemma/pali_gemma_image_converter.py,sha256=5yM_jUtrFsWIieiwfFBoP7mtPmQAwywkeLKbd7fhmzk,371
-keras_hub/src/models/pali_gemma/pali_gemma_presets.py,sha256=O648iwzs0wooiQCfDQ-n0wOtzIOEDGXRSwSb_Brx2Ck,8985
+keras_hub/src/models/pali_gemma/pali_gemma_presets.py,sha256=Ka1ChUUSKw-yY2th3QtmNtkeXt0krYfwhkHrScioMls,8979
 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
@@ -271,7 +271,7 @@ keras_hub/src/models/phi3/phi3_presets.py,sha256=sb2ce7Gq1OikFEf2KIYG69rFKHYKj8q
 keras_hub/src/models/phi3/phi3_rotary_embedding.py,sha256=wqiRn8nETNcLc5Vsm_d_8s11Ro6ibWZbWvODdLqIOo4,5013
 keras_hub/src/models/phi3/phi3_tokenizer.py,sha256=bOPH14wTVVHJHq8mgzXLjsgvKMNhfO8eayevAPpjYVA,1992
 keras_hub/src/models/resnet/__init__.py,sha256=C5UqlQ6apm8WSp1bnrxB6Bi3BGaknxRQs-r3b2wpaGA,257
-keras_hub/src/models/resnet/resnet_backbone.py,sha256=3acTjdWbnos8l_TPxYLgoV3Y4V_vJ_o1AqGhiQu459k,31274
+keras_hub/src/models/resnet/resnet_backbone.py,sha256=Q7nlqcTXZzjqd0e-DsjHC4ok58yOX7qxseotym3uZpM,31276
 keras_hub/src/models/resnet/resnet_image_classifier.py,sha256=nf35EKDzvBkfhHsK-s6Ks0nbhvKO7HEOYZm94YckyWE,510
 keras_hub/src/models/resnet/resnet_image_classifier_preprocessor.py,sha256=fM7gyQ0qB-RRuI4USJkRD6q9-HVfuC71e-BLTo-UhHQ,543
 keras_hub/src/models/resnet/resnet_image_converter.py,sha256=fgTxihJznGFss-y3Z-jp0JE3X1gaaB2y-f2KMwrT8Pk,342
@@ -279,7 +279,7 @@ keras_hub/src/models/resnet/resnet_presets.py,sha256=cryfXlC_FSEN_jrexKIh5aVbzp8
 keras_hub/src/models/retinanet/__init__.py,sha256=veWIFvMN6151M69l7FvTcI-IIEe_8dLmNO5NLOszQ1c,275
 keras_hub/src/models/retinanet/anchor_generator.py,sha256=0OgKSW3OKmbc0cOPHF6FYTAzn7fcHklg665PGSwAaDM,6504
 keras_hub/src/models/retinanet/box_matcher.py,sha256=l820r1R-ByqiyVgmZ0YFjjz0njchDda-wItzLn1X84o,10834
-keras_hub/src/models/retinanet/feature_pyramid.py,sha256=VxLcOEjJSXIDu30oMcZEYdVlpHaOP3IutZNwh0N3uHQ,17604
+keras_hub/src/models/retinanet/feature_pyramid.py,sha256=hbdrj6X-D2SlwOp2h1WcBlTdSAlLmFK43X7OrkJRoMA,17614
 keras_hub/src/models/retinanet/non_max_supression.py,sha256=PMOLlRw-EnyEmhlUhJjEbHf1xXiplN95pUxQbiJQbN4,20996
 keras_hub/src/models/retinanet/prediction_head.py,sha256=xWHt21-SS2t7vCmTONlR1lSbJXhml5jx68V8MGbGybg,7863
 keras_hub/src/models/retinanet/retinanet_backbone.py,sha256=BJBPJLxpOCOU0Br7b4JsgCZBHQHLAhxLqo9BHNIsl1g,5659
@@ -314,11 +314,11 @@ 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=poJlz-xt06hgOtn_Bw5YQDxZtDBc9L4Vo0ahhGwPly4,33340
-keras_hub/src/models/stable_diffusion_3/stable_diffusion_3_backbone.py,sha256=u0Wwtbl5b-1z_vn07TRw4jpkVYrReZeHbWqQIrZjyCA,23368
+keras_hub/src/models/stable_diffusion_3/mmdit.py,sha256=0gq2tcIqcbiGKKDDj3vrRsF67U3qE9g706XPs2BfCOY,40979
+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=aZMIC-GYjLhdU_yM7fJEznApCo1zwRAgwQbW0tCW0xY,6399
-keras_hub/src/models/stable_diffusion_3/stable_diffusion_3_presets.py,sha256=z6wrfv8rCqLBzn7_edRcKCIDQRTNUgLqyr-LLp55-IE,1680
+keras_hub/src/models/stable_diffusion_3/stable_diffusion_3_inpaint.py,sha256=t4uw920Jn1k80air3WRGimKf71aMVu6q73oWFH348vk,6384
+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=m5PdVSgTcYuqd7jOQ8wD4PAnMa7wY2WdhwpK3hdydhM,2756
 keras_hub/src/models/stable_diffusion_3/t5_encoder.py,sha256=oV7P1uwCKdGiD93zXq7kmqX0elMZQU4UvBa8wg6P1hs,5113
@@ -344,7 +344,7 @@ keras_hub/src/models/vit/vit_backbone.py,sha256=kGmRZO4u-1q4PBcbhJbiWVIEVYAcp2H4
 keras_hub/src/models/vit/vit_image_classifier.py,sha256=lMVxiD1_6drx7XQ7P7YzlqnFP7kT1zlMe84f-T3SDQI,6332
 keras_hub/src/models/vit/vit_image_classifier_preprocessor.py,sha256=wu6YcBlXMWB9sKCPvmNdGBZKTLQt_HyHWS6P9nyDwsk,504
 keras_hub/src/models/vit/vit_image_converter.py,sha256=5xVF04BzMcdTDc6aErAYj3_BuGmVd3zoJMcH1ho4T0g,2561
-keras_hub/src/models/vit/vit_layers.py,sha256=s4j3n3qnJnv6W9AdUkNsO3Vsi_BhxEGECYkaLVCU6XY,13238
+keras_hub/src/models/vit/vit_layers.py,sha256=Zsz-ARPY49S1nXLUtpFwtPfw31D-vCtKesEo_2JIKPA,13240
 keras_hub/src/models/vit/vit_presets.py,sha256=zZhxUleOom1ie3gn0Mi-_xhhdFEEsnqSQyKADV2L38k,4479
 keras_hub/src/models/vit_det/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 keras_hub/src/models/vit_det/vit_det_backbone.py,sha256=DOZ5J7c1t5PAZ6y0pMmBoQTMOUup7UoUrYVfCs69ltY,7697
@@ -383,13 +383,13 @@ keras_hub/src/tests/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSu
 keras_hub/src/tests/test_case.py,sha256=oGWoUhlKgjVMNIjvUVnQR-k5iKvodztHsFMOs669Trw,27402
 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
-keras_hub/src/tokenizers/byte_tokenizer.py,sha256=c1a41eVuLzGmBtscQ0RxPIqFi41m_604KJ9fdpPR7Sc,10437
+keras_hub/src/tokenizers/byte_tokenizer.py,sha256=GPIKaddXugbfckfhodADsBpaYb72DgFMs_xfXHnK4qU,10418
 keras_hub/src/tokenizers/sentence_piece_tokenizer.py,sha256=nOqkpa2nHitITpdowPHdwxiN87e8huLW8Dt2gozVnhI,9350
-keras_hub/src/tokenizers/sentence_piece_tokenizer_trainer.py,sha256=LhUxwcaDKt5V58DBzK9Sh4D-hOL80SHGpL4LavWbq74,4642
+keras_hub/src/tokenizers/sentence_piece_tokenizer_trainer.py,sha256=caqgV9N4lH97zBviFPdpwo_O95AaJBEJLQv6Icq3Hs8,4774
 keras_hub/src/tokenizers/tokenizer.py,sha256=v0Ka5ayrBwpsGBlkIadXK-b4RsMTbhV6BZrvKullbxY,9722
-keras_hub/src/tokenizers/unicode_codepoint_tokenizer.py,sha256=KxuVsUx3ntGsuqaQ-gnFWFfoVLsl5Hag7rBk6xfq-fQ,13572
+keras_hub/src/tokenizers/unicode_codepoint_tokenizer.py,sha256=hRv_XxoPIPDpHfO0ZttSOv_M89sMaFpvmllojvKz_ac,13553
 keras_hub/src/tokenizers/word_piece_tokenizer.py,sha256=vP6AZgbzsRiuPCt3W_n94nsF7XiERnagWcH_rqJHtVU,19943
-keras_hub/src/tokenizers/word_piece_tokenizer_trainer.py,sha256=Zz1SGgArykxBVWnS5YV-ViqyMOrw3j3i_i_jto96zCg,6610
+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/pipeline_model.py,sha256=jgzB6NQPSl0KOu08N-TazfOnXnUJbZjH2EXXhx25Ftg,9084
@@ -399,9 +399,9 @@ keras_hub/src/utils/tensor_utils.py,sha256=YVJesN91bk-OzJXY1mOKBppuY8noBU7zhPQNX
 keras_hub/src/utils/imagenet/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 keras_hub/src/utils/imagenet/imagenet_utils.py,sha256=MvIvv1WJo51ZXBxy4S7t_DsN3ZMtJWlC4cmRvKM2kIA,39304
 keras_hub/src/utils/timm/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-keras_hub/src/utils/timm/convert_densenet.py,sha256=V-GRjWuDnlh3b1EMxqahwZ3GMwSgOa3v0HOfb2ZZ-d0,3342
-keras_hub/src/utils/timm/convert_efficientnet.py,sha256=wkOKTLS_N_VKy1CQQGjSlD_TPSOOmCMMXQvbjravN6g,17098
-keras_hub/src/utils/timm/convert_resnet.py,sha256=ee8eTml0ffJKE8avzGoLFcpjPF63DsvoIUArAGa8Ngg,5832
+keras_hub/src/utils/timm/convert_densenet.py,sha256=fu8HBIQis5o3ib2tyI2qnmYScVrVIQySok8vTfa1qJ8,3393
+keras_hub/src/utils/timm/convert_efficientnet.py,sha256=SgEIlyyinS04qoQpEgh3WazHq544zNUCCpfmWh3EjSs,17100
+keras_hub/src/utils/timm/convert_resnet.py,sha256=8JFkVtdpy5z9h83LJ97rD-a8FRejXPZvMNksNuStqjM,5834
 keras_hub/src/utils/timm/convert_vgg.py,sha256=MT5jGnLrzenPpe66Af_Lp1IdR9KGtsSrcmn6_UPqHvQ,2419
 keras_hub/src/utils/timm/preset_loader.py,sha256=cdZDjthZdTD2myMOenQar4ACyi7VTuIzNRg24LuqS-4,3374
 keras_hub/src/utils/transformers/__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.dev202501080345.dist-info/METADATA,sha256=WWsYYpkd-P_ryoA3jId3bNDKaMQOJfy7eBeYQ7N_D6w,7260
-keras_hub_nightly-0.19.0.dev202501080345.dist-info/WHEEL,sha256=A3WOREP4zgxI0fKrHUG8DC8013e3dK3n7a6HDbcEIwE,91
-keras_hub_nightly-0.19.0.dev202501080345.dist-info/top_level.txt,sha256=N4J6piIWBKa38A4uV-CnIopnOEf8mHAbkNXafXm_CuA,10
-keras_hub_nightly-0.19.0.dev202501080345.dist-info/RECORD,,
+keras_hub_nightly-0.19.0.dev202501150344.dist-info/METADATA,sha256=FhbHeGMBpOfmdE1bEoJdl34xtvV3n85LqFH_5STUyUo,7498
+keras_hub_nightly-0.19.0.dev202501150344.dist-info/WHEEL,sha256=In9FTNxeP60KnTkGw7wk6mJPYd_dQSjEZmXdBdMCI-8,91
+keras_hub_nightly-0.19.0.dev202501150344.dist-info/top_level.txt,sha256=N4J6piIWBKa38A4uV-CnIopnOEf8mHAbkNXafXm_CuA,10
+keras_hub_nightly-0.19.0.dev202501150344.dist-info/RECORD,,

{keras_hub_nightly-0.19.0.dev202501080345.dist-info → keras_hub_nightly-0.19.0.dev202501150344.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (75.7.0)
+Generator: setuptools (75.8.0)
 Root-Is-Purelib: true
 Tag: py3-none-any