keras-hub-nightly 0.22.0.dev202507100418__py3-none-any.whl → 0.22.0.dev202507120419__py3-none-any.whl

keras_hub/src/models/hgnetv2/hgnetv2_presets.py

@@ -0,0 +1,58 @@
+# Metadata for loading pretrained model weights.
+backbone_presets = {
+    "hgnetv2_b4_ssld_stage2_ft_in1k": {
+        "metadata": {
+            "description": (
+                "HGNetV2 B4 model with 2-stage SSLD training, fine-tuned on "
+                "ImageNet-1K."
+            ),
+            "params": 13599072,
+            "path": "hgnetv2",
+        },
+        "kaggle_handle": "",
+    },
+    "hgnetv2_b5_ssld_stage1_in22k_in1k": {
+        "metadata": {
+            "description": (
+                "HGNetV2 B5 model with 1-stage SSLD training, pre-trained on "
+                "ImageNet-22K and fine-tuned on ImageNet-1K."
+            ),
+            "params": 33419680,
+            "path": "hgnetv2",
+        },
+        "kaggle_handle": "",
+    },
+    "hgnetv2_b5_ssld_stage2_ft_in1k": {
+        "metadata": {
+            "description": (
+                "HGNetV2 B5 model with 2-stage SSLD training, fine-tuned on "
+                "ImageNet-1K."
+            ),
+            "params": 33419680,
+            "path": "hgnetv2",
+        },
+        "kaggle_handle": "",
+    },
+    "hgnetv2_b6_ssld_stage1_in22k_in1k": {
+        "metadata": {
+            "description": (
+                "HGNetV2 B6 model with 1-stage SSLD training, pre-trained on "
+                "ImageNet-22K and fine-tuned on ImageNet-1K."
+            ),
+            "params": 69179888,
+            "path": "hgnetv2",
+        },
+        "kaggle_handle": "",
+    },
+    "hgnetv2_b6_ssld_stage2_ft_in1k": {
+        "metadata": {
+            "description": (
+                "HGNetV2 B6 model with 2-stage SSLD training, fine-tuned on "
+                "ImageNet-1K."
+            ),
+            "params": 69179888,
+            "path": "hgnetv2",
+        },
+        "kaggle_handle": "",
+    },
+}

keras_hub/src/models/qwen3/__init__.py

@@ -0,0 +1,5 @@
+from keras_hub.src.models.qwen3.qwen3_backbone import Qwen3Backbone
+from keras_hub.src.models.qwen3.qwen3_presets import backbone_presets
+from keras_hub.src.utils.preset_utils import register_presets
+
+register_presets(backbone_presets, Qwen3Backbone)

keras_hub/src/models/qwen3/qwen3_attention.py

@@ -303,7 +303,7 @@ class Qwen3Attention(keras.layers.Layer):
             attention_mask = self._mask_sliding_window(
                 attention_mask,
                 cache_update_index=cache_update_index
-                if cache_update_index
+                if cache_update_index is not None
                 else 0,
             )
         attention_scores = self._masked_softmax(

keras_hub/src/models/qwen3/qwen3_causal_lm.py

@@ -0,0 +1,390 @@
+import keras
+from keras import ops
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.causal_lm import CausalLM
+from keras_hub.src.models.qwen3.qwen3_backbone import Qwen3Backbone
+from keras_hub.src.models.qwen3.qwen3_causal_lm_preprocessor import (
+    Qwen3CausalLMPreprocessor,
+)
+from keras_hub.src.utils.tensor_utils import any_equal
+
+
+@keras_hub_export("keras_hub.models.Qwen3CausalLM")
+class Qwen3CausalLM(CausalLM):
+    """An end-to-end Qwen3 model for causal language modeling.
+
+    A causal language model (LM) predicts the next token based on previous
+    tokens. This task setup can be used to train the model unsupervised on plain
+    text input, or to autoregressively generate plain text similar to the data
+    used for training. This task can be used for pre-training or fine-tuning a
+    Qwen3 model, simply by calling `fit()`.
+
+    This model has a `generate()` method, which generates text based on a
+    prompt. The generation strategy used is controlled by an additional
+    `sampler` argument on `compile()`. You can recompile the model with
+    different `keras_hub.samplers` objects to control the generation.
+    By default, `"greedy"` sampling will be used.
+
+    This model can optionally be configured with a `preprocessor` layer, in
+    which case it will automatically apply preprocessing to string inputs during
+    `fit()`, `predict()`, `evaluate()`, and `generate()`. This is done by
+    default when creating the model with `from_preset()`.
+
+    Args:
+        backbone: A `keras_hub.models.Qwen3Backbone` instance.
+        preprocessor: A `keras_hub.models.Qwen3CausalLMPreprocessor` or
+            `None`. If `None`, this model will not apply preprocessing, and
+            inputs should be preprocessed before calling the model.
+
+    Examples:
+
+    Use `generate()` to do text generation.
+    ```python
+    qwen3_lm = keras_hub.models.Qwen3CausalLM.from_preset("qwen3_0.6b_en")
+    qwen3_lm.generate("I want to say", max_length=30)
+
+    # Generate with batched prompts.
+    qwen3_lm.generate(["This is a", "Where are you"], max_length=30)
+    ```
+
+    Compile the `generate()` function with a custom sampler.
+    ```python
+    qwen3_lm = keras_hub.models.Qwen3MoeCausalLM.from_preset("qwen3_0.6b_en")
+    qwen3_lm.compile(sampler="top_k")
+    qwen3_lm.generate("I want to say", max_length=30)
+
+    qwen3_lm.compile(sampler=keras_hub.samplers.BeamSampler(num_beams=2))
+    qwen3_lm.generate("I want to say", max_length=30)
+    ```
+
+    Use `generate()` without preprocessing.
+    ```python
+    prompt = {
+        # Token ids for "<bos> Qwen3 is".
+        "token_ids": np.array([[2, 12345, 678, 0, 0, 0, 0]] * 2),
+        # Use `"padding_mask"` to indicate values that should not be overridden.
+        "padding_mask": np.array([[1, 1, 1, 0, 0, 0, 0]] * 2),
+    }
+
+    qwen3_lm = keras_hub.models.Qwen3MoeCausalLM.from_preset(
+        "qwen3_0.6b_en",
+        preprocessor=None,
+    )
+    qwen3_lm.generate(prompt)
+    ```
+
+    Call `fit()` on a single batch.
+    ```python
+    features = ["The quick brown fox jumped.", "I forgot my homework."]
+    qwen3_lm = keras_hub.models.Qwen3MoeCausalLM.from_preset("qwen3_0.6b_en")
+    qwen3_lm.fit(x=features, batch_size=2)
+    ```
+
+    Call `fit()` with LoRA fine-tuning enabled.
+    ```python
+    features = ["The quick brown fox jumped.", "I forgot my homework."]
+    qwen3_lm = keras_hub.models.Qwen3MoeCausalLM.from_preset(
+        'qwen3_0.6b_en'
+    )
+    qwen3_lm.backbone.enable_lora(rank=4)
+    qwen3_lm.fit(x=features, batch_size=2)
+    ```
+
+    Call `fit()` without preprocessing.
+    ```python
+    x = {
+        # Token ids for "<bos> Qwen3 is a language model<eos>"
+        "token_ids": np.array([[2, 12345, 678, 543, 9876, 1, 0, 0]] * 2),
+        "padding_mask": np.array([[1, 1, 1, 1, 1, 1, 0, 0]] * 2),
+    }
+    y = np.array([[12345, 678, 543, 9876, 1, 0, 0, 0]] * 2)
+    sw = np.array([[1, 1, 1, 1, 1, 0, 0, 0]] * 2)
+
+    qwen3_lm = keras_hub.models.Qwen3MoeCausalLM.from_preset(
+        "qwen3_0.6b_en",
+        preprocessor=None,
+    )
+    qwen3_lm.fit(x=x, y=y, sample_weight=sw, batch_size=2)
+    ```
+
+    Custom backbone and vocabulary.
+    ```python
+    tokenizer = keras_hub.models.Qwen3MoeTokenizer(
+        proto="qwen3_moe_vocab.spm",
+    )
+    preprocessor = keras_hub.models.Qwen3MoeCausalLMPreprocessor(
+        tokenizer=tokenizer,
+        sequence_length=128,
+    )
+    backbone = keras_hub.models.Qwen3MoeBackbone(
+        vocabulary_size=151936,
+        num_layers=28,
+        num_query_heads=16,
+        num_key_value_heads=8,
+        hidden_dim=2048,
+        intermediate_dim=4096,
+        moe_intermediate_dim=128,
+        shared_expert_intermediate_dim=4096,
+        num_experts=60,
+        top_k=4,
+        max_sequence_length=4096,
+    )
+    qwen3_lm = keras_hub.models.Qwen3MoeCausalLM(
+        backbone=backbone,
+        preprocessor=preprocessor,
+    )
+    qwen3_lm.fit(x=features, batch_size=2)
+    ```
+    """
+
+    backbone_cls = Qwen3Backbone
+    preprocessor_cls = Qwen3CausalLMPreprocessor
+
+    def __init__(self, backbone, preprocessor=None, **kwargs):
+        # === Layers ===
+        self.backbone = backbone
+        self.preprocessor = preprocessor
+
+        # === Functional Model ===
+        # This must be "backbone.input" i.e. the full input structure,
+        # rather than "backbone.inputs" which is the flattened list of inputs.
+        inputs = backbone.input
+        hidden_states = backbone(inputs)
+        outputs = backbone.token_embedding(hidden_states, reverse=True)
+        super().__init__(
+            inputs=inputs,
+            outputs=outputs,
+            **kwargs,
+        )
+
+    def call_with_cache(
+        self,
+        token_ids,
+        cache,
+        cache_update_index,
+    ):
+        """Forward pass of `Qwen3CausalLM` with cache.
+
+        `call_with_cache` adds an additional forward pass for the model for
+        autoregressive inference. Unlike calling the model directly, this method
+        allows caching previous key/value Tensors in multi-head attention layer,
+        and avoids recomputing the outputs of seen tokens.
+
+        Args:
+            token_ids: a dense int Tensor with shape `(batch_size, max_length)`.
+            cache: a dense float Tensor, the cache of key and value.
+            cache_update_index: int, or int Tensor. The index of current inputs
+            in the whole sequence.
+
+        Returns:
+            A (logits, hidden_states, cache) tuple. Where `logits` is the
+            language model logits for the input token_ids, `hidden_states` is
+            the final hidden representation of the input tokens, and `cache` is
+            the decoding cache.
+        """
+        x = self.backbone.token_embedding(token_ids)
+        # Each decoder layer has a cache; we update them separately.
+        updated_cache = []
+        for i in range(self.backbone.num_layers):
+            current_cache = cache[:, i, ...]
+            x, next_cache = self.backbone.transformer_layers[i](
+                x,
+                self_attention_cache=current_cache,
+                self_attention_cache_update_index=cache_update_index,
+            )
+            updated_cache.append(next_cache)
+        cache = ops.stack(updated_cache, axis=1)
+        hidden_states = x = self.backbone.layer_norm(x)
+        logits = self.backbone.token_embedding(x, reverse=True)
+        return logits, hidden_states, cache
+
+    def _build_cache(self, token_ids):
+        """Build an empty cache for use with `call_with_cache()`."""
+        batch_size = ops.shape(token_ids)[0]
+        max_length = ops.shape(token_ids)[1]
+        num_layers = self.backbone.num_layers
+        num_key_value_heads = self.backbone.num_key_value_heads
+        head_dim = self.backbone.head_dim
+        shape = [
+            batch_size,
+            num_layers,
+            2,
+            max_length,
+            num_key_value_heads,
+            head_dim,
+        ]
+        cache = ops.zeros(shape, dtype=self.compute_dtype)
+        # Seed the cache.
+        _, hidden_states, cache = self.call_with_cache(token_ids, cache, 0)
+        return hidden_states, cache
+
+    def generate_step(
+        self,
+        inputs,
+        stop_token_ids=None,
+    ):
+        """A compilable generation function for a single batch of inputs.
+
+        This function represents the inner, XLA-compilable, generation function
+        for a single batch of inputs. Inputs should have the same structure as
+        model inputs, a dictionary with keys `"token_ids"` and `"padding_mask"`.
+
+        Args:
+            inputs: A dictionary with two keys `"token_ids"` and
+                `"padding_mask"` and batched tensor values.
+            stop_token_ids: Tuple of id's of the end token to stop on. If all
+                sequences have produced a new stop token, generation
+                will stop.
+        """
+        token_ids, padding_mask = inputs["token_ids"], inputs["padding_mask"]
+        # Create and seed cache with a single forward pass.
+        hidden_states, cache = self._build_cache(token_ids)
+        # Compute the lengths of all user inputted tokens ids.
+        row_lengths = ops.sum(ops.cast(padding_mask, "int32"), axis=-1)
+        # Start at the first index that has no user inputted id.
+        index = ops.min(row_lengths)
+
+        def next(prompt, cache, index):
+            # The cache index is the index of our previous token.
+            cache_update_index = index - 1
+            batch_size = ops.shape(prompt)[0]
+            prompt = ops.slice(prompt, [0, cache_update_index], [batch_size, 1])
+            logits, hidden_states, cache = self.call_with_cache(
+                prompt,
+                cache,
+                cache_update_index,
+            )
+            return (
+                ops.squeeze(logits, axis=1),
+                ops.squeeze(hidden_states, axis=1),
+                cache,
+            )
+
+        token_ids = self.sampler(
+            next=next,
+            prompt=token_ids,
+            cache=cache,
+            index=index,
+            mask=padding_mask,
+            stop_token_ids=stop_token_ids,
+            hidden_states=hidden_states,
+            model=self,
+        )
+        print("generated token ids = ", token_ids[0])
+
+        # Compute an output padding mask with the token ids we updated.
+        if stop_token_ids is not None:
+            # Build a mask of stop token locations not in the original
+            # prompt (not in locations where `padding_mask` is True).
+            end_locations = any_equal(
+                token_ids, stop_token_ids, ops.logical_not(padding_mask)
+            )
+            end_locations = ops.cast(end_locations, "int32")
+            # Use cumsum to get ones in all locations after end_locations.
+            cumsum = ops.cast(ops.cumsum(end_locations, axis=-1), "int32")
+            overflow = cumsum - end_locations
+            # Our padding mask is the inverse of these overflow locations.
+            padding_mask = ops.logical_not(ops.cast(overflow, "bool"))
+        else:
+            # Without early stopping, all locations will have been updated.
+            padding_mask = ops.ones_like(token_ids, dtype="bool")
+        return {
+            "token_ids": token_ids,
+            "padding_mask": padding_mask,
+        }
+
+    def score(
+        self,
+        token_ids,
+        padding_mask=None,
+        scoring_mode="logits",
+        layer_intercept_fn=None,
+        target_ids=None,
+    ):
+        """Score a generation represented by the provided token ids.
+
+        Args:
+            token_ids: A <int>[batch_size, num_tokens] tensor containing tokens
+                to score. Typically, this tensor captures the output from a call
+                to `Qwen3CausalLM.generate()`, i.e., tokens for both the input
+                text and the model-generated text.
+            padding_mask: A <bool>[batch_size, num_tokens] tensor indicating the
+                tokens that should be preserved during generation. This is an
+                artifact required by the `Qwen3Backbone` and isn't influential
+                on the computation of this function. If omitted, this function
+                uses `keras.ops.ones()` to create a tensor of the appropriate
+                shape.
+            scoring_mode: The type of scores to return, either "logits" or
+                "loss", both will be per input token.
+            layer_intercept_fn: An optional function for augmenting activations
+                with additional computation, for example, as part of
+                interpretability research. This function will be passed the
+                activations as its first parameter and a numeric index
+                associated with that backbone layer. _This index _is not_ an
+                index into `self.backbone.layers`_. The index -1 accompanies the
+                embeddings returned by calling `self.backbone.token_embedding()`
+                on `token_ids` in the forward direction. All subsequent indexes
+                will be 0-based indices for the activations returned by each of
+                the Transformers layers in the backbone. This function must
+                return a <float>[batch_size, num_tokens, hidden_dims] tensor
+                that can be passed as an input to the next layer in the model.
+            target_ids: An <bool>[batch_size, num_tokens] tensor containing the
+                predicted tokens against which the loss should be computed. If a
+                span of tokens is provided (sequential truthy values along
+                axis=1 in the tensor), the loss will be computed as the
+                aggregate across those tokens.
+
+        Raises:
+            ValueError: If an unsupported scoring_mode is provided, or if the
+                target_ids are not provided when using ScoringMode.LOSS.
+
+        Returns:
+            The per-token scores as a tensor of size
+            <float>[batch_size, num_tokens, vocab_size] in "logits" mode, or
+            <float>[batch_size, num_tokens] in "loss" mode.
+
+        ```
+        """
+        if scoring_mode not in ("logits", "loss"):
+            raise ValueError(
+                "Unsupported scoring_mode. Must be one of 'logits' or 'loss'."
+            )
+
+        if scoring_mode == "loss" and target_ids is None:
+            raise ValueError(
+                "Cannot compute loss without targets. Please provide target "
+                "token ids via the target_ids parameter."
+            )
+
+        batch_shape = ops.shape(token_ids)[:2]
+        assert len(batch_shape) == 2
+
+        if padding_mask is None:
+            padding_mask = ops.ones(shape=batch_shape)
+
+        if layer_intercept_fn is None:
+
+            def default_layer_intercept_fn(x, unused_i):
+                return x
+
+            layer_intercept_fn = default_layer_intercept_fn
+
+        token_embeddings = self.backbone.token_embedding(token_ids)
+        x = layer_intercept_fn(token_embeddings, -1)
+
+        for i, transformer_layer in enumerate(self.backbone.transformer_layers):
+            x = transformer_layer(x, decoder_padding_mask=padding_mask)
+            x = layer_intercept_fn(x, i)
+
+        x = self.backbone.layer_norm(x)
+        logits = self.backbone.token_embedding(x, reverse=True)
+
+        if scoring_mode == "logits":
+            return logits
+
+        per_token_loss_fn = keras.losses.SparseCategoricalCrossentropy(
+            from_logits=True, reduction="none"
+        )
+        per_token_loss = per_token_loss_fn(target_ids, logits)
+        return per_token_loss

keras_hub/src/models/qwen3/qwen3_presets.py

@@ -0,0 +1,73 @@
+"""Qwen3 model preset configurations."""
+
+backbone_presets = {
+    "qwen3_0.6b_en": {
+        "metadata": {
+            "description": (
+                "28-layer Qwen3 model with 596M parameters, optimized for "
+                "efficiency and fast inference on resource-constrained devices."
+            ),
+            "params": 596049920,
+            "path": "qwen-3",
+        },
+        "kaggle_handle": "kaggle://keras/qwen-3/keras/qwen3_0.6b_en/1",
+    },
+    "qwen3_1.7b_en": {
+        "metadata": {
+            "description": (
+                "28-layer Qwen3 model with 1.72B parameters, offering "
+                "a good balance between performance and resource usage."
+            ),
+            "params": 1720574976,
+            "path": "qwen-3",
+        },
+        "kaggle_handle": "kaggle://keras/qwen-3/keras/qwen3_1.7b_en/1",
+    },
+    "qwen3_4b_en": {
+        "metadata": {
+            "description": (
+                "36-layer Qwen3 model with 4.02B parameters, offering improved "
+                "reasoning capabilities and better performance than smaller "
+                "variants."
+            ),
+            "params": 4022468096,
+            "path": "qwen-3",
+        },
+        "kaggle_handle": "kaggle://keras/qwen-3/keras/qwen3_4b_en/1",
+    },
+    "qwen3_8b_en": {
+        "metadata": {
+            "description": (
+                "36-layer Qwen3 model with 8.19B parameters, featuring "
+                "enhanced reasoning, coding, and instruction-following "
+                "capabilities."
+            ),
+            "params": 8190735360,
+            "path": "qwen-3",
+        },
+        "kaggle_handle": "kaggle://keras/qwen-3/keras/qwen3_8b_en/1",
+    },
+    "qwen3_14b_en": {
+        "metadata": {
+            "description": (
+                "40-layer Qwen3 model with 14.77B parameters, featuring "
+                "advanced reasoning, coding, and multilingual capabilities."
+            ),
+            "params": 14768307200,
+            "path": "qwen-3",
+        },
+        "kaggle_handle": "kaggle://keras/qwen-3/keras/qwen3_14b_en/1",
+    },
+    "qwen3_32b_en": {
+        "metadata": {
+            "description": (
+                "64-layer Qwen3 model with 32.76B parameters, featuring "
+                "state-of-the-art performance across reasoning, coding, and "
+                "general language tasks."
+            ),
+            "params": 32762123264,
+            "path": "qwen-3",
+        },
+        "kaggle_handle": "kaggle://keras/qwen-3/keras/qwen3_32b_en/1",
+    },
+}

keras_hub/src/models/qwen_moe/qwen_moe_attention.py

@@ -67,6 +67,7 @@ class QwenMoeAttention(keras.layers.Layer):
         self.rope_scaling_factor = rope_scaling_factor
         self.use_sliding_window_attention = use_sliding_window_attention
         self.sliding_window_size = sliding_window_size
+        self.logit_soft_cap = None
 
     def build(self, inputs_shape):
         # Einsum variables:

keras_hub/src/utils/keras_utils.py

@@ -71,6 +71,23 @@ def fused_attention_op_available():
             )
             return False
         return True
+    elif (
+        hasattr(keras.config, "is_flash_attention_enabled")
+        and keras.config.backend() == "torch"
+    ):
+        try:
+            from torch.backends.cuda import SDPAParams as SDPAParams
+            from torch.backends.cuda import (
+                can_use_flash_attention as can_use_flash_attention,
+            )
+        except ImportError:
+            logging.warning(
+                "Flash attention is not supported in your current PyTorch "
+                "version. Please update it by following the official guide: "
+                "https://pytorch.org/get-started/locally/"
+            )
+            return False
+        return True
     else:
         return False
 

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.dev202507100418"
+__version__ = "0.22.0.dev202507120419"
 
 
 @keras_hub_export("keras_hub.version")

{keras_hub_nightly-0.22.0.dev202507100418.dist-info → keras_hub_nightly-0.22.0.dev202507120419.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: keras-hub-nightly
-Version: 0.22.0.dev202507100418
+Version: 0.22.0.dev202507120419
 Summary: Pretrained models for Keras.
 Author-email: Keras team <keras-users@googlegroups.com>
 License-Expression: Apache-2.0

{keras_hub_nightly-0.22.0.dev202507100418.dist-info → keras_hub_nightly-0.22.0.dev202507120419.dist-info}/RECORD

@@ -1,11 +1,11 @@
 keras_hub/__init__.py,sha256=bJbUZkqwhZvTb1Tqx1fbkq6mzBYiEyq-Hin3oQIkhdE,558
-keras_hub/layers/__init__.py,sha256=YQ4bW0_mI39Jqj2yoc8xcnynqoaXV2FBjHJviA9Ffas,5190
+keras_hub/layers/__init__.py,sha256=T1XBtpT0UH9-r0Jc-ljCxtZD_ccapf86ByvUgzdRbvg,5311
 keras_hub/metrics/__init__.py,sha256=KYalsMPBnfwim9BdGHFfJ5WxUKFXOQ1QoKIMT_0lwlM,439
-keras_hub/models/__init__.py,sha256=7MhCw7S-uIPcko-R6g5a-Jy1idKe7BwlI836PfekhHc,27076
+keras_hub/models/__init__.py,sha256=52UNIL7my_9g6ubPtOMDnGYeuGD4SOldfnGTVRMKTeE,27558
 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=OlA90iURRl_2HAopvzvHLCIWGESZFyt2k8pv4elg3I4,222
+keras_hub/src/version.py,sha256=rNXZuSfrhzhmJUw9OqMhah7oEZIE1yPYm31UkmVOLyU,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
@@ -193,7 +193,7 @@ keras_hub/src/models/flux/flux_presets.py,sha256=z7C_FbI1_F5YETXuWpc7Yh_0w-5N0eB
 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=iKSdBRkKEOnryXjz6K-thz70Dgp7LGXo5vYx8D-VMgY,10083
+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
 keras_hub/src/models/gemma/gemma_causal_lm.py,sha256=3OXaIXlrKqMIuUnBk-bUz-0SYFL-XkkQTWm8qRY2YII,16770
 keras_hub/src/models/gemma/gemma_causal_lm_preprocessor.py,sha256=bpKkEurWIfa6Kp9s4pz84-sBDSA6ZFNHP8nXG1fFQrg,2912
@@ -227,6 +227,14 @@ 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/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/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
@@ -320,14 +328,17 @@ keras_hub/src/models/qwen/qwen_decoder.py,sha256=utmAvZlU7_nP-6pjGPDinK4JaMzsQSw
 keras_hub/src/models/qwen/qwen_layernorm.py,sha256=DS35r3qd6g5ocL7Nhf_vNzLLMo1aI9VCSmL64dgNOYI,924
 keras_hub/src/models/qwen/qwen_presets.py,sha256=1FkKV6M3yqJz4EP1xa7bEvfIQ721xXT-_ikjWX0xvww,1992
 keras_hub/src/models/qwen/qwen_tokenizer.py,sha256=LCv3IyiDDHqVnM9N3lf5-BE3iwicIh0nKS1hjoPw9lE,1532
-keras_hub/src/models/qwen3/qwen3_attention.py,sha256=mq2Tak0PTItkNi5LnYVcNznS61D2Cdz1-9a-WC7Vw68,13038
+keras_hub/src/models/qwen3/__init__.py,sha256=fdndQouGmfNhB_Rj76A8my5FvpxOvRJ24DoUha-wlgw,251
+keras_hub/src/models/qwen3/qwen3_attention.py,sha256=9zjuzGZa6TzaFgO4ShNCEHMPVb3r6mFZW7vzutbwUGg,13050
 keras_hub/src/models/qwen3/qwen3_backbone.py,sha256=Ylpk_rRWWRxy8irlAPjJU-YrxYGpo8c9lSEO1zZl4gU,7456
+keras_hub/src/models/qwen3/qwen3_causal_lm.py,sha256=cn_4WFVxhlOArtIGAaqkNzIz9Rx8IEWwCVMRFKKk26k,15531
 keras_hub/src/models/qwen3/qwen3_causal_lm_preprocessor.py,sha256=H4g-bgvuhAUnDwjJovydK16Kes38ZFZWPvflrgHqZis,458
 keras_hub/src/models/qwen3/qwen3_decoder.py,sha256=68s9jQj53zFmXE4-SGXKYHu546fXOyi9LUbnKk-HGYY,11595
 keras_hub/src/models/qwen3/qwen3_layernorm.py,sha256=EJxjf7Pr6ufPQnNeuYQxkExzPjPk4PQxqMsoBeSEkDo,1073
+keras_hub/src/models/qwen3/qwen3_presets.py,sha256=ZcsmPFj3Z4TBoa7ZkJK4JN1D6iHHZ6kCrqXxQE8IH_k,2524
 keras_hub/src/models/qwen3/qwen3_tokenizer.py,sha256=LmPtg0vprMchDvYfTj8m5PraXI2QS3-YgdIIpIm5iAs,1448
 keras_hub/src/models/qwen_moe/__init__.py,sha256=5D8GUmVDsJs0J4sVZHcXOLkZf12U96l-WtwyVee4lu8,267
-keras_hub/src/models/qwen_moe/qwen_moe_attention.py,sha256=pE79_iHUm2LGkoWL6zMJw_pNfzIvmyq3yJaiq47W2TY,13242
+keras_hub/src/models/qwen_moe/qwen_moe_attention.py,sha256=o0mcVTDMtElMYq3NSYRCfuYVdF-W8YDSU5ogensrVJg,13277
 keras_hub/src/models/qwen_moe/qwen_moe_backbone.py,sha256=nrfELvIvRLmrgKrUNXci2CrecmeI6bWzJj7HH-RcWJA,15341
 keras_hub/src/models/qwen_moe/qwen_moe_causal_lm.py,sha256=MeP60v7GcN_SmH5_ULRpqgmFVgaYAosSecZiSQVlJvU,13256
 keras_hub/src/models/qwen_moe/qwen_moe_causal_lm_preprocessor.py,sha256=9P6TT7W_fqf4HsXcmlHF-DW_anR-XoDrRN2ZFGA7Ai4,3168
@@ -480,7 +491,7 @@ 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=2qrh4F-rqceVFSx0-cbsFBfWae5hBXFb_sEtPPcImf4,4628
+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/python_utils.py,sha256=N8nWeO3san4YnGkffRXG3Ix7VEIMTKSN21FX5TuL7G8,202
@@ -517,7 +528,7 @@ keras_hub/src/utils/transformers/preset_loader.py,sha256=K5FzDAtCuXS9rmZc0Zj7UCw
 keras_hub/src/utils/transformers/safetensor_utils.py,sha256=CYUHyA4y-B61r7NDnCsFb4t_UmSwZ1k9L-8gzEd6KRg,3339
 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.dev202507100418.dist-info/METADATA,sha256=3q4dYFBGZElOUYjNy4xHaOngkHOn7wbO09Gfv7VJvVw,7393
-keras_hub_nightly-0.22.0.dev202507100418.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-keras_hub_nightly-0.22.0.dev202507100418.dist-info/top_level.txt,sha256=N4J6piIWBKa38A4uV-CnIopnOEf8mHAbkNXafXm_CuA,10
-keras_hub_nightly-0.22.0.dev202507100418.dist-info/RECORD,,
+keras_hub_nightly-0.22.0.dev202507120419.dist-info/METADATA,sha256=FmJeWUJIafpgqRZRIC4nvRMeDHzdClq11rKbpHIffxQ,7393
+keras_hub_nightly-0.22.0.dev202507120419.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+keras_hub_nightly-0.22.0.dev202507120419.dist-info/top_level.txt,sha256=N4J6piIWBKa38A4uV-CnIopnOEf8mHAbkNXafXm_CuA,10
+keras_hub_nightly-0.22.0.dev202507120419.dist-info/RECORD,,