keras-hub-nightly 0.20.0.dev202503170356__py3-none-any.whl → 0.20.0.dev202503180354__py3-none-any.whl

keras_hub/src/models/qwen/qwen_decoder.py

@@ -0,0 +1,311 @@
+import keras
+from keras import ops
+
+from keras_hub.src.layers.modeling.transformer_layer_utils import (
+    compute_causal_mask,
+)
+from keras_hub.src.layers.modeling.transformer_layer_utils import (
+    merge_padding_and_attention_mask,
+)
+from keras_hub.src.models.qwen.qwen_attention import QwenAttention
+from keras_hub.src.models.qwen.qwen_layernorm import QwenLayerNorm
+from keras_hub.src.utils.keras_utils import clone_initializer
+
+
+class QwenTransformerDecoder(keras.layers.Layer):
+    """A Transformer decoder layer for the Qwen backbone.
+
+    This layer implements a Transformer decoder block that includes
+    self-attention with optional sliding window attention and a feed-forward
+    network.
+
+    Args:
+    intermediate_dim: Output dimension of the first dense layer in the
+        feed-forward network.
+    num_query_heads: Number of query attention heads.
+    num_key_value_heads: Number of key/value attention heads (for GQA).
+    rope_max_wavelength: Maximum wavelength for RoPE (Rotary Position
+        Embedding).
+    rope_scaling_factor: Scaling factor for RoPE, used for extending
+        context length.
+    activation: Activation function to use in the feed-forward network.
+    layer_norm_epsilon: Small float added to variance to avoid dividing
+        by zero in layer norm.
+    kernel_initializer: Initializer for the kernel weights.
+    dropout: Dropout rate for attention and hidden layers.
+    use_sliding_window_attention: Whether to use sliding window
+        attention.
+    sliding_window_size: Size of the sliding window for attention when
+        enabled.
+    **kwargs: Additional keyword arguments to pass to the Layer.
+    """
+
+    def __init__(
+        self,
+        intermediate_dim,
+        num_query_heads,
+        num_key_value_heads,
+        rope_max_wavelength=10000,
+        rope_scaling_factor=1.0,
+        activation="silu",
+        layer_norm_epsilon=1e-5,
+        kernel_initializer="glorot_uniform",
+        dropout=0,
+        use_sliding_window_attention=False,
+        sliding_window_size=4096,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.intermediate_dim = intermediate_dim
+        self.num_query_heads = num_query_heads
+        self.num_key_value_heads = num_key_value_heads
+
+        self.rope_max_wavelength = rope_max_wavelength
+        self.rope_scaling_factor = rope_scaling_factor
+
+        self.dropout = dropout
+
+        self.use_sliding_window_attention = use_sliding_window_attention
+        self.sliding_window_size = sliding_window_size
+
+        self.activation = keras.activations.get(activation)
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.kernel_initializer = keras.initializers.get(kernel_initializer)
+
+        self.supports_masking = True
+
+    def build(self, decoder_sequence_shape):
+        self._decoder_sequence_shape = decoder_sequence_shape
+        self.hidden_dim = decoder_sequence_shape[-1]
+
+        # Self attention layer.
+        self._self_attention_layer = QwenAttention(
+            num_query_heads=self.num_query_heads,
+            num_key_value_heads=self.num_key_value_heads,
+            rope_max_wavelength=self.rope_max_wavelength,
+            rope_scaling_factor=self.rope_scaling_factor,
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            dropout=self.dropout,
+            use_sliding_window_attention=self.use_sliding_window_attention,
+            sliding_window_size=self.sliding_window_size,
+            dtype=self.dtype_policy,
+            name="self_attention",
+        )
+        self._self_attention_layer.build(decoder_sequence_shape)
+
+        self._self_attention_layernorm = QwenLayerNorm(
+            epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
+            name="self_attention_layernorm",
+        )
+
+        self._self_attention_layernorm.build(decoder_sequence_shape)
+        self._self_attention_dropout = keras.layers.Dropout(
+            rate=self.dropout,
+            dtype=self.dtype_policy,
+            name="self_attention_dropout",
+        )
+
+        # Feedforward layers.
+        self._feedforward_intermediate_dense = keras.layers.Dense(
+            self.intermediate_dim,
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            use_bias=False,
+            dtype=self.dtype_policy,
+            name="feedforward_intermediate_dense",
+        )
+        self._feedforward_intermediate_dense.build(decoder_sequence_shape)
+
+        self._feedforward_gate_dense = keras.layers.Dense(
+            self.intermediate_dim,
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            use_bias=False,
+            dtype=self.dtype_policy,
+            name="feedforward_gate_dense",
+        )
+        self._feedforward_gate_dense.build(decoder_sequence_shape)
+
+        self._feedforward_output_dense = keras.layers.Dense(
+            self.hidden_dim,
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            use_bias=False,
+            dtype=self.dtype_policy,
+            name="feedforward_output_dense",
+        )
+
+        self._feedforward_output_dense.build(
+            self._feedforward_gate_dense.compute_output_shape(
+                decoder_sequence_shape
+            )
+        )
+
+        self._feedforward_layernorm = QwenLayerNorm(
+            epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
+            name="feedforward_layernorm",
+        )
+        self._feedforward_layernorm.build(decoder_sequence_shape)
+
+        self.built = True
+
+    def call(
+        self,
+        decoder_sequence,
+        decoder_padding_mask=None,
+        decoder_attention_mask=None,
+        self_attention_cache=None,
+        self_attention_cache_update_index=None,
+        training=None,
+    ):
+        """Forward pass for the decoder layer.
+
+        Args:
+            decoder_sequence: Input tensor of shape [batch_size, seq_length,
+                hidden_size].
+            decoder_padding_mask: Mask tensor for padding tokens.
+            decoder_attention_mask: Additional attention mask.
+            self_attention_cache: Optional cached key and value tensors for
+                self-attention.
+            self_attention_cache_update_index: Index at which to update the
+                cache.
+            training: Boolean indicating whether in training mode.
+
+        Returns:
+            decoder_output: Output tensor after applying transformer decoder
+                block.
+            self_attention_cache: Updated cache tensors (if cache is provided).
+        """
+        self_attention_mask = self._compute_self_attention_mask(
+            decoder_sequence=decoder_sequence,
+            decoder_padding_mask=decoder_padding_mask,
+            decoder_attention_mask=decoder_attention_mask,
+            self_attention_cache=self_attention_cache,
+            self_attention_cache_update_index=self_attention_cache_update_index,
+        )
+        residual = decoder_sequence
+
+        x = self._self_attention_layernorm(decoder_sequence)
+
+        # Self attention block.
+        x = self._self_attention_layer(
+            hidden_states=x,
+            attention_mask=self_attention_mask,
+            cache=self_attention_cache,
+            cache_update_index=self_attention_cache_update_index,
+        )
+
+        if self_attention_cache is not None:
+            x, self_attention_cache = x
+
+        x = self._self_attention_dropout(x, training=training)
+
+        x = x + residual
+        residual = x
+
+        x = self._feedforward_layernorm(x)
+        gate_output = self._feedforward_gate_dense(x)
+
+        # Note that we run the activation function in full 32-bit
+        # precision since this is what `torch.nn.functional.silu`
+        # does. Internally, `torch.nn.functional.silu` converts the
+        # inputs to float32, computes SiLU, and converts the outputs
+        # back to compute dtype.
+        # CPU Kernel: https://github.com/pytorch/pytorch/blob/35c493f2cf9b623bfdc7e6b34dc1cb39690a7919/aten/src/ATen/native/cpu/Activation.cpp#L1221-L1235  # noqa: E501
+        # CUDA Kernel: https://github.com/pytorch/pytorch/blob/35c493f2cf9b623bfdc7e6b34dc1cb39690a7919/aten/src/ATen/native/cuda/ActivationSiluKernel.cu  # noqa: E501
+        gate_output = ops.cast(gate_output, "float32")
+        gate_output = self.activation(gate_output)
+        gate_output = ops.cast(gate_output, self.compute_dtype)
+
+        x = self._feedforward_intermediate_dense(x)
+
+        x = self._feedforward_output_dense(ops.multiply(x, gate_output))
+
+        decoder_output = x + residual
+
+        if self_attention_cache is not None:
+            return decoder_output, self_attention_cache
+        return decoder_output
+
+    def _compute_self_attention_mask(
+        self,
+        decoder_sequence,
+        decoder_padding_mask,
+        decoder_attention_mask,
+        self_attention_cache,
+        self_attention_cache_update_index,
+    ):
+        """Computes the self-attention mask combining causal, padding and
+        attention masks.
+
+        Args:
+            decoder_sequence: Input tensor.
+            decoder_padding_mask: Mask tensor for padding tokens.
+            decoder_attention_mask: Additional attention mask.
+            self_attention_cache: Optional cached key and value tensors.
+            self_attention_cache_update_index: Index at which to update the
+                cache.
+
+        Returns:
+            Combined attention mask tensor.
+        """
+        decoder_mask = merge_padding_and_attention_mask(
+            decoder_sequence, decoder_padding_mask, decoder_attention_mask
+        )
+        batch_size = ops.shape(decoder_sequence)[0]
+        input_length = output_length = ops.shape(decoder_sequence)[1]
+        # We need to handle a rectangular causal mask when doing cached
+        # decoding. For generative inference, `decoder_sequence` will
+        # generally be length 1, and `cache` will be the full generation length.
+        if self_attention_cache is not None:
+            input_length = ops.shape(self_attention_cache)[2]
+
+        cache_update_index = (
+            0
+            if self_attention_cache_update_index is None
+            else self_attention_cache_update_index
+        )
+
+        causal_mask = compute_causal_mask(
+            batch_size, input_length, output_length, cache_update_index
+        )
+
+        return (
+            ops.minimum(decoder_mask, causal_mask)
+            if decoder_mask is not None
+            else causal_mask
+        )
+
+    def compute_output_shape(self, decoder_sequence_shape):
+        """Computes the output shape of the layer.
+
+        Args:
+            decoder_sequence_shape: Shape of the decoder sequence input.
+
+        Returns:
+            Output shape, which is the same as the input shape.
+        """
+        return decoder_sequence_shape
+
+    def get_config(self):
+        """Returns the config of the layer.
+
+        Returns:
+            Dictionary containing the parameters used to initialize this layer.
+        """
+        config = super().get_config()
+        config.update(
+            {
+                "intermediate_dim": self.intermediate_dim,
+                "num_query_heads": self.num_query_heads,
+                "rope_max_wavelength": self.rope_max_wavelength,
+                "rope_scaling_factor": self.rope_scaling_factor,
+                "num_key_value_heads": self.num_key_value_heads,
+                "activation": keras.activations.serialize(self.activation),
+                "layer_norm_epsilon": self.layer_norm_epsilon,
+                "kernel_initializer": keras.initializers.serialize(
+                    self.kernel_initializer
+                ),
+                "dropout": self.dropout,
+            }
+        )
+        return config

keras_hub/src/models/qwen/qwen_layernorm.py

@@ -0,0 +1,32 @@
+import keras
+from keras import ops
+
+
+class QwenLayerNorm(keras.layers.Layer):
+    """A normalization layer for Qwen that implements RMS normalization."""
+
+    def __init__(self, epsilon=1e-6, **kwargs):
+        super().__init__(**kwargs)
+        self.epsilon = epsilon
+
+    def build(self, input_shape):
+        dim = input_shape[-1]
+        self.scale = self.add_weight(
+            name="scale",
+            trainable=True,
+            shape=(dim,),
+            initializer="ones",
+            dtype=self.variable_dtype,
+        )
+        self.built = True
+
+    def call(self, x):
+        x = ops.cast(x, "float32")
+        var = ops.mean(ops.power(x, 2), axis=-1, keepdims=True)
+        x = x * ops.rsqrt(var + self.epsilon)
+        return ops.cast(x * self.scale, self.compute_dtype)
+
+    def get_config(self):
+        config = super().get_config()
+        config.update({"epsilon": self.epsilon})
+        return config

keras_hub/src/models/qwen/qwen_tokenizer.py

@@ -0,0 +1,51 @@
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.qwen.qwen_backbone import QwenBackbone
+from keras_hub.src.tokenizers.byte_pair_tokenizer import BytePairTokenizer
+
+
+@keras_hub_export(
+    [
+        "keras_hub.tokenizers.QwenTokenizer",
+        "keras_hub.tokenizers.Qwen2Tokenizer",
+        "keras_hub.models.QwenTokenizer",
+        "keras_hub.models.Qwen2Tokenizer",
+    ]
+)
+class QwenTokenizer(BytePairTokenizer):
+    """Tokenizer for Qwen models.
+
+    This tokenizer implements byte-pair encoding (BPE) for Qwen models,
+    handling special tokens like BOS (beginning of sequence) and EOS (end of
+    sequence).
+
+    Args:
+        vocabulary: Dictionary mapping tokens to token IDs, or path to
+            vocabulary file.
+        merges: List of BPE merges, or path to merges file.
+        bos_token: Beginning of sequence token. Defaults to None.
+        eos_token: End of sequence token. Defaults to "<|endoftext|>".
+        misc_special_tokens: Set of additional special tokens. Defaults to
+            empty set.
+    """
+
+    backbone_cls = QwenBackbone
+
+    def __init__(
+        self,
+        vocabulary=None,
+        merges=None,
+        **kwargs,
+    ):
+        # Add EOS token
+        eos_token = "<|endoftext|>"
+        self._add_special_token(eos_token, "end_token")
+
+        self.start_token_id = None
+        self.start_token = None
+        self.pad_token_id = 0
+
+        super().__init__(
+            vocabulary=vocabulary,
+            merges=merges,
+            **kwargs,
+        )

keras_hub/src/utils/transformers/convert_qwen.py

@@ -0,0 +1,148 @@
+import numpy as np
+
+from keras_hub.src.models.qwen.qwen_backbone import QwenBackbone
+from keras_hub.src.utils.preset_utils import load_json
+
+backbone_cls = QwenBackbone
+
+
+def convert_backbone_config(transformers_config):
+    return {
+        "vocabulary_size": transformers_config["vocab_size"],
+        "hidden_dim": transformers_config["hidden_size"],
+        "num_layers": transformers_config["num_hidden_layers"],
+        "num_query_heads": transformers_config["num_attention_heads"],
+        "num_key_value_heads": transformers_config["num_key_value_heads"],
+        "intermediate_dim": transformers_config["intermediate_size"],
+        "layer_norm_epsilon": transformers_config["rms_norm_eps"],
+        "rope_max_wavelength": transformers_config["rope_theta"],
+        "use_sliding_window": transformers_config["use_sliding_window"],
+        "sliding_window_size": transformers_config["sliding_window"],
+    }
+
+
+def convert_weights(backbone, loader, transformers_config):
+    loader.port_weight(
+        keras_variable=backbone.get_layer("token_embedding").embeddings,
+        hf_weight_key="model.embed_tokens.weight",
+    )
+    if not backbone.tie_word_embeddings:
+        loader.port_weight(
+            keras_variable=backbone.get_layer(
+                "token_embedding"
+            ).reverse_embeddings,
+            hf_weight_key="lm_head.weight",
+            # rearrange_pattern="b a -> a b",
+            hook_fn=lambda hf_tensor, _: np.transpose(hf_tensor, axes=(1, 0)),
+        )
+
+    def transpose_and_reshape(x, shape):
+        return np.reshape(np.transpose(x), shape)
+
+    for i in range(backbone.num_layers):
+        decoder_layer = backbone.get_layer(f"transformer_layer_{i}")
+
+        # Input layernorm
+        loader.port_weight(
+            keras_variable=decoder_layer._self_attention_layernorm.scale,
+            hf_weight_key=f"model.layers.{i}.input_layernorm.weight",
+        )
+
+        # Attention layers
+
+        ## Query
+        loader.port_weight(
+            keras_variable=decoder_layer._self_attention_layer._query_dense.kernel,
+            hf_weight_key=f"model.layers.{i}.self_attn.q_proj.weight",
+            hook_fn=transpose_and_reshape,
+        )
+        loader.port_weight(
+            keras_variable=decoder_layer._self_attention_layer._query_dense.bias,
+            hf_weight_key=f"model.layers.{i}.self_attn.q_proj.bias",
+            hook_fn=transpose_and_reshape,
+        )
+        ## Key
+        loader.port_weight(
+            keras_variable=decoder_layer._self_attention_layer._key_dense.kernel,
+            hf_weight_key=f"model.layers.{i}.self_attn.k_proj.weight",
+            hook_fn=transpose_and_reshape,
+        )
+        loader.port_weight(
+            keras_variable=decoder_layer._self_attention_layer._key_dense.bias,
+            hf_weight_key=f"model.layers.{i}.self_attn.k_proj.bias",
+            hook_fn=transpose_and_reshape,
+        )
+        ## Value
+        loader.port_weight(
+            keras_variable=decoder_layer._self_attention_layer._value_dense.kernel,
+            hf_weight_key=f"model.layers.{i}.self_attn.v_proj.weight",
+            hook_fn=transpose_and_reshape,
+        )
+        loader.port_weight(
+            keras_variable=decoder_layer._self_attention_layer._value_dense.bias,
+            hf_weight_key=f"model.layers.{i}.self_attn.v_proj.bias",
+            hook_fn=transpose_and_reshape,
+        )
+        ## Output
+        loader.port_weight(
+            keras_variable=decoder_layer._self_attention_layer._output_dense.kernel,
+            hf_weight_key=f"model.layers.{i}.self_attn.o_proj.weight",
+            # rearrange_patterns="c (a b) -> a b c",
+            # rearrange_dims={"a": backbone.num_query_heads},
+            hook_fn=transpose_and_reshape,
+        )
+
+        # MLP layers
+        loader.port_weight(
+            keras_variable=decoder_layer._feedforward_intermediate_dense.kernel,
+            hf_weight_key=f"model.layers.{i}.mlp.up_proj.weight",
+            # rearrange_patterns="b a -> a b",
+            hook_fn=lambda hf_tensor, _: np.transpose(hf_tensor, axes=(1, 0)),
+        )
+        loader.port_weight(
+            keras_variable=decoder_layer._feedforward_output_dense.kernel,
+            hf_weight_key=f"model.layers.{i}.mlp.down_proj.weight",
+            # rearrange_patterns="b a -> a b",
+            hook_fn=lambda hf_tensor, _: np.transpose(hf_tensor, axes=(1, 0)),
+        )
+        loader.port_weight(
+            keras_variable=decoder_layer._feedforward_gate_dense.kernel,
+            hf_weight_key=f"model.layers.{i}.mlp.gate_proj.weight",
+            # rearrange_patterns="b a -> a b",
+            hook_fn=lambda hf_tensor, _: np.transpose(hf_tensor, axes=(1, 0)),
+        )
+
+        # Feedforward layernorm
+        loader.port_weight(
+            keras_variable=decoder_layer._feedforward_layernorm.scale,
+            hf_weight_key=f"model.layers.{i}.post_attention_layernorm.weight",
+        )
+
+    # Final normalization layer
+    loader.port_weight(
+        keras_variable=backbone.get_layer("sequence_output_layernorm").scale,
+        hf_weight_key="model.norm.weight",
+    )
+
+    return backbone
+
+
+def convert_tokenizer(cls, preset, **kwargs):
+    tokenizer_config = load_json(preset, "tokenizer.json")
+    vocab = tokenizer_config["model"]["vocab"]
+    merges = tokenizer_config["model"]["merges"]
+
+    # Load all special tokens with the exception of "reserved" ones.
+    special_tokens = set()
+    for token in tokenizer_config["added_tokens"]:
+        if not token["content"].startswith("<|reserved_special_token_"):
+            vocab[token["content"]] = token["id"]
+            special_tokens.add(token["content"])
+
+    kwargs.update(
+        {
+            "unsplittable_tokens": list(special_tokens),
+        }
+    )
+
+    return cls(vocabulary=vocab, merges=merges, **kwargs)

keras_hub/src/utils/transformers/preset_loader.py

@@ -12,6 +12,7 @@ from keras_hub.src.utils.transformers import convert_gpt2
 from keras_hub.src.utils.transformers import convert_llama3
 from keras_hub.src.utils.transformers import convert_mistral
 from keras_hub.src.utils.transformers import convert_pali_gemma
+from keras_hub.src.utils.transformers import convert_qwen
 from keras_hub.src.utils.transformers import convert_vit
 from keras_hub.src.utils.transformers.safetensor_utils import SafetensorLoader
 
@@ -41,6 +42,8 @@ class TransformersPresetLoader(PresetLoader):
             self.converter = convert_pali_gemma
         elif model_type == "vit":
             self.converter = convert_vit
+        elif model_type == "qwen2":
+            self.converter = convert_qwen
         else:
             raise ValueError(
                 "KerasHub has no converter for huggingface/transformers models "

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__ = "0.20.0.dev202503170356"
+__version__ = "0.20.0.dev202503180354"
 
 
 @keras_hub_export("keras_hub.version")

{keras_hub_nightly-0.20.0.dev202503170356.dist-info → keras_hub_nightly-0.20.0.dev202503180354.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: keras-hub-nightly
-Version: 0.20.0.dev202503170356
+Version: 0.20.0.dev202503180354
 Summary: Industry-strength Natural Language Processing extensions for Keras.
 Home-page: https://github.com/keras-team/keras-hub
 Author: Keras team

{keras_hub_nightly-0.20.0.dev202503170356.dist-info → keras_hub_nightly-0.20.0.dev202503180354.dist-info}/RECORD

@@ -2,13 +2,13 @@ keras_hub/__init__.py,sha256=QGdXyHgYt6cMUAP1ebxwc6oR86dE0dkMxNy2eOCQtFo,855
 keras_hub/api/__init__.py,sha256=EzR6D-XWsm_gDrX5LDwKEmrah_gu3ffpj8GKBudE0yI,485
 keras_hub/api/layers/__init__.py,sha256=-yHyqsjWBhmFv9RSS2cMyPcieU1RkNzcNsq9IDXSVFE,3626
 keras_hub/api/metrics/__init__.py,sha256=So8Ec-lOcTzn_UUMmAdzDm8RKkPu2dbRUm2px8gpUEI,381
-keras_hub/api/models/__init__.py,sha256=vtOXBt8YNXvaDrfpKRpJ6MXjU5FzMv2a44Db_P8cGUg,17606
+keras_hub/api/models/__init__.py,sha256=U9LffuV0XchcdCWxl-I8qaOvYJ0bwdfq-6O_CTbb9Qc,18310
 keras_hub/api/samplers/__init__.py,sha256=n-_SEXxr2LNUzK2FqVFN7alsrkx1P_HOVTeLZKeGCdE,730
-keras_hub/api/tokenizers/__init__.py,sha256=lhvIqP8xqdkjmKSEBujHNxh5Tk5A3T0I7AUuMmKzx00,2597
+keras_hub/api/tokenizers/__init__.py,sha256=LsVLrAxTVe9YT9ixsGYnbtWuyfPW5-FW3Wt3xV08_LE,2758
 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=2VuoNZma825YX8T8OG2oe63hCK5X0pXyT0-5xLepnJQ,222
+keras_hub/src/version_utils.py,sha256=incPTPm_OMuYw4E1pHpSIS3-f26vWZoaY7Bys-vEw7E,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
@@ -29,7 +29,7 @@ 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=ILhb8sPusjdegKbc60PVdi7RbwNtqyv-Jm50tSUs6RI,15408
+keras_hub/src/layers/preprocessing/image_converter.py,sha256=E5EYbURAVNntnx0eQ886QOkOPAOc2TKrvZhe39SL5sU,15552
 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
@@ -271,6 +271,14 @@ keras_hub/src/models/phi3/phi3_layernorm.py,sha256=Oqu81tGd97Lzx3kG1QEtZ0S6gbfn3
 keras_hub/src/models/phi3/phi3_presets.py,sha256=sb2ce7Gq1OikFEf2KIYG69rFKHYKj8qhlN-Ea8d6J7k,1366
 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/qwen/__init__.py,sha256=hskG3tZUY_AYZPp0WVzbCtw37AIYENyp3DOnqHmdRBw,65
+keras_hub/src/models/qwen/qwen_attention.py,sha256=vBPGdNMRnfuETxxdwDzwpObOvt3zB2qqc9kbWRRKuQg,12951
+keras_hub/src/models/qwen/qwen_backbone.py,sha256=xBu2zEzFFAjKgaHOqPnxLU-j4oL3N2G2KT-uwL2zEM0,13018
+keras_hub/src/models/qwen/qwen_causal_lm.py,sha256=_f-UHaKHp0ncxknpkpEJiW3jlng3E4CmddjQfz2QzJo,12249
+keras_hub/src/models/qwen/qwen_causal_lm_preprocessor.py,sha256=Va-4TLJD3ycEnkS41rF3dVj4_6K0j-gxLTrREFRcyr0,609
+keras_hub/src/models/qwen/qwen_decoder.py,sha256=utmAvZlU7_nP-6pjGPDinK4JaMzsQSwOARG0ote-jAg,11771
+keras_hub/src/models/qwen/qwen_layernorm.py,sha256=DS35r3qd6g5ocL7Nhf_vNzLLMo1aI9VCSmL64dgNOYI,924
+keras_hub/src/models/qwen/qwen_tokenizer.py,sha256=LCv3IyiDDHqVnM9N3lf5-BE3iwicIh0nKS1hjoPw9lE,1532
 keras_hub/src/models/resnet/__init__.py,sha256=C5UqlQ6apm8WSp1bnrxB6Bi3BGaknxRQs-r3b2wpaGA,257
 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
@@ -424,10 +432,11 @@ keras_hub/src/utils/transformers/convert_gpt2.py,sha256=HCeHN_-GiQJRxLCM9OCJJ1wa
 keras_hub/src/utils/transformers/convert_llama3.py,sha256=zlg0yFscjytyOFymDwqnbuXkmYvb88qqYzAROKcpaPU,5250
 keras_hub/src/utils/transformers/convert_mistral.py,sha256=kVhN9h1ZFVhwkNW8p3wnS7eANJUXIsNy1RxWXy20Gqw,4760
 keras_hub/src/utils/transformers/convert_pali_gemma.py,sha256=B1leeDw96Yvu81hYumf66hIid07k5NLqoeWAJgPnaLs,10649
+keras_hub/src/utils/transformers/convert_qwen.py,sha256=I2bfwo8AQd_JfwFpiAuCQ3k_FC66J5lY7tYt99yMc9E,5811
 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/preset_loader.py,sha256=0Hi7R8HnATcwFVLsJwMMIMWTCXHNfep4IPiRpQXqM-w,3933
 keras_hub/src/utils/transformers/safetensor_utils.py,sha256=CYUHyA4y-B61r7NDnCsFb4t_UmSwZ1k9L-8gzEd6KRg,3339
-keras_hub_nightly-0.20.0.dev202503170356.dist-info/METADATA,sha256=Yt3QbpFldQbmp0Yu8fw1w0q1Z6zKh8WeqbRjWcAOTOM,7715
-keras_hub_nightly-0.20.0.dev202503170356.dist-info/WHEEL,sha256=52BFRY2Up02UkjOa29eZOS2VxUrpPORXg1pkohGGUS8,91
-keras_hub_nightly-0.20.0.dev202503170356.dist-info/top_level.txt,sha256=N4J6piIWBKa38A4uV-CnIopnOEf8mHAbkNXafXm_CuA,10
-keras_hub_nightly-0.20.0.dev202503170356.dist-info/RECORD,,
+keras_hub_nightly-0.20.0.dev202503180354.dist-info/METADATA,sha256=J801c04q7y8QuruHO3nIl4eGkEGz6EkldpjG82IhZ3w,7715
+keras_hub_nightly-0.20.0.dev202503180354.dist-info/WHEEL,sha256=beeZ86-EfXScwlR_HKu4SllMC9wUEj_8Z_4FJ3egI2w,91
+keras_hub_nightly-0.20.0.dev202503180354.dist-info/top_level.txt,sha256=N4J6piIWBKa38A4uV-CnIopnOEf8mHAbkNXafXm_CuA,10
+keras_hub_nightly-0.20.0.dev202503180354.dist-info/RECORD,,

{keras_hub_nightly-0.20.0.dev202503170356.dist-info → keras_hub_nightly-0.20.0.dev202503180354.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (76.0.0)
+Generator: setuptools (76.1.0)
 Root-Is-Purelib: true
 Tag: py3-none-any