keras-hub-nightly 0.21.0.dev202505200408__tar.gz → 0.21.0.dev202505230409__tar.gz

{keras_hub_nightly-0.21.0.dev202505200408 → keras_hub_nightly-0.21.0.dev202505230409}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: keras-hub-nightly
-Version: 0.21.0.dev202505200408
+Version: 0.21.0.dev202505230409
 Summary: Pretrained models for Keras.
 Author-email: Keras team <keras-users@googlegroups.com>
 License-Expression: Apache-2.0

{keras_hub_nightly-0.21.0.dev202505200408 → keras_hub_nightly-0.21.0.dev202505230409}/keras_hub/src/models/gemma/gemma_presets.py

@@ -61,7 +61,7 @@ backbone_presets = {
             "params": 8537680896,
             "path": "gemma",
         },
-        "kaggle_handle": "kaggle://keras/gemma/keras/gemma_7b_en/3",
+        "kaggle_handle": "kaggle://keras/gemma/keras/gemma_7b_en/4",
     },
     "gemma_instruct_7b_en": {
         "metadata": {
@@ -71,7 +71,7 @@ backbone_presets = {
             "params": 8537680896,
             "path": "gemma",
         },
-        "kaggle_handle": "kaggle://keras/gemma/keras/gemma_instruct_7b_en/3",
+        "kaggle_handle": "kaggle://keras/gemma/keras/gemma_instruct_7b_en/4",
     },
     "gemma_1.1_instruct_7b_en": {
         "metadata": {
@@ -82,7 +82,7 @@ backbone_presets = {
             "params": 8537680896,
             "path": "gemma",
         },
-        "kaggle_handle": "kaggle://keras/gemma/keras/gemma_1.1_instruct_7b_en/4",
+        "kaggle_handle": "kaggle://keras/gemma/keras/gemma_1.1_instruct_7b_en/5",
     },
     "code_gemma_7b_en": {
         "metadata": {
@@ -94,7 +94,7 @@ backbone_presets = {
             "params": 8537680896,
             "path": "gemma",
         },
-        "kaggle_handle": "kaggle://keras/codegemma/keras/code_gemma_7b_en/2",
+        "kaggle_handle": "kaggle://keras/codegemma/keras/code_gemma_7b_en/3",
     },
     "code_gemma_instruct_7b_en": {
         "metadata": {
@@ -106,7 +106,7 @@ backbone_presets = {
             "params": 8537680896,
             "path": "gemma",
         },
-        "kaggle_handle": "kaggle://keras/codegemma/keras/code_gemma_instruct_7b_en/2",
+        "kaggle_handle": "kaggle://keras/codegemma/keras/code_gemma_instruct_7b_en/3",
     },
     "code_gemma_1.1_instruct_7b_en": {
         "metadata": {
@@ -118,7 +118,7 @@ backbone_presets = {
             "params": 8537680896,
             "path": "gemma",
         },
-        "kaggle_handle": "kaggle://keras/codegemma/keras/code_gemma_1.1_instruct_7b_en/2",
+        "kaggle_handle": "kaggle://keras/codegemma/keras/code_gemma_1.1_instruct_7b_en/3",
     },
     "gemma2_2b_en": {
         "metadata": {
@@ -144,7 +144,7 @@ backbone_presets = {
             "params": 9241705984,
             "path": "gemma",
         },
-        "kaggle_handle": "kaggle://keras/gemma2/keras/gemma2_9b_en/3",
+        "kaggle_handle": "kaggle://keras/gemma2/keras/gemma2_9b_en/4",
     },
     "gemma2_instruct_9b_en": {
         "metadata": {
@@ -154,7 +154,7 @@ backbone_presets = {
             "params": 9241705984,
             "path": "gemma",
         },
-        "kaggle_handle": "kaggle://keras/gemma2/keras/gemma2_instruct_9b_en/3",
+        "kaggle_handle": "kaggle://keras/gemma2/keras/gemma2_instruct_9b_en/4",
     },
     "gemma2_27b_en": {
         "metadata": {
@@ -162,7 +162,7 @@ backbone_presets = {
             "params": 27227128320,
             "path": "gemma",
         },
-        "kaggle_handle": "kaggle://keras/gemma2/keras/gemma2_27b_en/2",
+        "kaggle_handle": "kaggle://keras/gemma2/keras/gemma2_27b_en/3",
     },
     "gemma2_instruct_27b_en": {
         "metadata": {
@@ -172,7 +172,7 @@ backbone_presets = {
             "params": 27227128320,
             "path": "gemma",
         },
-        "kaggle_handle": "kaggle://keras/gemma2/keras/gemma2_instruct_27b_en/2",
+        "kaggle_handle": "kaggle://keras/gemma2/keras/gemma2_instruct_27b_en/3",
     },
     "shieldgemma_2b_en": {
         "metadata": {

{keras_hub_nightly-0.21.0.dev202505200408 → keras_hub_nightly-0.21.0.dev202505230409}/keras_hub/src/models/gemma3/gemma3_presets.py

@@ -55,7 +55,7 @@ backbone_presets = {
             "params": 11765788416,
             "path": "gemma3",
         },
-        "kaggle_handle": "kaggle://keras/gemma3/keras/gemma3_12b_text/2",
+        "kaggle_handle": "kaggle://keras/gemma3/keras/gemma3_12b_text/3",
     },
     "gemma3_instruct_12b_text": {
         "metadata": {
@@ -66,7 +66,7 @@ backbone_presets = {
             "params": 11765788416,
             "path": "gemma3",
         },
-        "kaggle_handle": "kaggle://keras/gemma3/keras/gemma3_instruct_12b_text/2",
+        "kaggle_handle": "kaggle://keras/gemma3/keras/gemma3_instruct_12b_text/3",
     },
     "gemma3_27b_text": {
         "metadata": {
@@ -77,7 +77,7 @@ backbone_presets = {
             "params": 27009002240,
             "path": "gemma3",
         },
-        "kaggle_handle": "kaggle://keras/gemma3/keras/gemma3_27b_text/3",
+        "kaggle_handle": "kaggle://keras/gemma3/keras/gemma3_27b_text/4",
     },
     "gemma3_instruct_27b_text": {
         "metadata": {
@@ -88,7 +88,7 @@ backbone_presets = {
             "params": 27009002240,
             "path": "gemma3",
         },
-        "kaggle_handle": "kaggle://keras/gemma3/keras/gemma3_instruct_27b_text/2",
+        "kaggle_handle": "kaggle://keras/gemma3/keras/gemma3_instruct_27b_text/3",
     },
     "gemma3_4b": {
         "metadata": {
@@ -121,7 +121,7 @@ backbone_presets = {
             "params": 12187079280,
             "path": "gemma3",
         },
-        "kaggle_handle": "kaggle://keras/gemma3/keras/gemma3_12b/1",
+        "kaggle_handle": "kaggle://keras/gemma3/keras/gemma3_12b/2",
     },
     "gemma3_instruct_12b": {
         "metadata": {
@@ -132,7 +132,7 @@ backbone_presets = {
             "params": 12187079280,
             "path": "gemma3",
         },
-        "kaggle_handle": "kaggle://keras/gemma3/keras/gemma3_instruct_12b/1",
+        "kaggle_handle": "kaggle://keras/gemma3/keras/gemma3_instruct_12b/2",
     },
     "gemma3_27b": {
         "metadata": {
@@ -143,7 +143,7 @@ backbone_presets = {
             "params": 27432062576,
             "path": "gemma3",
         },
-        "kaggle_handle": "kaggle://keras/gemma3/keras/gemma3_27b/1",
+        "kaggle_handle": "kaggle://keras/gemma3/keras/gemma3_27b/2",
     },
     "gemma3_instruct_27b": {
         "metadata": {
@@ -154,6 +154,6 @@ backbone_presets = {
             "params": 27432062576,
             "path": "gemma3",
         },
-        "kaggle_handle": "kaggle://keras/gemma3/keras/gemma3_instruct_27b/1",
+        "kaggle_handle": "kaggle://keras/gemma3/keras/gemma3_instruct_27b/2",
     },
 }

{keras_hub_nightly-0.21.0.dev202505200408 → keras_hub_nightly-0.21.0.dev202505230409}/keras_hub/src/models/llama/llama_presets.py

@@ -8,7 +8,7 @@ backbone_presets = {
             "params": 6738415616,
             "path": "llama",
         },
-        "kaggle_handle": "kaggle://keras/llama2/keras/llama2_7b_en/2",
+        "kaggle_handle": "kaggle://keras/llama2/keras/llama2_7b_en/3",
     },
     "llama2_7b_en_int8": {
         "metadata": {
@@ -30,7 +30,7 @@ backbone_presets = {
             "params": 6738415616,
             "path": "llama",
         },
-        "kaggle_handle": "kaggle://keras/llama2/keras/llama2_instruct_7b_en/2",
+        "kaggle_handle": "kaggle://keras/llama2/keras/llama2_instruct_7b_en/3",
     },
     "llama2_instruct_7b_en_int8": {
         "metadata": {
@@ -52,6 +52,6 @@ backbone_presets = {
             "params": 6738415616,
             "path": "llama",
         },
-        "kaggle_handle": "kaggle://keras/vicuna/keras/vicuna_1.5_7b_en/2",
+        "kaggle_handle": "kaggle://keras/vicuna/keras/vicuna_1.5_7b_en/3",
     },
 }

{keras_hub_nightly-0.21.0.dev202505200408 → keras_hub_nightly-0.21.0.dev202505230409}/keras_hub/src/models/llama3/llama3_presets.py

@@ -8,7 +8,7 @@ backbone_presets = {
             "params": 8030261248,
             "path": "llama3",
         },
-        "kaggle_handle": "kaggle://keras/llama3/keras/llama3_8b_en/4",
+        "kaggle_handle": "kaggle://keras/llama3/keras/llama3_8b_en/5",
     },
     "llama3_8b_en_int8": {
         "metadata": {
@@ -30,7 +30,7 @@ backbone_presets = {
             "params": 8030261248,
             "path": "llama3",
         },
-        "kaggle_handle": "kaggle://keras/llama3/keras/llama3_instruct_8b_en/4",
+        "kaggle_handle": "kaggle://keras/llama3/keras/llama3_instruct_8b_en/5",
     },
     "llama3_instruct_8b_en_int8": {
         "metadata": {

{keras_hub_nightly-0.21.0.dev202505200408 → keras_hub_nightly-0.21.0.dev202505230409}/keras_hub/src/models/mistral/mistral_presets.py

@@ -8,7 +8,7 @@ backbone_presets = {
             "params": 7241732096,
             "path": "mistral",
         },
-        "kaggle_handle": "kaggle://keras/mistral/keras/mistral_7b_en/7",
+        "kaggle_handle": "kaggle://keras/mistral/keras/mistral_7b_en/8",
     },
     "mistral_instruct_7b_en": {
         "metadata": {
@@ -16,7 +16,7 @@ backbone_presets = {
             "params": 7241732096,
             "path": "mistral",
         },
-        "kaggle_handle": "kaggle://keras/mistral/keras/mistral_instruct_7b_en/7",
+        "kaggle_handle": "kaggle://keras/mistral/keras/mistral_instruct_7b_en/8",
     },
     "mistral_0.2_instruct_7b_en": {
         "metadata": {
@@ -24,6 +24,6 @@ backbone_presets = {
             "params": 7241732096,
             "path": "mistral",
         },
-        "kaggle_handle": "kaggle://keras/mistral/keras/mistral_0.2_instruct_7b_en/2",
+        "kaggle_handle": "kaggle://keras/mistral/keras/mistral_0.2_instruct_7b_en/3",
     },
 }

{keras_hub_nightly-0.21.0.dev202505200408 → keras_hub_nightly-0.21.0.dev202505230409}/keras_hub/src/models/mixtral/mixtral_attention.py

@@ -27,19 +27,19 @@ class CachedMixtralAttention(keras.layers.Layer):
         **kwargs,
     ):
         super().__init__(**kwargs)
-        self._num_query_heads = num_query_heads
-        self._num_key_value_heads = num_key_value_heads
-        self._sliding_window = sliding_window
-        self._dropout = dropout
+        self.num_query_heads = num_query_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.sliding_window = sliding_window
+        self.dropout = dropout
 
-        self._num_key_value_groups = num_query_heads // num_key_value_heads
-        self._rope_max_wavelength = rope_max_wavelength
+        self.num_key_value_groups = num_query_heads // num_key_value_heads
+        self.rope_max_wavelength = rope_max_wavelength
 
         self._kernel_initializer = keras.initializers.get(
             clone_initializer(kernel_initializer)
         )
 
-        self._rope_scaling_factor = rope_scaling_factor
+        self.rope_scaling_factor = rope_scaling_factor
 
     def build(self, inputs_shape):
         # Einsum variables:
@@ -51,12 +51,12 @@ class CachedMixtralAttention(keras.layers.Layer):
         # v = num key/value heads
         # h = head dim
         self._hidden_dim = inputs_shape[-1]
-        self._head_dim = self._hidden_dim // self._num_query_heads
+        self._head_dim = self._hidden_dim // self.num_query_heads
         self._inv_norm_factor = 1.0 / math.sqrt(self._head_dim)
 
         self.query_dense = keras.layers.EinsumDense(
             equation="bqm,muh->bquh",
-            output_shape=(None, self._num_query_heads, self._head_dim),
+            output_shape=(None, self.num_query_heads, self._head_dim),
             kernel_initializer=self._kernel_initializer,
             dtype=self.dtype_policy,
             name="query",
@@ -67,7 +67,7 @@ class CachedMixtralAttention(keras.layers.Layer):
             equation="bkm,mvh->bkvh",
             output_shape=(
                 None,
-                self._num_key_value_heads,
+                self.num_key_value_heads,
                 self._head_dim,
             ),
             kernel_initializer=self._kernel_initializer,
@@ -80,7 +80,7 @@ class CachedMixtralAttention(keras.layers.Layer):
             equation="bkm,mvh->bkvh",
             output_shape=(
                 None,
-                self._num_key_value_heads,
+                self.num_key_value_heads,
                 self._head_dim,
             ),
             kernel_initializer=self._kernel_initializer,
@@ -89,31 +89,31 @@ class CachedMixtralAttention(keras.layers.Layer):
         )
         self.value_dense.build(inputs_shape)
 
-        self._softmax = keras.layers.Softmax(
+        self.softmax = keras.layers.Softmax(
             axis=-1,
             dtype="float32",
             name="attention_softmax",
         )
 
-        self._dropout_layer = keras.layers.Dropout(
-            rate=self._dropout,
+        self.dropout_layer = keras.layers.Dropout(
+            rate=self.dropout,
             dtype=self.dtype_policy,
         )
 
-        self._output_dense = keras.layers.EinsumDense(
+        self.output_dense = keras.layers.EinsumDense(
             equation="bquh,uhm->bqm",
             output_shape=(None, self._hidden_dim),
             kernel_initializer=self._kernel_initializer,
             dtype=self.dtype_policy,
             name="attention_output",
         )
-        self._output_dense.build(
-            (None, None, self._num_query_heads, self._head_dim)
+        self.output_dense.build(
+            (None, None, self.num_query_heads, self._head_dim)
         )
 
         self.rotary_embedding_layer = RotaryEmbedding(
-            max_wavelength=self._rope_max_wavelength,
-            scaling_factor=self._rope_scaling_factor,
+            max_wavelength=self.rope_max_wavelength,
+            scaling_factor=self.rope_scaling_factor,
             dtype=self.dtype_policy,
         )
 
@@ -168,18 +168,18 @@ class CachedMixtralAttention(keras.layers.Layer):
 
         # [batch_shape, seq_len, num_key_value_heads, head_dim]
         # -> [batch_shape, seq_len, num_heads, head_dim]
-        key = ops.repeat(key, repeats=self._num_key_value_groups, axis=2)
-        value = ops.repeat(value, repeats=self._num_key_value_groups, axis=2)
+        key = ops.repeat(key, repeats=self.num_key_value_groups, axis=2)
+        value = ops.repeat(value, repeats=self.num_key_value_groups, axis=2)
 
         attention_output = self._compute_attention(
             query, key, value, attention_mask
         )
 
-        attention_output = self._dropout_layer(
+        attention_output = self.dropout_layer(
             attention_output, training=training
         )
 
-        attention_output = self._output_dense(attention_output)
+        attention_output = self.output_dense(attention_output)
 
         if cache is not None:
             return attention_output, cache
@@ -187,10 +187,8 @@ class CachedMixtralAttention(keras.layers.Layer):
 
     def _masked_softmax(self, attention_scores, attention_mask=None):
         if attention_mask is not None:
-            return self._softmax(
-                attention_scores, attention_mask[:, None, :, :]
-            )
-        return self._softmax(attention_scores)
+            return self.softmax(attention_scores, attention_mask[:, None, :, :])
+        return self.softmax(attention_scores)
 
     def _use_fused_attention_op(self):
         if not fused_attention_op_available():
@@ -198,9 +196,6 @@ class CachedMixtralAttention(keras.layers.Layer):
         if self.dropout > 0.0:
             return False
         if running_on_gpu():
-            # GPU never supports softcap in the fused op.
-            if self.logit_soft_cap is not None:
-                return False
             return gpu_supports_fused_attention_op()
         elif running_on_tpu():
             # TPU supports softcap with on keras >= 3.10.
@@ -215,18 +210,12 @@ class CachedMixtralAttention(keras.layers.Layer):
                 attention_mask = ops.expand_dims(attention_mask, axis=1)
                 attention_mask = ops.cast(attention_mask, dtype="bool")
 
-            if self.logit_soft_cap:
-                kwargs = {"attn_logits_soft_cap": self.logit_soft_cap}
-            else:
-                kwargs = {}
-
             attention_output = ops.dot_product_attention(
                 query,
                 key,
                 value,
                 mask=attention_mask,
                 scale=self._inv_norm_factor,
-                **kwargs,
             )
             return attention_output
 
@@ -249,15 +238,15 @@ class CachedMixtralAttention(keras.layers.Layer):
         config = super().get_config()
         config.update(
             {
-                "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,
+                "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": keras.initializers.serialize(
                     self._kernel_initializer
                 ),
-                "sliding_window": self._sliding_window,
-                "dropout": self._dropout,
+                "sliding_window": self.sliding_window,
+                "dropout": self.dropout,
             }
         )
         return config

{keras_hub_nightly-0.21.0.dev202505200408 → keras_hub_nightly-0.21.0.dev202505230409}/keras_hub/src/models/mixtral/mixtral_presets.py

@@ -10,7 +10,7 @@ backbone_presets = {
             "params": 46702792704,
             "path": "mixtral",
         },
-        "kaggle_handle": "kaggle://keras/mixtral/keras/mixtral_8_7b_en",
+        "kaggle_handle": "kaggle://keras/mixtral/keras/mixtral_8_7b_en/3",
     },
     "mixtral_8_instruct_7b_en": {
         "metadata": {
@@ -21,6 +21,6 @@ backbone_presets = {
             "params": 46702792704,
             "path": "mixtral",
         },
-        "kaggle_handle": "kaggle://keras/mixtral/keras/mixtral_8_instruct_7b_en",
+        "kaggle_handle": "kaggle://keras/mixtral/keras/mixtral_8_instruct_7b_en/3",
     },
 }

{keras_hub_nightly-0.21.0.dev202505200408 → keras_hub_nightly-0.21.0.dev202505230409}/keras_hub/src/models/moonshine/moonshine_presets.py

@@ -9,7 +9,7 @@ backbone_presets = {
             "params": 27092736,
             "path": "moonshine",
         },
-        "kaggle_handle": "kaggle://keras/moonshine/Keras/moonshine_tiny_en",
+        "kaggle_handle": "kaggle://keras/moonshine/Keras/moonshine_tiny_en/1",
     },
     "moonshine_base_en": {
         "metadata": {
@@ -20,6 +20,6 @@ backbone_presets = {
             "params": 61513920,
             "path": "moonshine",
         },
-        "kaggle_handle": "kaggle://keras/moonshine/Keras/moonshine_base_en",
+        "kaggle_handle": "kaggle://keras/moonshine/Keras/moonshine_base_en/1",
     },
 }

{keras_hub_nightly-0.21.0.dev202505200408 → keras_hub_nightly-0.21.0.dev202505230409}/keras_hub/src/models/pali_gemma/pali_gemma_presets.py

@@ -81,7 +81,7 @@ backbone_presets = {
             "path": "pali_gemma2",
             "model_card": "https://www.kaggle.com/models/google/paligemma-2",
         },
-        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_ft_docci_10b_448/2",
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_ft_docci_10b_448/3",
     },
     "pali_gemma2_mix_3b_224": {
         "metadata": {
@@ -126,7 +126,7 @@ backbone_presets = {
             "path": "pali_gemma2",
             "model_card": "https://www.kaggle.com/models/google/paligemma-2",
         },
-        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_mix_10b_224/2",
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_mix_10b_224/3",
     },
     "pali_gemma2_mix_10b_448": {
         "metadata": {
@@ -141,7 +141,7 @@ backbone_presets = {
             "path": "pali_gemma2",
             "model_card": "https://www.kaggle.com/models/google/paligemma-2",
         },
-        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_mix_10b_448/2",
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_mix_10b_448/3",
     },
     "pali_gemma2_mix_28b_224": {
         "metadata": {
@@ -156,7 +156,7 @@ backbone_presets = {
             "path": "pali_gemma2",
             "model_card": "https://www.kaggle.com/models/google/paligemma-2",
         },
-        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_28b_mix_224/2",
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_28b_mix_224/3",
     },
     "pali_gemma2_mix_28b_448": {
         "metadata": {
@@ -171,7 +171,7 @@ backbone_presets = {
             "path": "pali_gemma2",
             "model_card": "https://www.kaggle.com/models/google/paligemma-2",
         },
-        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_28b_mix_448/2",
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_28b_mix_448/3",
     },
     "pali_gemma2_pt_3b_224": {
         "metadata": {
@@ -231,7 +231,7 @@ backbone_presets = {
             "path": "pali_gemma2",
             "model_card": "https://www.kaggle.com/models/google/paligemma-2",
         },
-        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_10b_224/2",
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_10b_224/3",
     },
     "pali_gemma2_pt_10b_448": {
         "metadata": {
@@ -246,7 +246,7 @@ backbone_presets = {
             "path": "pali_gemma2",
             "model_card": "https://www.kaggle.com/models/google/paligemma-2",
         },
-        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_10b_448/2",
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_10b_448/3",
     },
     "pali_gemma2_pt_10b_896": {
         "metadata": {
@@ -261,7 +261,7 @@ backbone_presets = {
             "path": "pali_gemma2",
             "model_card": "https://www.kaggle.com/models/google/paligemma-2",
         },
-        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_10b_896/2",
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_10b_896/3",
     },
     "pali_gemma2_pt_28b_224": {
         "metadata": {
@@ -276,7 +276,7 @@ backbone_presets = {
             "path": "pali_gemma2",
             "model_card": "https://www.kaggle.com/models/google/paligemma-2",
         },
-        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_28b_224/3",
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_28b_224/4",
     },
     "pali_gemma2_pt_28b_448": {
         "metadata": {
@@ -291,7 +291,7 @@ backbone_presets = {
             "path": "pali_gemma2",
             "model_card": "https://www.kaggle.com/models/google/paligemma-2",
         },
-        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_28b_448/2",
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_28b_448/3",
     },
     "pali_gemma2_pt_28b_896": {
         "metadata": {
@@ -306,6 +306,6 @@ backbone_presets = {
             "path": "pali_gemma2",
             "model_card": "https://www.kaggle.com/models/google/paligemma-2",
         },
-        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_28b_896/2",
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_28b_896/3",
     },
 }

{keras_hub_nightly-0.21.0.dev202505200408 → keras_hub_nightly-0.21.0.dev202505230409}/keras_hub/src/models/qwen/qwen_presets.py

@@ -7,7 +7,7 @@ backbone_presets = {
             "params": 494032768,
             "path": "qwen",
         },
-        "kaggle_handle": "kaggle://keras/qwen/keras/qwen2.5_0.5b_en",
+        "kaggle_handle": "kaggle://keras/qwen/keras/qwen2.5_0.5b_en/1",
     },
     "qwen2.5_3b_en": {
         "metadata": {
@@ -15,7 +15,7 @@ backbone_presets = {
             "params": 3085938688,
             "path": "qwen",
         },
-        "kaggle_handle": "kaggle://keras/qwen/keras/qwen2.5_3b_en",
+        "kaggle_handle": "kaggle://keras/qwen/keras/qwen2.5_3b_en/1",
     },
     "qwen2.5_7b_en": {
         "metadata": {
@@ -23,7 +23,7 @@ backbone_presets = {
             "params": 6993420288,
             "path": "qwen",
         },
-        "kaggle_handle": "kaggle://keras/qwen/keras/qwen2.5_7b_en/2",
+        "kaggle_handle": "kaggle://keras/qwen/keras/qwen2.5_7b_en/3",
     },
     "qwen2.5_instruct_0.5b_en": {
         "metadata": {
@@ -34,7 +34,7 @@ backbone_presets = {
             "params": 494032768,
             "path": "qwen",
         },
-        "kaggle_handle": "kaggle://keras/qwen/keras/qwen2.5_instruct_0.5b_en",
+        "kaggle_handle": "kaggle://keras/qwen/keras/qwen2.5_instruct_0.5b_en/1",
     },
     "qwen2.5_instruct_32b_en": {
         "metadata": {
@@ -45,7 +45,7 @@ backbone_presets = {
             "params": 32763876352,
             "path": "qwen",
         },
-        "kaggle_handle": "kaggle://keras/qwen/keras/qwen2.5_instruct_32b_en",
+        "kaggle_handle": "kaggle://keras/qwen/keras/qwen2.5_instruct_32b_en/2",
     },
     "qwen2.5_instruct_72b_en": {
         "metadata": {
@@ -56,6 +56,6 @@ backbone_presets = {
             "params": 72706203648,
             "path": "qwen",
         },
-        "kaggle_handle": "kaggle://keras/qwen/keras/qwen2.5_instruct_72b_en",
+        "kaggle_handle": "kaggle://keras/qwen/keras/qwen2.5_instruct_72b_en/2",
     },
 }

{keras_hub_nightly-0.21.0.dev202505200408 → keras_hub_nightly-0.21.0.dev202505230409}/keras_hub/src/models/qwen_moe/qwen_moe_attention.py

@@ -256,9 +256,6 @@ class QwenMoeAttention(keras.layers.Layer):
         if self.dropout > 0.0:
             return False
         if running_on_gpu():
-            # GPU never supports softcap in the fused op.
-            if self.logit_soft_cap is not None:
-                return False
             return gpu_supports_fused_attention_op()
         elif running_on_tpu():
             # TPU supports softcap with on keras >= 3.10.
@@ -268,7 +265,13 @@ class QwenMoeAttention(keras.layers.Layer):
             return False
 
     def _compute_attention(
-        self, query, key, value, attention_mask=None, cache_update_index=None
+        self,
+        query,
+        key,
+        value,
+        attention_mask=None,
+        cache_update_index=None,
+        **kwargs,
     ):
         """Computes attention using query, key, and value tensors.
 
@@ -289,11 +292,6 @@ class QwenMoeAttention(keras.layers.Layer):
                 attention_mask = ops.expand_dims(attention_mask, axis=1)
                 attention_mask = ops.cast(attention_mask, dtype="bool")
 
-            if self.logit_soft_cap:
-                kwargs = {"attn_logits_soft_cap": self.logit_soft_cap}
-            else:
-                kwargs = {}
-
             attention_output = ops.dot_product_attention(
                 query,
                 key,

{keras_hub_nightly-0.21.0.dev202505200408 → keras_hub_nightly-0.21.0.dev202505230409}/keras_hub/src/models/qwen_moe/qwen_moe_presets.py

@@ -10,6 +10,6 @@ backbone_presets = {
             "params": 14315784192,
             "path": "qwen-1.5-moe",
         },
-        "kaggle_handle": "kaggle://keras/qwen-1.5-moe/Keras/qwen1.5_moe_2.7b_en",
+        "kaggle_handle": "kaggle://keras/qwen-1.5-moe/Keras/qwen1.5_moe_2.7b_en/3",
     },
 }