keras-hub-nightly 0.15.0.dev20240823171555__py3-none-any.whl

keras_hub/src/models/pali_gemma/pali_gemma_causal_lm.py

@@ -0,0 +1,313 @@
+# Copyright 2024 The KerasHub Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writingf, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+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.pali_gemma.pali_gemma_backbone import (
+    PaliGemmaBackbone,
+)
+from keras_hub.src.models.pali_gemma.pali_gemma_causal_lm_preprocessor import (
+    PaliGemmaCausalLMPreprocessor,
+)
+from keras_hub.src.utils.tensor_utils import any_equal
+
+
+@keras_hub_export("keras_hub.models.PaliGemmaCausalLM")
+class PaliGemmaCausalLM(CausalLM):
+    """An end-to-end multi modal PaliGemma 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
+    image and plain text input, or to autoregressively generate plain text
+    similar to the data used for training.
+
+    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.PaliGemmaBackbone` instance.
+        preprocessor: A `keras_hub.models.PaliGemmaCausalLMPreprocessor` 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
+    image = np.random.rand(224, 224, 3)
+    pali_gemma_lm = keras_hub.models.PaliGemmaCausalLM.from_preset(
+        "pali_gemma_3b_mix_224"
+    )
+    pali_gemma_lm.generate(
+      {
+        "images": image,
+        "text": ["answer en where is the cow standing?\\n"]
+      }
+    )
+
+    # Generate with batched prompts.
+    pali_gemma_lm.generate(
+      {
+        "images": [image, image],
+        "text": ["answer en where is the cow standing?\\n", "caption en\\n"]
+      }
+    )
+    ```
+
+    Use `generate()` without preprocessing.
+    ```python
+    image = np.random.rand(224, 224, 3)
+    inputs = {
+        "images": [image, image],
+        # Token ids for "<bos> Keras is".
+        "token_ids": np.array([[2, 214064, 603, 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),
+    }
+
+    pali_gemma_lm = keras_hub.models.PaliGemmaCausalLM.from_preset(
+        "pali_gemma_3b_mix_224",
+        preprocessor=None,
+    )
+    pali_gemma_lm.generate(inputs)
+    ```
+
+    Custom backbone and vocabulary.
+    ```python
+    tokenizer = keras_hub.models.PaliGemmaTokenizer(
+        proto="proto.spm",
+    )
+    preprocessor = keras_hub.models.PaliGemmaCausalLMPreprocessor(
+        tokenizer=tokenizer,
+        sequence_length=128,
+    )
+    backbone = keras_hub.models.PaliGemmaBackbone()
+    pali_gemma_lm = keras_hub.models.PaliGemmaCausalLM(
+        backbone=backbone,
+        preprocessor=preprocessor,
+    )
+    ```
+    """
+
+    backbone_cls = PaliGemmaBackbone
+    preprocessor_cls = PaliGemmaCausalLMPreprocessor
+
+    def __init__(
+        self,
+        preprocessor,
+        backbone,
+        **kwargs,
+    ):
+        # === Layers ===
+        self.preprocessor = preprocessor
+        self.backbone = backbone
+
+        # === Functional Model ===
+        inputs = backbone.inputs
+        hidden_state = backbone(inputs=inputs)
+        outputs = backbone.token_embedding(hidden_state, reverse=True)
+        outputs = outputs[:, backbone.image_sequence_length :, :]
+        super().__init__(
+            inputs=inputs,
+            outputs=outputs,
+            **kwargs,
+        )
+
+    def compile(
+        self,
+        optimizer="auto",
+        loss="auto",
+        *,
+        weighted_metrics="auto",
+        sampler="greedy",
+        **kwargs,
+    ):
+        super().compile(
+            optimizer=optimizer,
+            loss=loss,
+            weighted_metrics=weighted_metrics,
+            sampler=sampler,
+            **kwargs,
+        )
+
+    def call_with_cache(
+        self,
+        token_ids,
+        cache,
+        cache_update_index,
+        img_embeddings=None,
+        padding_mask=None,
+    ):
+        """Forward pass of `PaliGemmaCausalLM` 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.
+            img_embeddings: a dense float Tensor with shape
+                `(batch_size, image_sequence_length, hidden_dim)`.
+            padding_mask: a dense int Tensor with shape
+                `(batch_size, max_length)`.
+
+        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.
+        """
+        text_embeddings = self.backbone.token_embedding(token_ids)
+        text_embeddings = text_embeddings * ops.cast(
+            ops.sqrt(self.backbone.hidden_dim), text_embeddings.dtype
+        )
+
+        if img_embeddings is not None:
+            x = ops.concatenate((img_embeddings, text_embeddings), axis=1)
+        else:
+            x = text_embeddings
+
+        # Each decoder layer has a cache; we update them separately.
+        caches = []
+        for i, transformer_layer in enumerate(self.backbone.transformer_layers):
+            current_cache = cache[:, i, ...]
+            x, next_cache = transformer_layer(
+                x,
+                cache=current_cache,
+                cache_update_index=cache_update_index,
+                padding_mask=padding_mask,
+            )
+            caches.append(next_cache)
+        cache = ops.stack(caches, 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, img_embeddings, padding_mask):
+        """Build an empty cache for use with `call_with_cache()`."""
+        batch_size = ops.shape(token_ids)[0]
+        max_length = (
+            ops.shape(token_ids)[1] + self.backbone.image_sequence_length
+        )
+        num_layers = self.backbone.num_layers
+        num_heads = self.backbone.num_key_value_heads
+        head_dim = self.backbone.head_dim
+        shape = [batch_size, num_layers, 2, max_length, num_heads, head_dim]
+        cache = ops.zeros(shape, dtype=self.compute_dtype)
+        # Seed the cache.
+        logits, hidden_states, cache = self.call_with_cache(
+            token_ids=token_ids,
+            img_embeddings=img_embeddings,
+            cache=cache,
+            cache_update_index=0,
+            padding_mask=padding_mask,
+        )
+        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 end token's to stop on. If all
+                sequences have produced a new stop token, generation
+                will stop.
+        """
+        token_ids, padding_mask, images = (
+            inputs["token_ids"],
+            inputs["padding_mask"],
+            inputs["images"],
+        )
+        if len(ops.shape(images)) == 3:
+            # Handle an unbatched image. Unlike `token_ids` and `padding_mask`
+            # this will not automatically be upranked.
+            images = ops.expand_dims(images, axis=0)
+        img_embeddings = self.backbone.vit_encoder(images)
+
+        # Create and seed cache with a single forward pass.
+        hidden_states, cache = self._build_cache(
+            token_ids, img_embeddings, padding_mask
+        )
+        # 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 + self.backbone.image_sequence_length
+            batch_size = ops.shape(prompt)[0]
+            prompt = ops.slice(prompt, [0, index - 1], [batch_size, 1])
+            logits, hidden_states, cache = self.call_with_cache(
+                token_ids=prompt,
+                cache=cache,
+                cache_update_index=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,
+        )
+
+        # Compute an output padding mask with the token ids we updated.
+        if stop_token_ids is not None:
+            # Build a mask of `stop_token_ids` 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,
+            "images": images,
+        }

keras_hub/src/models/pali_gemma/pali_gemma_causal_lm_preprocessor.py

@@ -0,0 +1,147 @@
+# Copyright 2024 The KerasHub Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import keras
+from absl import logging
+from keras import ops
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.layers.preprocessing.multi_segment_packer import (
+    MultiSegmentPacker,
+)
+from keras_hub.src.models.gemma.gemma_causal_lm_preprocessor import (
+    GemmaCausalLMPreprocessor,
+)
+from keras_hub.src.models.pali_gemma.pali_gemma_tokenizer import (
+    PaliGemmaTokenizer,
+)
+from keras_hub.src.utils.keras_utils import (
+    convert_inputs_to_list_of_tensor_segments,
+)
+
+
+@keras_hub_export("keras_hub.models.PaliGemmaCausalLMPreprocessor")
+class PaliGemmaCausalLMPreprocessor(GemmaCausalLMPreprocessor):
+    tokenizer_cls = PaliGemmaTokenizer
+
+    def __init__(
+        self,
+        tokenizer,
+        sequence_length=512,
+        add_start_token=True,
+        add_end_token=True,
+        **kwargs,
+    ):
+        super().__init__(
+            tokenizer, sequence_length, add_start_token, add_end_token, **kwargs
+        )
+
+    def build(self, input_shape):
+        # Defer packer creation to `build()` so that we can be sure tokenizer
+        # assets have loaded when restoring a saved model.
+        self.packer = MultiSegmentPacker(
+            start_value=self.tokenizer.start_token_id,
+            end_value=self.tokenizer.end_token_id,
+            pad_value=self.tokenizer.pad_token_id,
+            sep_value=[],
+            sequence_length=self.sequence_length,
+        )
+        self.built = True
+
+    def call(
+        self,
+        x,
+        y=None,
+        sample_weight=None,
+        sequence_length=None,
+    ):
+        if y is not None or sample_weight is not None:
+            logging.warning(
+                "`PaliGemmaCausalLMPreprocessor` generates `y` and `sample_weight` "
+                "based on your input data, but your data already contains `y` "
+                "or `sample_weight`. Your `y` and `sample_weight` will be "
+                "ignored."
+            )
+        sequence_length = sequence_length or self.sequence_length
+
+        images, prompts, responses = x["images"], x["prompts"], x["responses"]
+        if keras.config.backend() == "tensorflow":
+            # Tensorflow backend needs uniform ouput types.
+            images = ops.convert_to_tensor(images)
+        prompts = convert_inputs_to_list_of_tensor_segments(prompts)[0]
+        prompts = self.tokenizer(prompts)
+        responses = convert_inputs_to_list_of_tensor_segments(responses)[0]
+        responses = self.tokenizer(responses)
+        # Pad with one extra token to account for the truncation below.
+        token_ids, segment_ids = self.packer(
+            (prompts, responses),
+            sequence_length=sequence_length + 1,
+            add_start_value=self.add_start_token,
+            add_end_value=self.add_end_token,
+        )
+        padding_mask = token_ids != self.tokenizer.pad_token_id
+        response_mask = segment_ids == 1
+        # The last token does not have a next token, so we truncate it out.
+        x = {
+            "token_ids": token_ids[..., :-1],
+            "response_mask": response_mask[..., :-1],
+            "padding_mask": padding_mask[..., :-1],
+            "images": images,
+        }
+        # Target `y` will be the next token.
+        y = token_ids[..., 1:]
+        # Only compute the loss for labels in the response.
+        sample_weight = response_mask[..., 1:]
+        return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)
+
+    def generate_preprocess(
+        self,
+        x,
+        sequence_length=None,
+    ):
+        """Convert strings to integer token input for generation.
+
+        Similar to calling the layer for training, this method takes in strings
+        or tensor strings, tokenizes and packs the input, and computes a padding
+        mask masking all inputs not filled in with a padded value.
+
+        Unlike calling the layer for training, this method does not compute
+        labels and will never append a `tokenizer.end_token_id` to the end of
+        the sequence (as generation is expected to continue at the end of the
+        inputted prompt).
+        """
+        if not self.built:
+            self.build(None)
+        sequence_length = sequence_length or self.sequence_length
+
+        images, prompts = x["images"], x["prompts"]
+        prompts = convert_inputs_to_list_of_tensor_segments(prompts)[0]
+        prompts = self.tokenizer(prompts)
+        segments = [prompts]
+        if "responses" in x:
+            responses = x["responses"]
+            responses = convert_inputs_to_list_of_tensor_segments(responses)[0]
+            segments.append(self.tokenizer(responses))
+        token_ids, segment_ids = self.packer(
+            segments,
+            sequence_length=sequence_length,
+            add_end_value=False,
+        )
+        padding_mask = token_ids != self.tokenizer.pad_token_id
+        response_mask = segment_ids == 1
+        return {
+            "images": images,
+            "token_ids": token_ids,
+            "response_mask": response_mask,
+            "padding_mask": padding_mask,
+        }

keras_hub/src/models/pali_gemma/pali_gemma_decoder_block.py

@@ -0,0 +1,160 @@
+# Copyright 2024 The KerasHub Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import keras
+from keras import ops
+
+from keras_hub.src.layers.modeling.transformer_layer_utils import (
+    compute_causal_mask,
+)
+from keras_hub.src.models.gemma.gemma_decoder_block import GemmaDecoderBlock
+
+
+class PaliGemmaDecoderBlock(GemmaDecoderBlock):
+    """PaliGemma mixed decoder block.
+
+    This class implements a decoder block of the PaliGemma Architecture: a
+    mixed transformer decoder block. Intended to be used with an input
+    sequence comprised of both embedded image and text data, this block
+    functions largely identically to the `GemmaDecoderBlock` class, with a
+    notable exception in the computation of attention masks.
+
+    Specifically, this decoder block will use causal self-attention on the
+    text portion of the input, while using full self-attention for image
+    data. It is expected that any image data occurs before text data in the
+    input.
+
+    Args:
+        hidden_dim: int. The size of the transformer hidden state at the end
+            of the block.
+        intermediate_dim: int. The output dimension of the first Dense layer in
+            the two-layer feedforward network.
+        head_dim: int. The size of each attention head.
+        num_query_heads: int. The number of heads for the query projections in
+            the attention layer.
+        num_key_value_heads: int. The number of heads for the key and value
+            projections in the attention layer.
+        layer_norm_epsilon: float. The epsilon hyperparameter used for layer
+            normalization.
+        dropout: float. The dropout rate for the transformer attention layer.
+    """
+
+    def __init__(
+        self,
+        hidden_dim,
+        intermediate_dim,
+        head_dim,
+        num_query_heads,
+        num_key_value_heads,
+        layer_norm_epsilon=1e-6,
+        dropout=0,
+        **kwargs,
+    ):
+        super().__init__(
+            hidden_dim=hidden_dim,
+            intermediate_dim=intermediate_dim,
+            head_dim=head_dim,
+            num_query_heads=num_query_heads,
+            num_key_value_heads=num_key_value_heads,
+            layer_norm_epsilon=layer_norm_epsilon,
+            dropout=dropout,
+            **kwargs,
+        )
+
+    def call(
+        self,
+        x,
+        padding_mask=None,
+        response_mask=None,
+        cache=None,
+        cache_update_index=0,
+    ):
+        normalized_x = self.pre_attention_norm(x)
+        attention_mask = self._compute_attention_mask(
+            normalized_x, padding_mask, cache, cache_update_index, response_mask
+        )
+        if cache is not None:
+            attention, new_cache = self.attention(
+                normalized_x,
+                attention_mask=attention_mask,
+                cache=cache,
+                cache_update_index=cache_update_index,
+            )
+        else:
+            attention = self.attention(
+                normalized_x,
+                attention_mask=attention_mask,
+            )
+
+        if self.dropout:
+            attention = self.attention_dropout(attention)
+
+        attention_x = x + attention
+        normalized_x = self.pre_ffw_norm(attention_x)
+
+        x1 = self.gating_ffw(normalized_x)
+        x2 = self.gating_ffw_2(normalized_x)
+        x = keras.activations.gelu(x1, approximate=True) * x2
+        x = self.ffw_linear(x)
+
+        x = x + attention_x
+
+        if cache is not None:
+            return x, new_cache
+        return x
+
+    def _compute_attention_mask(
+        self,
+        x,
+        padding_mask,
+        cache,
+        cache_update_index,
+        response_mask=None,
+    ):
+        batch_size = ops.shape(x)[0]
+        input_length = output_length = ops.shape(x)[1]
+        if cache is not None:
+            input_length = ops.shape(cache)[2]
+
+        causal_mask = compute_causal_mask(
+            batch_size=batch_size,
+            input_length=input_length,
+            output_length=output_length,
+            cache_index=cache_update_index,
+        )
+
+        if padding_mask is None:
+            # We should only hit this case during generative decoding.
+            # Just the causal mask is fine in this case.
+            return causal_mask
+
+        def token_to_attention_mask(mask, fill_value):
+            """Reshape token mask -> attention mask padding for image tokens."""
+            mask = ops.cast(mask, "int32")
+            pad = input_length - ops.shape(mask)[1]
+            mask = ops.pad(mask, ((0, 0), (pad, 0)), constant_values=fill_value)
+            return ops.expand_dims(mask, axis=1)
+
+        padding_mask = token_to_attention_mask(padding_mask, 1)
+        if response_mask is not None:
+            response_mask = token_to_attention_mask(response_mask, 0)
+            not_response_mask = ops.logical_not(response_mask)
+            # Only apply the causal mask to the response tokens.
+            causal_mask = ops.logical_and(causal_mask, response_mask)
+            # Only apply block attention to the non-response tokens.
+            padding_mask = ops.logical_and(padding_mask, not_response_mask)
+
+        # Use block attention for the padding mask,
+        # which marks all image and prompt tokens.
+        return ops.logical_or(padding_mask, causal_mask)

keras_hub/src/models/pali_gemma/pali_gemma_presets.py

@@ -0,0 +1,78 @@
+# Copyright 2024 The KerasHub Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PaliGemma model preset configurations."""
+
+# Metadata for loading pretrained model weights.
+backbone_presets = {
+    "pali_gemma_3b_mix_224": {
+        "metadata": {
+            "description": (
+                "image size 224, mix fine tuned, text sequence " "length is 256"
+            ),
+            "params": 2923335408,
+            "official_name": "PaliGemma",
+            "path": "pali_gemma",
+            "model_card": "https://www.kaggle.com/models/google/paligemma",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_mix_224/1",
+    },
+    "pali_gemma_3b_mix_448": {
+        "metadata": {
+            "description": (
+                "image size 448, mix fine tuned, text sequence length is 512"
+            ),
+            "params": 2924220144,
+            "official_name": "PaliGemma",
+            "path": "pali_gemma",
+            "model_card": "https://www.kaggle.com/models/google/paligemma",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_mix_448/1",
+    },
+    "pali_gemma_3b_224": {
+        "metadata": {
+            "description": (
+                "image size 224, pre trained, text sequence length is 128"
+            ),
+            "params": 2923335408,
+            "official_name": "PaliGemma",
+            "path": "pali_gemma",
+            "model_card": "https://www.kaggle.com/models/google/paligemma",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_224/1",
+    },
+    "pali_gemma_3b_448": {
+        "metadata": {
+            "description": (
+                "image size 448, pre trained, text sequence length is 512"
+            ),
+            "params": 2924220144,
+            "official_name": "PaliGemma",
+            "path": "pali_gemma",
+            "model_card": "https://www.kaggle.com/models/google/paligemma",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_448/1",
+    },
+    "pali_gemma_3b_896": {
+        "metadata": {
+            "description": (
+                "image size 896, pre trained, text sequence length " "is 512"
+            ),
+            "params": 2927759088,
+            "official_name": "PaliGemma",
+            "path": "pali_gemma",
+            "model_card": "https://www.kaggle.com/models/google/paligemma",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_896/1",
+    },
+}