Add PaliGemma

docs/source/en/_toctree.yml

@@ -784,6 +784,8 @@
         title: OWL-ViT
       - local: model_doc/owlv2
         title: OWLv2
+      - local: model_doc/paligemma
+        title: PaliGemma
       - local: model_doc/perceiver
         title: Perceiver
       - local: model_doc/pix2struct

docs/source/en/index.md

@@ -230,6 +230,7 @@ Flax), PyTorch, and/or TensorFlow.
 |                           [OPT](model_doc/opt)                           |       ✅        |         ✅         |      ✅      |
 |                       [OWL-ViT](model_doc/owlvit)                        |       ✅        |         ❌         |      ❌      |
 |                         [OWLv2](model_doc/owlv2)                         |       ✅        |         ❌         |      ❌      |
+|                     [PaliGemma](model_doc/paligemma)                     |       ✅        |         ❌         |      ❌      |
 |                  [PatchTSMixer](model_doc/patchtsmixer)                  |       ✅        |         ❌         |      ❌      |
 |                      [PatchTST](model_doc/patchtst)                      |       ✅        |         ❌         |      ❌      |
 |                       [Pegasus](model_doc/pegasus)                       |       ✅        |         ✅         |      ✅      |

docs/source/en/model_doc/paligemma.md

@@ -0,0 +1,38 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+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
+
+http://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.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# PaliGemma
+
+## Overview
+
+The PaliGemma model was proposed by Google. It is a 3B VLM composed by a Siglip-400m vision encoder and a Gemma-2B decoder linked by a multimodal linear projection. It is not a chat model with images. It cuts an image into a fixed number of VIT tokens and prepends it to an optional prompt. One particularity is that the model uses full block attention on all the image tokens plus the input text tokens. It comes in 3 resolutions, 224x224, 448x448 and 896x896 with 3 base models, with 55 fine-tuned versions for different tasks, and 2 mix models.
+
+
+This model was contributed by [Molbap](https://huggingface.co/Molbap).
+
+
+## PaliGemmaConfig
+
+[[autodoc]] PaliGemmaConfig
+
+## PaliGemmaProcessor
+
+[[autodoc]] PaliGemmaProcessor
+
+## PaliGemmaForConditionalGeneration
+
+[[autodoc]] PaliGemmaForConditionalGeneration
+    - forward

docs/source/en/perf_infer_gpu_one.md

@@ -203,6 +203,7 @@ For now, Transformers supports SDPA inference and training for the following arc
 * [Jamba](https://huggingface.co/docs/transformers/model_doc/jamba#transformers.JambaModel)
 * [Llama](https://huggingface.co/docs/transformers/model_doc/llama#transformers.LlamaModel)
 * [OLMo](https://huggingface.co/docs/transformers/model_doc/olmo#transformers.OlmoModel)
+* [PaliGemma](https://huggingface.co/docs/transformers/model_doc/paligemma#transformers.PaliGemmaForConditionalGeneration)
 * [Phi](https://huggingface.co/docs/transformers/model_doc/phi#transformers.PhiModel)
 * [Idefics](https://huggingface.co/docs/transformers/model_doc/idefics#transformers.IdeficsModel)
 * [Whisper](https://huggingface.co/docs/transformers/model_doc/whisper#transformers.WhisperModel)

src/transformers/__init__.py

@@ -582,6 +582,7 @@
         "OwlViTTextConfig",
         "OwlViTVisionConfig",
     ],
+    "models.paligemma": ["PaliGemmaConfig"],
     "models.patchtsmixer": ["PatchTSMixerConfig"],
     "models.patchtst": ["PatchTSTConfig"],
     "models.pegasus": [
@@ -2651,6 +2652,13 @@
             "OwlViTVisionModel",
         ]
     )
+    _import_structure["models.paligemma"].extend(
+        [
+            "PaliGemmaForConditionalGeneration",
+            "PaliGemmaPreTrainedModel",
+            "PaliGemmaProcessor",
+        ]
+    )
     _import_structure["models.patchtsmixer"].extend(
         [
             "PatchTSMixerForPrediction",
@@ -5126,6 +5134,9 @@
         OwlViTTextConfig,
         OwlViTVisionConfig,
     )
+    from .models.paligemma import (
+        PaliGemmaConfig,
+    )
     from .models.patchtsmixer import (
         PatchTSMixerConfig,
     )
@@ -6956,6 +6967,11 @@
             OwlViTTextModel,
             OwlViTVisionModel,
         )
+        from .models.paligemma import (
+            PaliGemmaForConditionalGeneration,
+            PaliGemmaPreTrainedModel,
+            PaliGemmaProcessor,
+        )
         from .models.patchtsmixer import (
             PatchTSMixerForPrediction,
             PatchTSMixerForPretraining,

src/transformers/models/__init__.py

@@ -173,6 +173,7 @@
     opt,
     owlv2,
     owlvit,
+    paligemma,
     patchtsmixer,
     patchtst,
     pegasus,

src/transformers/models/auto/configuration_auto.py

@@ -182,6 +182,7 @@
         ("opt", "OPTConfig"),
         ("owlv2", "Owlv2Config"),
         ("owlvit", "OwlViTConfig"),
+        ("paligemma", "PaliGemmaConfig"),
         ("patchtsmixer", "PatchTSMixerConfig"),
         ("patchtst", "PatchTSTConfig"),
         ("pegasus", "PegasusConfig"),
@@ -464,6 +465,7 @@
         ("opt", "OPT"),
         ("owlv2", "OWLv2"),
         ("owlvit", "OWL-ViT"),
+        ("paligemma", "PaliGemma"),
         ("patchtsmixer", "PatchTSMixer"),
         ("patchtst", "PatchTST"),
         ("pegasus", "Pegasus"),

src/transformers/models/auto/image_processing_auto.py

@@ -93,6 +93,7 @@
         ("oneformer", "OneFormerImageProcessor"),
         ("owlv2", "Owlv2ImageProcessor"),
         ("owlvit", "OwlViTImageProcessor"),
+        ("paligemma", "CLIPImageProcessor"),
         ("perceiver", "PerceiverImageProcessor"),
         ("pix2struct", "Pix2StructImageProcessor"),
         ("poolformer", "PoolFormerImageProcessor"),

src/transformers/models/auto/modeling_auto.py

@@ -313,6 +313,7 @@
         ("nezha", "NezhaForPreTraining"),
         ("nllb-moe", "NllbMoeForConditionalGeneration"),
         ("openai-gpt", "OpenAIGPTLMHeadModel"),
+        ("paligemma", "PaliGemmaForConditionalGeneration"),
         ("retribert", "RetriBertModel"),
         ("roberta", "RobertaForMaskedLM"),
         ("roberta-prelayernorm", "RobertaPreLayerNormForMaskedLM"),
@@ -697,6 +698,7 @@
         ("kosmos-2", "Kosmos2ForConditionalGeneration"),
         ("llava", "LlavaForConditionalGeneration"),
         ("llava_next", "LlavaNextForConditionalGeneration"),
+        ("paligemma", "PaliGemmaForConditionalGeneration"),
         ("pix2struct", "Pix2StructForConditionalGeneration"),
         ("vipllava", "VipLlavaForConditionalGeneration"),
         ("vision-encoder-decoder", "VisionEncoderDecoderModel"),

src/transformers/models/auto/processing_auto.py

@@ -74,6 +74,7 @@
         ("oneformer", "OneFormerProcessor"),
         ("owlv2", "Owlv2Processor"),
         ("owlvit", "OwlViTProcessor"),
+        ("paligemma", "PaliGemmaProcessor"),
         ("pix2struct", "Pix2StructProcessor"),
         ("pop2piano", "Pop2PianoProcessor"),
         ("sam", "SamProcessor"),

src/transformers/models/auto/tokenization_auto.py

@@ -331,6 +331,7 @@
             ("opt", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
             ("owlv2", ("CLIPTokenizer", "CLIPTokenizerFast" if is_tokenizers_available() else None)),
             ("owlvit", ("CLIPTokenizer", "CLIPTokenizerFast" if is_tokenizers_available() else None)),
+            ("paligemma", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)),
             (
                 "pegasus",
                 (

src/transformers/models/gemma/modeling_gemma.py

@@ -794,7 +794,6 @@ def _init_weights(self, module):
     "The bare Gemma Model outputting raw hidden-states without any specific head on top.",
     GEMMA_START_DOCSTRING,
 )
-# Copied from transformers.models.llama.modeling_llama.LlamaModel with LLAMA->GEMMA,Llama->Gemma
 class GemmaModel(GemmaPreTrainedModel):
     """
     Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`GemmaDecoderLayer`]
@@ -988,8 +987,6 @@ def _update_causal_mask(
 
         if attention_mask is not None and attention_mask.dim() == 4:
             # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            if attention_mask.max() != 0:
-                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
             causal_mask = attention_mask
         else:
             causal_mask = torch.full(

src/transformers/models/paligemma/__init__.py

@@ -0,0 +1,54 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# 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
+#
+#     http://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.
+from typing import TYPE_CHECKING
+
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
+
+
+_import_structure = {"configuration_paligemma": ["PaliGemmaConfig"]}
+
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_paligemma"] = [
+        "PaliGemmaForConditionalGeneration",
+        "PaliGemmaPreTrainedModel",
+    ]
+    _import_structure["processing_paligemma"] = ["PaliGemmaProcessor"]
+
+
+if TYPE_CHECKING:
+    from .configuration_paligemma import PaliGemmaConfig
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_paligemma import (
+            PaliGemmaForConditionalGeneration,
+            PaliGemmaPreTrainedModel,
+        )
+        from .processing_paligemma import PaliGemmaProcessor
+
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure)

src/transformers/models/paligemma/configuration_paligemma.py

@@ -0,0 +1,130 @@
+# coding=utf-8
+# Copyright 2024 Microsoft Research & University of Wisconsin-Madison and the HuggingFace Inc. team. All rights reserved.
+# 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
+#
+#     http://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.
+""" PaliGemmamodel configuration"""
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+from ..auto import CONFIG_MAPPING
+
+
+logger = logging.get_logger(__name__)
+
+
+class PaliGemmaConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`PaliGemmaForConditionalGeneration`]. It is used to instantiate an
+    PaliGemmamodel according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the PaliGemma-2B.
+
+    e.g. [paligemma-hf/paligemma-2b](https://huggingface.co/paligemma-hf/paligemma-2b)
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vision_config (`PaliGemmaVisionConfig`,  *optional*):
+            Custom vision config or dict
+        text_config (`Union[AutoConfig, dict]`, *optional*):
+            The config object of the text backbone. Can be any of `LlamaConfig` or `MistralConfig`.
+        ignore_index (`int`, *optional*, defaults to -100):
+            The ignore index for the loss function.
+        image_token_index (`int`, *optional*, defaults to 256000):
+            The image token index to encode the image prompt.
+        vocab_size (`int`, *optional*, defaults to 257152):
+            Vocabulary size of the PaliGemmamodel. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`~PaliGemmaForConditionalGeneration`]
+        projection_dim (`int`, *optional*, defaults to 2048):
+            Dimension of the multimodal projection space.
+        hidden_size (`int`, *optional*, defaults to 2048):
+            Dimension of the hidden layer of the Language model.
+
+    Example:
+
+    ```python
+    >>> from transformers import PaliGemmaForConditionalGeneration, PaliGemmaConfig, SiglipVisionConfig, GemmaConfig
+
+    >>> # Initializing a Siglip-like vision config
+    >>> vision_config = SiglipVisionConfig()
+
+    >>> # Initializing a PaliGemma config
+    >>> text_config = GemmaConfig()
+
+    >>> # Initializing a PaliGemma paligemma-3b-224 style configuration
+    >>> configuration = PaliGemmaConfig(vision_config, text_config)
+
+    >>> # Initializing a model from the paligemma-3b-224 style configuration
+    >>> model = PaliGemmaForConditionalGeneration(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "paligemma"
+    is_composition = False
+
+    def __init__(
+        self,
+        vision_config=None,
+        text_config=None,
+        ignore_index=-100,
+        image_token_index=256000,
+        vocab_size=257152,
+        projection_dim=2048,
+        hidden_size=2048,
+        **kwargs,
+    ):
+        self.ignore_index = ignore_index
+        self.image_token_index = image_token_index
+        self.vocab_size = vocab_size
+        self.projection_dim = projection_dim
+        self.hidden_size = hidden_size
+        self.vision_config = vision_config
+        self.is_encoder_decoder = False
+
+        if isinstance(self.vision_config, dict):
+            vision_config["model_type"] = (
+                vision_config["model_type"] if "model_type" in vision_config else "siglip_vision_model"
+            )
+            self.vision_config = CONFIG_MAPPING[vision_config["model_type"]](**vision_config)
+        elif vision_config is None:
+            self.vision_config = CONFIG_MAPPING["siglip_vision_model"](
+                intermediate_size=4096,
+                hidden_size=1152,
+                patch_size=14,
+                image_size=224,
+                num_hidden_layers=27,
+                num_attention_heads=16,
+                vocab_size=257152,
+                vision_use_head=False,
+            )
+        self.vocab_size = self.vocab_size
+
+        self.text_config = text_config
+
+        if isinstance(self.text_config, dict):
+            text_config["model_type"] = text_config["model_type"] if "model_type" in text_config else "gemma"
+            self.text_config = CONFIG_MAPPING[text_config["model_type"]](**text_config)
+            self.vocab_size = self.text_config.vocab_size
+        elif text_config is None:
+            self.text_config = CONFIG_MAPPING["gemma"](
+                hidden_size=2048,
+                num_hidden_layers=18,
+                intermediate_size=16384,
+                num_attention_heads=8,
+                num_key_value_heads=1,
+                is_encoder_decoder=False,
+            )
+        self.text_config.num_image_tokens = (self.vision_config.image_size // self.vision_config.patch_size) ** 2
+        self.vision_config.projection_dim = projection_dim
+        super().__init__(**kwargs)