Add Aria (#34157)

* Add Aria
---------

Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

src/transformers/models/aria/__init__.py

@@ -0,0 +1,30 @@
+# 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 _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_aria import *
+    from .image_processing_aria import *
+    from .modeling_aria import *
+    from .processing_aria import *
+
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/aria/configuration_aria.py

@@ -0,0 +1,299 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/aria/modular_aria.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_aria.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2024 The Rhymes-AI Teams Authors 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.
+from typing import Dict
+
+from ...configuration_utils import PretrainedConfig
+from ...modeling_rope_utils import rope_config_validation
+from ..auto import CONFIG_MAPPING, AutoConfig
+
+
+class AriaTextConfig(PretrainedConfig):
+    r"""
+    This class handles the configuration for the text component of the Aria model.
+    Instantiating a configuration with the defaults will yield a similar configuration to that of the model of the Aria
+    [rhymes-ai/Aria](https://huggingface.co/rhymes-ai/Aria) architecture.
+    This class extends the LlamaConfig to include additional parameters specific to the Mixture of Experts (MoE) architecture.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 32000):
+            Vocabulary size of the LLaMA model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`LlamaModel`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 4096):
+            The size of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer decoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        num_key_value_heads (`int`, *optional*):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
+            `num_attention_heads`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 2048):
+            The maximum sequence length that this model might ever be used with. Llama 1 supports up to 2048 tokens,
+            Llama 2 up to 4096, CodeLlama up to 16384.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        pad_token_id (`int`, *optional*, defaults to 2):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 2):
+            End of stream token id.
+        pretraining_tp (`int`, *optional*, defaults to 1):
+            Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
+            document](https://huggingface.co/docs/transformers/main/perf_train_gpu_many#tensor-parallelism) to
+            understand more about it. This value is necessary to ensure exact reproducibility of the pretraining
+            results. Please refer to [this issue](https://github.com/pytorch/pytorch/issues/76232).
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+            accordingly.
+            Expected contents:
+                `rope_type` (`str`):
+                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+                    'llama3'], with 'default' being the original RoPE implementation.
+                `factor` (`float`, *optional*):
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (`float`, *optional*):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation. If unspecified, it defaults to value recommended by the implementation, using the
+                    `factor` field to infer the suggested value.
+                `beta_fast` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `long_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+        attention_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        mlp_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use a bias in up_proj, down_proj and gate_proj layers in the MLP layers.
+        head_dim (`int`, *optional*):
+            The attention head dimension. If None, it will default to hidden_size // num_heads
+        moe_num_experts (`int`, *optional*, defaults to 8):
+            The number of experts in the MoE layer.
+        moe_topk (`int`, *optional*, defaults to 2):
+            The number of top experts to route to for each token.
+        moe_num_shared_experts (`int`, *optional*, defaults to 2):
+            The number of shared experts.
+    """
+
+    model_type = "aria_text"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    # Default tensor parallel plan for base model `AriaTextModel`
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+    base_config_key = "text_config"
+
+    def __init__(
+        self,
+        vocab_size=32000,
+        hidden_size=4096,
+        intermediate_size: int = 4096,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=None,
+        hidden_act="silu",
+        max_position_embeddings=2048,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=2,
+        bos_token_id=1,
+        eos_token_id=2,
+        pretraining_tp=1,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        attention_dropout=0.0,
+        mlp_bias=False,
+        head_dim=None,
+        moe_num_experts: int = 8,
+        moe_topk: int = 2,
+        moe_num_shared_experts: int = 2,
+        **kwargs,
+    ):
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.pretraining_tp = pretraining_tp
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.mlp_bias = mlp_bias
+        self.head_dim = head_dim if head_dim is not None else self.hidden_size // self.num_attention_heads
+        # Validate the correctness of rotary position embeddings parameters
+        # BC: if there is a 'type' field, copy it it to 'rope_type'.
+        if self.rope_scaling is not None and "type" in self.rope_scaling:
+            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
+        rope_config_validation(self)
+        self.moe_num_experts = moe_num_experts
+        self.moe_topk = moe_topk
+        self.moe_num_shared_experts = moe_num_shared_experts
+
+
+class AriaConfig(PretrainedConfig):
+    r"""
+    This class handles the configuration for both vision and text components of the Aria model,
+    as well as additional parameters for image token handling and projector mapping.
+    Instantiating a configuration with the defaults will yield a similar configuration to that of the model of the Aria
+    [rhymes-ai/Aria](https://huggingface.co/rhymes-ai/Aria) architecture.
+
+    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 (`AriaVisionConfig` or `dict`, *optional*):
+            Configuration for the vision component.
+        vision_feature_layer (`int`, *optional*, defaults to -1):
+            The index of the layer to select the vision feature.
+        text_config (`AriaTextConfig` or `dict`, *optional*):
+            Configuration for the text component.
+        projector_patch_to_query_dict (`dict`, *optional*):
+            Mapping of patch sizes to query dimensions.
+        image_token_index (`int`, *optional*, defaults to 9):
+            Index used to represent image tokens.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated normal initializer for initializing all weight matrices.
+
+    Attributes:
+        model_type (`str`):
+            Type of the model, set to `"aria"`.
+        image_token_index (`int`):
+            Index used to represent image tokens.
+        projector_patch_to_query_dict (`dict`):
+            Mapping of patch sizes to query dimensions.
+        vision_config (`AriaVisionConfig`):
+            Configuration for the vision component.
+        text_config (`AriaTextConfig`):
+            Configuration for the text component.
+    """
+
+    model_type = "aria"
+    sub_configs = {"text_config": AriaTextConfig, "vision_config": AutoConfig}
+
+    def __init__(
+        self,
+        vision_config=None,
+        vision_feature_layer: int = -1,
+        text_config: AriaTextConfig = None,
+        projector_patch_to_query_dict: Dict = None,
+        image_token_index: int = 9,
+        initializer_range: float = 0.02,
+        **kwargs,
+    ):
+        self.image_token_index = image_token_index
+
+        # Convert the keys and values of projector_patch_to_query_dict to integers
+        # This ensures consistency even if they were provided as strings
+        if projector_patch_to_query_dict is None:
+            projector_patch_to_query_dict = {
+                1225: 128,
+                4900: 256,
+            }
+        self.projector_patch_to_query_dict = {int(k): int(v) for k, v in projector_patch_to_query_dict.items()}
+        self.max_value_projector_patch_to_query_dict = max(self.projector_patch_to_query_dict.values())
+        self.vision_feature_layer = vision_feature_layer
+        if isinstance(vision_config, dict):
+            vision_config["model_type"] = "idefics3_vision"
+            vision_config = CONFIG_MAPPING[vision_config["model_type"]](**vision_config)
+        elif vision_config is None:
+            vision_config = CONFIG_MAPPING["idefics3_vision"]()
+
+        self.vision_config = vision_config
+        self.initializer_range = initializer_range
+
+        if isinstance(text_config, dict) and "model_type" in text_config:
+            text_config = AriaTextConfig(**text_config)
+        elif text_config is None:
+            text_config = AriaTextConfig()
+
+        self.text_config = text_config
+
+        super().__init__(**kwargs)
+
+
+__all__ = ["AriaConfig", "AriaTextConfig"]

src/transformers/models/aria/convert_aria_weights_to_hf.py

@@ -0,0 +1,162 @@
+# Copyright 2024 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.
+import argparse
+import glob
+
+import torch
+from huggingface_hub import snapshot_download
+from safetensors import safe_open
+
+from transformers import (
+    AddedToken,
+    AriaForConditionalGeneration,
+    AriaProcessor,
+    AutoConfig,
+    AutoTokenizer,
+)
+
+
+EPILOG_TXT = """Example:
+    python transformers/src/transformers/models/aria/convert_aria_weights_to_hf.py --text_model_id rhymes-ai/Aria --vision_model_id rhymes-ai/Aria --output_hub_path m-ric/Aria_hf_2 --old_state_dict_id rhymes-ai/Aria
+
+Example for creating the old state dict file with Python:
+
+    import torch
+    from aria.model.language_model.aria_llama import AriaTextForCausalLM
+
+    # load model
+    kwargs = {"device_map": "auto", "torch_dtype": torch.float16}
+    model = AriaTextForCausalLM.from_pretrained("rhymes-ai/Aria", low_cpu_mem_usage=True, **kwargs)
+
+    # load vision tower
+    model.get_vision_tower().load_model()
+
+    # Save state dict
+    torch.save(model.state_dict(), "tmp/hf_models/aria/model_state_dict.bin")
+"""
+
+KEYS_TO_MODIFY_MAPPING = {
+    "vision_tower.vision_model": "vision_tower",
+    "ln_ffn": "layer_norm",
+    "ffn": "feed_forward",
+    "ln_kv": "layer_norm_kv",
+}
+
+
+def load_original_state_dict(model_id):
+    directory_path = snapshot_download(repo_id=model_id, allow_patterns=["*.safetensors"])
+
+    original_state_dict = {}
+    for path in glob.glob(f"{directory_path}/*"):
+        if path.endswith(".safetensors"):
+            with safe_open(path, framework="pt", device="cpu") as f:
+                for key in f.keys():
+                    original_state_dict[key] = f.get_tensor(key)
+
+    return original_state_dict
+
+
+def convert_state_dict_to_hf(state_dict):
+    new_state_dict = {}
+    for key, value in state_dict.items():
+        if key.endswith(".inv_freq"):
+            continue
+        for key_to_modify, new_key in KEYS_TO_MODIFY_MAPPING.items():
+            if key_to_modify in key:
+                key = key.replace(key_to_modify, new_key)
+
+        new_state_dict[key] = value
+    new_state_dict["vision_tower.post_layernorm.weight"] = torch.zeros((1152,))
+    new_state_dict["vision_tower.post_layernorm.bias"] = torch.zeros((1152,))
+
+    return new_state_dict
+
+
+def convert_aria_llama_to_hf(text_model_id, vision_model_id, output_hub_path, old_state_dict_id):
+    torch.set_default_dtype(torch.float16)
+
+    tokenizer = AutoTokenizer.from_pretrained(
+        text_model_id,
+        extra_special_tokens={
+            "image_token": "<|img|>",
+            "pad_token": "<pad>",
+        },
+    )
+    tokenizer.add_tokens(AddedToken("<|img|>", special=True, normalized=False), special_tokens=True)
+    tokenizer.add_special_tokens({"pad_token": "<pad>"})
+    tokenizer.chat_template = "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}<|im_start|>{{ message['role'] }}\n{% if message['content'] is string %}{{ message['content'] }}{% elif message['content'] is iterable %}{% for item in message['content'] %}{% if item['type'] == 'text' %}{{ item['text'] }}{% elif item['type'] == 'image' %}<fim_prefix><|img|><fim_suffix>{% endif %}{% endfor %}{% endif %}<|im_end|>\n{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}"
+
+    processor = AriaProcessor.from_pretrained(
+        text_model_id,
+        tokenizer=tokenizer,
+    )
+
+    config = AutoConfig.from_pretrained(text_model_id)
+    config.vision_config.hidden_size = 1152
+    config.vision_config.attention_heads = 16
+    config.pad_token_id = 2
+    config.image_token_index = 9
+    config.intermediate_size = config.moe_intermediate_size
+    config.auto_map = {
+        "AutoConfig": "modeling_aria.AriaConfig",
+        "AutoModelForCausalLM": "modeling_aria.AriaForConditionalGeneration",
+    }
+
+    with torch.device("meta"):
+        model = AriaForConditionalGeneration(config)
+
+    state_dict = load_original_state_dict(old_state_dict_id)
+
+    state_dict = convert_state_dict_to_hf(state_dict)
+    model.load_state_dict(state_dict, strict=False, assign=True)
+
+    # print("Saving models")
+    # model.save_pretrained("local_aria", safe_serialization=False)
+    # processor.save_pretrained("local_aria")
+    print("Pushing to hub")
+    model.push_to_hub(output_hub_path, create_pr=True)
+    processor.push_to_hub(output_hub_path, create_pr=True)
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        epilog=EPILOG_TXT,
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    parser.add_argument(
+        "--text_model_id",
+        default="rhymes-ai/Aria",
+        help="Hub location of the text model",
+    )
+    parser.add_argument(
+        "--vision_model_id",
+        default="rhymes-ai/Aria",
+        help="Hub location of the vision model",
+    )
+    parser.add_argument(
+        "--output_hub_path",
+        default="rhymes-ai/Aria",
+        help="Location on the hub of the converted model",
+    )
+    parser.add_argument(
+        "--old_state_dict_id",
+        default="rhymes-ai/Aria",
+        help="Location on the hub of the raw state dict of the original model. The filename needs to be `model_state_dict.bin`",
+    )
+    args = parser.parse_args()
+    convert_aria_llama_to_hf(args.text_model_id, args.vision_model_id, args.output_hub_path, args.old_state_dict_id)
+
+
+if __name__ == "__main__":
+    main()

src/transformers/models/aria/image_processing_aria.py

@@ -0,0 +1,504 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/aria/modular_aria.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_aria.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2024 The Rhymes-AI Teams Authors 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.
+import math
+from typing import Iterable, List, Optional, Tuple, Union
+
+import numpy as np
+
+from ...image_processing_utils import BaseImageProcessor, BatchFeature, select_best_resolution
+from ...image_transforms import PaddingMode, convert_to_rgb, pad, resize, to_channel_dimension_format
+from ...image_utils import (
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    get_image_size,
+    infer_channel_dimension_format,
+    is_valid_image,
+    to_numpy_array,
+    valid_images,
+    validate_preprocess_arguments,
+)
+from ...utils import TensorType
+
+
+def make_batched_images(images) -> List[List[ImageInput]]:
+    """
+    Accepts images in list or nested list format, and makes a list of images for preprocessing.
+
+    Args:
+        images (`Union[List[List[ImageInput]], List[ImageInput], ImageInput]`):
+            The input image.
+
+    Returns:
+        list: A list of images.
+    """
+    if isinstance(images, (list, tuple)) and isinstance(images[0], (list, tuple)) and is_valid_image(images[0][0]):
+        return [img for img_list in images for img in img_list]
+
+    elif isinstance(images, (list, tuple)) and is_valid_image(images[0]):
+        return images
+
+    elif is_valid_image(images):
+        return [images]
+
+    raise ValueError(f"Could not make batched video from {images}")
+
+
+def divide_to_patches(image: np.array, patch_size: int, input_data_format) -> List[np.array]:
+    """
+    Divides an image into patches of a specified size.
+
+    Args:
+        image (`np.array`):
+            The input image.
+        patch_size (`int`):
+            The size of each patch.
+        input_data_format (`ChannelDimension` or `str`):
+            The channel dimension format of the input image.
+
+    Returns:
+        list: A list of np.array representing the patches.
+    """
+    patches = []
+    height, width = get_image_size(image, channel_dim=input_data_format)
+    for i in range(0, height, patch_size):
+        for j in range(0, width, patch_size):
+            if input_data_format == ChannelDimension.LAST:
+                patch = image[i : i + patch_size, j : j + patch_size]
+            else:
+                patch = image[:, i : i + patch_size, j : j + patch_size]
+            patches.append(patch)
+
+    return patches
+
+
+def _get_patch_output_size(image, target_resolution, input_data_format):
+    original_height, original_width = get_image_size(image, channel_dim=input_data_format)
+    target_height, target_width = target_resolution
+
+    scale_w = target_width / original_width
+    scale_h = target_height / original_height
+
+    if scale_w < scale_h:
+        new_width = target_width
+        new_height = min(math.ceil(original_height * scale_w), target_height)
+    else:
+        new_height = target_height
+        new_width = min(math.ceil(original_width * scale_h), target_width)
+
+    return new_height, new_width
+
+
+class AriaImageProcessor(BaseImageProcessor):
+    """
+    A vision processor for the Aria model that handles image preprocessing.
+    Initialize the AriaImageProcessor.
+
+    Args:
+        image_mean (`list`, *optional*, defaults to [0.5, 0.5, 0.5]):
+            Mean values for normalization.
+        image_std (`list`, *optional*, defaults to [0.5, 0.5, 0.5]):
+            Standard deviation values for normalization.
+        max_image_size (`int`, *optional*, defaults to 980):
+            Maximum image size.
+        min_image_size (`int`, *optional*, defaults to 336):
+            Minimum image size.
+        split_resolutions (`list`, *optional*, defaults to a list of optimal,resolutions as tuples):
+            The optimal resolutions for splitting the image.
+        split_image (`bool`, *optional*, defaults to `False`):
+            Whether to split the image.
+        do_convert_rgb (`bool`, *optional*, defaults to `True`):
+            Whether to convert the image to RGB.
+        do_normalize (`bool`, *optional*, defaults to `True`):
+            Whether to normalize the image.
+        resample (PILImageResampling, *optional*, defaults to `BICUBIC`):
+            The resampling filter to use if resizing the image.
+    """
+
+    def __init__(
+        self,
+        image_mean: List[float] = None,
+        image_std: List[float] = None,
+        max_image_size: int = 980,
+        min_image_size: int = 336,
+        split_resolutions: Optional[List[Tuple[int, int]]] = None,
+        split_image: Optional[bool] = False,
+        do_convert_rgb: Optional[bool] = True,
+        do_normalize: Optional[bool] = True,
+        resample: PILImageResampling = PILImageResampling.BICUBIC,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        if image_mean is None:
+            image_mean = [0.5, 0.5, 0.5]
+        if image_std is None:
+            image_std = [0.5, 0.5, 0.5]
+        self.max_image_size = max_image_size
+        self.min_image_size = min_image_size
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.split_image = split_image
+        if split_resolutions is None:
+            split_resolutions = [(1, 2), (1, 3), (1, 4), (1, 5), (1, 6), (1, 7), (1, 8), (2, 4), (2, 3), (2, 2), (2, 1), (3, 1), (3, 2), (4, 1), (4, 2), (5, 1), (6, 1), (7, 1), (8, 1)]  # fmt: skip
+            split_resolutions = [(el[0] * 490, el[1] * 490) for el in split_resolutions]
+        self.split_resolutions = split_resolutions
+        self.do_convert_rgb = do_convert_rgb
+        self.do_normalize = do_normalize
+        self.resample = resample
+
+    def preprocess(
+        self,
+        images: Union[ImageInput, List[ImageInput]],
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        max_image_size: Optional[int] = None,
+        min_image_size: Optional[int] = None,
+        split_image: Optional[bool] = None,
+        do_convert_rgb: Optional[bool] = None,
+        do_normalize: Optional[bool] = None,
+        resample: PILImageResampling = None,
+        return_tensors: Optional[Union[str, TensorType]] = "pt",
+        data_format: Optional[ChannelDimension] = ChannelDimension.FIRST,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ):
+        """
+        Process a list of images.
+
+        Args:
+            images (ImageInput or list of ImageInput):
+                The input image or a list of images.
+            image_mean (`list`, *optional*, defaults to [0.5, 0.5, 0.5]):
+                Mean values for normalization.
+            image_std (`list`, *optional*, defaults to [0.5, 0.5, 0.5]):
+                Standard deviation values for normalization.
+            max_image_size (`int`, *optional*, defaults to `self.max_image_size` (980)):
+                Maximum image size.
+            min_image_size (`int`, *optional*, defaults to `self.min_image_size` (336)):
+                Minimum image size.
+            split_image (`bool`, *optional*, defaults to `self.split_image` (False)):
+                Whether to split the image.
+            do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb` (True)):
+                Whether to convert the image to RGB.
+            do_normalize (`bool`, *optional*, defaults to `self.do_normalize` (True)):
+                Whether to normalize the image.
+            resample (PILImageResampling, *optional*, defaults to `self.resample` (BICUBIC)):
+                The resampling filter to use if resizing the image.
+            return_tensors (`str` or `TensorType`, *optional*, defaults to "pt"):
+                The type of tensor to return.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format for the output image. Can be one of:
+                    - `"channels_first"` or `ChannelDimension.FIRST`:
+                        image in (num_channels, height, width) format.
+                    - `"channels_last"` or `ChannelDimension.LAST`:
+                        image in (height, width, num_channels) format.
+                If unset, will use same as the input image.
+            input_data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format for the input image. Can be one of:
+                    - `"channels_first"` or `ChannelDimension.FIRST`:
+                        image in (num_channels, height, width) format.
+                    - `"channels_last"` or `ChannelDimension.LAST`:
+                        image in (height, width, num_channels) format.
+                If unset, will use the inferred format of the input image.
+
+        Returns:
+            BatchFeature:
+                A BatchFeature object containing:
+                - 'pixel_values':
+                    Tensor of processed image pixel values.
+                - 'pixel_mask':
+                    Boolean pixel mask. This mask is a 2D tensor of shape (max_image_size, max_image_size) where:
+                    - True (1) values indicate pixels that belong to the original resized image.
+                    - False (0) values indicate pixels that are part of the padding.
+                  The mask helps distinguish between actual image content and padded areas in subsequent processing steps.
+                - 'num_crops':
+                    The maximum number of crops across all images.
+        """
+        image_mean = image_mean if image_mean is not None else self.image_mean
+        image_std = image_std if image_std is not None else self.image_std
+        max_image_size = max_image_size if max_image_size is not None else self.max_image_size
+        min_image_size = min_image_size if min_image_size is not None else self.min_image_size
+        split_image = split_image if split_image is not None else self.split_image
+        do_convert_rgb = do_convert_rgb if do_convert_rgb is not None else self.do_convert_rgb
+        do_normalize = do_normalize if do_normalize is not None else self.do_normalize
+        resample = resample if resample is not None else self.resample
+
+        if max_image_size not in [490, 980]:
+            raise ValueError("max_image_size must be either 490 or 980")
+
+        images = make_batched_images(images)
+
+        if not valid_images(images):
+            raise ValueError(
+                "Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
+                "torch.Tensor, tf.Tensor or jax.ndarray."
+            )
+
+        validate_preprocess_arguments(
+            do_normalize=do_normalize,
+            image_mean=image_mean,
+            image_std=image_std,
+            resample=resample,
+        )
+
+        if do_convert_rgb:
+            images = [convert_to_rgb(image) for image in images]
+
+        # All transformations expect numpy arrays.
+        images = [to_numpy_array(image) for image in images]
+
+        if input_data_format is None:
+            # We assume that all images have the same channel dimension format.
+            input_data_format = infer_channel_dimension_format(images[0])
+
+        pixel_values = []
+        pixel_masks = []
+        num_crops = None
+
+        for image in images:
+            if split_image:
+                crop_images = self.get_image_patches(
+                    image,
+                    self.split_resolutions,
+                    max_image_size,
+                    resample,
+                    data_format=input_data_format,
+                    input_data_format=input_data_format,
+                )
+            else:
+                crop_images = [image]
+            if num_crops is None or len(crop_images) > num_crops:
+                num_crops = len(crop_images)
+
+            for crop_image in crop_images:
+                # At this point the scale is the rescaling factor that would bring the image to max_size in its larger dimension
+                h, w = get_image_size(crop_image)
+                scale = max_image_size / max(h, w)
+                if w >= h:
+                    new_size = (max(int(h * scale), min_image_size), max_image_size)  # h, w
+                else:
+                    new_size = (max_image_size, max(int(w * scale), min_image_size))  # h, w
+
+                crop_image_resized = resize(
+                    crop_image,
+                    new_size,
+                    resample=resample,
+                    data_format=input_data_format,
+                    input_data_format=input_data_format,
+                )
+
+                padding_bottom, padding_right = max_image_size - new_size[0], max_image_size - new_size[1]
+                crop_image_padded = pad(
+                    crop_image_resized,
+                    ((0, padding_bottom), (0, padding_right)),
+                    data_format=input_data_format,
+                    input_data_format=input_data_format,
+                )
+
+                # Create a pixel mask
+                pixel_mask = np.zeros((max_image_size, max_image_size), dtype=bool)
+                pixel_mask[: new_size[0], : new_size[1]] = 1
+                pixel_masks.append(pixel_mask)
+
+                if do_normalize:
+                    crop_image_padded = self.normalize(
+                        crop_image_padded / 255.0,
+                        self.image_mean,
+                        self.image_std,
+                        data_format=input_data_format,
+                        input_data_format=input_data_format,
+                    )
+                    crop_image_padded = (
+                        to_channel_dimension_format(crop_image_padded, data_format, input_data_format)
+                        if data_format is not None
+                        else crop_image_padded
+                    )
+
+                pixel_values.append(crop_image_padded)
+        return BatchFeature(
+            data={
+                "pixel_values": np.stack(pixel_values, axis=0),
+                "pixel_mask": np.stack(pixel_masks, axis=0),
+                "num_crops": num_crops,
+            },
+            tensor_type=return_tensors,
+        )
+
+    def _resize_for_patching(
+        self, image: np.array, target_resolution: tuple, resample, input_data_format: ChannelDimension
+    ) -> np.array:
+        """
+        Resizes an image to a target resolution while maintaining aspect ratio.
+
+        Args:
+            image (np.array):
+                The input image.
+            target_resolution (tuple):
+                The target resolution (height, width) of the image.
+            resample (`PILImageResampling`):
+                Resampling filter to use if resizing the image.
+            input_data_format (`ChannelDimension` or `str`):
+                The channel dimension format of the input image.
+
+        Returns:
+            np.array: The resized and padded image.
+        """
+        new_height, new_width = _get_patch_output_size(image, target_resolution, input_data_format)
+
+        # Resize the image
+        resized_image = resize(image, (new_height, new_width), resample=resample, input_data_format=input_data_format)
+
+        return resized_image
+
+    def _pad_for_patching(
+        self, image: np.array, target_resolution: tuple, input_data_format: ChannelDimension
+    ) -> np.array:
+        """
+        Pad an image to a target resolution while maintaining aspect ratio.
+        """
+        target_height, target_width = target_resolution
+        new_height, new_width = _get_patch_output_size(image, target_resolution, input_data_format)
+
+        paste_x = (target_width - new_width) // 2
+        paste_y = (target_height - new_height) // 2
+
+        padded_image = self.pad(image, padding=((paste_y, paste_y), (paste_x, paste_x)))
+
+        return padded_image
+
+    def pad(
+        self,
+        image: np.ndarray,
+        padding: Union[int, Tuple[int, int], Iterable[Tuple[int, int]]],
+        mode: PaddingMode = PaddingMode.CONSTANT,
+        constant_values: Union[float, Iterable[float]] = 0.0,
+        data_format: Optional[Union[str, ChannelDimension]] = None,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ) -> np.ndarray:
+        """
+        Pads the `image` with the specified `padding` and `mode`. Padding can be in the (`height`, `width`)
+        dimension of in the (`num_patches`) dimension. In the second case an iterable if tuples is expected
+        as input.
+
+        Args:
+            image (`np.ndarray`):
+                The image to pad.
+            padding (`int` or `Tuple[int, int]` or `Iterable[Tuple[int, int]]`):
+                Padding to apply to the edges of the height, width axes. Can be one of three formats:
+                - `((before_height, after_height), (before_width, after_width))` unique pad widths for each axis.
+                - `((before, after),)` yields same before and after pad for height and width.
+                - `(pad,)` or int is a shortcut for before = after = pad width for all axes.
+            mode (`PaddingMode`):
+                The padding mode to use. Can be one of:
+                    - `"constant"`: pads with a constant value.
+                    - `"reflect"`: pads with the reflection of the vector mirrored on the first and last values of the
+                    vector along each axis.
+                    - `"replicate"`: pads with the replication of the last value on the edge of the array along each axis.
+                    - `"symmetric"`: pads with the reflection of the vector mirrored along the edge of the array.
+            constant_values (`float` or `Iterable[float]`, *optional*):
+                The value to use for the padding if `mode` is `"constant"`.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format for the output image. Can be one of:
+                    - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                    - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                If unset, will use same as the input image.
+            input_data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format for the input image. Can be one of:
+                    - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                    - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                If unset, will use the inferred format of the input image.
+
+        Returns:
+            `np.ndarray`: The padded image.
+
+        """
+
+        # call the general `pad` if padding on `height/width`, otherwise it's the `num_patched` dim
+        if isinstance(padding, int) or len(padding) != 4:
+            return pad(image, padding, mode, constant_values, data_format, input_data_format)
+
+        if input_data_format is None:
+            input_data_format = infer_channel_dimension_format(image)
+
+        padding_mode_mapping = {
+            PaddingMode.CONSTANT: "constant",
+            PaddingMode.REFLECT: "reflect",
+            PaddingMode.REPLICATE: "edge",
+            PaddingMode.SYMMETRIC: "symmetric",
+        }
+        image = np.pad(image, padding, mode=padding_mode_mapping[mode], constant_values=constant_values)
+        image = (
+            to_channel_dimension_format(image, data_format, input_data_format) if data_format is not None else image
+        )
+        return image
+
+    def get_image_patches(
+        self,
+        image: np.array,
+        grid_pinpoints: List[Tuple[int, int]],
+        patch_size: int,
+        resample: PILImageResampling,
+        data_format: ChannelDimension,
+        input_data_format: ChannelDimension,
+    ) -> List[np.array]:
+        """
+        Process an image with variable resolutions by dividing it into patches.
+
+        Args:
+            image (`np.array`):
+                The input image to be processed.
+            grid_pinpoints (List[Tuple[int, int]]):
+                A list of possible resolutions as tuples.
+            patch_size (`int`):
+                Size of the patches to divide the image into.
+            resample (`PILImageResampling`):
+                Resampling filter to use if resizing the image.
+            data_format (`ChannelDimension` or `str`):
+                The channel dimension format for the output image.
+            input_data_format (`ChannelDimension` or `str`):
+                The channel dimension format of the input image.
+
+        Returns:
+            `List[np.array]`: A list of NumPy arrays containing the processed image patches.
+        """
+        if not isinstance(grid_pinpoints, list):
+            raise TypeError("grid_pinpoints must be a list of possible resolutions.")
+
+        possible_resolutions = grid_pinpoints
+
+        image_size = get_image_size(image, channel_dim=input_data_format)
+        best_resolution = select_best_resolution(image_size, possible_resolutions)
+        resized_image = self._resize_for_patching(
+            image, best_resolution, resample=resample, input_data_format=input_data_format
+        )
+        padded_image = self._pad_for_patching(resized_image, best_resolution, input_data_format=input_data_format)
+
+        patches = divide_to_patches(padded_image, patch_size=patch_size, input_data_format=input_data_format)
+
+        # make sure that all patches are in the input data format
+        patches = [
+            to_channel_dimension_format(patch, channel_dim=data_format, input_channel_dim=input_data_format)
+            for patch in patches
+        ]
+        return patches
+
+
+__all__ = ["AriaImageProcessor"]

src/transformers/models/aria/processing_aria.py

@@ -0,0 +1,164 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/aria/modular_aria.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_aria.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2024 The Rhymes-AI Teams Authors 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.
+from typing import Dict, List, Optional, Union
+
+from ...image_processing_utils import BatchFeature
+from ...image_utils import ImageInput
+from ...processing_utils import ProcessingKwargs, ProcessorMixin, Unpack
+from ...tokenization_utils import PreTokenizedInput, TextInput
+from ...utils import TensorType
+from ..auto import AutoTokenizer
+
+
+class AriaProcessorKwargs(ProcessingKwargs, total=False):
+    _defaults = {
+        "text_kwargs": {
+            "padding": False,
+        },
+        "images_kwargs": {
+            "max_image_size": 980,
+            "split_image": False,
+        },
+        "return_tensors": TensorType.PYTORCH,
+    }
+
+
+class AriaProcessor(ProcessorMixin):
+    """
+    AriaProcessor is a processor for the Aria model which wraps the Aria image preprocessor and the LLama slow tokenizer.
+
+    Args:
+        image_processor (`AriaImageProcessor`, *optional*):
+            The AriaImageProcessor to use for image preprocessing.
+        tokenizer (`PreTrainedTokenizerBase`, *optional*):
+            An instance of [`PreTrainedTokenizerBase`]. This should correspond with the model's text model. The tokenizer is a required input.
+        chat_template (`str`, *optional*):
+            A Jinja template which will be used to convert lists of messages in a chat into a tokenizable string.
+        size_conversion (`Dict`, *optional*):
+            A dictionary indicating size conversions for images.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+    valid_kwargs = ["chat_template", "size_conversion"]
+    image_processor_class = "AriaImageProcessor"
+    tokenizer_class = "AutoTokenizer"
+
+    def __init__(
+        self,
+        image_processor=None,
+        tokenizer: Union[AutoTokenizer, str] = None,
+        chat_template: Optional[str] = None,
+        size_conversion: Optional[Dict[Union[float, int], int]] = None,
+    ):
+        if size_conversion is None:
+            size_conversion = {490: 128, 980: 256}
+        self.size_conversion = {int(k): v for k, v in size_conversion.items()}
+
+        if tokenizer is not None and tokenizer.pad_token is None:
+            tokenizer.pad_token = tokenizer.unk_token
+
+        super().__init__(image_processor, tokenizer, chat_template=chat_template)
+
+    def __call__(
+        self,
+        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]],
+        images: Optional[ImageInput] = None,
+        audio=None,
+        videos=None,
+        **kwargs: Unpack[AriaProcessorKwargs],
+    ) -> BatchFeature:
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s).
+
+        Args:
+            text (`TextInput`, `PreTokenizedInput`, `List[TextInput]`, `List[PreTokenizedInput]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            images (`ImageInput`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. Both channels-first and channels-last formats are supported.
+
+
+        Returns:
+            [`BatchFeature`]: A [`BatchFeature`] with the following fields:
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+            `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+            `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+            - **pixel_mask** -- Pixel mask to be fed to a model. Returned when `images` is not `None`.
+        """
+        output_kwargs = self._merge_kwargs(
+            AriaProcessorKwargs,
+            tokenizer_init_kwargs=self.tokenizer.init_kwargs,
+            **kwargs,
+        )
+        if isinstance(text, str):
+            text = [text]
+        elif not isinstance(text, list) and not isinstance(text[0], str):
+            raise ValueError("Invalid input text. Please provide a string, or a list of strings")
+        if images is not None:
+            image_inputs = self.image_processor(
+                images,
+                **output_kwargs["images_kwargs"],
+            )
+            # expand the image_token according to the num_crops and tokens per image
+            tokens_per_image = self.size_conversion[image_inputs.pixel_values.shape[2]]
+            prompt_strings = []
+            num_crops = image_inputs.pop("num_crops") * tokens_per_image
+            for sample in text:
+                sample = sample.replace(self.tokenizer.image_token, self.tokenizer.image_token * num_crops)
+                prompt_strings.append(sample)
+
+        else:
+            image_inputs = {}
+            prompt_strings = text
+
+        text_inputs = self.tokenizer(
+            prompt_strings,
+            **output_kwargs["text_kwargs"],
+        )
+
+        return BatchFeature(data={**text_inputs, **image_inputs})
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    @property
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
+
+
+__all__ = ["AriaProcessor"]