Aluode/PerceptionLabPortable

1# coding=utf-82# Copyright 2025 The HuggingFace Inc. team. All rights reserved.3#4# Licensed under the Apache License, Version 2.0 (the "License");5# you may not use this file except in compliance with the License.6# You may obtain a copy of the License at7#8#     http://www.apache.org/licenses/LICENSE-2.09#10# Unless required by applicable law or agreed to in writing, software11# distributed under the License is distributed on an "AS IS" BASIS,12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.13# See the License for the specific language governing permissions and14# limitations under the License.15 16from ...configuration_utils import PretrainedConfig17from ..qwen2.configuration_qwen2 import Qwen2Config18 19 20class Ovis2VisionConfig(PretrainedConfig):21    r"""22    This is the configuration class to store the configuration of a [`Ovis2VisionModel`]. It is used to instantiate a23    Ovis2VisionModel model according to the specified arguments, defining the model architecture. Instantiating a configuration24    with the defaults will yield a similar configuration to that of Ovis2.25 26    Args:27        hidden_size (`int`, *optional*, defaults to 1024):28            Dimensionality of the encoder layers and the pooler layer.29        intermediate_size (`int`, *optional*, defaults to 2816):30            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.31        num_hidden_layers (`int`, *optional*, defaults to 24):32            Number of hidden layers in the Transformer encoder.33        num_attention_heads (`int`, *optional*, defaults to 8):34            Number of attention heads for each attention layer in the Transformer encoder.35        num_channels (`int`, *optional*, defaults to 3):36            Number of channels in the input images.37        image_size (`int`, *optional*, defaults to 224):38            The size (resolution) of each image.39        patch_size (`int`, *optional*, defaults to 14):40            The size (resolution) of each patch.41        rms_norm_eps (`float`, *optional*, defaults to 1e-05):42            The epsilon used by the RMSNorm layers.43        attention_dropout (`float`, *optional*, defaults to 0.0):44            The dropout ratio for the attention probabilities.45        qkv_bias (`bool`, *optional*, defaults to `False`):46            Whether to add a learnable bias to the query, key, and value sequences at each attention head.47        mlp_bias (`bool`, *optional*, defaults to `False`):48            Whether to add a learnable bias to the MLP layers.49        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):50            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,51            `"relu"`, `"selu"` and `"gelu_new"` `"quick_gelu"` are supported.52        vocab_size (`int`, *optional*, defaults to 16384):53            Vocabulary size of the Vision Transformer.54        hidden_stride (`int`, *optional*, defaults to 1):55            The stride of the hidden layer in the Vision Transformer.56        num_visual_indicator_tokens (`int`, *optional*, defaults to 5):57            Number of visual indicator tokens.58        initializer_range (`float`, *optional*, defaults to 0.02):59            The standard deviation of the truncated normal initializer for initializing all weight matrices.60        tokenize_function (`str`, *optional*, defaults to `"softmax"`):61            The function used to tokenize the visual indicator tokens.62    """63 64    base_config_key = "vision_config"65 66    def __init__(67        self,68        hidden_size: int = 1024,69        intermediate_size: int = 2816,70        num_hidden_layers: int = 24,71        num_attention_heads: int = 8,72        num_channels: int = 3,73        image_size: int = 224,74        patch_size: int = 14,75        rms_norm_eps: float = 1e-5,76        attention_dropout: float = 0.0,77        qkv_bias: bool = False,78        mlp_bias: bool = False,79        hidden_act="silu",80        vocab_size=16384,81        hidden_stride=1,82        num_visual_indicator_tokens=5,83        initializer_range=0.02,84        tokenize_function="softmax",85        **kwargs,86    ):87        super().__init__(**kwargs)88 89        self.hidden_size = hidden_size90        self.intermediate_size = intermediate_size91        self.num_hidden_layers = num_hidden_layers92        self.num_attention_heads = num_attention_heads93        self.num_channels = num_channels94        self.patch_size = patch_size95        self.image_size = image_size96 97        self.attention_dropout = attention_dropout98        self.hidden_act = hidden_act99        self.qkv_bias = qkv_bias100        self.mlp_bias = mlp_bias101        self.rms_norm_eps = rms_norm_eps102        self.vocab_size = vocab_size103        self.hidden_stride = hidden_stride104        self.num_visual_indicator_tokens = num_visual_indicator_tokens105        self.tokenize_function = tokenize_function106        self.initializer_range = initializer_range107 108 109class Ovis2Config(PretrainedConfig):110    r"""111    This is the configuration class to store the configuration of a [`Ovis2ForConditionalGeneration`]. It is used to instantiate a112    Ovis2 model according to the specified arguments, defining the model architecture. Instantiating a configuration113    with the defaults will yield a similar configuration to that of Ovis2.114 115    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the116    documentation from [`PretrainedConfig`] for more information.117 118    e.g. [thisisiron/Ovis2-1B-hf](https://huggingface.co/thisisiron/Ovis2-1B-hf)119 120    Args:121        vision_config (`Union[AutoConfig, dict]`,  *optional*, defaults to `Ovis2VisionConfig`):122            The config object or dictionary of the vision backbone.123        text_config (`Union[AutoConfig, dict]`, *optional*, defaults to `Qwen2Config`):124            The config object or dictionary of the text backbone.125        image_token_id (`int`, *optional*, defaults to 151665):126            The image token id to encode the image prompt.127        visual_indicator_token_ids (`List[int]`, *optional*, defaults to `[151666, 151667, 151668, 151669, 151670]`):128            The visual indicator token ids to encode the image prompt.129        vocab_size (`int`, *optional*, defaults to 151643):130            Vocabulary size of the text model.131        hidden_size (`int`, *optional*, defaults to 1536):132            Dimensionality of the encoder layers and the pooler layer.133 134    ```python135    >>> from transformers import Ovis2ForConditionalGeneration, Ovis2Config136 137    >>> # Initializing a Ovis2 style configuration138    >>> configuration = Ovis2Config()139 140    >>> # Initializing a model from the Ovis2-2B style configuration141    >>> model = Ovis2ForConditionalGeneration(configuration)142 143    >>> # Accessing the model configuration144    >>> configuration = model.config145    ```146    """147 148    model_type = "ovis2"149    sub_configs = {"text_config": Qwen2Config, "vision_config": Ovis2VisionConfig}150 151    def __init__(152        self,153        vision_config=None,154        text_config=None,155        image_token_id=151665,156        visual_indicator_token_ids=[151666, 151667, 151668, 151669, 151670],157        vocab_size=151643,158        hidden_size=1536,159        **kwargs,160    ):161        if isinstance(vision_config, dict):162            self.vision_config = Ovis2VisionConfig(**vision_config)163        elif isinstance(vision_config, Ovis2VisionConfig):164            self.vision_config = vision_config165        if vision_config is None:166            self.vision_config = Ovis2VisionConfig(num_visual_indicator_tokens=len(visual_indicator_token_ids))167 168        if isinstance(text_config, dict):169            self.text_config = Qwen2Config(**text_config)170        elif isinstance(text_config, Qwen2Config):171            self.text_config = text_config172        elif text_config is None:173            self.text_config = Qwen2Config()174 175        self.vocab_size = vocab_size176        self.hidden_size = hidden_size177        self.image_token_id = image_token_id178        self.visual_indicator_token_ids = visual_indicator_token_ids179        super().__init__(**kwargs)180 181 182__all__ = ["Ovis2VisionConfig", "Ovis2Config"]183