OpenGVLab/VisualPRM-8B

1# --------------------------------------------------------2# InternVL3# Copyright (c) 2024 OpenGVLab4# Licensed under The MIT License [see LICENSE for details]5# --------------------------------------------------------6 7import os8from typing import Union9 10from transformers.configuration_utils import PretrainedConfig11from transformers.utils import logging12 13logger = logging.get_logger(__name__)14 15 16class InternVisionConfig(PretrainedConfig):17    r"""18    This is the configuration class to store the configuration of a [`InternVisionModel`]. It is used to19    instantiate a vision encoder according to the specified arguments, defining the model architecture.20 21    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the22    documentation from [`PretrainedConfig`] for more information.23 24    Args:25        num_channels (`int`, *optional*, defaults to 3):26            Number of color channels in the input images (e.g., 3 for RGB).27        patch_size (`int`, *optional*, defaults to 14):28            The size (resolution) of each patch.29        image_size (`int`, *optional*, defaults to 224):30            The size (resolution) of each image.31        qkv_bias (`bool`, *optional*, defaults to `False`):32            Whether to add a bias to the queries and values in the self-attention layers.33        hidden_size (`int`, *optional*, defaults to 3200):34            Dimensionality of the encoder layers and the pooler layer.35        num_attention_heads (`int`, *optional*, defaults to 25):36            Number of attention heads for each attention layer in the Transformer encoder.37        intermediate_size (`int`, *optional*, defaults to 12800):38            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.39        qk_normalization (`bool`, *optional*, defaults to `True`):40            Whether to normalize the queries and keys in the self-attention layers.41        num_hidden_layers (`int`, *optional*, defaults to 48):42            Number of hidden layers in the Transformer encoder.43        use_flash_attn (`bool`, *optional*, defaults to `True`):44            Whether to use flash attention mechanism.45        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):46            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,47            `"relu"`, `"selu"` and `"gelu_new"` ``"gelu"` are supported.48        layer_norm_eps (`float`, *optional*, defaults to 1e-6):49            The epsilon used by the layer normalization layers.50        dropout (`float`, *optional*, defaults to 0.0):51            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.52        drop_path_rate (`float`, *optional*, defaults to 0.0):53            Dropout rate for stochastic depth.54        attention_dropout (`float`, *optional*, defaults to 0.0):55            The dropout ratio for the attention probabilities.56        initializer_range (`float`, *optional*, defaults to 0.02):57            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.58        initializer_factor (`float`, *optional*, defaults to 0.1):59            A factor for layer scale.60    """61 62    model_type = 'intern_vit_6b'63 64    def __init__(65            self,66            num_channels=3,67            patch_size=14,68            image_size=224,69            qkv_bias=False,70            hidden_size=3200,71            num_attention_heads=25,72            intermediate_size=12800,73            qk_normalization=True,74            num_hidden_layers=48,75            use_flash_attn=True,76            hidden_act='gelu',77            norm_type='rms_norm',78            layer_norm_eps=1e-6,79            dropout=0.0,80            drop_path_rate=0.0,81            attention_dropout=0.0,82            initializer_range=0.02,83            initializer_factor=0.1,84            **kwargs,85    ):86        super().__init__(**kwargs)87 88        self.hidden_size = hidden_size89        self.intermediate_size = intermediate_size90        self.dropout = dropout91        self.drop_path_rate = drop_path_rate92        self.num_hidden_layers = num_hidden_layers93        self.num_attention_heads = num_attention_heads94        self.num_channels = num_channels95        self.patch_size = patch_size96        self.image_size = image_size97        self.initializer_range = initializer_range98        self.initializer_factor = initializer_factor99        self.attention_dropout = attention_dropout100        self.layer_norm_eps = layer_norm_eps101        self.hidden_act = hidden_act102        self.norm_type = norm_type103        self.qkv_bias = qkv_bias104        self.qk_normalization = qk_normalization105        self.use_flash_attn = use_flash_attn106 107    @classmethod108    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> 'PretrainedConfig':109        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)110 111        if 'vision_config' in config_dict:112            config_dict = config_dict['vision_config']113 114        if 'model_type' in config_dict and hasattr(cls, 'model_type') and config_dict['model_type'] != cls.model_type:115            logger.warning(116                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "117                f'{cls.model_type}. This is not supported for all configurations of models and can yield errors.'118            )119 120        return cls.from_dict(config_dict, **kwargs)121