OpenGVLab/InternVL3_5-30B-A3B-Instruct

1# --------------------------------------------------------2# InternVL3# Copyright (c) 2024 OpenGVLab4# Licensed under The MIT License [see LICENSE for details]5# --------------------------------------------------------6 7import copy8from typing import Dict, Any, Optional9 10from transformers.configuration_utils import PretrainedConfig11from transformers.utils import logging12 13from .configuration_intern_vit import InternVisionConfig14 15logger = logging.get_logger(__name__)16 17 18class InternVLChatConfig(PretrainedConfig):19    model_type = 'internvl_chat'20    is_composition = True21 22    def __init__(23        self,24        vision_config: Optional[Dict[str, Any]] = None,25        llm_config: Optional[Dict[str, Any]] = None,26        use_backbone_lora=0,27        use_llm_lora=0,28        select_layer=-1,29        force_image_size=None,30        downsample_ratio=0.5,31        template=None,32        dynamic_image_size=False,33        use_thumbnail=False,34        ps_version="v1",35        min_dynamic_patch=1,36        max_dynamic_patch=6,37        **kwargs,38    ):39        super().__init__(**kwargs)40 41        if vision_config is None:42            vision_config = {'architectures': ['InternVisionModel']}43            logger.info('vision_config is None. Initializing the InternVisionConfig with default values.')44 45        if llm_config is None:46            llm_config = {'architectures': ['Qwen2ForCausalLM']}47            logger.info('llm_config is None. Initializing the LlamaConfig config with default values (`LlamaConfig`).')48        assert 'architectures' in llm_config, "Should specify architecture in llm_config"49 50        if isinstance(vision_config, dict):51            self.vision_config = InternVisionConfig(**vision_config)52        else:53            self.vision_config = vision_config54 55        if isinstance(llm_config, dict):56            architecture: str = llm_config['architectures'][0]57            if architecture == 'LlamaForCausalLM':58                from transformers import LlamaConfig59                self.llm_config = LlamaConfig(**llm_config)60            elif architecture == 'Qwen2ForCausalLM':61                from transformers import Qwen2Config62                self.llm_config = Qwen2Config(**llm_config)63            elif architecture == 'Qwen3MoeForCausalLM':64                from transformers import Qwen3MoeConfig65                self.llm_config = Qwen3MoeConfig(**llm_config)66            elif architecture == 'Qwen3ForCausalLM':67                from transformers import Qwen3Config68                self.llm_config = Qwen3Config(**llm_config)69            else:70                raise ValueError('Unsupported architecture: {}'.format(architecture))71        else:72            self.llm_config = llm_config73 74        self.use_backbone_lora = use_backbone_lora75        self.use_llm_lora = use_llm_lora76        self.select_layer = select_layer77        self.force_image_size = force_image_size78        self.downsample_ratio = downsample_ratio79        self.template = template80        self.dynamic_image_size = dynamic_image_size81        self.use_thumbnail = use_thumbnail82        self.ps_version = ps_version  # pixel shuffle version83        self.min_dynamic_patch = min_dynamic_patch84        self.max_dynamic_patch = max_dynamic_patch85        self.tie_word_embeddings = self.llm_config.tie_word_embeddings86 87        logger.info(f'vision_select_layer: {self.select_layer}')88        logger.info(f'ps_version: {self.ps_version}')89        logger.info(f'min_dynamic_patch: {self.min_dynamic_patch}')90        logger.info(f'max_dynamic_patch: {self.max_dynamic_patch}')91 92    def to_dict(self):93        """94        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].95 96        Returns:97            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,98        """99        output = copy.deepcopy(self.__dict__)100        output['vision_config'] = self.vision_config.to_dict()101        output['llm_config'] = self.llm_config.to_dict()102        output['model_type'] = self.__class__.model_type103        output['use_backbone_lora'] = self.use_backbone_lora104        output['use_llm_lora'] = self.use_llm_lora105        output['select_layer'] = self.select_layer106        output['force_image_size'] = self.force_image_size107        output['downsample_ratio'] = self.downsample_ratio108        output['template'] = self.template109        output['dynamic_image_size'] = self.dynamic_image_size110        output['use_thumbnail'] = self.use_thumbnail111        output['ps_version'] = self.ps_version112        output['min_dynamic_patch'] = self.min_dynamic_patch113        output['max_dynamic_patch'] = self.max_dynamic_patch114 115        return output116