OpenGVLab

/

InternVL-Chat-V1-5

like 416

Follow

OpenGVLab 1.71k

When loading InternVideo2_5_Chat_8B model or InternVL2_5 series models:
It raises AttributeError: 'InternVLChatConfig' object has no attribute 'llm_config'
Please refer: https://huggingface.co/OpenGVLab/InternVideo2_5_Chat_8B/discussions/14

Simple Solution for InternVL Configuration Issue

(Tested with transformers v4.52.4)

Required Modifications

1. Add Initialization (configuration_internvl_chat.py:49)

   self.vision_config = InternVisionConfig(**vision_config)
   self.llm_config = None  # Initialize llm_config to prevent AttributeError
   

2. Add Null Check (configuration_internvl_chat.py:85)

   output['llm_config'] = self.llm_config.to_dict() if self.llm_config is not None else {}
   

Root Cause Analysis

When executing:

model = AutoModel.from_pretrained(model_path, trust_remote_code=True).half().cuda().to(torch.bfloat16)

The following occurs:

1. The Hugging Face framework downloads and parses configuration_internvl_chat.py
2. During config initialization (transformers/configuration_utils.py:816-822):

   config_dict = self.to_dict()
   
   # Get the default config dict (from a fresh PreTrainedConfig instance)
   default_config_dict = PretrainedConfig().to_dict()
   
   # get class specific config dict
   class_config_dict = self.__class__().to_dict() if not self.has_no_defaults_at_init else {}
   

3. Key Issue:
   • self.llm_config is None during class_config_dict generation because llm_config is None
   • Without explicit initialization, this triggers an AttributeError when .to_dict() is called

Why the Fix Works

1. The initialization ensures self.llm_config always exists (even as None)
2. The null check prevents method calls on None while maintaining expected dictionary structure