Confusion about `adapter_dir` usage in `from_pretrained('conf/model.yaml', args.adapter_dir)`

[tahsirmunna]

Hello,

I'm trying to understand how to correctly use the following line from your codebase:

model, tokenizer = from_pretrained('conf/model.yaml', args.adapter_dir)

Specifically, I’m confused about the adapter_dir argument. I couldn't find clear documentation or examples that explain:

• What kind of directory this should be?
• What files it must contain?
• Whether it refers to LoRA adapters, PEFT adapters, or something else?

Could you please:

1. Provide an example of a valid adapter_dir path?
2. List the required files (e.g., adapter_config.json, adapter_model.bin, etc.) that should be inside that directory?
3. Clarify whether the adapters are format-specific (e.g., Hugging Face PEFT-compatible)?

Thank you for your great work! Looking forward to your response to better integrate this with my project.

[tahsirmunna]

I found the adapter_config.json file in your github page. but still I have error as like below,

python demo.py "adapter_config.json"

usage: demo.py [-h] [--model.help [CLASS_PATH_OR_NAME]]
               [--model CONFIG | CLASS_PATH_OR_NAME | .INIT_ARG_NAME VALUE]
error: Parser key "model":
  Does not validate against any of the Union subtypes
  Subtypes: [<class 'NoneType'>, <class 'mmmm.models.mmmm.MMMMForCausalLM'>]
  Errors:
    - Expected a <class 'NoneType'>
    - Problem with given class_path 'mmmm.models.mmmm.MMMMForCausalLM__build_class':
        Parser key "tokenizer":
          With fail_untyped=True, all mandatory parameters must have a supported type. Parameter 'vocab_file' from 'transformers.LlamaTokenizer.__init__' does not specify a type.
  Given value type: <class 'dict'>
  Given value: {'class_path': 'mmmm.models.build', 'init_args': {'pretrained_model_name_or_path': 'THUDM/cogvlm-chat-hf', 'vision_override': {'patch_size': 16, 'pos_embed_shape': [8, 32, 32], 'pt_pos_embed_shape': [35, 35]}, 'tokenizer': 'tokenizer.yaml'}}

[function2-llx]

Hi! The adapter_dir should be the directory path where the LoRA adapter is saved. You should be able to obtain the pretrained adapter checkpoint from the release of this repository, downloading and placing them under some directory of your filesystem.