THUDM

/

cogvlm-chat-hf

like 174

Recent changes to Huggingface transformers package will break this implementation. The inv_freq is being stored in the state_dict but not being in _parameters will cause a failure that the key is missing.

The other models that are part of the core transformers packages that use rotary embeding were all modified here: https://github.com/huggingface/transformers/pull/28837/files

However since this model uses the remote loading it will simply break upon updating transformers if you use load_in_4bit or pass in the bnb_config to turn quantized loading on.

It can be monkey patched like so:

from transformers.quantizers.quantizer_bnb_4bit import Bnb4BitHfQuantizer, get_module_from_name
from transformers.modeling_utils import PreTrainedModel
# CogVLM stores inv_freq in the state dictionary but it is not in models._parameters so it cannot be quantized
# was patched in transformers for other models here: https://github.com/huggingface/transformers/pull/28837/files but cog is not part of transformers
def patched_check_quantized_param(
        self, model: "PreTrainedModel", param_value: "torch.Tensor", param_name: str, state_dict: Dict[str, Any]
    ) -> bool:
    module, tensor_name = get_module_from_name(model, param_name)
    if ("inv_freq" == tensor_name):  # PATCH
        return False  # PATCH
    if isinstance(module._parameters[tensor_name], bnb.nn.Params4bit):  # KEY ERROR on inv_freq
        return True
    elif isinstance(module, bnb.nn.Linear4bit) and tensor_name == "bias":
        return True
    else:
        return False

Bnb4BitHfQuantizer.check_quantized_param = patched_check_quantized_param