Skip to content

State_Dict Size Mismatch Error #28

@mmdrahmani

Description

@mmdrahmani

Hi Again
I used float16 in the parameter setting and I can load the model into memory. However, when I run the demo, I get this error:
Error(s) in loading state_dict for LLaMA: size mismatch for transformer.h.0.attn.c_attn.lora_A: copying a param with shape torch.Size([128, 5120]) from checkpoint, the shape in current model is torch.Size([128, 4096]).

I think I get this error because the checkpoint has a parameter shape of [128, 5120], but the current model expects a shape of [128, 4096]. Maybe during training, the Llama 7B model structure was different compared to the model that I got from HuggingFace.

Do you have any ideas how to solve this issue?
Thanks a lot for your support.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions