llama.cpp icon indicating copy to clipboard operation
llama.cpp copied to clipboard
Published 2 months ago • verified publisher icon ggml-org

Why skip rope_freqs, attn_rot_embd modules when serialization to gguf?

Open lifelongeeek opened this issue 1 year ago • 0 comments

During convert llama-2-7b-hf & vicuna-13b, I found that convert script skip rope_freqs, attn_rot_embd modules in each layer. These are defined in MODEL_TENSOR_SKIP as well.

What are they and why can we skip those modules?

lifelongeeek avatar Feb 21 '24 08:02 lifelongeeek