llama-2-70b.ggmlv3.q4_K_S.bin : wrong shape errors

#2
by mikeee - opened
error loading model: llama.cpp: tensor 'layers.0.attention.wk.weight' has wrong shape; expected  8192 x  8192, got  8192 x1024
llama_init_from_file: failed to load model
Traceback (most recent call last):
  File "/home/mu2018/github/llama2-7b-chat-ggml/app.py", line 126, in <module>
    LLM = AutoModelForCausalLM.from_pretrained(
  File "/home/mu2018/github/llama2-7b-chat-ggml/.venv/lib/python3.10/site-packages/ctransformers/hub.py", line 157, in from_pretrained
    return LLM(
  File "/home/mu2018/github/llama2-7b-chat-ggml/.venv/lib/python3.10/site-packages/ctransformers/llm.py", line 214, in __init__
    raise RuntimeError(
RuntimeError: Failed to create LLM 'llama' from 'models/llama-2-70b.ggmlv3.q4_K_S.bin'.

Thanks very much for providing all these ggml models. Really awesome!

I tried some of the llama-2-7b, llama-2-13b ggml models and all run without a problem.

llama-2-70b.ggmlv3.q4_K_S.bin however seems to have some problem. I tried it with Python 3.10.6 and ctransformers in Ubuntu 22. I'll probably give another try with llama-2-70b.ggmlv3.q4_0.bin.

This is what I fonud out and the way I understand it: the 70b model is slightly different, needs to set -gqa 8. But ctransformers 0.2.15 is able to handle 70b model now.

mikeee changed discussion status to closed

Sign up or log in to comment