Conversion process

#1
by AlfredWALLACE - opened

Thanks for the quantized model which allows us to test this great AI.
Would you share your conversion method as I was not able to do it myself with llama.cpp scripts and would like to quantize more versions ?

Sure @AlfredWALLACE
You have to download and compile llama.cpp from github

git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
make LLAMA_CUBLAS=1 

then you need to create python env and install requirements of llama.cpp

 pip install -r requirements.txt

then run the convert script to make the f16 format

python ~/dev/llama.cpp/convert.py ./Magicoder-S-CL-7B --outtype f16

then run the compiled app quantize which will be generated after compiling llama.cpp

quantize ./Magicoder-S-CL-7B/ggml-model-f16.gguf q5_k_m

Good Luck.

Thanks! I had no luck with loading the model quantized with the same commands, previous to my post, but with a S-DS model.

Thanks! I had no luck with loading the model quantized with the same commands, previous to my post, but with a S-DS model.

try this fork , will work for sure
https://github.com/akhil3417/llama.cpp

Owner

Thanks! I had no luck with loading the model quantized with the same commands, previous to my post, but with a S-DS model.

try this fork , will work for sure
https://github.com/akhil3417/llama.cpp

Could you please explain what is the changes or features in your fork ?

merged '417884e regex_gpt2_preprocess pr'

Thanks! I'll try! in the mean time, the GPTQ version works really well and also loads on low VRAM.

Sign up or log in to comment