Great work! Can the FP16 weights be released?
#6
by
digitous
- opened
If you don't mind. The 13b ones are released, for parity sake and model merging it would be a great boon.
Otherwise I'm going to have to figure out how to adjust this script to reverse a q_5
https://github.com/ductai199x/llama.cpp/blob/84ba1fd25b1a0f12d11a655fecf90c7cbb0babb0/convert_ggml_to_pth.py