param_generator should be in original precision
#1
by Fgdfgfthgr - opened
This weight group appears particularly susceptible to quantisation error, producing visible noise even at q_8 precision.
This weight group appears particularly susceptible to quantisation error, producing visible noise even at q_8 precision.
Hi. Thanks for the notice. I added quant support for chroma_x0 using a custom llama.cpp build, so there might be some issues. I've made new quants for a newer version of x0 with some fixes — hopefully without problems this time