llama.cpp

f45b59a5 - Revert "quantize : assume the neutral prior is equal imatrix weights"

Revert "quantize : assume the neutral prior is equal imatrix weights" This reverts commit 46a8601140d3e1891316c93f8ea4ed4c73fe7f55. Going towards 1 is more stable than towards the mean, which can otherwise still be 0 and cause problems.