mirror of
https://github.com/ikawrakow/ik_llama.cpp.git
synced 2026-07-01 07:50:16 -05:00
I always wanted to know if transposing the model tensors may improve quantization. If for whatever reason there was a correlation between weights in different rows but at the same position within a row, a transposed version of the tensor would quantize better. This commit tried and, nope, no luck.