llama.cpp

3b1c4e76 - cuda : speed-up reduce part of the kernel