llama.cpp

e0a40022 - CUDA: fixed row rounding for 0 tensor splits (#4594)