llama.cpp

6272b676 - use stride=128 if built for tensor cores