llama.cpp

Improve cuBLAS performance by using a memory pool

#1094

Merged