llama.cpp

e9095e60 - async direct io per tensor test