llama.cpp

57819b8d - llama : disable graph reuse with pipeline parallelism (#20463)