llama.cpp

a0d9dd20 - ggml: backend-agnostic tensor parallelism