llama.cpp

39b96f8f - ggml: backend-agnostic tensor parallelism