llama.cpp

e57f5233 - ggml-cuda: fixes for concurrent streams (#18496)