llama.cpp

f6711cef - CUDA: determine FA parallel blocks at runtime