llama.cpp

Avoid unnecessarily disabling CUDA graphs

#7302

Merged