llama.cpp

f4837d3e - CUDA: tune GLM 4.7 Flash FA kernel selection logic (DGX Spark)