llama.cpp

66d65ec2 - cuda: cap grid.y at 65535 in non-contiguous dequantize/convert kernels (#19999)