llama.cpp

Gemma 4: move some computations to BF16

#21451

Closed