b1299
f5ef5cfb · ggml-cuda : perform cublas mat mul of quantized types as f16 (#3412) ·