view article Article One model, two accelerators: faster prefill and decode on the M6 with the Neural Engine and the GPU together anemll • 2 days ago