I assume those don't just work automatically with an off-the-shelf gguf. What do you need in your local inference stack to take advantage of M5's neural accelerators?
replyApple muddied the waters by calling them "neural accelerators" but it seems like what they actually added in the M5 generation is tensor instructions for the existing GPU cores. It's not a separate accelerator like the ANE.
replyllama.cpp's Metal backend does use them when they're available.
They do work with llama.cpp and MLX automatically.
replyApple Mac Studio (M3 Ultra Chip/28 CPU, 60 GPU/96 GB/1 TB
replyHow is this config?