[anvuong]

You mean autotune? I think 10 minutes is pretty normal, torch.compile('max-autotune') can be much slower than that for large models.

[Mars008]

Add to that it can be done only once by developers before distribution for major hardware. Configs saved. Then on client side selected.