3 ms·
You mean autotune? I think 10 minutes is pretty normal, torch.compile('max-autotune') can be much slower than that for large models.
by anvuong 11mo ago
You mean autotune? I think 10 minutes is pretty normal, torch.compile('max-autotune') can be much slower than that for large models.
- Mars008 11mo agoAdd to that it can be done only once by developers before distribution for major hardware. Configs saved. Then on client side selected.