4 ms·

Is it normal to spend 10minutes on tuning nowadays? Do we need to spend another 10 minutes upon changing the code?

• anvuong 10mo 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 10mo ago
    Add to that it can be done only once by developers before distribution for major hardware. Configs saved. Then on client side selected.