fix(training): scaler.unscale_() called after step() on optim_d

scaler.update() was only called after optim_g.step(), which resets
the scaler's internal state. On the next iteration, optim_d tried
to call scaler.unscale_() on an already-updated scaler -> crash.

Fix: add scaler.update() right after scaler.step(optim_d) so each
optimizer properly completes its own scaler cycle before the other
one runs.

Author: Z User

advanced_rvc_inference/engine/training/runner/train.py

@@ -762,6 +762,7 @@ def train_and_evaluate(rank, epoch, hps, nets, optims, scaler, train_loader, wri
                 scaler.unscale_(optim_d)
                 grad_norm_d = commons.clip_grad_value(net_d.parameters(), None)
                 scaler.step(optim_d)
+                scaler.update()
             else:
                 loss_disc.backward()
                 grad_norm_d = commons.clip_grad_value(net_d.parameters(), None)