question on training scheme

[xianshunw]

It seems that the network doesn't use the previous hidden state in training phase:

with autocast(enabled=not self.args.disable_mixed_precision):
            pred_fgr, pred_pha = self.model_ddp(true_src, downsample_ratio=downsample_ratio)[:2]
            loss = matting_loss(pred_fgr, pred_pha, true_fgr, true_pha)

self.scaler.scale(loss['total']).backward()
self.scaler.step(self.optimizer)
self.scaler.update()
self.optimizer.zero_grad()

But it is fed into the network in the test phase.

src = src.to(device, dtype, non_blocking=True).unsqueeze(0) # [B, T, C, H, W]
fgr, pha, *rec = model(src, *rec, downsample_ratio)

Why does the network use different feedforward scheme in these two stage. Will it be better to take the hidden state as input during training stage?