What's the correct way of using AMP with multiple losses?

[mcarilli (Michael Carilli)]

@Zequn_Qin

scaler = torch.cuda.amp.GradScaler()
with autocast():
    loss0 = some_loss
    loss1 = another_loss
    loss_all = w0 * loss0 + w1 * loss1

scaler.scale(loss_all).backward()

yes that would work, and you could do the same without amp.

There’s no need for retain_graph=True on the first backward unless the two losses share backward graph. scaler.scale(loss0).backward(retain_graph=True) in the example only uses retain_graph=True because the two losses share some backward graph, it has nothing to do with amp.