AMP Mixed-Precision Backward
Spot the mistake in this automatic mixed-precision training step.
Codepython
scaler = torch.cuda.amp.GradScaler()
optimizer.zero_grad()
with torch.cuda.amp.autocast():
output = model(x)
loss = criterion(output, y)
loss.backward()
scaler.step(optimizer)
scaler.update()What is the bug in this mixed-precision training step?