scaler.scale(loss).backward pytorch 1.6+

RuntimeError: value cannot be converted to type at::Half without overflow:

Could you post a minimal code snippet to reproduce this issue, please?