diff --git a/python/paddle/fluid/backward.py b/python/paddle/fluid/backward.py index 7af6ed1463ab737e871da487f2a687301652ef2d..32b1b65bd97ef1e512a5880843509611b606f52d 100644 --- a/python/paddle/fluid/backward.py +++ b/python/paddle/fluid/backward.py @@ -480,6 +480,8 @@ def append_backward(loss, parameter_list=None, no_grad_set=None, program.current_block_idx = current_block_idx program.sync_with_cpp() + # FIXME(zcd): prevent loss.grad optimized by mem_opt. + loss.block.var(_append_grad_suffix_(loss.name)).persistable = True if parameter_list is not None: parameters = parameter_list