paddle 版本：1.5.2
问题：
prob = layers.softmax(predict, use_cudnn=True)
prob_tec = layers.log(prob)

p, q = layers.split(prob, num_or_sections=2, dim=0)
p_tec, q_tec = layers.split(prob_tec, num_or_sections=2, dim=0)

其中，prob shape是[batch, len, dim]，目的是要把prob和prob_tec按照batch都平分，
p, q = layers.split(prob, num_or_sections=2, dim=0) 正常
p_tec, q_tec = layers.split(prob_tec, num_or_sections=2, dim=0) 报错：
in split
'axis': dim
File "src/model.py", line 589, in transformer
p_tec, q_tec = layers.split(prob_tec, num_or_sections=2, dim=0)
File "src/train.py", line 641, in train
is_test=False)
File "src/train.py", line 745, in
train(args)
C++ Call stacks:
Cannot find variable split_1.tmp_0@GRAD at [/paddle/paddle/fluid/framework/op_desc.cc:195]