refine eager backward

[wanghuancoder]

PR Category

Execute Infrastructure

PR Types

Improvements

Description

Paddle 对于如下代码会报错：

import paddle

a = paddle.ones([100])
a.stop_gradient = False
b = paddle.ones([100])
b.stop_gradient = False

c = a + b
d = c + c
e = d + d
# paddle_out_grads = paddle.autograd.backward([e], [paddle.ones([100])]) # rigth
paddle_out_grads = paddle.autograd.backward([e,c,d], [paddle.ones([100]),paddle.ones([100]),paddle.ones([100])])

报错内容为：

Traceback (most recent call last):
  File "/host_home/wanghuan29/Paddle/test_grad.py", line 13, in <module>
    paddle_out_grads = paddle.autograd.backward([e,c,d], [paddle.ones([100]),paddle.ones([100]),paddle.ones([100])])
  File "/usr/local/lib/python3.9/dist-packages/decorator.py", line 232, in fun
    return caller(func, *(extras + args), **kw)
  File "/host_home/wanghuan29/Paddle2/build/python/paddle/base/wrapped_decorator.py", line 40, in __impl__
    return wrapped_func(*args, **kwargs)
  File "/host_home/wanghuan29/Paddle2/build/python/paddle/base/framework.py", line 726, in __impl__
    return func(*args, **kwargs)
  File "/host_home/wanghuan29/Paddle2/build/python/paddle/autograd/backward_mode.py", line 140, in backward
    core.eager.run_backward(tensors, grad_tensors, retain_graph)
SystemError: (Fatal) Unable to find next node in the GradTensorHolder 
Trying to run Node without configuring its GradTensorHolder.
  [Hint: Expected node_input_buffer_iter != node_input_buffers_dict.end(), but received node_input_buffer_iter == node_input_buffers_dict.end().] (at /host_home/wanghuan29/Paddle2/paddle/fluid/eager/backward.cc:272)

如下代码有同样的问题：

import paddle

a = paddle.ones([100])
a.stop_gradient = False
b = paddle.ones([100])
b.stop_gradient = False

c = a + b
d = c + c
e = d + d
f, g = paddle.split(e, num_or_sections=2, axis=0)
h = f + f

paddle_out_grads = paddle.autograd.backward([h,g], [paddle.ones([50]),paddle.ones([50])])

第一种情况，用户提供了c和d的梯度，那么通过反向回传获得的梯度该怎么办。实测torch.autograd.grad和paddle.grad都是将用户提供的梯度以及回传算出的梯度加和。因此，paddle.autograd.backward也有必要这么做。
第二种情况，是paddle.autograd.backward存在bug。

这两种错误在paddle.autograd.backward时会发生，在paddle.grad时不会发生，因为paddle.grad会走PreparedForGeneralGrad，对反向图做剪枝等优化。

Pcard-67164

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Please upload report for BASE (develop@9fd40f9). Learn more about missing BASE report.

Additional details and impacted files

@@             Coverage Diff             @@
##             develop    #72816   +/-   ##
===========================================
  Coverage           ?   100.00%           
===========================================
  Files              ?         1           
  Lines              ?         5           
  Branches           ?         0           
===========================================
  Hits               ?         5           
  Misses             ?         0           
  Partials           ?         0           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[JiabinYang]

LGTM