add IndexPutGradInfermeta to fix backward error in static-mode

[zoooo0820]

PR types

Bug fixes

PR changes

OPs

Description

Card-73285
Since the auto-generated code is different between dynamic and static mode, Grad OP index_put_grad works fine in dynamic mode but raises error in static mode.

I0720 12:47:08.784021  8805 op_call_stack.cc:62] InvalidArgumentError: shape of value can't not be broadcast to shape of x[indices] (at /paddle/Paddle/paddle/phi/kernels/funcs/index_put_utils.h:286)
Traceback (most recent call last):
  File "test_index_put_grad.py", line 50, in <module>
    res = exe.run(fetch_list=[z, x.grad_name, value.grad_name])
  File "/usr/local/lib/python3.7/dist-packages/paddle/fluid/executor.py", line 1444, in run
    use_prune=use_prune,
  File "/usr/local/lib/python3.7/dist-packages/paddle/fluid/executor.py", line 1669, in _run_impl
    ret = new_exe.run(list(feed.keys()), return_numpy)
  File "/usr/local/lib/python3.7/dist-packages/paddle/fluid/executor.py", line 684, in run
    tensors = self._new_exe.run(feed_names)._move_to_list()
ValueError: In user code:

    File "test_index_put_grad.py", line 44, in <module>
      z  = paddle.index_put(y, (index1,), value)
    File "/usr/local/lib/python3.7/dist-packages/paddle/tensor/manipulation.py", line 4918, in index_put
      attrs={'accumulate': accumulate},
    File "/usr/local/lib/python3.7/dist-packages/paddle/fluid/layer_helper.py", line 45, in append_op
      return self.main_program.current_block().append_op(*args, **kwargs)
    File "/usr/local/lib/python3.7/dist-packages/paddle/fluid/framework.py", line 4033, in append_op
      attrs=kwargs.get("attrs", None),
    File "/usr/local/lib/python3.7/dist-packages/paddle/fluid/framework.py", line 2795, in __init__
      for frame in traceback.extract_stack():

    InvalidArgumentError: shape of value can't not be broadcast to shape of x[indices] (at /paddle/Paddle/paddle/phi/kernels/funcs/index_put_utils.h:286)
      [operator < index_put_grad > error]

The essence of the problem is in GeneralBinaryGradInferMeta , the second input (indices[0]) is used incorrectly in the static graph mode, resulting in error size of value_grad. This PR fixes this by adding IndexPutGradInferMeta.

[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.

[Courtesy-Xs]

LGTM

[jeff41404]

LGTM