fix RecordStreamForGC nullptr exception

[engineer1109]

PR types

Bug fixes

PR changes

Others

Description

某种情况下，同时打开GPU和CUSTOM_DEVICE的编译选项

cmake .. -DWITH_GPU=ON -DWITH_CUSTOM_DEVICE=ON -DWITH_TENSORRT=ON

使用Custom Device进行推理。
会出现空指针的异常。

I0613 13:41:04.066522 3527536 interpretercore.cc:237] New Executor is Running.
I0613 13:41:04.432148 3527536 conditional_block_op.cc:98] [ControlFlow][ConditionalBlock] New Executor is Running.
W0613 13:41:04.549906 3528349 interpretercore.cc:1022] memcpy_h2d raises an exception paddle::PD_Exception, the gpu stream is nullptr.
  [/media/wjl/D2/github/fork/Paddle/paddle/phi/backends/gpu/gpu_context.cc:324]
W0613 13:41:04.552162 3527536 operator.cc:804] while raises an exception std::exception, std::exception
terminate called after throwing an instance of 'std::exception'
  what():  std::exception

原因是 paddle/fluid/framework/new_executor/interpretercore.cc
RecordStreamForGC 这个函数里
有这样一行代码

 gpuStream_t stream =
      reinterpret_cast<const phi::GPUContext&>(instr.DeviceContext()).stream();

不管是GPUContext还是CustomContext 都会执行。

目前仅在这个模型下发现这个问题。
链接: https://pan.baidu.com/s/1CSqbKXVlT1gwIyCkAkDBJg?pwd=twk5 提取码: twk5
修复补丁确认可以修复。

[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.

[engineer1109]

@luotao1 有空能否review一下

[luotao1]

@engineer1109 后续可以直接 @qili93 和 @ronny1996 review哈

[engineer1109]

@ronny1996 How about this PR？

[ronny1996]

LGTM