distribute training crashed on develop branch
Created by: Yancey1989
Run the demo code test_word2vec.py
with 2pservers + 2trainers, other demos can also reproduce this bug:
I0514 07:31:17.266929 11136 operator.cc:546] expected_kernel_key:data_type[float32]:data_layout[ANY_LAYOUT]:place[CPUPlace]:library_type[PLAIN]
PC: @ 0x0 (unknown)
*** SIGSEGV (@0x0) received by PID 10770 (TID 0x7f8132309700) from PID 0; stack trace: ***
@ 0x7f83817d6390 (unknown)
@ 0x7f8334340dc7 std::_Hashtable<>::find()
@ 0x7f833433eb69 paddle::framework::Scope::FindVarLocally()
@ 0x7f833433eba4 paddle::framework::Scope::FindVar()
@ 0x7f833432aa1f paddle::framework::ExecutionContext::Input<>()
@ 0x7f8334230240 paddle::operators::SGDOpKernel<>::Compute()
@ 0x7f833432dac6 paddle::framework::OperatorWithKernel::RunImpl()
@ 0x7f833432a078 paddle::framework::OperatorBase::Run()
@ 0x7f8333b83ace paddle::framework::Executor::RunPreparedContext()