ppc64le build breaks after #10478.
I am building with :

export USE_SYSTEM_NCCL=TRUE
export NCCL_ROOT_DIR=/usr/local/cuda/
export NCCL_LIB_DIR=/usr/local/cuda/lib64
export NCCL_INCLUDE_DIR=/usr/local/cuda/include

I have NCCL2 from nvidia
cuda 9.2
cudnn 7

here is the log :
/pytorch/torch/lib/c10d/test/ProcessGroupNCCLTest.cpp: In member function ‘std::vectorat::Tensor NCCLTest::getTensors()’:
/pytorch/torch/lib/c10d/test/ProcessGroupNCCLTest.cpp:99:49: error: no matching function for call to ‘at::Tensor::toBackend(const at::DeviceType&)’
outputs[i] = inputs_[i].toBackend(at::kCPU);
^
/pytorch/torch/lib/c10d/test/ProcessGroupNCCLTest.cpp:99:49: note: candidate is:
In file included from /pytorch/torch/lib/tmp_install/include/ATen/NativeFunctions.h:7:0,
from /pytorch/torch/lib/tmp_install/include/ATen/Functions.h:11,
from /pytorch/torch/lib/tmp_install/include/ATen/ATen.h:13,
from /pytorch/torch/lib/c10d/../c10d/ProcessGroup.hpp:7,
from /pytorch/torch/lib/c10d/../c10d/ProcessGroupNCCL.hpp:8,
from /pytorch/torch/lib/c10d/test/ProcessGroupNCCLTest.cpp:5:
/pytorch/torch/lib/tmp_install/include/ATen/TensorMethods.h:38:15: note: at::Tensor at::Tensor::toBackend(at::Backend) const
inline Tensor Tensor::toBackend(Backend b) const {
^
/pytorch/torch/lib/tmp_install/include/ATen/TensorMethods.h:38:15: note: no known conversion for argument 1 from ‘const at::DeviceType’ to ‘at::Backend’
/pytorch/torch/lib/c10d/test/ProcessGroupNCCLTest.cpp: In member function ‘std::vector<std::vectorat::Tensor > NCCLTest::getOutputTensors()’:
/pytorch/torch/lib/c10d/test/ProcessGroupNCCLTest.cpp:118:58: error: no matching function for call to ‘at::Tensor::toBackend(const at::DeviceType&)’
outputs[i][j] = outputs_[i][j].toBackend(at::kCPU);
^

I noticed there is another reported build break pointing to backend() == kCPU as well ( #10678)