Can't install PyTorch0.4.1 From Source

1

I am trying to build Pytorch0.4.1 from the source on my Ubuntu18.04 desktop PC. The environment details are as follows,

  • OS: Ubuntu 18.04
  • GPU: RTX 2080 Supper
  • CUDA: 10.2
  • Anaconda Environment with Python 3.6

I used the following commands to install Pytorch from source,

  1. conda install numpy ninja pyyaml mkl mkl-include setuptools cmake cffi
  2. conda install -c pytorch magma-cuda102
  3. git clone --recursive https://github.com/pytorch/pytorch
  4. cd pytorch
  5. git checkout v0.4.1
  6. git submodule sync
  7. git submodule update --init --recursive
  8. export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/…/"}
  9. python setup.py install

I am getting the following error
CMakeFiles/Makefile2:923: recipe for target ‘caffe2/CMakeFiles/caffe2_gpu.dir/all’ failed
make[1]: *** [caffe2/CMakeFiles/caffe2_gpu.dir/all] Error 2
Makefile:140: recipe for target ‘all’ failed
make: *** [all] Error 2
Failed to run ‘bash tools/build_pytorch_libs.sh --use-cuda --use-nnpack nccl caffe2 nanopb libshm gloo THD c10d’

Any workaround? Thanks

2

Could you check the complete log for another preceding error message?
The posted log does only show the failure, but usually you should see some errors before that.

3

Hi @ptrblck,

Here are the complete logs.

[ 71%] Building NVCC (Device) object caffe2/CMakeFiles/caffe2_gpu.dir/__/aten/src/THC/generated/caffe2_gpu_generated_THCTensorMathCompareTByte.cu.o
/home/maaz/mmaaz60/pytorch_from_source/pytorch-0.4.1/aten/src/THC/THCAtomics.cuh(100): error: cannot overload functions distinguished by return type alone

/home/maaz/mmaaz60/pytorch_from_source/pytorch-0.4.1/aten/src/THC/THCAtomics.cuh(123): error: return value type does not match the function type

2 errors detected in the compilation of “/tmp/tmpxft_00006653_00000000-6_THCTensorScatterGather.cpp1.ii”.
CMake Error at caffe2_gpu_generated_THCTensorScatterGather.cu.o.Release.cmake:279 (message):
Error generating file
/home/maaz/mmaaz60/pytorch_from_source/pytorch-0.4.1/build/caffe2/CMakeFiles/caffe2_gpu.dir/__/aten/src/THC/./caffe2_gpu_generated_THCTensorScatterGather.cu.o

caffe2/CMakeFiles/caffe2_gpu.dir/build.make:9994: recipe for target 'caffe2/CMakeFiles/caffe2_gpu.dir//aten/src/THC/caffe2_gpu_generated_THCTensorScatterGather.cu.o’ failed
make[2]: *** [caffe2/CMakeFiles/caffe2_gpu.dir//aten/src/THC/caffe2_gpu_generated_THCTensorScatterGather.cu.o] Error 1
make[2]: *** Waiting for unfinished jobs…
/home/maaz/mmaaz60/pytorch_from_source/pytorch-0.4.1/aten/src/THC/THCAtomics.cuh(100): error: cannot overload functions distinguished by return type alone

/home/maaz/mmaaz60/pytorch_from_source/pytorch-0.4.1/aten/src/THC/THCAtomics.cuh(123): error: return value type does not match the function type

2 errors detected in the compilation of “/tmp/tmpxft_00006699_00000000-6_THCTensorIndex.cpp1.ii”.
CMake Error at caffe2_gpu_generated_THCTensorIndex.cu.o.Release.cmake:279 (message):
Error generating file
/home/maaz/mmaaz60/pytorch_from_source/pytorch-0.4.1/build/caffe2/CMakeFiles/caffe2_gpu.dir/__/aten/src/THC/./caffe2_gpu_generated_THCTensorIndex.cu.o

caffe2/CMakeFiles/caffe2_gpu.dir/build.make:8536: recipe for target 'caffe2/CMakeFiles/caffe2_gpu.dir//aten/src/THC/caffe2_gpu_generated_THCTensorIndex.cu.o’ failed
make[2]: *** [caffe2/CMakeFiles/caffe2_gpu.dir//aten/src/THC/caffe2_gpu_generated_THCTensorIndex.cu.o] Error 1
CMakeFiles/Makefile2:923: recipe for target ‘caffe2/CMakeFiles/caffe2_gpu.dir/all’ failed
make[1]: *** [caffe2/CMakeFiles/caffe2_gpu.dir/all] Error 2
Makefile:140: recipe for target ‘all’ failed
make: *** [all] Error 2
Failed to run ‘bash tools/build_pytorch_libs.sh --use-cuda --use-nnpack nccl caffe2 nanopb libshm gloo THD c10d’

4

Moreover, I also tried building PyTorch0.4.1 with Cuda10.0, and same error occurred. Looking forward to hearing from you.

5

Thanks for the logs. The issue seems to have been solved in this cross-post.