Error while building PyTorch with nccl codes

[topcuburak]

Issue description

As I shared in the terminal screenshot, I observed an error while building PyTorch to run with GPGPU-Sim. The error states that cudaPointerAttributes does not have a memoryType member. This member is not available for the newest versions of CUDA.

• https://docs.nvidia.com/cuda/cuda-runtime-api/structcudaPointerAttributes.html#structcudaPointerAttributes
• Current version does not build NVIDIA/NvPipe#29

Code example

System Info

• PyTorch or Caffe2: PyTorch
• How you installed PyTorch (conda, pip, source): pip3 torchvision=0.2.2 (as mentioned in the manual)
• Build command you used (if compiling from source): At one step, we need to run setup.py to build related packages (as far as I understand)
• OS: ubuntu 18.04 LTS
• PyTorch version:
• Python version: 2.7
• CUDA/cuDNN version: v11.4 / v8.2.2
• GPU models and configuration: a simple intel GPU, namely Intel Corporation HD Graphics 530 :(
• GCC version (if compiling from source): GCC version is 5.5.
• CMake version: 3.10.2
• Versions of any other relevant libraries: If you need any additional configuration info, I can share as fast as I can do.

[topcuburak]

1. [email protected] +cuda build fails: class "cudaPointerAttributes" has no member "memoryType" spack/spack#27180

2. Cuda 10 does not support built with PyTorch 0.4.1. Here is the issue: Has this code been tested with cuda 10.0? open-mmlab/mmdetection#53

[KangkangStu]

Did you run it successfully？