无法使用 CUDA 11.0、Python 3.8、Torch 1.8 编译项目

Question

我正在尝试编译 DiffDVR，一个可微分的渲染器。这首先需要运行ning cmake，所以我必须安装一些依赖项。我想使用与 repo 中的依赖项相同的依赖项。我正在使用 Linux Mint 20.3，它基于 Ubuntu 20.04。按照顺序，我安装了：

CUDA 11.0（使用官方runfile）
cuDNN v8.1.1（与 11.0 兼容的最高版本，根据 cuDNN archive）
使用 Torch 1.8.0 创建了一个 Python 3.8 环境（基于回购中的 environment.yml 文件）
编译需要的 libglm-dev、libglfw3-dev 和 libglew-dev。

进入构建目录并运行ning cmake .. 时出现的错误是：

CMake Error: The following variables are used in this project, but they are set to NOTFOUND.
Please set them or make sure they are set and tested correctly in the CMake files:
TORCH_LIB_c10
    linked by target "Renderer" in directory /home/andrei/PycharmProjects/DiffDVR/renderer
    linked by target "Gui" in directory /home/andrei/PycharmProjects/DiffDVR/gui
TORCH_LIB_c10_cuda
    linked by target "Renderer" in directory /home/andrei/PycharmProjects/DiffDVR/renderer
    linked by target "Gui" in directory /home/andrei/PycharmProjects/DiffDVR/gui
TORCH_LIB_torch_cpu
    linked by target "Renderer" in directory /home/andrei/PycharmProjects/DiffDVR/renderer
    linked by target "Gui" in directory /home/andrei/PycharmProjects/DiffDVR/gui
TORCH_LIB_torch_cuda
    linked by target "Renderer" in directory /home/andrei/PycharmProjects/DiffDVR/renderer
    linked by target "Gui" in directory /home/andrei/PycharmProjects/DiffDVR/gui
TORCH_LIB_torch_python
    linked by target "Renderer" in directory /home/andrei/PycharmProjects/DiffDVR/renderer
    linked by target "Gui" in directory /home/andrei/PycharmProjects/DiffDVR/gui

在输出中向上滚动显示以下问题： Torch: full library list: /usr/lib/libtorch.so;TORCH_LIB_c10-NOTFOUND;TORCH_LIB_c10_cuda-NOTFOUND;/usr/lib/libtorch.so;TORCH_LIB_torch_cpu-NOTFOUND;TORCH_LIB_torch_cuda-NOTFOUND;TORCH_LIB_torch_python-NOTFOUND

所以好像找不到这些库。我尝试 sudo apt-get install libtorch3-dev 但 apt 说它已经安装（我假设它发生在 conda 安装 pytorch 包时？），并且没有可用的库与 c10、c10_cuda 等后缀。我如何确保找到这些库，以便我可以编译项目？

我通过以下（hacky）解决方案取得了一些进展：

sudo ln -s /home/andrei/miniconda3/envs/py38torch18/lib/python3.8/site-packages/torch/lib/libc10.so libc10.so
sudo ln -s /home/andrei/miniconda3/envs/py38torch18/lib/python3.8/site-packages/torch/lib/libc10_cuda.so libc10_cuda.so
sudo ln -s /home/andrei/miniconda3/envs/py38torch18/lib/python3.8/site-packages/torch/lib/libtorch_cpu.so libtorch_cpu.so
sudo ln -s /home/andrei/miniconda3/envs/py38torch18/lib/python3.8/site-packages/torch/lib/libtorch_cuda.so libtorch_cuda.so
sudo ln -s /home/andrei/miniconda3/envs/py38torch18/lib/python3.8/site-packages/torch/lib/libtorch_python.so libtorch_python.so

修复了之前的错误。

很遗憾，我现在遇到以下错误：

CMake Error at renderer/CMakeLists.txt:92 (add_library):
  CUDA_STANDARD is set to invalid value '17'

更多进展：通过 apt 卸载并通过 snap 重新安装（因为 apt 版本太旧），成功管理运行 cmake。不幸的是，make 现在失败了，错误是：

[ 46%] Building CXX object renderer/CMakeFiles/Renderer.dir/volume.cpp.o
In file included from /opt/project/renderer/volume.cpp:1:
/opt/project/renderer/volume.h:9:10: fatal error: torch/types.h: No such file or directory
    9 | #include <torch/types.h>
      |          ^~~~~~~~~~~~~~~
compilation terminated.
make[2]: *** [renderer/CMakeFiles/Renderer.dir/build.make:76: renderer/CMakeFiles/Renderer.dir/volume.cpp.o] Error 1
make[1]: *** [CMakeFiles/Makefile2:244: renderer/CMakeFiles/Renderer.dir/all] Error 2
make: *** [Makefile:91: all] Error 2

Answer 1

更新：手动指定所有路径。就我而言，这是：

cmake .. -DTORCH_PATH=/home/andrei/miniconda3/envs/py38torch18/lib/python3.8/site-packages/torch -DTorch_DIR=/home/andrei/miniconda3/envs/py38torch18/lib/python3.8/site-packages/torch/share/cmake/Torch -DPYTHON_LIBRARY=/home/andrei/miniconda3/envs/py38torch18/lib/libpython3.8.so -DPYTHON_EXECUTABLE=/home/andrei/miniconda3/envs/py38torch18/bin/python

无法使用 CUDA 11.0、Python 3.8、Torch 1.8 编译项目

Failing to compile project using CUDA 11.0, Python 3.8, Torch 1.8

c++

python

cuda

cmake

pytorch