Error using outside docker environment

[ShaoxiongYao]

I have built PyFlex in the docker environment and the outputs are like
-- Configuring done -- Generating done -- Build files have been written to: /workspace/PyFleX/bindings/build
after cmake .. and
[100%] Built target pyflex
after make -j

However, when I moved out of the docker environment and try to use PyFleX, I first executed commands to export environment paths and I can import pyflex in python. When I try to run pyflex.init(), this error happens,
Reshaping Flex: Could not open CUDA driver, ensure you have the latest driver (nvcuda.dll) installed and an NVIDIA GPU present on the system - ./../../../src/cuda/util.cpp:124 Flex: Could not initialize CUDA driver - ../../../src/cuda/flex.cu:4624 Could not initialize Flex, exiting.

But I checked I have the CUDA driver installed by running cat /proc/driver/nvidia/version
NVRM version: NVIDIA UNIX x86_64 Kernel Module 440.82 Wed Apr 1 20:04:33 UTC 2020 GCC version: gcc version 7.5.0 (Ubuntu 7.5.0-3ubuntu1~18.04)

I wonder if you had a similar experience before and I am looking forward to your help.

[Chen-Cai-OSU]

Hello,

I was wondering do you figure this out? In my case, I had

libGL error: No matching fbConfigs or visuals found
libGL error: failed to load driver: swrast
Could not initialize GL extensions

when run pyflex.init().

Do you happen to know how to resolve this? Many thanks!

[ShaoxiongYao]

Sorry I did not work further on this problem...

[YunzhuLi]

I haven't encountered this issue myself, and the issue seems to relate to the OpenGL setup.

Xingyu from CMU built on top of this repo and developed SoftGym: https://github.com/Xingyu-Lin/softgym. The compilation process and the supported working environments should be very similar, and people have been documenting their experience in using the package https://danieltakeshi.github.io/2021/02/20/softgym/, which may be a helpful resource.

[Chen-Cai-OSU]

Thank a lot @YunzhuLi . I am able to make it work now. One quick question: I can use SoftGym and PyFlex in a remote machine using Teamviewer. Do you know how to forward the graphics in the remote machine to the local machine, so that I can run the simulation in the remote machine but see the animation on my local computer (which has no gpu)?

Many thanks!

[DanielTakeshi]

@Chen-Cai-OSU if memory serves (and if my blog post is accurate :) ) then I believe SoftGym appends frames from the animation, so you can save the frames and then just replay the animation locally. There may be ways to automate this process as well.

[adamhkan]

@ShaoxiongYao
I had this same issue when running Flingbot, which uses the PyFlex simulator. My solution was to uninstall my local CUDA packages (even though I'm running PyFlex in a conda environment), and locally install the CUDA version identical to the one I was using in my conda env.

I was running cudatoolkit version 11.0.221 in my conda environment, with all the correct conda packages installed. My local CUDA version was 10.1, but I had several local versions installed, and I think this could have been the issue.

I first uninstalled nvidia-cuda-toolkit using apt. I then followed the official NVIDIA CUDA instalation guide. I only did sections 2, 3, and 13.1. Specifically, in section 2.9, I made sure that I deleted all cuda versions I had installed using sudo apt-get --purge remove cuda, sudo apt-get --purge remove cuda-10.1, etc, as well as sudo apt-get autoremove. Finally, I reinstalled nvidia-cuda-toolkit. After this process, running nvcc -V in terminal displayed that I had CUDA version 11.0, just like my conda env.