pytorch3d
pytorch3d copied to clipboard
facebookresearch
Pytorch3D install error
❓ Pytorch3D Install Error
Hi everybody,
My pytorch version is 1.12.1 cuda 11.6 cudnn is 8.4.
While installing pytorch3d I got an error like in (https://stackoverflow.com/questions/62304087/installing-pytorch3d-fails-with-anaconda-and-pip-on-windows-10).
Then, I ran pip install "git+https://github.com/facebookresearch/pytorch3d.git" with my newly added environment variables.
I got this error.
Thank you...
Can you all confirm whether this is happening on Windows? A few days ago we enabled C++17 features and used some in iou_utils code. The error is complaining because the CUDA compiler doesn't understand structured bindings. It worked on Linux but maybe something's wrong with the Windows compiler. We'll look into it next week, pls let me know if any of you is using windows, also your pytorch/cuda/cudnn versions.
@mehdizade It looks like nvcc 10.2 is the last version that doesn't support C++17 (see https://gist.github.com/ax3l/9489132#device-side-c-standard-support). We'd like to be able to use C++17 features, but I'm not sure how long we want to support CUDA < 11. Let me get back to you on that on Monday.
But 11.x should work. @ilyasyalcin can you do nvcc.exe --version and share your nvcc version? (Apparently "starting with CUDA Toolkit 11.0.1, nvcc and CUDA Toolkit versions are not equivalent anymore")
First of all, thank you your contributions. (@kjchalup )
I have windows platform.
My nvcc version is shown below as screenshot.
NVCC
@mehdizade It looks like nvcc 10.2 is the last version that doesn't support C++17 (see https://gist.github.com/ax3l/9489132#device-side-c-standard-support). We'd like to be able to use C++17 features, but I'm not sure how long we want to support CUDA < 11. Let me get back to you on that on Monday.
But 11.x should work. @ilyasyalcin can you do nvcc.exe --version and share your nvcc version? (Apparently "starting with CUDA Toolkit 11.0.1, nvcc and CUDA Toolkit versions are not equivalent anymore")
I am using linux with CUDA=11.4, it never worked for me. I have tried all the CUDA and gcc version. Is there some suggestion?
The main branch has just been updated with 95771985b7ac14dae4606d799e82703b530d4e0c to remove the use of C++17 features. This specific error should not happen any more.
For a few days recently, the main branch using C++17 features but not setting that with a flag on the compiler, which is likely to have broken linux builds. @keloee and @mehdizade hopefully that should fix
The latest release v0.7.0 is still fine. Anyone having problems with the actual released code or new build problems should open a new issue.
@czy2014hust That commit doesn't affect installation of PyTorch3D.
I have a amd gpu (Radeon RX Vega 11 Graphics) with windows 10. I didn't installed cuda, rather i installed pytorch with cpu option. Next, when i try to install pytorch3d via github (as instructed on official website). I also see kinda same error, and i can't see a specific error to solve. Can anyone help?
