using pytorch1.11, I still can not compile DCNv2 because of #include <ATen/ceil_div.h>

[xiao-pengyou]

gpu 3080ti

nvidia-smi show Driver version 510.47.03 CUDA Version: 11.6 nvcc -V show Cuda compilation tools, release 11.4, V11.4.120

when I create conda env I use create -n FairMOT python=3.8 when I install pytorch ,use conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch then I check the version , input python, import torch torch.version 1.11 torch.version.cuda 11.3 torch.backends.cudnn.version() 8200

then input sudo python3 setup.py build develop fatal error: ATen/ceil_div.h: 没有那个文件或目录

include <ATen/ceil_div.h>

Is there something I do wrong?

[Natsu-Akatsuki]

try python3 setup.py build develop without sudo...

[xiao-pengyou]

try python3 setup.py build develop without sudo...

Thank you very much. I find when add sudo，the compiler use the default environment(python3.6) instead of python3.8 which I bulid for the project(python3.8)

[JehanJay]

I do have the same error as @xiao-pengyou fatal error: ATen/ceil_div.h: No such file or directory In my case, I am using a conda environment developed with Pytorch LTS 1.8.2 and Cudatoolkit 11.1 Nvidia 3050Ti

try python3 setup.py build develop without sudo...

@Natsu-Akatsuki Tried this way, but still, the issue remains.

What exactly would be the bug causing this error?

[Natsu-Akatsuki]

Try to use the previous version. i.e. git checkout fa9b2fd740ced2a22e0e7e913c3bf3934fd08098 Please read the README carefully.

[JehanJay]

I do have the same error as @xiao-pengyou fatal error: ATen/ceil_div.h: No such file or directory In my case, I am using a conda environment developed with Pytorch LTS 1.8.2 and Cudatoolkit 11.1 Nvidia 3050Ti

try python3 setup.py build develop without sudo...

@Natsu-Akatsuki Tried this way, but still, the issue remains.

What exactly would be the bug causing this error?

However, DCnv2 got successfully compiled with Cudatoolkit 11.3 with PyTorch stable 1.11.0.

[Natsu-Akatsuki]

What I mean is that the lateset version of DCNv2 only support pytorch 1.11.+. On the other hand, when you use pytorch 1.8.2 you need to use previous version.

[xiao-pengyou]

What I mean is that the lateset version of DCNv2 only support pytorch 1.11.+. On the other hand, when you use pytorch 1.8.2 you need to use previous version.

You can find the latest DCNv2 in the issue ,which support Pytorch 1.12, or other version support 1.6, 1.7, 1.8.

[JehanJay]

What I mean is that the lateset version of DCNv2 only support pytorch 1.11.+. On the other hand, when you use pytorch 1.8.2 you need to use previous version.

For the previous version of DCNv2, PyTorch 1.8.2 was not supported and supported only versions <=1.7.0. The issue with that was, PyTorch 1.7.0 is not supported with Cudatoolkit 11.1 which is the oldest compatible version with CUDA SDK 8.6. So as far as I understood the issue, conceptually the previous DCNv2 is no longer supported with newer GPU architectures.

[DhruvKoolRajamani]

Just add #ifdef macro flags to check for the version of PyTorch before the ATen library include.

[Albert337]

What I mean is that the lateset version of DCNv2 only support pytorch 1.11.+. On the other hand, when you use pytorch 1.8.2 you need to use previous version.

For the previous version of DCNv2, PyTorch 1.8.2 was not supported and supported only versions <=1.7.0. The issue with that was, PyTorch 1.7.0 is not supported with Cudatoolkit 11.1 which is the oldest compatible version with CUDA SDK 8.6. So as far as I understood the issue, conceptually the previous DCNv2 is no longer supported with newer GPU architectures.

my docker environment is pytorch 1.8.0 and cuda 11.1, i met the same error as you mean,i need change pytorch to 1.7.0 version?

[cvcvcccq0912]

What I mean is that the lateset version of DCNv2 only support pytorch 1.11.+. On the other hand, when you use pytorch 1.8.2 you need to use previous version.

For the previous version of DCNv2, PyTorch 1.8.2 was not supported and supported only versions <=1.7.0. The issue with that was, PyTorch 1.7.0 is not supported with Cudatoolkit 11.1 which is the oldest compatible version with CUDA SDK 8.6. So as far as I understood the issue, conceptually the previous DCNv2 is no longer supported with newer GPU architectures.

my docker environment is pytorch 1.8.0 and cuda 11.1, i met the same error as you mean,i need change pytorch to 1.7.0 version?

I'm having the same problem as you, as are the pytorch and cuda versions. But I can't lower my pytoch version any more. Have you solved your problem? How to solve it?