support pytorch1.11

[QingYuan-L]

according to pytorch/pytorch#65492
and pytorch/pytorch#65472

[hukaixuan19970627]

/home/test/Persons/hukaixuan/yolov5/utils/nms_rotated/src/poly_nms_cuda.cu:5:27: fatal error: ATen/ceil_div.h: No such file or directory
compilation terminated.
ninja: build stopped: subcommand failed.

It seems ’ATen/ceil_div.h‘ is missing in Torch1.10.0.

[QingYuan-L]

we can wait for pytoch1.11(stable) to test this pr, if you use the newest NGC image, maybe the pr work. my dev environment : cuda11.5 pytorch1.11 ,nvcr.io/nvidia/pytorch:21.12-py3 from NGC

[Prabhakaran-Gokul]

Thanks for the PR @QingYuan-L. In addition to the changes that you have made, I had to comment out THC/THC.h and include c10/cuda/CUDACachingAllocator.h in the poly_nms_cuda.cu file to make it compile. For example:

//#include <THC/THC.h>
#include <c10/cuda/CUDACachingAllocator.h>

CUDA version: 11.7
Nvcc version: 11.7
PyTorch version: 1.11.0
PyTorch cuda version (ie. torch.version.cuda): 11.3

[lcw99]

Same environment, but I need to change THCudaCheck to C10_CUDA_CHECK to compile.

[Ni-Wo]

I'm running into the same problem, I'm quite new to this field and don't understand everything here. Could someone help me out with a more detailed explanation what to do?
My environment:

• Windows10
• Python 3.9
• nvcc 11.3, nvidia-smi: 11.6
• torch: 1.12
  I already commented out the "#include <THC/THC.h>" and added "#include <c10/cuda/CUDACachingAllocator.h>". But I didn't understand the comments before that.

[LUO77123]

目前支持pytorch1.11吗

[tinly00]

have you solved this problem？

[LUO77123]

I'm running into the same problem, I'm quite new to this field and don't understand everything here. Could someone help me out with a more detailed explanation what to do? My environment:

• Windows10
• Python 3.9
• nvcc 11.3, nvidia-smi: 11.6
• torch: 1.12
  I already commented out the "#include <THC/THC.h>" and added "#include <c10/cuda/CUDACachingAllocator.h>". But I didn't understand the comments before that.

have you solved this problem？

[shotyme]

My solution to compile successfully was to downgrade torch and change the requirements.txt to:

torch==1.9.1
torchvision==0.10.1

[LUO77123]

我改了算子，采用mmcv的，不用编译，直接可以用最新的torch版本1.12

…

---Original--- From: ***@***.***&gt; Date: Tue, Jan 10, 2023 21:51 PM To: ***@***.***&gt;; Cc: ***@***.******@***.***&gt;; Subject: Re: [hukaixuan19970627/yolov5_obb] support pytorch1.11 (PR #152) My solution to compile successfully was to downgrade torch and change the requirements.txt to: torch==1.9.1 torchvision==0.10.1 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***&gt;

[myknowntime]

你好，请问方便提供一下相关代码或者指点一下迷津吗？

[yjoh25]

hi?
Has the above error been resolved?