Windows torchvision-0.11.1 CUDA-10.2 wheel packages are not bundled with nvjpeg dll

[malfet]

🐛 Describe the bug

Steps to reproduce:

1. Download TorchVision 0.11.1 windows wheel from https://download.pytorch.org/whl/cu102/torchvision/
2. Unpack the binary
3. Run dumpbin /dependents image.pyd to make sure it depends on nvjpeg64_10.dll:

C:\Users\circleci\Downloads\torchvision>dumpbin /dependents image.pyd
Microsoft (R) COFF/PE Dumper Version 14.00.24245.0
Copyright (C) Microsoft Corporation.  All rights reserved.


Dump of file image.pyd

File Type: DLL

  Image has the following dependencies:

    api-ms-win-crt-heap-l1-1-0.dll
    api-ms-win-crt-runtime-l1-1-0.dll
    api-ms-win-crt-filesystem-l1-1-0.dll
    api-ms-win-crt-stdio-l1-1-0.dll
    libpng16.dll
    nvjpeg64_10.dll
    c10.dll
    torch_cpu.dll
    cudart64_102.dll
    c10_cuda.dll
    MSVCP140.dll
    KERNEL32.dll
    VCRUNTIME140_1.dll
    VCRUNTIME140.dll
    api-ms-win-crt-string-l1-1-0.dll
    api-ms-win-crt-environment-l1-1-0.dll

  Summary

        2000 .data
        3000 .pdata
       2E000 .rdata
        1000 .reloc
        1000 .rsrc
       3C000 .text

4. Noticesthat nvjpeg64_10.dll is missing from the archive

This causes torchvision.io.read_image fail on Windows

Versions

0.11.1 win+cu102

cc @peterjc123 @nbcsm @guyang3532 @maxluk @gunandrose4u @mszhanyi

[mega-optimus]

Hi,
I just installed torchvision==0.11.1+cu102 on Windows 7 and get this "nvjpeg64_10.dll is missing" error window. Where can I find this dll manually and fix the blocking problem? Thank you!

[malfet]

nvjpeg64_10.dll can be downloaded from https://developer.nvidia.com/cuda-10.2-download-archive

[mega-optimus]

nvjpeg64_10.dll can be downloaded from https://developer.nvidia.com/cuda-10.2-download-archive

Thanks!
A bit curious, I've installed CUDA 10.2 already.

[puyebu]

I had the same problem and this is the only page I found,

Python/Pytorch does not search the path Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2\bin for some reason, despite the path being specified in the environment, therefore causing dll error.
I couldn't find a proper solution and instead made hardlinks to the missing dlls in the Python\Lib\site-packages\torch\lib folder using mklink /H cmd command.

I am sure there must be a better way but I couldn't figure it out.

[you74674]

I also have this problem, and I fix it by manually add cuda path using os.add_dll_directory.
I think it's related to #4787? Although #4893 claimed to fix it.
Looks like the problem is _get_extension_path only add the package folder as dll directory but not any other path, but adding all path to dll directory doesn't seem to be the right way.

[jwdunn1]

Same problem. Thanks to @puyebu, I was able to figure out a temporary solution. The complete command I used (running as administrator) was:
> mklink /H nvjpeg64_10.dll "c:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2\bin\nvjpeg64_10.dll"

[panicwq]

I had this problem today.
After installing cuda10.2 on windows server2012 r2 with pip install torch==1.10.1+cu102 torchvision==0.11.2+cu102 torchaudio==0.10.1, it comes up with the missing nvjpeg64_10.dll popup.

My temporary solution is to copy nvjpeg64_10.dll to "Python\Lib\site-packages\torch\lib".
I'm not sure if there is a better solution.