nvJPEG is missing from Windows GPU runners

[pmeier]

According to the NVIDIA website, nvJPEG should come with the CUDA installation:

• For the most current version of nvJPEG, download the CUDA Toolkit.
• If you are using CUDA Toolkit 10.0 or 9.0, please download the nvJPEG installer.

However, this does not seem to be the case: https://github.com/pytorch/vision/actions/runs/4594756756/jobs/8114184733#step:9:234

The images torchvision uses for CircleCI have an explicit check for this: https://github.com/pytorch/vision/blob/781f512b01bc2324d7fdd11f0901f60571fc476f/.circleci/unittest/windows/scripts/set_cuda_envs.sh#L32-L35 Doing the same on the windows_job, fails (pytorch/vision#7475)

This currently blocks migrating Windows GPU workflows for torchvision.

[malfet]

CUDA installer can choose whether or not to include various subcomponents and I had to fix missing nvjpeg in torchvision many times, most recent in pytorch/vision#7186

I.e. looks like all one has to do is to add nvjpeg to

test-infra/aws/ami/windows/scripts/Installers/Install-CUDA-Tools.ps1

Line 13 in 2fec0c9

$installerArgs = "nvcc_$cudaVersion cuobjdump_$cudaVersion nvprune_$cudaVersion nvprof_$cudaVersion cupti_$cudaVersion cublas_$cudaVersion cublas_dev_$cudaVersion cudart_$cudaVersion cufft_$cudaVersion cufft_dev_$cudaVersion curand_$cudaVersion curand_dev_$cudaVersion cusolver_$cudaVersion cusolver_dev_$cudaVersion cusparse_$cudaVersion cusparse_dev_$cudaVersion npp_$cudaVersion npp_dev_$cudaVersion nvrtc_$cudaVersion nvrtc_dev_$cudaVersion nvml_dev_$cudaVersion"

and the deploy new AMI. @atalman can you please take care of that?

[pmeier]

Why do we install only parts of the toolkit and not all of it? Is storage on the runners a concern? If yes, I saw that there are older versions, i.e. < CUDA 11.7, present on the runner that could potentially be removed.

[atalman]

this is completed AMI was deployed