Create a GitHub Action to run Packer builds for Windows AMI

[zxiiro]

Currently the Windows AMI is being manually created and uploaded to both Meta and PyTorch Foundation AWS accounts. Instead we should utilize GitHub Actions to create a publish job that can handle this automatically for us.

[zxiiro]

This task will require more infrastructure setup than I initially thought. So the current ALI system deployed in PyTorch Foundation's AWS Account is a repository level ALI deployment so it only works with pytorch/pytorch meaning we currently don't have a way to launch instances from the LF fleet against any other repository.

[malfet]

Looks like we need to create some sort of policies thru AWS accounts to allow both account push AMIs
@zxiiro , @ZainRizvi , @atalman do you mind leaving the comment what is blocking you at the moment?

[zxiiro]

Originally I was thinking that we'd run 2 jobs one on Meta and one on LF side to push the AMI automatically. But on the CI Sync meeting today there were a few other suggestions.

1. Allow Meta runners to deploy the Windows AMI in LF Account
2. Meta account shares AMI to the LF Account
3. LF account shares AMI to Meta account and Allow Meta runners to deploy the Windows AMI to LF Account

I think all of these options will work however option 2 is probably the simplest solution since Meta runners are already setup to work on pytorch/test-infra and would only require the sharing of AMI to the LF Account.

If we agree to go with option 2 then I think I can start working on this task to create a job to build the AMI in Meta's AWS Account and someone from Meta can setup the AMI sharing permissions with the LF account.

@atalman @ZainRizvi any opinions on which option you want to go with?

[zxiiro]

Okay so I realized I could create the job on pytorch/pytorch and have it fetch test-infra before running packer build. This will allow us to just build the AMI directly in the PyTorch AWS Account. I think we should build there so that it's done on the community side and we can share the AMI back to Meta's CI account.

This ci-infra PR pytorch/ci-infra#312 is ready to go. I've confirmed the permissions work to allow GHA to run packer build.

I'm currently testing pytorch/pytorch#143132 which adds the job in pytorch/pytorch and will update once I've confirmed it to be working and can successfully create the AMI.

[zxiiro]

I was able to create a job that uses ubuntu-latest GitHub hosted runner and no longer requires a self-hosted runner. We can likely deploy this directly in test-infra now. Will investigate this afternoon.