Bitbucket GPU demo

[casperdcl]

self-hosted Bitbucket Pipelines don't support GPU runners iterative/cml#1015. A work-around is using TPI task directly in BB Pipelines.

Proposal

Put together an example repo/tuorial:

BB Pipelines -> TPI task (exporting env vars) -> AWS ec2 -> DVC repro && CML report

• "export env vars" means¹ environment = { "BITBUCKET_*" = "", CI = "" }
  • this would mean CI commands (e.g. cml send-comment) would "just work" out-of-the-box
• BB equivalent of easy GH ssh debugging: https://registry.terraform.io/providers/iterative/iterative/latest/docs/guides/getting-started#debugging
• and/or maybe uploading tfstate as an artifact?
• probably fixing TPI#630
• image = "nvidia"
• tags = { Name = "Foo" }?
• basic readme (explaining how to use)
• post link on docs (TPI & CML?)
• blog (devrel)
  • part 1: python train.py (using https://github.com/pytorch/examples/tree/main/mnist)
  • part 2: docker run

Alternatives

• deconstruct, patch & rebuild BB java runner
• monitor BB java runner releases & repeat the above whenever there's a new release

Footnotes

1. https://support.atlassian.com/bitbucket-cloud/docs/variables-and-secrets ↩