Bitbucket Pipelines — GPU support

[0x2b3bfa0]

Continuation of https://github.com/iterative/cml/pull/798#pullrequestreview-982578801

As per the official documentation, self-hosted Bitbucket Pipelines runners for Linux don't “provide access to the host file system” although that limitation is “subject to change in the future”

Linux-only

Can’t access the local files on the runner host from a build script: In order to provide isolation and reproducibility, a runner doesn’t provide access directly to the host file system, we recommend using SCP / Rsync, if necessary.

The limits described above are subject to change in the future.

In practice, this means that every single step is containerized, and users don't have any control over container options like docker run --gpus, even when the NVIDIA runtime is installed.

• https://github.com/iterative/cml/pull/798#issuecomment-1129155399
• https://github.com/iterative/cml/pull/798#issuecomment-1131728075

[0x2b3bfa0]

Related

• https://jira.atlassian.com/browse/BCLOUD-21459 (community question)

[casperdcl]

potential work-around: use (TP)I task (BB Pipelines -> TPI, exporting/forwarding BB env vars -> AWS EC2 -> user script && CML report). ~Likely depends on resolving https://github.com/iterative/terraform-provider-iterative/issues/630.~

moved to https://github.com/iterative/terraform-provider-iterative/issues/633