Skip to main content

New large runners available for GitHub Actions

· 3 min read
Jaime Rodríguez-Guerra
Jaime Rodríguez-Guerra
Member of conda-forge/core

In March, we were notified that the colocation where the open-gpu-server CI was running was going to be decommissioned. This meant no GPU CI jobs and no long-running CPU jobs. Fortunately, we could get in touch with several companies that kindly agreed to sponsor the long-running CPU CI. We have not found a replacement for the GPU jobs, but we are still searching!

Decommissioning the open-gpu-server CI​

After almost three years, the open-gpu-server powering the long-running GPU and CPU CI jobs was decommissioned on March 13th. We started searching for suitable providers that could provide equivalent CI services and decided to give Cirrus Runners (CPU and GPU) a try. Our timing was unlucky, because Cirrus Labs was acquired by OpenAI on April 7th and their Runners service is being wound down.

New providers for long-running CPU jobs​

Immediately after, we reached out to a few companies offering external runners for GitHub Actions. Maybe someone out there would be able to help us with our most resource-demanding feedstocks... Fortunately, three different providers agreed to sponsor large CPU runners for Linux, macOS and Windows! We are incredibly grateful to these three companies:

Namespace.so logoNamespace.so logo

Namespace.so

Blacksmith.sh logoBlacksmith.sh logo

Blacksmith.sh

Depot.dev logoDepot.dev logo

Depot.dev

If you maintain a feedstock with special CI requirements (e.g. it cannot be built with the Microsoft-hosted Azure Pipelines or GitHub Actions runners within the allocated six hours, or it runs out of RAM or disk despite all possible workarounds), learn how to opt-in at How to> Advanced> Self-hosted runners.

What about the GPU CI jobs?​

Sadly, we don't have GPU CI jobs yet. Despite our attempts, we have not been able to find a GPU CI service that meets our requirements, which include:

  • Fixed costs or org-wide capped concurrency to estimate an upper bound for the costs.
  • Runs on GitHub Actions or equivalent integration effort in conda-smithy.
  • Based on ephemeral Linux and/or Windows VMs that can (ideally) run Docker.
  • (Ideally) VM images are maintained by the service provider, not us.

If you know of a service that can cover these points, please reach out! You can also email us at [email protected].