Cog does not always load image into docker after build

[danfairs]

Cog does not always load images into docker after build. It sometimes does, though - but only if running on Apple Silicon:

https://github.com/replicate/cog/blob/main/pkg/docker/docker_command.go#L273

It's not entirely clear to me why Apple is special-cased here - it's actually to do with the build driver.

Docker's behaviour here depends on the build driver in use:

https://docs.docker.com/build/builders/drivers/

It looks like the intent of the code was to always load the image (the commend just above refers to an output type which actually isn't specified in the command), in which case Cog could always pass --load, and/or specify the builder as the comment suggests.

However, it would also be helpful to just allow the user to pass extra arguments to the build command, and this would be a way of doing so. That would be helpful to me, building a CI pipeline. (Actually, I'd prefer Cog just to give me the dockerfile and let me do the build, but that's another discussion!)

I should be OK to create a PR for whatever is appropriate here.

[michaeldwan]

Thanks for the issue and PR! I'm midway through some large-ish changes to the build code -- I'll rebase your changes and review within the next couple of days.

It's not entirely clear to me why Apple is special-cased here - it's actually to do with the build driver.

Good catch! Circa 2021 dockerd+buildx would only load images into the image store if they match the host platform. Since arm64 macOS was building linux/amd64 the --load flag was needed. Today that isn't always the case as you pointed out, and it's safe to just force --load for all builds.

That would be helpful to me, building a CI pipeline. (Actually, I'd prefer Cog just to give me the dockerfile and let me do the build, but that's another discussion!)

I'm curious about your CI pipeline... What are you trying to do? What's been a challenge? We're quite interested in CI/CD workflows for cog models.