[cregouby]

And fix #1440

[sebffischer]

@cregouby did you test that this actually works?

[cregouby]

Actually, i can't test. Just trying to make it work through CI, but with a lot of doubt on the underlying logic Le 14 octobre 2026 09:11:56 GMT+02:00, Sebastian Fischer ***@***.***> a écrit :

…

sebffischer left a comment (mlverse/torch#1483) @cregouby did you test that this actually works? -- Reply to this email directly or view it on GitHub: #1483 (comment) You are receiving this because you were mentioned. Message ID: ***@***.***>

[sebffischer]

Yeah, I think the GPU build is broken when creating PRs through branches. @dfalbel is this current design intended?

[dfalbel]

It seems to have broken recently. I don't know much how to fix, ideally we would have images built by the github runner and pushed somwhere, before being picked up by the expensive AWS runners. We may need to set some config in the build-image workflow that allows it to push to the mlverse image registry.

[cregouby]

It looks like the workflow "Build-gpu-image" shall be run with specific access right :

16 [auth] mlverse/torch-ci:pull,push token for ghcr.io
#16 DONE 0.0s

#15 exporting to image
#15 pushing layers 0.6s done
#15 ERROR: failed to push ghcr.io/mlverse/torch-ci:489987d1acd369a4e5f818457905289a3789288a: denied: installation not allowed to Write organization package
------
 > exporting to image:
------
ERROR: failed to build: failed to solve: failed to push ghcr.io/mlverse/torch-ci:489987d1acd369a4e5f818457905289a3789288a: denied: installation not allowed to Write organization package