Skip to content

ci: pull CI images from the domestic mirror on self-hosted runners - #7762

Closed
Critsium-xy wants to merge 1 commit into
deepmodeling:developfrom
Critsium-xy:ci/use-domestic-registry-mirror
Closed

ci: pull CI images from the domestic mirror on self-hosted runners#7762
Critsium-xy wants to merge 1 commit into
deepmodeling:developfrom
Critsium-xy:ci/use-domestic-registry-mirror

Conversation

@Critsium-xy

Copy link
Copy Markdown
Collaborator

No description provided.

The self-hosted runner pools give every job a fresh pod, so no Docker
layer cache is ever reused: in 37 consecutive CUDA jobs, 37 distinct
runner pods pulled all 17 layers cold ("Already exists" never appears).
Every GPU run therefore pays a full 3.98 GB pull from ghcr.io, which
takes 8.7 min at best.

When the link degrades, the two nvidia/cuda base layers (2343 MB and
1099 MB, together 86% of the image) hit connection failures and restart
from scratch, since a failed blob download is not resumed. Job
91378883548 retried them 5 and 4 times respectively and spent 39m09s in
docker pull alone, which is the whole of its 41-minute
"Initialize containers" step.

registry.dp.tech is the Aliyun (cn-zhangjiakou) mirror already published
by devcontainer.yml. It is anonymously pullable and serves an identical
digest, so no credentials are needed and the images are unchanged:
abacus-cuda resolves to sha256:499d7bdc7f9f..., the same digest ghcr.io
returned in the pull log above.

Only workflows on self-hosted runners are switched. Those on
ubuntu-latest (ase_plugin_test, build_test_cmake, build_test_makefile,
pytest) keep ghcr.io, which is the fast path from GitHub-hosted runners.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@Critsium-xy Critsium-xy closed this Aug 2, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant