Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
Publish Docker on the daily schedule, and move layer cache off the Actions quota (#10429)
Daniel Han583191a35766
diff --git a/.github/workflows/docker-publish.yml b/.github/workflows/docker-publish.ymlindex c083fc07549..ae5c2c51ff3 100644--- a/.github/workflows/docker-publish.yml+++ b/.github/workflows/docker-publish.yml@@ -11,7 +11,12 @@ name: Publish Blackwell Docker imageon:push:- branches: [main]+ # Deliberately NOT branches: [main]. Publishing on every merge meant 23 runs on+ # 2026-09-06 and 14 on 2026-09-05 against one scheduled run each day, and since+ # each is a two-arch build of both images it was the bulk of the repo's Actions+ # cache and a large share of its runner minutes. The daily cron below is the+ # release cadence for main now. Tags still publish immediately: a release that+ # produced no image until the next morning would be worse than the spend.tags: ['v*']schedule:# Daily. Actions cron is UTC with no DST handling, so this is 11:17 in San@@ -50,14 +55,17 @@ env:# Not a secret, and named once so the four login steps cannot drift apart.REGISTRY_USERNAME: unsloth-# Serialise per-ref runs, EXCEPT on main where each commit gets its own group: a-# shared group does not queue a main burst, it DISCARDS it (GitHub cancels any PENDING-# run the moment a newer one is queued), so the earlier commit would publish no image-# at all. The accepted cost is that two main commits can race to retag :core/:latest,-# which is self-correcting on the next push, where a cancelled run leaves that commit-# with no image at all. Each published image is still internally coherent: the digest-# handoff and the smoke test resolve this run's own tag, not tags[0], and verify it.-# tests/studio/test_main_runs_survive_merge_bursts.py enforces this shape.+# Serialise per-ref runs, EXCEPT on main where each run gets its own group: a shared+# group does not queue a burst, it DISCARDS it (GitHub cancels any PENDING run the+# moment a newer one is queued), so the earlier one would publish no image at all.+# Merges no longer trigger this workflow, so the burst it now guards against is a+# scheduled run overlapping a workflow_dispatch, rather than a merge burst. Keeping+# the per-sha key costs nothing and still prevents a dispatch being discarded.+# Each published image is internally coherent either way: the digest handoff and the+# smoke test resolve this run's own tag, not tags[0], and verify it.+# Dropping the main push trigger also drops this workflow out of the scan in+# tests/studio/test_main_runs_survive_merge_bursts.py, which selects on+# `push.branches` containing main, so no exemption is needed there.concurrency:group: docker-publish-${{ github.ref }}-${{ github.ref == 'refs/heads/main' && github.sha || '' }}cancel-in-progress: false@@ -254,8 +262,14 @@ jobs:file: ./docker/Dockerfileplatforms: ${{ matrix.platform }}labels: ${{ steps.meta.outputs.labels }}- cache-from: type=gha,scope=build-${{ matrix.platform }}- cache-to: type=gha,scope=build-${{ matrix.platform }},mode=max+ # Registry cache, not type=gha. mode=max exports every intermediate layer,+ # which on the GHA backend put 127 blob entries and 32GB into a 50GB repo+ # cache in a single morning and starved the GGUF and pip caches sharing it.+ # A cache tag on the image's own repo has no such budget, so mode=max keeps+ # its rebuild speed without the eviction pressure. The Docker Hub login+ # above covers both reading and writing this ref.+ cache-from: type=registry,ref=${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:buildcache-core-${{ matrix.platform == 'linux/amd64' && 'amd64' || 'arm64' }}+ cache-to: type=registry,ref=${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:buildcache-core-${{ matrix.platform == 'linux/amd64' && 'amd64' || 'arm64' }},mode=maxoutputs: type=image,name=${{ env.REGISTRY }}/${{ env.IMAGE_NAME }},push-by-digest=true,name-canonical=true,push=true# Keep prose OUT of build-args: build-push-action forwards every non-empty# line verbatim, so a #-line becomes a bogus --build-arg.@@ -439,9 +453,12 @@ jobs:file: ./docker/Dockerfile.studioplatforms: ${{ matrix.platform }}labels: ${{ steps.meta.outputs.labels }}- # mode=max here would blow the 10GB GHA quota and evict the base build's cache- cache-from: type=gha,scope=studio-${{ matrix.platform }}- cache-to: type=gha,scope=studio-${{ matrix.platform }},mode=min+ # Kept at mode=min. The original reason was the GHA cache quota, which no+ # longer applies now that this is a registry ref on the image's own repo,+ # but the remaining cost is Docker Hub storage rather than a hard budget,+ # so raising it is a separate decision from moving the transport.+ cache-from: type=registry,ref=${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:buildcache-studio-${{ matrix.platform == 'linux/amd64' && 'amd64' || 'arm64' }}+ cache-to: type=registry,ref=${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:buildcache-studio-${{ matrix.platform == 'linux/amd64' && 'amd64' || 'arm64' }},mode=minoutputs: type=image,name=${{ env.REGISTRY }}/${{ env.IMAGE_NAME }},push-by-digest=true,name-canonical=true,push=true# the same values the base build baked, so Studio matches it even if upstream# moved mid-run (build-args must be KEY=VALUE only)