ollama37-ci-build

name	ollama37-ci-build
description	Build the ollama37 CI image on a self-hosted GPU runner — never with a local make
user-invocable	true

Build the ollama37 CI image with CI

Build the fork's runtime image through the GitHub Actions build workflow on a self-hosted runner — gh workflow run, never make build-runtime-local on a dev box. To apply a built image to a testbed, see ollama37-ci-apply.

$ARGUMENTS

Never build by hand — a local build is version-crippled

A local make build defaults OLLAMA_VERSION=0.0.0 (docker/Makefile), and the ollama registry rejects every gated model pull from an old version with 412: requires a newer version — so the image looks fine and silently cannot pull modern models (qwen3.5, gemma3, ministral-3, …). The CI build injects OLLAMA_VERSION from the repo variable (check it: gh variable list), which is > any 0.x minimum and clears the gate. Only the CI image can pull those models. It also leaves logs; a hand make leaves nothing.

The build

TC-BUILD-002 IS the build — test-build.yml runs make build-runtime-local-no-cache, a full no-cache compile (preset "CUDA 11 K80", native CUBIN sm_37…sm_86, no GPU needed). It produces the local tag ollama37:latest on the runner it ran on.

gh workflow run test-build.yml --ref <branch> -f runner_label=<sm37|sm75>
gh run watch <run-id> --exit-status        # block until done; GitHub holds the logs

• runner_label — which box: sm37 (rocky9-k80-cicd-1, 4× K80) or sm75 (rocky9-2060-cicd-1, RTX 2060). Default sm37. Never a bare runs-on: self-hosted.
• --ref <branch> — which code. Omit to build the default branch.
• The build normally runs on sm37 (its ~15 GB builder image is cached there; on a fresh host ensure-builder rebuilds it, ~90 min). It can run on sm75 only if ollama37-builder:latest is already present there.

The CI image lands where compose reads it — same host, no push

The build produces the local tag ollama37:latest, but docker/docker-compose.yml reads dogkeeper886/ollama37:latest. TC-BUILD-004 (in the build suite) bridges that gap: it retags the fresh ollama37:latest → dogkeeper886/ollama37:latest locally, so every later docker compose up (runtime/inference/models/perf) exercises the build we just made — no compose edit, no Docker Hub push.

A single-test build skips the retag. TC-BUILD-004 only runs as part of the full build suite. If you build one test (-f test_id=TC-BUILD-002), dogkeeper886/ollama37:latest still points at the previous image — so apply + every downstream test silently exercises stale code. Run the full test-build.yml (no test_id), or run the retag too (-f test_id=TC-BUILD-004). ollama37-ci-apply restates this as a precondition to confirm before applying.

This is single-host: build and test on the same runner. The old cross-host path — publish to dogkeeper886/ollama37-ci:ci-<sha> and pull by digest on a second card — was retired (#441): the push was slow and nothing ever consumed the pushed image. To restore the published image on a runner (e.g. for a published-image baseline), docker pull dogkeeper886/ollama37:latest.

Once built, hand off to ollama37-ci-apply to roll it onto the testbed.