Build from source → GHCR #135

Workflow file for this run

.github/workflows/build-from-source.yml at 2191164

	# UNIVERSAL build-from-source workflow — BYTE-IDENTICAL across every Model B fork.
	#
	# DO NOT customize this file per-fork. If a fork needs different behavior, use
	# GitHub Actions repository Variables (Settings → Secrets and variables →
	# Actions → Variables):
	#
	# DOCKERFILE override auto-detected path (default: root Dockerfile, then shallowest)
	# BUILD_CONTEXT override docker build context (default: . — repo root)
	# PLATFORMS override platform list (default: linux/amd64,linux/arm64)
	# BUILD_ARGS multi-line KEY=VALUE pairs passed to docker build (default: none).
	# Used by forks whose upstream Dockerfile declares an ARG with no
	# default that the upstream's own CI passes externally (e.g.
	# tika-docker's TIKA_VERSION via republish-images.sh).
	# IS_SOURCE_BUILT job-level gate: build runs only when set to 'true'. Plain-mirror
	# forks set 'false' so the workflow file stays installed (uniform
	# scaffolding) but the job no-ops; fork-publish.yml owns :latest.
	# FREE_DISK_SPACE opt-in (set 'true') for forks whose images blow past the
	# 14GB default disk on ubuntu-latest (vllm-class CUDA images).
	# Adds ~3 min per build; reclaims ~30GB by removing pre-installed
	# Android SDK / .NET / Haskell / large npm caches.
	# RUNS_ON per-fork runner-label override (default: "ubuntu-latest").
	# JSON-encoded — single string for one label, or JSON array for
	# multiple. Set to '["self-hosted","lan-docker","big-build"]'
	# for forks that exceed hosted-runner capacity (vllm).
	# BUILD_TIMEOUT_MINUTES per-fork job-timeout override (default 360, the
	# GitHub-hosted-runner ceiling). Raise as a ONE-SHOT bootstrap
	# on forks whose first cold-cache build exceeds 360min (vllm
	# single-arch CUDA: ~5-6h first run, ~30-45m once :buildcache
	# is populated). Revert to default after first green run.
	# Honest-fact #97.
	# PRECOMPILED_WHEEL_BASE_URL master switch for the upstream-prebuilt-wheel
	# build path. Set to e.g. `https://wheels.vllm.ai` on forks
	# whose upstream publishes per-commit prebuilt wheels AND
	# whose carry stack is 100% non-compiled (Python/MD only).
	# When set, the build passes ${PREFIX}_USE_PRECOMPILED=1 + the
	# resolved commit + variant as build-args, and the fork's
	# Dockerfile carry can early-out the local C++/CUDA compile.
	# When unset (default for all forks), build path is unchanged.
	# sync-upstream.yml writes PRECOMPILED_WHEEL_COMMIT after
	# curl-verifying the wheel exists; if no variant in
	# PRECOMPILED_WHEEL_VARIANT_CHAIN publishes, sync CLEARS
	# PRECOMPILED_WHEEL_COMMIT to empty so the build falls back
	# to source compile rather than serving stale-ABI binaries
	# (the only safe semantics — a wheel built against SHA-N-1
	# with Python carries against SHA-N's APIs segfaults at
	# runtime with no build-time signal). Honest-fact #99.
	# PRECOMPILED_WHEEL_COMMIT SHA of the upstream commit whose wheel to fetch.
	# Managed by sync-upstream — operator should NOT set this
	# directly UNLESS PRECOMPILED_WHEEL_COMMIT_FROZEN=true (then
	# sync respects the manual pin). Empty when no wheel is
	# currently available; build falls back to source compile.
	# PRECOMPILED_WHEEL_VARIANT wheel build variant (e.g. `cu130`, `cu128`).
	# Operator-set initial value; sync-upstream updates in-place
	# when the variant chain falls through to a different one
	# (unless PRECOMPILED_WHEEL_COMMIT_FROZEN=true).
	# PRECOMPILED_WHEEL_VARIANT_CHAIN optional. Space-separated fallback list
	# (e.g. `cu130 cu128`). Sync-upstream tries each in order;
	# first one whose wheel exists wins. Defaults to the single
	# PRECOMPILED_WHEEL_VARIANT value.
	# PRECOMPILED_WHEEL_BUILD_ARG_PREFIX prefix for the synthesized build-arg
	# env names. Default `VLLM` — synthesizes VLLM_USE_PRECOMPILED,
	# VLLM_PRECOMPILED_WHEEL_COMMIT, etc. For forks whose upstream
	# uses a different convention (MinerU, torchao, future), set
	# to that project's prefix (e.g. `MINERU`) → MINERU_USE_PRECOMPILED.
	# Keeps the template generic across vendor opt-ins.
	# PRECOMPILED_WHEEL_COMMIT_FROZEN set to `true` to pin
	# PRECOMPILED_WHEEL_COMMIT/_VARIANT manually — sync-upstream
	# will NOT overwrite them when set. Use for testing a specific
	# wheel SHA, or for rolling back from a bad upstream commit.
	# Default unset (sync auto-manages).
	#
	# If a fork's Dockerfile needs out-of-context pre-build steps, the right fix is
	# to cherry-pick a multi-stage self-contained Dockerfile from upstream into our
	# intarweb-dev — NOT to add pre-build logic here.
	#
	# Codified in oss-contributing:ghcr-fork-mirror skill honest-fact #52.

	name: Build from source → GHCR

	on:
	push:
	branches: [intarweb-dev]
	workflow_dispatch:

	# Single-flight per repo: schedule-driven sync push + Heal K dispatch + manual
	# workflow_dispatch can all converge. cancel-in-progress: false because builds
	# are expensive (vllm 30-45 min, bifrost 45 min) — let the in-flight one finish,
	# queue the next behind it. Honest-fact #74.
	concurrency:
	group: build-from-source-${{ github.repository }}
	cancel-in-progress: false

	permissions:
	contents: read
	packages: write

	jobs:
	build:
	# IS_SOURCE_BUILT Variable is the authoritative source-vs-mirror switch
	# for the fleet. Setting it to anything other than 'true' (or leaving it
	# unset) skips the build entirely. This keeps the WORKFLOW FILE itself
	# byte-identical across every fork (uniform infra scaffolding) while
	# giving operators a per-fork Variable to control source-build vs
	# plain-mirror behavior without removing files or disabling workflows.
	# Plain-mirror forks (docker-neo4j etc) still have build-from-source.yml
	# installed but the job no-ops; fork-publish.yml owns :latest for them.
	#
	# Branch gate: when triggered by a push, only run on intarweb-dev. The
	# `on.push.branches: [intarweb-dev]` filter above is supposed to handle
	# this, but in practice GitHub still records 0-second false-failure runs
	# against build-from-source.yml when sync-upstream's bot pushes to main
	# (empty jobs[], shows red in Actions UI, head_branch=main, name=workflow
	# path rather than display name). Belt-and-suspenders branch gate here
	# ensures the job no-ops cleanly to a green "skipped" rather than a
	# mysterious red failure if the trigger ever evaluates against main /
	# master / any other ref. Manual workflow_dispatch and workflow_call
	# paths remain unaffected. Honest-fact #109.
	if: >-
	vars.IS_SOURCE_BUILT == 'true'
	&& (
	github.event_name == 'workflow_dispatch'
	\|\| github.event_name == 'workflow_call'
	\|\| (github.event_name == 'push' && github.ref == 'refs/heads/intarweb-dev')
	)
	# Per-fork override via repo Variable RUNS_ON (default: ubuntu-latest hosted).
	# Set to '["self-hosted","lan-docker","big-build"]' (JSON array, single line)
	# for forks whose builds exceed hosted-runner capacity (16GB RAM / 14GB disk
	# — vllm's CUDA csrc-build is the canonical case). Honest-fact #89.
	runs-on: ${{ fromJSON(vars.RUNS_ON \|\| '"ubuntu-latest"') }}
	# Per-fork override via repo Variable BUILD_TIMEOUT_MINUTES (default 360, the
	# GitHub-hosted-runner hard ceiling). Raise to 720 as a ONE-SHOT bootstrap on
	# forks whose first build needs to populate :buildcache from scratch (vllm
	# at single-arch CUDA: ~5-6h cold-cache vs ~30-45m warm). Revert to 360 once
	# warm. Codified after 27340784996 / 27373307060 cancelled at 360min on the
	# vllm fork's first cold-cache run. Honest-fact #91.
	timeout-minutes: ${{ fromJSON(vars.BUILD_TIMEOUT_MINUTES \|\| '360') }}
	# Read-once capture of precompiled-wheel Variables at job start. Subsequent
	# steps reference $PRECOMPILED_* env vars (resolved once at job-init) rather
	# than ${{ vars.* }} (re-resolved per-step expression). This makes an
	# in-flight build immune to mid-job Variable mutations from a concurrent
	# sync-upstream writing PRECOMPILED_WHEEL_COMMIT/_VARIANT. Eliminates the
	# read/write race architect-review #1 flagged. Honest-fact #99.
	env:
	PRECOMPILED_BASE_URL: ${{ vars.PRECOMPILED_WHEEL_BASE_URL }}
	PRECOMPILED_COMMIT: ${{ vars.PRECOMPILED_WHEEL_COMMIT }}
	PRECOMPILED_VARIANT: ${{ vars.PRECOMPILED_WHEEL_VARIANT }}
	PRECOMPILED_PREFIX: ${{ vars.PRECOMPILED_WHEEL_BUILD_ARG_PREFIX \|\| 'VLLM' }}
	OPERATOR_BUILD_ARGS: ${{ vars.BUILD_ARGS }}
	steps:
	- uses: actions/checkout@v4
	with:
	persist-credentials: false

	# Optional disk reclamation for forks whose images blow past the 14GB
	# default disk on ubuntu-latest hosted runners (vllm being the canonical
	# offender — its CUDA build artifacts + intermediate layers exceed 20GB
	# easily). Frees ~30GB by removing pre-installed Android SDK, .NET,
	# Haskell, large npm caches, and codeql databases. Costs ~3 min — gated
	# behind FREE_DISK_SPACE Variable so non-vllm-class forks don't pay it.
	- name: 🧹 Free disk space (opt-in via FREE_DISK_SPACE Variable)
	if: vars.FREE_DISK_SPACE == 'true'
	uses: jlumbroso/free-disk-space@v1.3.1
	with:
	tool-cache: true
	android: true
	dotnet: true
	haskell: true
	large-packages: true
	docker-images: false
	swap-storage: false

	- name: 🔍 Resolve Dockerfile
	id: df
	env:
	DOCKERFILE_VAR: ${{ vars.DOCKERFILE }}
	run: \|
	set -euo pipefail
	# 1. Explicit override (repo Variable) always wins.
	if [ -n "${DOCKERFILE_VAR}" ]; then
	echo " ✓ using DOCKERFILE override: ${DOCKERFILE_VAR}"
	echo "path=${DOCKERFILE_VAR}" >> "$GITHUB_OUTPUT"
	echo "skip=false" >> "$GITHUB_OUTPUT"
	exit 0
	fi
	# 2. Prefer a root Dockerfile.
	if [ -f Dockerfile ]; then
	echo " ✓ found root Dockerfile"
	echo "path=Dockerfile" >> "$GITHUB_OUTPUT"
	echo "skip=false" >> "$GITHUB_OUTPUT"
	exit 0
	fi
	# 3. Deterministic discovery: shallowest non-test Dockerfile.
	cand="$(find . -maxdepth 4 -type f -name Dockerfile \
	-not -path './node_modules/*' \
	-not -path './.git/*' \
	-not -path '/test/' \
	-not -path '/tests/' \
	-not -path '/example/*' \
	-not -path './vendor/*' \
	-not -path './third_party/*' \
	\| sed 's\|^\./\|\|' \
	\| awk '{print gsub(/\//,"/"), $0}' \
	\| sort -n \| head -n1 \| cut -d' ' -f2- \|\| true)"
	if [ -z "$cand" ]; then
	echo " ✗ no Dockerfile found → plain-mirror fallback (no image built)"
	echo "skip=true" >> "$GITHUB_OUTPUT"
	exit 0
	fi
	echo " ✓ auto-found: $cand"
	echo "path=$cand" >> "$GITHUB_OUTPUT"
	echo "skip=false" >> "$GITHUB_OUTPUT"

	- name: Set up QEMU
	if: steps.df.outputs.skip != 'true'
	uses: docker/setup-qemu-action@v3

	- name: Set up Buildx
	if: steps.df.outputs.skip != 'true'
	uses: docker/setup-buildx-action@v3

	- name: Log in to ghcr.io
	if: steps.df.outputs.skip != 'true'
	uses: docker/login-action@v3
	with:
	registry: ghcr.io
	username: ${{ github.actor }}
	password: ${{ secrets.GITHUB_TOKEN }}

	- name: 🔡 Compute image (lowercase — GHCR requires lowercase repo names)
	id: imgname
	run: \|
	# vars.IMAGE_NAME may already be lowercase (set as such manually), but
	# the github.event.repository.name fallback can be CamelCase if the fork
	# repo name has uppercase letters (e.g. RetroSaveManager). docker/build-push
	# rejects ANY uppercase in the image path. fork-publish.yml does the same
	# via tr; we do it here. Honest-fact #63.
	NAME="${{ vars.IMAGE_NAME \|\| github.event.repository.name }}"
	NAME_LOWER=$(echo "$NAME" \| tr '[:upper:]' '[:lower:]')
	echo "image=ghcr.io/${{ github.repository_owner }}/${NAME_LOWER}" >> "$GITHUB_OUTPUT"

	- name: 🏷️ Docker metadata
	id: meta
	if: steps.df.outputs.skip != 'true'
	uses: docker/metadata-action@v5
	with:
	# Per-fork override via repo Variable IMAGE_NAME — for forks where the
	# published image name differs from the repo name (docker-autoheal repo →
	# autoheal image). Falls back to repo name. Same as fork-publish.yml so
	# ONE image per fork. Lowercased above (metadata-action's lowercase only
	# applies to tags, not to the images input — uppercase repo names slip
	# through and break build-push with "repository name must be lowercase").
	images: ${{ steps.imgname.outputs.image }}
	tags: \|
	type=raw,value=latest,enable=${{ github.ref == 'refs/heads/intarweb-dev' }}
	type=sha,prefix=sha-,format=short

	- name: 🧩 Synthesize build-args (BUILD_ARGS + optional precompiled-wheel envs)
	# Combines operator-set BUILD_ARGS with auto-derived precompiled-wheel
	# build-args when PRECOMPILED_WHEEL_BASE_URL is set on the fork. For
	# forks WITHOUT that Variable (the default), output is byte-identical to
	# the pre-existing `build-args: ${{ vars.BUILD_ARGS \|\| '' }}` behavior.
	# Uses $PRECOMPILED_* env vars captured at job start (job-level env
	# block) — immune to mid-build Variable mutations. Synthesizes envs with
	# the configured PREFIX (default `VLLM`), so non-vllm forks adopting this
	# pattern (MinerU, torchao, future) just set
	# PRECOMPILED_WHEEL_BUILD_ARG_PREFIX without template surgery.
	# Honest-fact #99/#104.
	id: bargs
	if: steps.df.outputs.skip != 'true'
	run: \|
	set -euo pipefail
	BARGS="${OPERATOR_BUILD_ARGS:-}"
	if [ -n "${PRECOMPILED_BASE_URL:-}" ] && [ -n "${PRECOMPILED_COMMIT:-}" ]; then
	# Build the precompiled-wheel envs via printf so the multi-line
	# KEY=VALUE block doesn't dedent the surrounding YAML run block
	# (honest-fact #104 — broken-yaml-from-bash-multiline-assignment
	# 0-second-failed all 22 forks 2026-06-12).
	#
	# Emit BOTH ${PREFIX}_PRECOMPILED_WHEEL_COMMIT and
	# ${PREFIX}_MERGE_BASE_COMMIT to the same SHA — different
	# Dockerfile carries may key off either name (vllm-bot's carry
	# uses MERGE_BASE_COMMIT; theoretical future carries may use
	# the more-explicit PRECOMPILED_WHEEL_COMMIT). Both work.
	# Honest-fact #105.
	PREFIX="${PRECOMPILED_PREFIX:-VLLM}"
	EXTRA=$(printf '%s_USE_PRECOMPILED=1\n%s_PRECOMPILED_WHEEL_COMMIT=%s\n%s_MERGE_BASE_COMMIT=%s\n%s_PRECOMPILED_WHEEL_VARIANT=%s\n%s_PRECOMPILED_WHEEL_BASE_URL=%s' "$PREFIX" "$PREFIX" "$PRECOMPILED_COMMIT" "$PREFIX" "$PRECOMPILED_COMMIT" "$PREFIX" "$PRECOMPILED_VARIANT" "$PREFIX" "$PRECOMPILED_BASE_URL")
	if [ -n "$BARGS" ]; then
	BARGS="${BARGS}"$'\n'"${EXTRA}"
	else
	BARGS="${EXTRA}"
	fi
	echo " ✓ precompiled-wheel build-args appended (prefix=${PREFIX} commit=${PRECOMPILED_COMMIT} variant=${PRECOMPILED_VARIANT})"
	elif [ -n "${PRECOMPILED_BASE_URL:-}" ] && [ -z "${PRECOMPILED_COMMIT:-}" ]; then
	# Master switch on, but no resolved commit. Sync-upstream cleared it
	# because no wheel published for the current upstream SHA. Build will
	# fall back to source compile — degraded perf, but correct binaries
	# (vs the silent-ABI-break risk of serving a stale wheel).
	echo "::warning::PRECOMPILED_WHEEL_BASE_URL set but PRECOMPILED_WHEEL_COMMIT empty — sync-upstream couldn't find a wheel for current upstream SHA. Falling back to source compile."
	fi
	{
	echo "build_args<<__EOB__"
	echo "$BARGS"
	echo "__EOB__"
	} >> "$GITHUB_OUTPUT"

	- name: 🚀 Build & push
	if: steps.df.outputs.skip != 'true'
	uses: docker/build-push-action@v6
	with:
	context: ${{ vars.BUILD_CONTEXT \|\| '.' }}
	file: ${{ steps.df.outputs.path }}
	platforms: ${{ vars.PLATFORMS \|\| 'linux/amd64,linux/arm64' }}
	# build-args synthesized above. Combines operator-set BUILD_ARGS with
	# auto-derived precompiled-wheel envs when PRECOMPILED_WHEEL_BASE_URL
	# is set. For forks without it, equivalent to BUILD_ARGS alone.
	# Honest-facts #88, #99.
	build-args: ${{ steps.bargs.outputs.build_args }}
	# BuildKit registry-cache. Reuses prior build's intermediate layers
	# across runs to dramatically cut wall time on warm builds (vllm:
	# ~30min → ~10-12min target). cache-from is read-only — first run on
	# a fresh fork misses everything (cold cache, no penalty). cache-to
	# `mode=max` exports ALL intermediate stages (not just the final
	# image), so the next run can resume mid-stage. `compression=zstd`
	# for faster decompression on cache pulls.
	# ignore-error=true on cache-to: if the push to :buildcache fails
	# (perms, network), the build still succeeds; just warm cache misses
	# on next run. Honest-fact #107.
	cache-from: type=registry,ref=${{ steps.imgname.outputs.image }}:buildcache
	cache-to: type=registry,ref=${{ steps.imgname.outputs.image }}:buildcache,mode=max,compression=zstd,ignore-error=true
	push: true
	tags: ${{ steps.meta.outputs.tags }}
	labels: ${{ steps.meta.outputs.labels }}
	provenance: false

Navigation Menu

Uh oh!

Build from source → GHCR #135

Workflow file

Build from source → GHCR #135

Uh oh!

Workflow file for this run