test: repair three pre-existing failures under lane-select latching #130

Workflow file for this run

	# gridbook CI -- no GPU runners exist, so this workflow proves exactly the
	# things a GPU-less machine can prove, and says where it stops.
	#
	# What it gates:
	# * the sdist contains every gridbook/csrc/.cu \| .hpp in the checkout, and
	# the wheel contains every one of them EXCEPT the declared sdist-only set
	# (standalone developer binaries under csrc/tools/, and the pristine
	# cutlass_fork/*_orig.hpp diff baselines). check_dist.py gates that split in
	# both directions -- missing from the wheel is a broken install, present in
	# the wheel is dead weight in every user's site-packages -- and
	# tests/test_release_metadata.py re-checks it from inside the wheel that
	# cpu-tests actually installs
	# * a NON-EDITABLE install of the wheel resolves those sources from
	# site-packages (the historical bug was repo-root `os.pardir` arithmetic,
	# which resolved to <site-packages>/csrc; old releases then selected a
	# Triton path, while the current runtime fails closed)
	# * `import gridbook` needs no torch / vLLM or Triton package
	# * the `vllm.general_plugins` entry point is discoverable and loadable
	# * the GPU-free part of the test suite passes on every supported Python
	#
	# What it does NOT gate, and cannot here:
	# * that the CUDA extension compiles. That needs nvcc; free GitHub runners
	# have none. It is a MANUAL pre-tag gate -- the exact docker command is in
	# docs/RELEASING.md and takes ~30 s. Steps 1-3 above pass happily on a
	# wheel whose resolver is still wrong, so the manual gate is not optional.
	# * the Dockerfile. Nothing here builds it -- hosted runners have no nvcc and
	# the base image alone is ~32 GB on disk. Its manual gate is
	# `bash scripts/verify-image.sh` (build + 10 run-time checks, no GPU); see
	# docs/CONTAINER.md, "Keeping this image from rotting". Run it before
	# tagging and whenever VLLM_TAG moves, or the image rots silently.
	#
	# Action pinning: actions are pinned to exact patch tags (latest releases
	# checked 2026-07-28), never to a floating major like `@v7` and never to a
	# branch. .github/dependabot.yml bumps them weekly so the pins do not rot.
	# SHA pinning is stricter still (git tags are mutable) -- see docs/RELEASING.md.

	name: CI

	on:
	push:
	branches: [master]
	pull_request:
	workflow_dispatch:

	permissions:
	contents: read

	concurrency:
	group: ci-${{ github.ref }}
	cancel-in-progress: true

	jobs:
	build:
	name: build sdist + wheel
	runs-on: ubuntu-24.04
	steps:
	- uses: actions/checkout@v7.0.1

	- uses: actions/setup-python@v7.0.0
	with:
	python-version: "3.12" # reference interpreter (matches vllm-node)

	- name: Install build tooling
	run: python -m pip install --upgrade pip build twine

	# `python -m build` defaults to sdist-then-wheel-FROM-the-sdist, which is
	# the property we want: if package-data / MANIFEST.in drop the .cu files,
	# the wheel built from that sdist drops them too, and check_dist.py sees it.
	- name: Build
	run: python -m build

	- name: twine check
	run: python -m twine check --strict dist/*

	- name: Assert native sources are packaged
	run: python .github/scripts/check_dist.py dist .

	- uses: actions/upload-artifact@v7.0.1
	with:
	name: dist
	path: dist/
	if-no-files-found: error

	install:
	name: install wheel (py${{ matrix.python }})
	needs: build
	runs-on: ubuntu-24.04
	strategy:
	fail-fast: false
	matrix:
	# requires-python = ">=3.10". Nothing had ever executed this code on
	# 3.10 or 3.11 before this workflow existed -- that floor was an
	# assumption in published metadata, and this matrix is what turns it
	# into a measurement. 3.14 stays out until torch ships wheels for it.
	python: ["3.10", "3.11", "3.12", "3.13"]
	steps:
	- uses: actions/checkout@v7.0.1

	- uses: actions/setup-python@v7.0.0
	with:
	python-version: ${{ matrix.python }}

	- uses: actions/download-artifact@v8.0.1
	with:
	name: dist
	path: dist

	# --no-deps on purpose: this job asserts the package imports and resolves
	# its own packaged resources with NO torch / triton / vLLM present. The
	# full dependency set is installed for real in the cpu-tests job below.
	- name: Install the wheel (no deps)
	run: python -m pip install --no-deps dist/*.whl

	# Run from runner.temp, never the checkout: ./gridbook would shadow
	# site-packages and the check would prove nothing (check_installed.py
	# asserts this itself and fails loudly if it is ever run from the repo).
	- name: Post-install checks
	working-directory: ${{ runner.temp }}
	run: python "${GITHUB_WORKSPACE}/.github/scripts/check_installed.py"

	cpu-tests:
	name: cpu tests (py${{ matrix.python }})
	needs: build
	runs-on: ubuntu-24.04
	strategy:
	fail-fast: false
	matrix:
	python: ["3.10", "3.11", "3.12", "3.13"]
	steps:
	- uses: actions/checkout@v7.0.1

	- uses: actions/setup-python@v7.0.0
	with:
	python-version: ${{ matrix.python }}

	- uses: actions/download-artifact@v8.0.1
	with:
	name: dist
	path: dist

	# CPU torch from PyTorch's own index. The default PyPI torch wheel drags
	# in ~3 GB of nvidia-* CUDA runtime packages this runner can never use;
	# the +cpu build is a fraction of that. Verified 2026-07-28 that
	# torch 2.13.0+cpu ships manylinux_2_28_x86_64 wheels for cp310-cp315.
	- name: Install CPU torch
	run: \|
	python -m pip install --upgrade pip
	python -m pip install torch \
	--index-url https://download.pytorch.org/whl/cpu

	# The real `pip install gridbook`: torch is already satisfied by the +cpu
	# build above, so pip resolves safetensors / the rest from PyPI and never
	# pulls the CUDA torch.
	- name: Install the wheel with its dependencies
	run: python -m pip install pytest dist/*.whl

	# Stage the tests outside the checkout so `./gridbook` cannot shadow
	# site-packages -- the point is to exercise the installed wheel.
	- name: Stage tests outside the checkout
	run: cp -r tests "${RUNNER_TEMP}/gbtests"

	- name: Confirm the installed package is the one under test
	working-directory: ${{ runner.temp }}
	run: python "${GITHUB_WORKSPACE}/.github/scripts/check_installed.py"

	- name: Run GPU-free tests
	working-directory: ${{ runner.temp }}
	run: \|
	bash "${GITHUB_WORKSPACE}/.github/scripts/run_cpu_tests.sh" \
	"${RUNNER_TEMP}/gbtests"

Navigation Menu

test: repair three pre-existing failures under lane-select latching #130

Workflow file

test: repair three pre-existing failures under lane-select latching #130

Uh oh!

Workflow file for this run