GeometricMachineLearning.jl/.github/workflows/Scripts.yml at main · JuliaGNI/GeometricMachineLearning.jl · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
# The gate for `scripts/`, in its own workflow rather than as a second job in CI.yml.
#
# CI.yml is byte-identical in every repository under Packages/ and Experiments/ — that is what
# keeps the required-status-check list static and stops branch protection going stale. A repository
# that needs an extra job therefore gets an extra file, never a local edit to CI.yml. See
# ~/Research/Environment/Harness/githooks/install-workflows.sh.
#
# WHY THIS EXISTS.
#
# Nothing ran `scripts/` until this job did. The directory was repaired by hand several times —
# "Bring the scripts to the new AdamOptimizerWithDecay", "`tspan` -> `timespan`" — and each repair
# regressed in silence, because the package's own suite never loads a single file from it. The
# figures in the manual and the committed network weights under `docs/src/tutorials/` were produced
# by these scripts and by nothing else, so a script that stops running takes the only executable
# record of a result with it.
#
# WHAT IT RUNS.
#
# `scripts/runscripts.jl` runs every `.jl` file under `verification/` and under `reproduction/`,
# each in its own process. `verification/` runs in full: those scripts establish a mathematical
# claim and are written to finish in seconds. `reproduction/` runs with GML_SMOKE set, which is
# what `utilities/smoke.jl` reads to return the small size of every constant that costs time. A
# smoke run proves that the script still executes end to end; it says nothing about the result,
# which is the honest limit of what CI can check for a run that takes hours on a GPU.
#
# One entry, not a matrix: this job asks whether the scripts still run, and that question does not
# have a different answer on a second operating system. The scripts environment is large — CairoMakie,
# CUDA, Lux — and six copies of it buy nothing.
#
# Deliberately not a required status check: its name does not begin with `Julia `, and the required
# list in ci-protection.sh is static across the tree.

name: Scripts

on:
    push:
        branches:
            - main
            - master
        tags: ["*"]
    pull_request:
    workflow_dispatch:

concurrency:
    group: ${{ github.workflow }}-${{ github.ref }}
    cancel-in-progress: ${{ startsWith(github.ref, 'refs/pull/') }}

jobs:
    scripts:
        name: Scripts - ubuntu-latest
        runs-on: ubuntu-latest
        timeout-minutes: 90
        permissions:
            # julia-actions/cache proactively deletes the caches it created.
            actions: write
            contents: read
        steps:
            - uses: actions/checkout@v7
            - uses: julia-actions/setup-julia@v3
              with:
                  version: "1"
            - uses: julia-actions/cache@v3
            - name: Instantiate the scripts environment
              # `scripts/Project.toml` reaches the package through `[sources] = {path = ".."}`, so
              # `Pkg.develop` is neither needed nor wanted here: it would write a machine-local
              # absolute path into that table.
              run: julia --startup-file=no --project=scripts -e 'using Pkg; Pkg.instantiate()'
            - name: Verification scripts
              run: julia --startup-file=no --project=scripts scripts/runscripts.jl verification
            - name: Reproduction scripts, smoke sizes
              run: julia --startup-file=no --project=scripts scripts/runscripts.jl reproduction