Use the default shallow checkout for jobs that only read files or test prebuilt
wheels. Source builds need complete commit ancestry and release tags so
setuptools-scm can derive each package's version.
For those jobs, use fetch-depth: 0 with filter: blob:none. The checkout
action fetches complete ancestry for all branches and tags, while the blob
filter skips historical file contents. This avoids downloading the large
historical gh-pages contents discussed in issue #2197. Git retrieves file
contents as needed when checking out the requested commit.
CI change detection uses this full-history checkout for PR mirrors, so the
actual PR base branch is available when computing its merge base. Non-PR runs
need only a shallow checkout. PR-preview cleanup checks out its script
shallowly and fetches gh-pages separately when needed. Pathfinder release
preparation fetches its exact tag without requesting every other tag.
Git documents blob filtering in its
clone reference.
The pinned checkout action's
README
documents fetch-depth: 0 for fetching all history, branches, and tags.
Jobs that build cuda.core fetch its pinned CCCL headers in a separate step:
set submodules: false on actions/checkout, then run:
git submodule sync --recursive
git submodule update --init --force --depth=1 --filter=blob:none --recursiveThe checkout action's filter input applies only to the root repository, so
letting it fetch submodules downloads CCCL's unfiltered history. The separate
step limits submodule history to one commit and skips historical blobs, while
checking out the exact revision pinned by the root repository. Syncing URLs
and forcing checkout preserve the action's behavior on reused workspaces.
Keep fetch-depth: 0 and filter: blob:none on the root checkout: package
versioning still needs its complete ancestry and release tags. See
issue #3351.
Two pre-commit hooks check links with one pinned native wheel (lychee-bin,
see .pre-commit-config.yaml) through the wrapper ci/tools/run_lychee.py:
lychee checks authored Markdown and reST (fragments unchecked), and the
manual-stage lychee-rendered checks the generated HTML in artifacts/docs,
including fragments. Developers run the same hooks locally. Public repository
main links are checked against the proposed checkout, as described under
Repository Links.
CI runs a hook through ci/tools/run_lychee_precommit.py <authored|rendered>,
wrapped with the shared cache and report handling in
.github/actions/lychee-links. The driver runs the real hook and merges the
native JSON report of every invocation (pre-commit may split long file lists).
It makes up to three attempts with 60/120-second cooldowns, but only while all
remaining failures are transient HTTP trouble: timeouts, 408, 429, or 5xx.
Missing, malformed, or inconsistent reports, and scans that check no links,
fail explicitly. Per-attempt reports and a summary are kept for seven days.
| Where | What runs | Final result |
|---|---|---|
| Linux/Windows pre-commit | Offline hook smoke check, then authored link warm-up from the shared cache, then the full suite once | The strict pre-commit result, including unresolved HTTP 429s |
| Full-CI documentation build | Rendered link check of the HTML it just built | Unresolved links fail, except that up to ten distinct URLs may remain at HTTP 429 as a warning |
Nightly (lychee.yml) |
Afresh authored and rendered checks | Unresolved links are warnings; setup, execution, and report errors fail |
The warm-up only fills the cache so the full suite, whose formatter changes must stay failures, mostly reuses successful checks. With the maintainer override, the driver's report-based result becomes the final link gate, and the full suite skips its duplicate Lychee invocation. Every other hook still runs once.
When rate limiting persists, maintainers can apply ci: lychee-non-blocking
to the PR and rerun the affected Linux/Windows pre-commit jobs or rendered
documentation job. Applying or removing the label does not itself start a
new workflow run. The shared link-check action queries the PR's current labels,
so a rerun sees the current decision rather than the original event's snapshot.
The lookup makes up to three attempts, waiting 5 and 10 seconds between them.
If every attempt fails or returns an invalid response, the action warns and
keeps the caller's default policy without enabling the override: authored
checks remain strict in the full pre-commit suite, and rendered checks keep
their allowance for at most ten distinct HTTP 429 URLs. Rerun the affected job
to retry the lookup and apply the label; other pre-commit hooks still run.
The label selects the driver's allow-429 policy for both authored and
rendered checks. Lychee still runs and retries normally; only a final report
whose remaining failures are all HTTP 429 responses from external HTTP(S)
URLs can pass with a warning. This explicit override has no URL-count limit.
The warning and retained reports identify the URLs that remain unverified.
Successful checks retain their usual cache behavior; failures are never cached.
Any other failure still blocks, including a mixture of HTTP 429s and other errors, broken local files or fragments, timeouts, and HTTP 403, 404, or 5xx responses. Hook setup, invalid or missing reports, and the offline smoke checks remain blocking. Other exceptional failures require the normal admin-merge decision.
The label applies only to PR checks. Local pre-commit runs, default-branch checks, and nightly cache production keep their existing policies. Without the label, authored checks remain strict and rendered checks retain their existing allowance for at most ten distinct HTTP 429 URLs.
Nightly and manual refreshes skip restoring, check afresh, and publish successful results; ordinary checks restore the newest baseline and never publish. Failed URLs and filesystem targets are never cached. Successes keep their original timestamps and expire after three days, so a missed nightly refresh does not break reuse, but a link that breaks after a nightly success can stay unnoticed for up to 72 hours.
The action keys each baseline by checker kind and a hash of
.pre-commit-config.yaml (hook options and wheel pin), lychee.toml, and
ci/tools/run_lychee.py, run_lychee_precommit.py, retry_lychee.py, and
checkout_source_links.py, so editing any of them starts a cold cache.
Authored and rendered caches stay apart because only rendered checks validate
fragments and Lychee's cache does not record which validation ran. URL results
are shared across operating systems; pre-commit's installed environments keep
their OS- and Python-specific caches.
To rebuild both shared caches manually, open
CI: Lychee checks and caches
(.github/workflows/lychee.yml), select the default branch, and enable
refresh-cache before choosing Run workflow. This runs the cache refresh
without the rest of nightly CI.
Checks that restore no matching baseline add a notice and a job-summary hint. They still check links, with potentially more runtime and network traffic. Checker changes in a PR intentionally start with a cold cache; refreshing the default branch cannot seed a cache for settings that exist only in that PR. Refreshing rebuilds cached successes; broken links and upstream rate limits still need their usual investigation.
Documentation can link to a path added or moved in the same PR using its
intended public URL, for example
https://github.com/NVIDIA/cuda-python/tree/main/new-directory. Lychee
validates the proposed target before it exists on public main. Only the
validation target changes; source files and built documentation retain their URLs.
ci/tools/checkout_source_links.py matches the complete public repository
prefix followed by blob/main/ or tree/main/:
- Authored Markdown and reST map to file URIs in the working tree, including uncommitted and unpushed paths. Missing files and directories fail even after an earlier success. These checks need no repository metadata and retain the existing policy of omitting fragments.
- Rendered HTML from the public repository maps source links to GitHub's file or directory view at the actual checked-out commit, retaining fragment checks. That commit must be available on GitHub.
- Rendered HTML from other repositories maps the same public source links to local files and directories. Only those source links omit fragments, since local files do not reproduce every GitHub source-view anchor. Generated documentation pages and their anchors still receive full validation.
Tags, explicit commits, other branches, unrelated repositories, and URLs embedded
inside other URLs retain their original targets. For rendered checks and checking
builds, CI supplies the checkout identity through GITHUB_REPOSITORY. Local runs
use git config lychee.repository, or otherwise the upstream remote followed by
origin. A checkout whose remotes only name a fork can select public-repository
validation explicitly:
git config lychee.repository NVIDIA/cuda-pythonFor rendered checks of an unpublished local commit, select your fork's identity to validate source links against local files:
git config lychee.repository YOUR_GITHUB_USER/cuda-pythonUse your actual fork's owner and repository. Set this before both the checking
documentation build and the rendered hook; rebuild existing checking output if
its example links contain an unpublished SHA. Clear any explicit
CUDA_PYTHON_DOCS_GITHUB_REF override to use this selection. Only GitHub
source-view fragments are omitted; generated HTML anchors remain checked.
CI uses GITHUB_REPOSITORY and retains its normal validation policy.
The checking build uses the public checkout's actual commit for generated
example links. In other repositories it uses main, allowing those public URLs
to receive the same local target checks without inserting another repository's
commit into public links. An explicit CUDA_PYTHON_DOCS_GITHUB_REF override is
retained; release builds continue to use their release tags.
These rules validate the proposed source tree. Before adopting them on a maintenance branch, choose an explicit policy for links that intentionally point to a different tree from the maintenance checkout.
Pixi lockfile freshness enforcement and automated refreshes are temporarily suspended because of Pixi #7215 and #7000. Repeated lockfile operations can change the committed solution without converging, so a general refresh is not a reliable repair.
The CI: pixi lockfile freshness check workflow still runs on every PR with its
existing pixi lock --check (all workspaces) check name. It succeeds with a
warning annotation and job summary explicitly reporting the suspension;
success does not mean freshness passed. The refresh workflow has no schedule.
Manual dispatch emits the same notice without solving dependencies, rewriting
lockfiles, pushing a branch, or creating a PR.
Manifest/lock consistency is not automatically enforced during this pause.
Necessary dependency changes require manual review and validation of the
affected manifests and lockfiles. Real source builds and tests continue using
PIXI_FROZEN=true, including nested pixi run calls, to build the checkout's
source packages against the committed dependency solution. The pinned Pixi
version was raised to v0.82.1 and the affected certificate-related packages
(certifi, pip, uv, ca-certificates) were updated individually in the
committed lockfiles to address the eMudhra/emSign CA incident; no general
lockfile refresh was performed.
Re-enable maintenance only after a released Pixi passes both upstream issue
reproducers and repeated byte-stable update/lock/check cycles across every
workspace reported by python ci/tools/list_pixi_workspaces.py, including both
bindings roots. Verify pixi update --no-install and pixi lock, followed by
repeated pixi lock --check commands, converge to unchanged lockfile bytes.
Then restore strict freshness checks, scheduled refreshes, and PIXI_LOCKED=true
in source-build CI.
The workflows in this repository use the CI_CUSTOMIZATIONS_* namespace for
GitHub Actions configuration variables that opt an alternative synchronized
repository into repository-specific CI behavior. This keeps the workflow logic
shared without hard-coding the names of private repositories into the public
source tree.
These variables are non-secret strings configured under
Settings > Secrets and variables > Actions > Variables.
An unset variable, or any value other than the literal string true, leaves
the customization disabled. Do not store credentials or other secret values in
these variables.
| Variable | Default | Purpose |
|---|---|---|
CI_CUSTOMIZATIONS_SECURITY_SUITE_ENABLED |
Disabled | Enables the NVIDIA Security Suite after its runner, Actions variables, and OIDC/Vault authorization have been provisioned for the repository. |
The canonical NVIDIA/cuda-python repository does not need this variable
because its standard workflow behavior is enabled directly. Before enabling a
customization elsewhere, document the repository-specific prerequisites and
verification procedure in that repository's own documentation.
ci/versions.yml is the authoritative public registry for CUDA bindings
package roots. Each mapping key is a repository-relative package root; its
record declares an exact toolkit build/test pin and a release_status of
current or maintenance. Release routing uses the tag's CUDA major/minor
and the toolkit pin to select a root. Bindings and metapackage builds use
setuptools-scm's default tag parser. Each root's git_describe_command
selects its CUDA ABI major for source builds (v12.* or v13.*),
independently of the toolkit minor used to build it.
Release workflows on main require schema_version: 2 in the tagged source
too. For older layouts, use an appropriate historical workflow ref and its
own inputs; see older release layouts.
The repository-private cuda-python-ci-tools package provides the shared
configuration readers and their dependencies. Workflows use these helpers
from their control revision when inspecting another source revision. Install
the package locally from the repository root:
python -m pip install -e ./ciUse ci.tools.bindings_config instead of reading the YAML directly. It emits
JSON containing each package directory, toolkit version, and release role,
plus the derived toolkit major/minor (ctk_target), CUDA major (cuda_major),
and wheel variant (cuda_variant, such as cu12). Workflows share this
normalized configuration:
python -m ci.tools.bindings_config
python -m ci.tools.bindings_config --package-roots
python -m ci.tools.bindings_config --release-status currentInstall the test extra to run the CI-tool tests:
python -m pip install -e './ci[test]'
python -m pytest ci/tools/testsThe public wheel builder requires one current package root and one
maintenance package root with different CUDA ABI majors.
cuda_bindings_12 names a stable CUDA-major source directory, while current
and maintenance describe release roles that can change at the next major.
The standalone cuda-python source distribution carries its selectors in
cuda_python/pyproject.toml, without depending on repository CI tooling. The
pre-commit metadata check keeps those selectors, bindings packaging, and Core
toolkit floors aligned with the registry:
python -m ci.tools.bindings_config --check-package-metadataSource builds need the matching tag and the history connecting it to the checkout; fetching tags alone does not repair shallow ancestry. No version override or local preview-tag preparation is needed. Follow the Git-derived version lifecycle.
ci-nightly.yml downloads wheels from the latest successful main CI run and
reads the bindings registry and test matrix from that run's source commit.
The source must use the bindings registry in ci/versions.yml and supply
toolkit-qualified artifacts for both bindings directories. Pre-migration
snapshots are unsupported.
Custom-driver rows require DRIVER_SHA256 in the selected source matrix.
An explicit run-id selects both the wheel artifacts and their source
revision with the same requirements; those artifacts must still be available.
A name such as pathfinder+bindings12+core12+meta12 / py3.11, CTK12.9.1 (wheels; bindings PR) identifies the enabled test suites, Python version,
and toolkit. meta12 means the CUDA 12 cuda-python metapackage; core12
means Core tested with CUDA 12. The corresponding CUDA 13 suites use 13.
pathfinder is shared across majors.
The two source descriptions answer different questions:
wheelsorlocal CTKdescribes how the toolkit libraries are supplied.bindings PRorbindings floordescribes the bindings source.bindings PRuses the selected root's artifact from this CI run. Unaffected packages can reuse exact-baseline wheels, so this label alone does not mean the bindings were rebuilt for the PR.bindings flooruses the oldest published bindings supported by thecuda-corewheel for that CUDA major, as declared in its metadata. These rows test Core or the selected nightly integration against published bindings, without running the bindings or metapackage suites.
Open the change-detection job's CI workplan summary and expand its JSON.
The component build/test flags identify selected work; the baseline records
the original run and exact source SHA for reusable artifacts. Reuse requires a
successful push CI run for the exact merge-base commit on the PR's base branch,
with the complete required artifact set. If that baseline is unavailable, CI
builds and tests everything. A bindings-only change reuses the unchanged
merged cuda-core wheel while testing cuda-core against the affected
cuda-bindings major.
The platform matrix-setup jobs' Selected wheel test coverage summaries show the final suites, toolkit source, bindings root or published floor, and GPU. Shared infrastructure changes, regular branch runs, and scheduled CI cover both majors; a bindings tag deliberately selects one bindings/metapackage line and published Pathfinder. Windows ARM64 omits CUDA 12 because that toolkit has no Windows ARM64 build.
ci/tools/select_test_matrix.py applies shared selection and local-wheel
coverage rules after each platform reads its source matrix and applies an
optional jq filter. Matrix data comes from the artifact
source revision; the helper comes from the workflow control revision.
cuda_bindings_12/ and cuda_bindings_13/ are the CUDA 12 and CUDA 13 source
directories on main. The registry identifies CUDA 12 as the maintenance
line. Routine CUDA 12 fixes, tests, and documentation belong in
cuda_bindings_12/. The old 12.9.x branch stays frozen except for emergency
fixes using compatible historical tooling. Release old-layout tags from a
validated workflow on the appropriate maintenance branch or historical tag,
rather than from main. Reconcile any emergency fix into the CUDA 12 directory
on main and assess the CUDA 13 directory too.
For independent releases from a shared source commit, see two tags at one commit.
For generated bindings, change the generator first and regenerate each affected
package root against its intended toolkit inputs. Review the generated diff,
record the generator revision, toolkit inputs, and command in the pull request,
and check modified files with
python toolshed/check_generated_file_seals.py <paths...>. The seal verifies
the content of an individual generated file; it does not identify the generator
revision or prove that the two roots should have identical output. Generated
output differs across toolkit targets.
For a handwritten bindings fix, inspect the corresponding code in the other root. In the pull request, identify the matching change or explain concretely why the fix does not apply there. A reviewer must verify that judgment: there is no general deterministic check for semantic equivalence of handwritten code across the two roots. Where files are intentionally identical, a narrowly scoped byte-for-byte check can be added after establishing that invariant.
Keep generator development and continuous toolkit qualification independent of this source layout. A generated change should be exercised against both supported consumers before its output is imported. File seals establish content integrity, not generation provenance.
When support moves from CUDA 12/13 to CUDA 13/14, retain cuda_bindings_13/
as the maintenance directory, add a CUDA 14 directory, and update
ci/versions.yml to replace CUDA 12 with CUDA 14, assign the release roles,
and pin both toolkits.
Align each directory's SCM tag selector with its CUDA major and update its
packaging, environments, documentation, and dependent package constraints.
Build and test both lines and their dependents, validate release selection,
and run release dry runs for tags from both lines before publishing.