Skip to content

Document CUDA library target families #21

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
cerisier wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Document CUDA library target families #21

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
29 changes: 29 additions & 0 deletions AGENTS.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -115,6 +115,29 @@ In short:
- upstream component boundaries may drift
- this repository’s public surface should drift as little as possible

## Library Target Families

New public library capabilities should use a consistent target family when the corresponding upstream artifacts exist:
- `foo_shared_library`: raw packaged shared-library import
- `foo`: semantic shared wrapper with its complete shared dependency closure
- `foo_interface_library`: raw system-provided link interface
- `foo_system`: semantic system-provided wrapper with the same logical dependency closure as `foo`
- `foo_static_library`: raw static archive import
- `foo_static`: semantic static wrapper with its complete hermetic static dependency closure
- `headers`: compile-time header surface and header dependency closure
- `shared_library_files`: packaged runtime shared-library files only

Target-family rules:
- a new `*_shared_library` should have a matching `*_interface_library` and `_system` wrapper unless upstream provides no usable link-time interface; document and test intentional absence
- interface imports use `cc_import(interface_library = ..., system_provided = True)`
- prefer NVIDIA linker stubs under `lib/stubs/`; otherwise use the unversioned shared library as the link interface
- raw imports stay lightweight; cross-library semantic composition belongs in `foo`, `foo_system`, and `foo_static`
- `foo_system` mirrors `foo` dependencies using system variants and must not reach packaged `*_shared_library` targets
- `foo_static` must not reach shared or system variants and must include all required hermetic static side dependencies
- compile-time transitive dependencies belong on `headers` so shared, system, and static variants see the same header closure
- existing public labels may predate this naming convention; preserve compatibility rather than renaming them mechanically
- add every public family member to `cuda/redist_proxy_targets.bzl` and exercise it through the stable proxy namespace

## Scope

Current scope is the CUDA redistribution model implemented by this repository.
Expand Down Expand Up @@ -153,6 +176,7 @@ Key files:
- Do not invent fake components when upstream does not ship one.
- Keep broad version selection in registries; keep fine layout differences in BUILD templates.
- Add public targets through `redist_proxy_targets.bzl` plus matching templates.
- For new library capabilities, follow the target-family and dependency-purity rules above.
- Public API changes require README/docs review.

## Verification
Expand All @@ -172,3 +196,8 @@ CI matrix mirrors:
- `13_3_1`

For compatibility-sensitive changes, run the relevant e2e platform for every affected CUDA major/minor.

For library target-family changes:
- build the public shared, system, and static variants that exist upstream
- inspect the configured dependency graph and confirm `_system` closures contain no packaged `*_shared_library` targets
- inspect link actions or output artifacts to confirm system targets use only the intended link interface and do not package redistribution DSOs
Loading