Refactor cuda-bindings setup.py to defer CUDA_HOME check to build time #1635

Problem

cuda_bindings/setup.py requires CUDA_HOME/CUDA_PATH at module import time (line 29-31) and imports Cython/pyclibrary at the top level (lines 16-17). This means any tool that needs to read package metadata — including uv lock, uv sync, and pip install --dry-run — must have the CUDA Toolkit installed and CUDA_HOME set, even when no actual build is happening.

This creates friction with modern Python tooling (uv workspaces, lockfile generation, metadata introspection) and prevents patterns like uv sync --package cuda-bindings --group test from working without a full CUDA environment.

Proposed solution

Convert cuda_bindings to use a custom PEP 517 build backend (similar to what cuda_core already does with build_hooks.py):

1. Defer CUDA_HOME check — Move it from module-level into the actual build functions (build_wheel, build_editable). get_requires_for_build_wheel should not need CUDA.
2. Defer heavy imports — Cython, pyclibrary, and other build-only imports should happen inside build functions, not at module import.
3. Static metadata in pyproject.toml — Anything that can be declared statically (package name, deps, classifiers) should stay in pyproject.toml.
4. Custom build backend — A build_hooks.py (or similar) that wraps setuptools.build_meta and only invokes CUDA-dependent logic when actually building.

Context

This came up during the uv migration (#1615). Currently uv lock requires CUDA_HOME=/usr/local/cuda as a workaround. The cuda_core package already follows the deferred pattern via its build_hooks.py.

References

• cuda_core/build_hooks.py — existing example of the deferred pattern
• cuda_bindings/setup.py — current implementation (~400 lines)
• Switch CI to uv and use build[uv] frontend for cibuildwheel #1615 — uv workspace migration PR where this friction was encountered