test rapids build backend (#639)

* test rapids

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* remove free-threading

* refactor release note

* fix docs

* clean up build from source

* update release note

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

Author: Intron7

.github/workflows/publish.yml

@@ -58,46 +58,6 @@ jobs:
           echo "version=$VERSION" >> $GITHUB_OUTPUT
           echo "Detected version: $VERSION"
 
-      - name: Set package name, extras, and CUDA archs for CUDA ${{ matrix.cuda_major }}
-        run: |
-          python3 - ${{ matrix.cuda_major }} "${{ matrix.cuda_archs }}" <<'SCRIPT'
-          import sys, pathlib
-
-          cuda = sys.argv[1]
-          cuda_archs = sys.argv[2]
-          other = "13" if cuda == "12" else "12"
-          path = pathlib.Path("pyproject.toml")
-          text = path.read_text()
-
-          # Rename package
-          text = text.replace(
-              'name = "rapids-singlecell"',
-              f'name = "rapids-singlecell-cu{cuda}"',
-          )
-          # Rename matching extra to "rapids", remove the other
-          text = text.replace(f'rapids-cu{cuda} =', 'rapids =')
-          # Remove the other CUDA extra line entirely
-          lines = text.splitlines(keepends=True)
-          text = "".join(l for l in lines if f'rapids-cu{other}' not in l)
-
-          # Set CUDA architectures (replace "native" with CI target archs)
-          text = text.replace(
-              'CMAKE_CUDA_ARCHITECTURES = "native"',
-              f'CMAKE_CUDA_ARCHITECTURES = "{cuda_archs}"',
-          )
-
-          path.write_text(text)
-
-          # Verify
-          for line in path.read_text().splitlines():
-              if "name" in line or "rapids" in line.lower() or "CUDA_ARCH" in line:
-                  print(line)
-          SCRIPT
-
-      - name: Sanity check pyproject.toml
-        run: |
-          grep -E "name|rapids|CUDA_ARCH" pyproject.toml
-
       - name: Build CUDA manylinux image
         run: |
           docker build -t "${{ matrix.cibw_image }}" -f "${{ matrix.dockerfile }}" docker
@@ -106,17 +66,19 @@ jobs:
         uses: pypa/cibuildwheel@v3.1.4
         env:
           SETUPTOOLS_SCM_PRETEND_VERSION: ${{ steps.version.outputs.version }}
+          CIBW_BUILD: 'cp312-*'
           CIBW_SKIP: '*-musllinux*'
           CIBW_MANYLINUX_X86_64_IMAGE: ${{ matrix.os == 'linux-intel' && matrix.cibw_image || '' }}
           CIBW_MANYLINUX_AARCH64_IMAGE: ${{ matrix.os == 'linux-arm' && matrix.cibw_image || '' }}
           CIBW_ENVIRONMENT_PASS_LINUX: SETUPTOOLS_SCM_PRETEND_VERSION
+          CIBW_CONFIG_SETTINGS: "cmake.define.CMAKE_CUDA_ARCHITECTURES=${{ matrix.cuda_archs }}"
           CIBW_ENVIRONMENT: >
             CUDA_PATH=/usr/local/cuda
             LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
             PATH=/usr/local/cuda/bin:$PATH
+            RAPIDS_DISABLE_CUDA=false
           CIBW_BEFORE_BUILD: >
             python -m pip install -U pip
-            scikit-build-core cmake ninja nanobind
           CIBW_TEST_SKIP: "*"
           CIBW_TEST_COMMAND: ""
           CIBW_REPAIR_WHEEL_COMMAND: "auditwheel repair --exclude libcublas.so.${{ matrix.cuda_major }} --exclude libcublasLt.so.${{ matrix.cuda_major }} --exclude libcudart.so.${{ matrix.cuda_major }} -w {dest_dir} {wheel}"
@@ -132,8 +94,10 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v5
+        with:
+          fetch-depth: 0
       - name: Build sdist
-        run: pipx run build --sdist
+        run: RAPIDS_DISABLE_CUDA=true pipx run build --sdist
       - uses: actions/upload-artifact@v4
         with:
           name: cibw-sdist

ci/environment.yml

@@ -5,6 +5,6 @@ channels:
  - conda-forge
 dependencies:
  - rapids=26.04
- - python=3.13
+ - python=3.14
  - cuda-version=12.9
  - cudnn

ci/environment_alpha.yml

@@ -5,7 +5,7 @@ channels:
  - conda-forge
 dependencies:
  - rapids=26.04
- - python=3.13
+ - python=3.14
  - cuda-version=12.9
  - zarr=3.0.6
  - cudnn

conda/rsc_rapids_26.04_cuda12.yml

@@ -6,7 +6,7 @@ channels:
  - bioconda
 dependencies:
  - rapids=26.04
- - python=3.13
+ - python=3.14
  - cuda-version=12.9
  - cudnn
  - cutensor

conda/rsc_rapids_26.04_cuda13.yml

@@ -6,7 +6,7 @@ channels:
  - bioconda
 dependencies:
  - rapids=26.04
- - python=3.13
+ - python=3.14
  - cuda-version=13.1
  - cudnn
  - cutensor

dependencies.yaml

@@ -0,0 +1,33 @@
+files:
+  pyproject:
+    output: pyproject
+    pyproject_dir: .
+    extras:
+      table: project.optional-dependencies
+      key: rapids
+    includes:
+      - rapids
+
+dependencies:
+  rapids:
+    specific:
+      - output_types: pyproject
+        matrices:
+          - matrix:
+              cuda: "12.*"
+            packages:
+              - cupy-cuda12x
+              - cudf-cu12>=25.10
+              - cuml-cu12>=25.10
+              - cugraph-cu12>=25.10
+              - cuvs-cu12>=25.10
+          - matrix:
+              cuda: "13.*"
+            packages:
+              - cupy-cuda13x
+              - cudf-cu13>=25.10
+              - cuml-cu13>=25.10
+              - cugraph-cu13>=25.10
+              - cuvs-cu13>=25.10
+          - matrix:
+            packages: []

docs/contributing.md

@@ -156,14 +156,14 @@ The `add_nb_cuda_module` helper automatically handles:
 
 The project uses [hatch](https://hatch.pypa.io/) to manage test environments. The test matrix is defined in `hatch.toml` with two axes:
 
-- **`cuda`**: `12` or `13` — selects the matching RAPIDS/CuPy packages
+- **`cuda`**: `12` or `13` — selects the matching RAPIDS and CuPy packages
 - **`deps`**: `stable`, `dev`, or `rapids_prerelease` — controls Python version and dependency sources
 
 | `deps` | Python | Description |
 |---|---|---|
 | `stable` | 3.12 | Released versions of all dependencies |
-| `dev` | 3.13 | Upstream `main` branches of anndata and scanpy |
-| `rapids_prerelease` | 3.13 | RAPIDS nightly wheels |
+| `dev` | 3.14 | Upstream `main` branches of anndata and scanpy |
+| `rapids_prerelease` | 3.14 | RAPIDS nightly wheels |
 
 To run the test suite against a specific matrix combination:
 
@@ -230,7 +230,7 @@ The project publishes three separate packages:
 ### Wheel builds
 
 Wheels are built via [cibuildwheel](https://cibuildwheel.pypa.io/) in GitHub Actions using custom manylinux Docker images with CUDA toolkit pre-installed.
-The CI renames the package and adjusts optional dependencies per CUDA version using an inline Python script in `publish.yml`.
+The RAPIDS build backend applies the CUDA package suffix and generates the `rapids` extra from `dependencies.yaml`.
 
 Each wheel contains:
 - Compiled `.abi3.so` modules (stable ABI, one wheel per platform for all Python 3.12+ versions)

docs/installation.md

@@ -10,15 +10,15 @@ conda env create -f conda/rsc_rapids_26.04_cuda13.yml
 # or
 mamba env create -f conda/rsc_rapids_26.04_cuda13.yml
 ```
-*Python 3.13, CUDA 13.1*
+*Python 3.14, CUDA 13.1*
 ````
 ````{tab-item} CUDA 12
 ```bash
 conda env create -f conda/rsc_rapids_26.04_cuda12.yml
 # or
 mamba env create -f conda/rsc_rapids_26.04_cuda12.yml
 ```
-*Python 3.13, CUDA 12.9*
+*Python 3.14, CUDA 12.9*
 ````
 `````
 
@@ -82,6 +82,7 @@ This is the recommended approach for **conda/mamba users** who already have RAPI
 ### Prebuilt wheels with RAPIDS dependencies
 
 To also install the RAPIDS stack via pip, use the `rapids` extra.
+This is the only RAPIDS dependency extra; select CUDA 12 or CUDA 13 with the package name (`rapids-singlecell-cu12` or `rapids-singlecell-cu13`), not with a CUDA-specific extra.
 This requires the `--extra-index-url` flag for the NVIDIA PyPI index:
 
 `````{tab-set}
@@ -107,14 +108,17 @@ pip install rapids-singlecell
 ```
 
 The CUDA kernels will be compiled during installation for your local GPU architecture.
-You can select RAPIDS dependencies with the `rapids-cu12` or `rapids-cu13` extras:
+Install the RAPIDS stack separately before building from source, or use the single `rapids` extra:
 
 ```bash
-pip install 'rapids-singlecell[rapids-cu12]' --extra-index-url=https://pypi.nvidia.com
+pip install 'rapids-singlecell[rapids]' --extra-index-url=https://pypi.nvidia.com
 ```
 
+The `rapids` extra is the only RAPIDS dependency extra; there are no CUDA-specific extras such as `rapids-cu12` or `rapids-cu13`.
+
 ```{note}
 Building from source requires the CUDA toolkit (nvcc) and CMake >= 3.24 to be available in your environment.
+The nvcc/CUDAToolkit found during the build should match the RAPIDS/CuPy CUDA major runtime version in or linked to the environment.
 ```
 
 ### Install from GitHub

docs/release-notes/0.15.0.md

@@ -1,9 +1 @@
 ### 0.15.0 {small}`the-future`
-
-```{rubric} Features
-```
-* Add Harmony2 support with stabilized diversity penalty, dynamic per-cluster-per-batch ridge regularization, and automatic batch pruning {cite:p}`Patikas2026` {pr}`625` {smaller}`S Dicks`
-
-```{rubric} Misc
-```
-* Refactor ``tl.rank_genes_groups`` internals to use categorical integer codes instead of boolean mask matrices {pr}`570` {smaller}`S Dicks`

docs/release-notes/0.15.0rc6.md

@@ -0,0 +1,11 @@
+### 0.15.0rc6 {small}`2026-04-27`
+
+```{rubric} Features
+```
+* Add Harmony2 support with stabilized diversity penalty, dynamic per-cluster-per-batch ridge regularization, and automatic batch pruning {cite:p}`Patikas2026` {pr}`625` {smaller}`S Dicks`
+
+```{rubric} Misc
+```
+* Switch wheel packaging to ``rapids-build-backend`` so CUDA-suffixed wheel names and RAPIDS dependency extras are generated from a single dependency specification {pr}`639` {smaller}`S Dicks`
+* Align RAPIDS 26.04 conda and CI environments with Python 3.14 {pr}`639` {smaller}`S Dicks`
+* Refactor ``tl.rank_genes_groups`` internals to use categorical integer codes instead of boolean mask matrices {pr}`570` {smaller}`S Dicks`