backport PR #986

Author: leofang

.github/workflows/build-wheel.yml

@@ -30,6 +30,7 @@ jobs:
           - "3.11"
           - "3.12"
           - "3.13"
+          - "3.13t"
     name: py${{ matrix.python-version }}
     runs-on: ${{ (inputs.host-platform == 'linux-64' && 'linux-amd64-cpu8') ||
                  (inputs.host-platform == 'linux-aarch64' && 'linux-arm64-cpu8') ||
@@ -116,6 +117,7 @@ jobs:
           CIBW_REPAIR_WHEEL_COMMAND_WINDOWS: "delvewheel repair --namespace-pkg cuda -w {dest_dir} {wheel}"
           CIBW_ENVIRONMENT: >
             CUDA_PYTHON_PARALLEL_LEVEL=${{ env.CUDA_PYTHON_PARALLEL_LEVEL }}
+          CIBW_ENABLE: "cpython-freethreading"
         with:
           package-dir: ./cuda_core/
           output-dir: ${{ env.CUDA_CORE_ARTIFACTS_DIR }}
@@ -163,6 +165,7 @@ jobs:
             CUDA_PYTHON_PARALLEL_LEVEL=${{ env.CUDA_PYTHON_PARALLEL_LEVEL }}
           CIBW_BEFORE_BUILD_WINDOWS: "pip install delvewheel"
           CIBW_REPAIR_WHEEL_COMMAND_WINDOWS: "delvewheel repair --namespace-pkg cuda -w {dest_dir} {wheel}"
+          CIBW_ENABLE: "cpython-freethreading"
         with:
           package-dir: ./cuda_bindings/
           output-dir: ${{ env.CUDA_BINDINGS_ARTIFACTS_DIR }}
@@ -223,6 +226,10 @@ jobs:
           # workaround for actions/runner-images#12377 (the cached 3.13.4 is buggy on Windows)
           python-version: ${{ matrix.python-version == '3.13' && '3.13.5' || matrix.python-version }}
 
+      - name: verify free-threaded build
+        if: endsWith(matrix.python-verison, 't')
+        run: python -c 'import sys; assert not sys._is_gil_enabled()'
+
       - name: Set up Python include paths
         run: |
           if [[ "${{ inputs.host-platform }}" == linux* ]]; then

ci/tools/run-tests

@@ -75,12 +75,18 @@ elif [[ "${test_module}" == "core" ]]; then
   echo "Installing core wheel"
   pwd
   ls
+
+  FREE_THREADING=""
+  if python -c 'import sys; assert not sys._is_gil_enabled()' 2> /dev/null; then
+    FREE_THREADING+="-ft"
+  fi
+
   if [[ "${LOCAL_CTK}" == 1 ]]; then
     # We already installed cuda-bindings, and all CTK components exist locally,
     # so just install the test dependencies.
-    pip install $(ls *.whl)["test-cu${TEST_CUDA_MAJOR}"]
+    pip install $(ls *.whl)["test-cu${TEST_CUDA_MAJOR}${FREE_THREADING}"]
   else
-    pip install $(ls *.whl)["cu${TEST_CUDA_MAJOR}","test-cu${TEST_CUDA_MAJOR}"]
+    pip install $(ls *.whl)["cu${TEST_CUDA_MAJOR}","test-cu${TEST_CUDA_MAJOR}${FREE_THREADING}"]
   fi
   popd
   pushd ./cuda_core

cuda_bindings/pyproject.toml

@@ -2,7 +2,7 @@
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 
 [build-system]
-requires = ["setuptools>=77.0.0", "cython>=3.0,<3.1.0", "pyclibrary>=0.1.7"]
+requires = ["setuptools>=77.0.0", "cython>=3.1,<3.2", "pyclibrary>=0.1.7"]
 build-backend = "setuptools.build_meta"
 
 [project]
@@ -40,7 +40,7 @@ all = [
 ]
 
 test = [
-    "cython>=3.0,<3.1.0",
+    "cython>=3.1,<3.2",
     "setuptools>=77.0.0",
     "numpy>=1.21.1",
     "pytest>=6.2.4",