CI: GCC 14 on Linux, and the CUDA wheel loaded from every Python

.github/workflows/copilot-setup-steps.yml

@@ -20,7 +20,7 @@ jobs:
           submodules: recursive
 
       - name: Install compiler and CMake
-        run: sudo apt-get install -y cmake clang
+        run: sudo apt-get install -y cmake gcc-14 g++-14
 
       - name: Install torch (CPU) and the build backend
         run: |
@@ -37,6 +37,6 @@ jobs:
 
       - name: Build and install bartorch
         env:
-          CC: clang
-          CXX: clang++
+          CC: gcc-14
+          CXX: g++-14
         run: pip install -e . --no-build-isolation

.github/workflows/docs.yml

@@ -107,9 +107,10 @@ jobs:
           python-version: "3.12"
           cache: pip
 
-      # libbartorch is a clang build: BART's nested functions become Blocks.
-      - name: Install the compiler, OpenMP and CMake
-        run: sudo apt-get install -y cmake clang libomp-dev
+      # GCC 14, as upstream BART builds on Linux: its heap trampolines are
+      # what BART's nested functions need.
+      - name: Install the compiler and CMake
+        run: sudo apt-get install -y cmake gcc-14 g++-14
 
       - name: Install torch (CPU) and the build backend
         run: |
@@ -119,8 +120,8 @@ jobs:
 
       - name: Build and install bartorch
         env:
-          CC: clang
-          CXX: clang++
+          CC: gcc-14
+          CXX: g++-14
         run: pip install -e . --no-build-isolation
 
       - name: Install the documentation and example requirements

.github/workflows/publish.yml

@@ -208,19 +208,25 @@ jobs:
             -w wheelhouse/ --plat manylinux_2_28_x86_64 \
             --exclude libcudart.so.12 --exclude libcufft.so.11 \
             --exclude libcublas.so.12 --exclude libcublasLt.so.12
-      - name: Check that it loads without a card
+      # The wheel carries no interpreter, so it is loaded from each Python the
+      # package supports.
+      - name: Check that it loads without a card, on every Python
         run: |
-          /opt/python/cp312-cp312/bin/python -m pip install wheelhouse/*.whl \
-            torch --index-url https://download.pytorch.org/whl/cpu \
-            --extra-index-url https://pypi.org/simple
-          /opt/python/cp312-cp312/bin/python -c "
+          for py in cp310-cp310 cp311-cp311 cp312-cp312 cp313-cp313 cp314-cp314; do
+            python=/opt/python/$py/bin/python
+            $python -m pip install wheelhouse/*.whl \
+              torch --index-url https://download.pytorch.org/whl/cpu \
+              --extra-index-url https://pypi.org/simple
+            $python -c "
+          import sys
           import bartorch
           from bartorch import _cuda
           info = bartorch.build_info()
-          print(info)
+          print(sys.version.split()[0], info)
           assert 'cuda=ON' in info, info
           print('devices:', _cuda.device_count())
           "
+          done
       - uses: actions/upload-artifact@v4
         with:
           name: cuda-wheel

.github/workflows/tests.yml

@@ -10,8 +10,9 @@ env:
   FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
 
 jobs:
-  # BART's nested functions become Blocks under clang and heap trampolines
-  # under GCC 14.  Each platform runs the oldest and the newest Python the
+  # The toolchains are upstream BART's: GCC 14 on Linux, whose heap
+  # trampolines carry BART's nested functions, and clang on macOS, where they
+  # become Blocks.  Each platform runs the oldest and the newest Python the
   # package supports; the wheel workflow tests the wheel on both as well.
   # Windows is clang from MSYS2's CLANG64 environment, linked against the
   # Universal C Runtime that official CPython uses.
@@ -26,7 +27,7 @@ jobs:
             extras: "[mkl]"
           - os: ubuntu-latest
             python-version: "3.14"
-            cc: clang
+            cc: gcc-14
             # One job measures coverage for Codecov; the rest only test.
             coverage: true
           - os: macos-latest
@@ -210,8 +211,8 @@ jobs:
           python-version: "3.12"
           cache: pip
 
-      - name: Install the CUDA toolkit, clang and OpenMP
-        run: sudo apt-get update -qq && sudo apt-get install -y --no-install-recommends cmake clang libomp-dev nvidia-cuda-toolkit
+      - name: Install the CUDA toolkit and GCC
+        run: sudo apt-get update -qq && sudo apt-get install -y --no-install-recommends cmake gcc-14 g++-14 g++-12 nvidia-cuda-toolkit
 
       - name: Install torch (CPU) and the build backend
         run: |
@@ -226,10 +227,13 @@ jobs:
       - name: Install torchvision from the same index as torch
         run: pip install torchvision --index-url https://download.pytorch.org/whl/cpu
 
+      # BART is GCC 14's, for its heap trampolines; the host side of the
+      # kernels is GCC 12's, the newest this distribution's nvcc accepts.
       - name: Build with BART's CUDA kernels
         env:
-          CC: clang
-          CXX: clang++
+          CC: gcc-14
+          CXX: g++-14
+          CUDAHOSTCXX: g++-12
         run: pip install -e . --no-build-isolation --config-settings=cmake.define.BARTORCH_CUDA=ON
 
       - name: The CUDA build must report its device code and run on the host

AGENTS.md

@@ -624,7 +624,9 @@ the stack needs an executable stack, which glibc 2.41 refuses to `dlopen`; so
 GCC 14's `-ftrampoline-impl=heap` is required and older GCC is rejected at
 configure time. BART's own `NOEXEC_STACK` workaround does not help here: it
 parses a trampoline layout GCC emits only for non-PIC executables, not for a
-shared library. Both compilers are built and tested in CI.
+shared library. CI builds and tests with upstream BART's toolchains, GCC 14 on
+Linux and clang on macOS, and with clang on Windows, whose Blocks come from the
+vendored runtime.
 
 **Windows** is LLP64: `long` is 32 bits, so BART keeps every extent, stride
 and flag set in the fixed-width types of `misc/dimtypes.h` (`bart_dim_t`,

docs/guides/developer/pull-requests.md

@@ -11,8 +11,7 @@ template:
 | Validation | The commands run and their results; for a numerical change, the reference and the tolerance; hardware or optional dependencies that were not available |
 | Public API and documentation | Changes to public behaviour and the documentation updated for them |
 
-CI runs the test matrix (Linux with GCC 14 and clang, macOS and Windows with
-clang, each on Python 3.10 and 3.14, and a CUDA build without a device), the lint and spelling checks, and two documentation builds:
+CI runs the test matrix (Linux with GCC 14, macOS and Windows with clang, each on Python 3.10 and 3.14, and a CUDA build without a device), the lint and spelling checks, and two documentation builds:
 the reference without the compiled library, and the executed gallery, which is
 published from `main` and from release tags.  A pull request is merged when CI passes and a
 maintainer has approved it.