feat: add cuda buffer backend

.github/workflows/testpr.yml

@@ -161,12 +161,6 @@ jobs:
       run: |
         # Uncomment to force-rebuild a package with a corrupted cache entry.
         # rm -rf ${{ matrix.folder_cache }}/<package>* 2>/dev/null || true
-        rm -rf ${{ matrix.folder_cache }}/ros2-roboplan* 2>/dev/null || true
-        rm -rf ${{ matrix.folder_cache }}/ros-rolling-roboplan* 2>/dev/null || true
-        rm -rf ${{ matrix.folder_cache }}/ros-rolling-rosidl-generator-py* ${{ matrix.folder_cache }}/ros2-rosidl-generator-py* 2>/dev/null || true
-        rm -rf ${{ matrix.folder_cache }}/ros-rolling-rosidl-core-generators* ${{ matrix.folder_cache }}/ros2-rosidl-core-generators* 2>/dev/null || true
-        rm -rf ${{ matrix.folder_cache }}/ros-rolling-rosidl-default-generators* ${{ matrix.folder_cache }}/ros2-rosidl-default-generators* 2>/dev/null || true
-        rm -rf ${{ matrix.folder_cache }}/ros-rolling-std-msgs* ${{ matrix.folder_cache }}/ros2-std-msgs* 2>/dev/null || true
         mkdir -p ${{ matrix.folder_cache }}
         pixi run rattler-index fs ${{ matrix.folder_cache }}/.. --force
 

README.md

@@ -39,3 +39,49 @@ If you use RoboStack in your academic work, please refer to the following paper:
 ## Installation, FAQ, and Contributing Instructions
 
 Please see our instructions [here](https://robostack.github.io/GettingStarted.html).
+
+## CUDA buffer builds
+
+On Linux, `cuda_buffer` and `cuda_buffer_backend` build two variants. The
+`cuda_buffer_cuda_version` matrix in `conda_build_config.yaml` is preserved by
+the matching override in `vinca_pinning.yaml`.
+
+| Build SDK | Runtime constraint | CUDA runtime ABI |
+| --- | --- | --- |
+| 12.9 | `cuda-version >=12.9,<13` | `libcudart.so.12` |
+| 13.0 | `cuda-version >=13.0,<14` | `libcudart.so.13` |
+
+`pixi run build` builds both variants. To debug the dependency chain, run
+`pixi run build-one ros2-cuda-buffer` followed by
+`pixi run build-one ros2-cuda-buffer-backend`. Select the installed variant by
+constraining `cuda-version` in the consuming environment; the package build
+hashes distinguish the two variants. Match that CUDA major with GPU LibTorch
+when using both libraries in the same environment.
+
+CUDA buffer and GPU LibTorch builds use GCC 14 on Linux, to satisfy the CUDA
+tool packages' compiler constraints. Native build tools live in the build
+prefix; target CUDA runtime,
+driver stubs and CRT headers live in the host prefix. The installed packages
+depend on the CUDA runtime, without bundling driver stubs or requiring nvcc at
+runtime. The NVIDIA driver supplies `libcuda.so.1` on the target machine.
+macOS and Windows skip these packages because the sources use POSIX IPC and
+`librt`. LibTorch and the message-only packages retain their existing platform
+support.
+
+## LibTorch CPU and GPU builds
+
+`libtorch_vendor` reuses conda-forge's `libtorch` through `find_package(Torch)`.
+The default builds use its CPU variant. Set `CF_CUDA_ENABLED=True` to additionally
+build CUDA 12.9 and 13.0 variants on `linux-64`, `linux-aarch64`, and `win-64`:
+
+```bash
+CF_CUDA_ENABLED=True pixi run build
+```
+
+This uses the existing `cuda_compiler_version` matrix in `vinca_pinning.yaml`;
+macOS keeps CPU builds even with the flag enabled. LibTorch and
+`torch_conversions` runtime dependencies select the same CPU/CUDA variant.
+Linux GPU `torch_conversions` also enables `cuda_buffer`, with the matching
+CUDA major. Builds use the CUDA SDK and driver stubs, so an NVIDIA driver and
+GPU are not required for compilation. Running CUDA tensor operations requires
+a compatible NVIDIA driver and GPU.

conda_build_config.yaml

@@ -9,6 +9,7 @@ c_compiler_version:            # [unix]
   - 15                           # [linux]
   - 21                           # [osx]
   - 14                           # [linux and (x86_64 or aarch64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
+  - 14                           # [linux and (x86_64 or aarch64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
 c_stdlib:
   - sysroot                    # [linux]
   - macosx_deployment_target   # [osx]
@@ -21,6 +22,7 @@ c_stdlib_version:              # [unix]
   - 2.28                         # [linux and not riscv64]
   - 2.39                         # [linux and riscv64]
   - 2.28                         # [linux and (x86_64 or aarch64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
+  - 2.28                         # [linux and (x86_64 or aarch64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
   - 14.0                         # [osx]
 cxx_compiler:
   - gxx                        # [linux]
@@ -31,6 +33,7 @@ cxx_compiler_version:          # [unix]
   - 15                           # [linux]
   - 21                           # [osx]
   - 14                           # [linux and (x86_64 or aarch64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
+  - 14                           # [linux and (x86_64 or aarch64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
 llvm_openmp:                   # [osx]
   - 21                         # [osx]
 fortran_compiler:              # [unix or win]
@@ -41,6 +44,7 @@ fortran_compiler_version:      # [unix or win]
   - 5                            # [win64]
   - 22                           # [win and arm64]
   - 14                           # [linux and (x86_64 or aarch64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
+  - 14                           # [linux and (x86_64 or aarch64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
 m2w64_c_compiler:                 # [win]
   - gcc                           # [win]
 m2w64_c_compiler_version:         # [win]
@@ -81,6 +85,7 @@ cuda_compiler:
 cuda_compiler_version:
   - None
   - 12.9                         # [((linux and (x86_64 or aarch64)) or win64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
+  - 13.0                         # [((linux and (x86_64 or aarch64)) or win64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
 cuda_compiler_version_min:
   - None                       # [not ((linux and (x86_64 or aarch64)) or win64)]
   - 12.9                       # [((linux and (x86_64 or aarch64)) or win64)]
@@ -979,7 +984,6 @@ python:
   - 3.14.* *_cp314
 python_impl:
   - cpython
-
 python_min:
   # minimum supported python version per CFEP-25
   # bump to next minor version when we drop python versions
@@ -1172,6 +1176,11 @@ zlib_ng:
   - '2.3'
 zstd:
   - '1.5'
+cuda_buffer_cuda_version:
+  - '12.9'
+  - '13.0'
+  # Build commands provide their channels with `-c`, so omit the inherited
+  # conda-forge-pinning channel_sources value from the rendered configuration.
 libzenohc:
   - 1.10.1
 libzenohcxx:

patch/dependencies.yaml

@@ -32,6 +32,68 @@ sdl2_vendor:
 libcurl_vendor:
   add_host: ["libcurl"]
   add_run: ["libcurl"]
+libtorch_vendor:
+  # Torch_FOUND skips the upstream download. CPU/cu12/cu13 must stay consistent.
+  add_build: &libtorch_build
+    - "${{ compiler('cuda') if cuda_compiler_version != 'None' else '' }}"
+  add_host:
+    - "libtorch * ${{ 'cpu' if cuda_compiler_version == 'None' else 'cuda' ~ (cuda_compiler_version | replace('.', '')) }}*"
+    - "${{ 'cuda-version ' ~ cuda_compiler_version ~ '.*' if cuda_compiler_version != 'None' else '' }}"
+    - "${{ 'cuda-cudart-dev' if cuda_compiler_version != 'None' else '' }}"
+    - "${{ 'cuda-driver-dev' if linux and cuda_compiler_version != 'None' else '' }}"
+    - "${{ 'cuda-nvrtc-dev' if cuda_compiler_version != 'None' else '' }}"
+    - "${{ 'nvtx-c' if cuda_compiler_version != 'None' else '' }}"
+    - "${{ 'arm-variant * sbsa' if aarch64 and cuda_compiler_version != 'None' else '' }}"
+  add_run:
+    - "libtorch * ${{ 'cpu' if cuda_compiler_version == 'None' else 'cuda' ~ (cuda_compiler_version | replace('.', '')) }}*"
+    - "${{ 'cuda-version >=' ~ cuda_compiler_version ~ ',<' ~ ((cuda_compiler_version | split('.'))[0] | int + 1) if cuda_compiler_version != 'None' else '' }}"
+torch_conversions:
+  add_build: *libtorch_build
+  add_host:
+    - "libtorch * ${{ 'cpu' if cuda_compiler_version == 'None' else 'cuda' ~ (cuda_compiler_version | replace('.', '')) }}*"
+    - "${{ 'cuda-version ' ~ cuda_compiler_version ~ '.*' if cuda_compiler_version != 'None' else '' }}"
+    - "${{ 'cuda-cudart-dev' if cuda_compiler_version != 'None' else '' }}"
+    - "${{ 'cuda-driver-dev' if linux and cuda_compiler_version != 'None' else '' }}"
+    - "${{ 'cuda-nvrtc-dev' if cuda_compiler_version != 'None' else '' }}"
+    - "${{ 'nvtx-c' if cuda_compiler_version != 'None' else '' }}"
+    - "${{ 'arm-variant * sbsa' if aarch64 and cuda_compiler_version != 'None' else '' }}"
+    - "${{ 'ros2-cuda-buffer' if linux and cuda_compiler_version != 'None' else '' }}"
+  add_run:
+    - "libtorch * ${{ 'cpu' if cuda_compiler_version == 'None' else 'cuda' ~ (cuda_compiler_version | replace('.', '')) }}*"
+    - "${{ 'cuda-version >=' ~ cuda_compiler_version ~ ',<' ~ ((cuda_compiler_version | split('.'))[0] | int + 1) if cuda_compiler_version != 'None' else '' }}"
+    - "${{ 'ros2-cuda-buffer' if linux and cuda_compiler_version != 'None' else '' }}"
+cuda_buffer:
+  # No .cu sources: nvcc is used only by FindCUDAToolkit for discovery/versioning.
+  # The CUDA 12.9/13.0 tool packages constrain GCC to <15, even for --version.
+  remove_build: ["${{ compiler('c') }}", "${{ compiler('cxx') }}"]
+  add_build:
+    - "gcc_${{ target_platform }} 14.*"
+    - "gxx_${{ target_platform }} 14.*"
+    - "cuda-nvcc-tools ${{ cuda_buffer_cuda_version }}.*"
+  add_host:
+    - "cuda-version ${{ cuda_buffer_cuda_version }}.*"
+    - cuda-cudart-dev
+    - cuda-driver-dev
+    - "cuda-crt-dev_${{ target_platform }} ${{ cuda_buffer_cuda_version }}.*"
+    - if: aarch64
+      then: "arm-variant * sbsa"
+  add_run:
+    - "cuda-version >=${{ cuda_buffer_cuda_version }},<${{ (cuda_buffer_cuda_version | split('.'))[0] | int + 1 }}"
+cuda_buffer_backend:
+  remove_build: ["${{ compiler('c') }}", "${{ compiler('cxx') }}"]
+  add_build:
+    - "gcc_${{ target_platform }} 14.*"
+    - "gxx_${{ target_platform }} 14.*"
+    - "cuda-nvcc-tools ${{ cuda_buffer_cuda_version }}.*"
+  add_host:
+    - "cuda-version ${{ cuda_buffer_cuda_version }}.*"
+    - cuda-cudart-dev
+    - cuda-driver-dev
+    - "cuda-crt-dev_${{ target_platform }} ${{ cuda_buffer_cuda_version }}.*"
+    - if: aarch64
+      then: "arm-variant * sbsa"
+  add_run:
+    - "cuda-version >=${{ cuda_buffer_cuda_version }},<${{ (cuda_buffer_cuda_version | split('.'))[0] | int + 1 }}"
 intra_process_demo:
   add_host: ["${{ 'libgl-devel' if linux }}", "${{ 'libopengl-devel' if linux }}"]
   # Depends on opencv with gui support, see https://github.com/ros2/demos/blob/8f361a1ece2daae54016959f3ea9ab2e7692ddbb/intra_process_demo/include/image_pipeline/image_view_node.hpp#L58

patch/ros-rolling-torch-conversions.patch

@@ -0,0 +1,12 @@
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -46,3 +46,3 @@
+ if(_torch_has_cuda AND CUDAToolkit_FOUND)
+-  target_link_libraries(${PROJECT_NAME} INTERFACE CUDA::cudart)
++  target_link_libraries(${PROJECT_NAME} INTERFACE CUDA::cudart CUDA::cuda_driver)
+   target_compile_definitions(${PROJECT_NAME} INTERFACE
+@@ -68,3 +68,3 @@
+ ament_export_dependencies(
+-  libtorch_vendor rosidl_buffer tensor_msgs rmw rcutils)
++  libtorch_vendor Torch rosidl_buffer tensor_msgs rmw rcutils)
+ if(_torch_has_cuda AND CUDAToolkit_FOUND)

pkg_additional_info.yaml

@@ -17,6 +17,18 @@ compressed_depth_image_transport:
   additional_cmake_args: "-DCMAKE_WINDOWS_EXPORT_ALL_SYMBOLS=ON"
 console_bridge_vendor:
   additional_cmake_args: "-DAMENT_VENDOR_POLICY=NEVER_VENDOR_IGNORE_SATISFIED_CHECK"
+cuda_buffer:
+  # nvcc runs in the build environment; headers/libraries come from the host SDK.
+  # BIN_DIR prevents FindCUDAToolkit from replacing host includes with nvcc's build paths.
+  additional_cmake_args:
+    if: aarch64
+    then: "-DCUDAToolkit_BIN_DIR=$PREFIX/bin -DCUDAToolkit_NVCC_EXECUTABLE=$BUILD_PREFIX/bin/nvcc -DCUDAToolkit_INCLUDE_DIRECTORIES=$PREFIX/targets/sbsa-linux/include"
+    else: "-DCUDAToolkit_BIN_DIR=$PREFIX/bin -DCUDAToolkit_NVCC_EXECUTABLE=$BUILD_PREFIX/bin/nvcc -DCUDAToolkit_INCLUDE_DIRECTORIES=$PREFIX/targets/x86_64-linux/include"
+cuda_buffer_backend:
+  additional_cmake_args:
+    if: aarch64
+    then: "-DCUDAToolkit_BIN_DIR=$PREFIX/bin -DCUDAToolkit_NVCC_EXECUTABLE=$BUILD_PREFIX/bin/nvcc -DCUDAToolkit_INCLUDE_DIRECTORIES=$PREFIX/targets/sbsa-linux/include"
+    else: "-DCUDAToolkit_BIN_DIR=$PREFIX/bin -DCUDAToolkit_NVCC_EXECUTABLE=$BUILD_PREFIX/bin/nvcc -DCUDAToolkit_INCLUDE_DIRECTORIES=$PREFIX/targets/x86_64-linux/include"
 cyclonedds:
   generate_dummy_package_with_run_deps:
     dep_name: cyclonedds
@@ -202,6 +214,10 @@ gz_sim_vendor:
   additional_cmake_args: "-DAMENT_VENDOR_POLICY=NEVER_VENDOR_IGNORE_SATISFIED_CHECK"
   # Rebuild with qt6-main 6.11 (b29 was built against qt6-main 6.12, which pyqt6 6.11 can't install with).
   build_number: 30
+libtorch_vendor:
+  # CUDA architecture is only for Torch's compiler probe; no kernels are built here.
+  build_number: 31
+  additional_cmake_args: "-DFORCE_BUILD_VENDOR_PKG=OFF -DCMAKE_CUDA_ARCHITECTURES=75"
 rqt_plot:
   # Rebuild with qt6-main 6.11 (b29 was built against qt6-main 6.12, which pyqt6 6.11 can't install with).
   build_number: 30
@@ -232,6 +248,10 @@ rviz_satellite:
 rviz_visual_testing_framework:
   # Rebuild with qt6-main 6.11 (b29 was built against qt6-main 6.12, which pyqt6 6.11 can't install with).
   build_number: 30
+torch_conversions:
+  # Rebuild the exported Torch/CUDA targets and the variant-specific dependencies.
+  build_number: 31
+  additional_cmake_args: "-DCMAKE_CUDA_ARCHITECTURES=75"
 turtlesim:
   # Rebuild with qt6-main 6.11 (b29 was built against qt6-main 6.12, which pyqt6 6.11 can't install with).
   build_number: 30

vinca.yaml

@@ -52,6 +52,8 @@ packages_skip_by_deps:
 
   - if: not linux
     then:
+      - cuda_buffer
+      - cuda_buffer_backend
       - pendulum_control
       - rttest
       - tlsf
@@ -66,6 +68,8 @@ packages_remove_from_deps:
 
   - if: not linux
     then:
+      - cuda_buffer
+      - cuda_buffer_backend
       - pendulum_control
       - rttest
       - tlsf
@@ -312,6 +316,7 @@ packages_select_by_deps:
   - teleop
   - teleop_tools
   - toppra
+  - torch_conversions
   - tracetools
   - tracetools_launch
   - tracetools_trace
@@ -333,6 +338,7 @@ packages_select_by_deps:
       - apriltag_ros  # Depends on camera_ros
       - avt_vimba_camera
       - camera_ros  # Depends on libcamera that is only available on linux
+      - cuda_buffer_backend  # Depends on CUDA
       - libcamera
       - livox_ros_driver2
       - mavros  # libmavconn currently fails to build on macOS and windows

vinca_pinning.yaml

@@ -9,6 +9,10 @@ migrations:
   - urdfdom6
   - vtk970
 pinning_overrides:
+  # Only CUDA buffer recipes use this key; do not expand the global compiler matrix.
+  cuda_buffer_cuda_version:
+    - '12.9'
+    - '13.0'
   # Build commands provide their channels with `-c`, so omit the inherited
   # conda-forge-pinning channel_sources value from the rendered configuration.
   channel_sources: null
@@ -22,23 +26,28 @@ pinning_overrides:
     - 2.28                       # [linux and not riscv64]
     - 2.39                       # [linux and riscv64]
     - 2.28                       # [linux and (x86_64 or aarch64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
+    - 2.28                       # [linux and (x86_64 or aarch64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
     - 14.0                       # [osx]
   c_compiler_version:
     - 15                         # [linux]
     - 21                         # [osx]
     - 14                         # [linux and (x86_64 or aarch64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
+    - 14                         # [linux and (x86_64 or aarch64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
   cxx_compiler_version:
     - 15                         # [linux]
     - 21                         # [osx]
     - 14                         # [linux and (x86_64 or aarch64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
+    - 14                         # [linux and (x86_64 or aarch64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
   fortran_compiler_version:
     - 15                         # [unix]
     - 5                          # [win64]
     - 22                         # [win and arm64]
     - 14                         # [linux and (x86_64 or aarch64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
+    - 14                         # [linux and (x86_64 or aarch64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
   cuda_compiler_version:
     - None
     - 12.9                       # [((linux and (x86_64 or aarch64)) or win64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
+    - 13.0                       # [((linux and (x86_64 or aarch64)) or win64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
   libzenohc:
     - 1.10.1
   libzenohcxx:
@@ -67,4 +76,3 @@ pinning_overrides:
     - false
   python_impl:
     - cpython
-