Vendor-specific storage backends for rosidl::Buffer, enabling ROS 2 publishers and subscribers to share accelerator-accessible memory without copying when supported, with automatic fallback to the default CPU path.

• cuda_buffer_backend -- CUDA buffer backend packages for zero-copy GPU memory sharing using CUDA VMM IPC.
• qc_buffer_backend -- Qualcomm buffer backend packages for HTP-CPU zero-copy sharing using dma-buf memory.
• libtorch_vendor -- Vendor package that downloads and installs the pre-built LibTorch C++ distribution.
• tensor_msgs -- DLPack-aligned ExperimentalTensor.msg definition.
• torch_conversions -- Header-only helper library that converts between tensor_msgs/ExperimentalTensor and at::Tensor and exposes DLPack import / export. Replaces the older torch_buffer_backend plugin approach with a plain message + bridge library that rides on top of whichever rosidl::Buffer backend is registered (CUDA when available, CPU otherwise).

Package	Source deb	Binary deb (amd64)	Binary deb (arm64)
cuda_buffer
cuda_buffer_backend
cuda_buffer_backend_msgs
libtorch_vendor
tensor_msgs
torch_conversions

Package	Source deb	Binary deb (amd64)	Binary deb (arm64)
cuda_buffer
cuda_buffer_backend
cuda_buffer_backend_msgs
libtorch_vendor
tensor_msgs
torch_conversions

• A ROS 2 Rolling development environment. See the upstream Building ROS 2 on Ubuntu guide for the canonical source-build flow, or use the pixi workflow shipped by the ros2/ros2 meta-repo.

• Each backend depends on its vendor-specific platform support. Refer to the backend documentation for detailed prerequisites, setup, build, test, and usage instructions:

• cuda_buffer_backend/README.md

• qc_buffer_backend/README.md

• torch_conversions/README.md

Apache-2.0