NVIDIA Drops Container Toolkit v1.19.0, Killing Legacy GPU Hook Model
NVIDIA Container Toolkit v1.19.0 makes CDI the default GPU access method, adding IGX 2.0, Tegra CUDA Forward Compatibility, and read-only rootfs support
Takeaways
- v1.19.0 released: Latest feature release adds IGX 2.0 Thor support, CUDA Forward Compatibility on Tegra, and read-only rootfs support.
- CDI is now default: The legacy hook-based GPU injection mode is deprecated; Container Device Interface (CDI) is the new standard across Docker, containerd, CRI-O, and Podman.
- JIT CDI mode: CDI specs are auto-generated at container creation time, removing the need for manual pre-generation steps.
- Kubernetes default: GPU Operator v25.10.0+ uses CDI by default for all GPU workloads, making runtime-specific plugins unnecessary.
- Non-root & read-only support: Containers can now run as unprivileged users without extra group config, and on read-only root filesystems.
- Free & open source: Apache 2.0 license; available via GitHub and official docs.
NVIDIA Container Toolkit is the foundational layer that makes GPU-accelerated containers possible. Without it, Docker and Kubernetes have no way to see, allocate, or pass through NVIDIA hardware to workloads. The project just pushed v1.19.0 as its latest feature release, and the changes reflect a broader architectural shift that has been building across the last several versions: the old hook-based GPU injection model is being replaced by a cleaner, standards-based approach called CDI.
What the toolkit actually does
The NVIDIA Container Toolkit allows users to build and run GPU-accelerated containers, including a container runtime library and utilities to automatically configure containers to leverage NVIDIA GPUs. In practice, it sits between your container engine (Docker, containerd, CRI-O, Podman) and the host NVIDIA driver, handling the injection of GPU device nodes, driver libraries, and firmware into each container at startup.

The stack has four main components that work together:
- libnvidia-container1 , the low-level C library that does the actual device injection using kernel primitives
- nvidia-container-cli , a CLI wrapper around the library that runtimes call to configure containers
- nvidia-container-runtime , a thin wrapper around
runcthat injects the NVIDIA hook before a container starts - nvidia-ctk , a utility CLI for configuring runtimes and generating CDI specifications
You do not need to install the CUDA Toolkit on the host system, but the NVIDIA driver needs to be installed. The toolkit handles everything else, including mounting the right driver libraries into the container at runtime.