nvidia-cdi-hook

NVIDIA CDI Hook

The CLI nvidia-cdi-hook provides container device runtime hook capabilities when called by a container runtime, as specific in a Container Device Interface file.

Generating a CDI

The CDI itself is created for an NVIDIA-capable device using the nvidia-ctk cdi generate command.

When nvidia-ctk cdi generate is run, the CDI specification is generated as a yaml file. The CDI specification provides instructions for a container runtime to set up devices, files and other resources for the container prior to starting it. Those instructions may include executing command-line tools to prepare the filesystem. The execution of such command-line tools is called a hook.

nvidia-cdi-hook is the CLI tool that is expected to be called by the container runtime, when specified by the CDI file.

See the nvidia-ctk documentation for more information on generating a CDI file.

Functionality

The nvidia-cdi-hook CLI provides the following functionality:

• chmod - Change the permissions of a file or directory inside the directory path to be mounted into a container.
• create-symlinks - Create symlinks inside the directory path to be mounted into a container.
• update-ldcache - Update the dynamic linker cache inside the directory path to be mounted into a container.
• enable-cuda-compat - Ensure that the directory containing the CUDA compat libraries is added to the ldconfig search path if required.
• disable-device-node-modification - Ensure that the /proc/driver/nvidia/params file present in the container does not allow device node modifications.
• update-application-profile - Update driver settings through "application profiles". Currently, this hook sets EGLVisibleDGPUDevices to restrict EGL/Vulkan GPU visibility inside the container.
• apply-cuda-memory-limits - Set the soft and hard limits of CUDA memory usage on GPU device(s) in the container.