NOTE: This release is a unified release of the NVIDIA Container Toolkit that consists of the following packages:
The packages for this release are published to the libnvidia-container package repositories.
Fixes and Features
- Added support for Azure Linux 4.
- Fixed an issue where volume-mount requests could add GPUs to a container in addition to the GPUs selected through CDI. This issue affected configurations with
accept-nvidia-visible-devices-as-volume-mountsset totrue. The setting defaults tofalse. For more information, refer to PR #2184. - Management CDI specifications now include
libnvsdm.so.1and its dependencies,libibumad.so.3andlibibmad.so.5, when available on the host. Management applications such as NVIDIA Data Center GPU Manager (DCGM) can initialize NVSDM for NVSwitch monitoring on systems that require these libraries. For more information, refer to PR #2162. - The
disable-device-node-modificationhook now mounts the modified driver parameter file as read-only in the container. The read-only mount prevents writes to the driver parameter file through the bind mount. For more information, refer to PR #2160. - Commands run through
chrootin the toolkit container now execute host binaries, such asmodprobe, so kernel modules and their dependencies load correctly. For more information, refer to PR #2189 and issue #299 in /NVIDIA/k8s-driver-manager. - The
update-application-profilehook now writes the GPU visibility profile atomically, replacing an existing profile without writing through a symbolic link at the profile path. The hook also ignores busy-file errors, in addition to read-only filesystem errors, so these errors do not prevent container startup. For more information, refer to PR #2201.