v0.14.0-rc.2

Full Changelog: v0.14.0-rc.1...v0.14.0-rc.2

Changes

• Fix bug from v0.14.0-rc.1 when using cdi-enabled=false

v0.14.0-rc.1

Full Changelog: v0.13.0...v0.14.0-rc.1

Changes

• Added --cdi-enabled flag to GPU Device Plugin. With this enabled, the device plugin will generate CDI specifications for available NVIDIA devices. Allocation will add CDI anntiations (cdi.k8s.io/*) to the response. These are read by a CDI-enabled runtime to make the required modifications to a container being created.
• Updated GFD subchard to version 0.8.0-rc.1
• Bumped Golang version to 1.20.1
• Bumped CUDA base images version to 12.1.0
• Switched to klog for logging
• Added a static deployment file for Microshift

v0.13.0

Full Changelog: v0.12.2...v0.13.0

Changes

• Skip NVIDIA DGX Display devices when generating labels.
• Fail on startup if no valid resources are detected
• Bump GFD subchart to version 0.7.0

Changes from v0.13.0-rc.3

• Use nodeAffinity instead of nodeSelector by default in daemonsets
• Add machine-file-path option to GFD config flags
• Mount /sys instead of /sys/class/dmi/id/product_name in GPU Feature Discovery daemonset
• Bump GFD subchard to version 0.7.0-rc.3

Changes from v0.13.0-rc.2

• Bump cuda base image to 11.8.0
• Use consistent indendation in YAML manifests
• Fix bug from v0.13.0-rc.1 when using mig-strategy="mixed"
• Add logged error message if setting up health checks fails
• Support MIG devices with 1g.10gb+me profile
• Distribute replicas evenly across GPUs during allocation
• Bump GFD subchart to version 0.7.0-rc.2

Changes from v0.13.0-rc.1

• Improve health checks to detect errors when waiting on device events
• Log ECC error events detected during health check
• Add the GIT sha to version information for the CLI and container images
• Use NVML interfaces from go-nvlib to query devices
• Refactor plugin creation from resources
• Add a CUDA-based resource manager that can be used to expose integrated devices on Tegra-based systems
• Bump GFD subchart to version 0.7.0-rc.1

Note:

The container image nvcr.io/nvidia/k8s-device-plugin:v0.13.0-ubi8 contains the following high-severity CVEs:

• CVE-2022-42898 - Vulnerability found in os package type (rpm) - krb5-libs

v0.13.0-rc.3

• Use nodeAffinity instead of nodeSelector by default in daemonsets
• Add machine-file-path option to GFD config flags
• Mount /sys instead of /sys/class/dmi/id/product_name in GPU Feature Discovery daemonset
• Bump GFD subchard to version 0.7.0-rc.3

Full Changelog: v0.13.0-rc.2...v0.13.0-rc.3

v0.13.0-rc.2

• Bump cuda base image to 11.8.0
• Use consistent indendation in YAML manifests
• Fix bug from v0.13.0-rc.1 when using mig-strategy="mixed"
• Add logged error message if setting up health checks fails
• Support MIG devices with 1g.10gb+me profile
• Distribute replicas evenly across GPUs during allocation
• Bump GFD subchart to version 0.7.0-rc.2

v0.13.0-rc.1

• Improve health checks to detect errors when waiting on device events
• Log ECC error events detected during health check
• Add the GIT sha to version information for the CLI and container images
• Use NVML interfaces from go-nvlib to query devices
• Refactor plugin creation from resources
• Add a CUDA-based resource manager that can be used to expose integrated devices on Tegra-based systems
• Bump GFD subchart to version 0.7.0-rc.1

v0.12.3

• Bump cuda base image to 11.7.1
• Remove CUDA compat libs from the device-plugin image in favor of libs installed by the driver
• Fix securityContext.capabilities indentation
• Add namespace override for multi-namespace deployments

v0.12.2

• Fix example configmap settings in values.yaml file
• Fix assertions for panicking on uniformity with migStrategy=single
• Make priorityClassName configurable through helm
• Move NFD servicAccount info under 'master' in helm chart
• Bump GFD subchart to version 0.6.1
• Allow an empty config file and default to "version: v1"
• Make config fallbacks for config-manager a configurable, ordered list
• Add an 'empty' config fallback (but don't apply it by default)

v0.12.1

• Exit the plugin and GFD sidecar containers on error instead of logging and continuing
• Only force restart of daemonsets when using config files and allow overrides
• Fix bug in calculation for GFD security context in helm chart
• Fix bug prohibiting GFD from being started from the plugin helm chart

v0.12.0

This release is a promotion of v0.12.0-rc.6 to v0.12.0

v0.12.0-rc.6

• Send SIGHUP from GFD sidecar to GFD main container on config change
• Reuse main container's securityContext in sidecar containers
• Update GFD subchart to v0.6.0-rc.1
• Bump CUDA base image version to 11.7.0
• Add a flag called FailRequestsGreaterThanOne for TimeSlicing resources

v0.12.0-rc.5

• Allow either an external ConfigMap name or a set of configs in helm
• Handle cases where no default config is specified to config-manager
• Update API used to pass config files to helm to use map instead of list
• Fix bug that wasn't properly stopping plugins across a soft restart

v0.12.0-rc.4

• Disable support for resource-renaming in the config (will no longer be part of this release)
• Add field for TimeSlicing.RenameByDefault to rename all replicated resources to .shared
• Refactor main to allow configs to be reloaded across a (soft) restart
• Add support to helm to provide multiple config files for the config map
• Add new config-manager binary to run as sidecar and update the plugin's configuration via a node label
• Make GFD and NFD (optional) subcharts of the device plugin's helm chart

v0.12.0-rc.3

• Add ability to parse Duration fields from config file
• Omit either the Plugin or GFD flags from the config when not present
• Fix bug when falling back to none strategy from single strategy

v0.12.0-rc.2

• Move MigStrategy from Sharing.Mig.Strategy back to Flags.MigStrategy
• Remove TimeSlicing.Strategy and any allocation policies built around it
• Add support for specifying a config file to the helm chart

v0.12.0-rc.1

• Add API for specifying time-slicing parameters to support GPU sharing
• Add API for specifying explicit resource naming in the config file
• Update config file to be used across plugin and GFD
• Stop publishing images to dockerhub (now only published to nvcr.io)
• Add NVIDIA_MIG_MONITOR_DEVICES=all to daemonset envvars when mig mode is enabled
• Print the plugin configuration at startup
• Add the ability to load the plugin configuration from a file
• Remove deprecated tolerations for critical-pod
• Drop critical-pod annotation(removed from 1.16+) in favor of priorityClassName
• Pass all parameters as env in helm chart and example daemonset.yamls files for consistency