CCE AI Suite (NVIDIA GPU)

basic

Yes

object

Basic add-on configuration parameters

custom

Yes

Table 3 object

Custom parameters

cluster_version

No

String

CCE cluster version

device_version

Yes

String

Add-on version

driver_version

Yes

String

Image tag of an add-on pod where a driver is installed. Generally, the value is the same as that of device_version.

obs_url

Yes

String

When a GPU driver is downloaded from the default driver address, the value is the GPU driver address.

swr_addr

Yes

String

Image repository address

swr_user

Yes

String

Tenant path of an image repository

compatible_with_legacy_api

No

Bool

API compatibility switch

Default value: false

true: The add-on supports the GPU native mode and xGPU virtualization.

component_schedulername

Yes

String

Name of the scheduler used by the add-on.

Default value: default-scheduler

disable_mount_path_v1

No

Bool

Default value: false

true: /opt/cloud/cce/nvidia is not mounted to the /usr/lib/nvidia directory of a GPU container.

disable_nvidia_gsp

No

Bool

Default value: true

true: The GPU GSP firmware is disabled.

driver_mount_paths

No

String

Driver file directory that needs to be automatically mounted to a GPU container

Default value: "bin,lib64"

enable_fault_isolation

No

Bool

Default value: true

true: The add-on detects hardware faults or driver issues of a GPU and then sets the GPU to be unavailable.

enable_health_monitoring

No

Bool

Default value: true

true: The add-on detects hardware faults or driver issues of a GPU.

enable_metrics_monitoring

No

Bool

Default value: true

true: The add-on collects GPU metrics and reports these metrics to Prometheus.

enable_simple_lib64_mount

No

Bool

Default value: true

true: Only the libxxx.so.x file is mounted to a container.

enable_xgpu

No

Bool

Default value: false

Whether to enable xGPU virtualization.

gpu_driver_config

No

Map

Configurations of the GPU driver for a single node pool

Default value: {}

health_check_xids_v2

No

String

GPU error range for the add-on health checks

Default value: "74,79"

inject_ld_Library_path

No

String

Value of the LD_LIBRARY_PATH environment variable automatically injected by the add-on to a GPU container

Default value: ""

lib64_container_paths

No

String

Mount path of NVIDIA lib64 in a GPU container

Default value: "/usr/lib64,/usr/lib/x86_64-linux-gnu"

metrics_delete_interval

No

int

Timeout threshold for deleting a metric when the metric cannot be obtained. The unit is millisecond.

Default value: 30000

metrics_monitor_interval

No

int

Interval for obtaining metrics, in milliseconds.

Default value: 15000

nvidia_driver_download_url

Yes

String

Path for downloading the NVIDIA driver

Default value: ""