Help Center/ ModelArts/ ModelArts User Guide (Lite Cluster)/ Before You Start/ Software Versions Required by Different Models
Updated on 2025-08-15 GMT+08:00

Software Versions Required by Different Models

A resource pool for elastic clusters can use either Elastic Bare Metal Servers (BMSs) or Elastic Cloud Servers (ECSs) as nodes. Each node model has its own operating system (OS) and compatible CCE cluster versions. This document outlines the necessary software versions for each model to simplify image creation and software upgrades.

CCE Cluster Maintenance Policy

The CCE cluster used by ModelArts Lite Cluster belongs to the user who has full control over the cluster.

  • If your Lite Cluster uses the EOS CCE cluster, upgrade the cluster to the version recommended by ModelArts in time according to the lifecycle policy released by CCE.

    For details about how to upgrade a CCE cluster, see Cluster Upgrade Overview.

    For details about CCE cluster version policies, see Cluster Version Release Notes.

  • If you have any technical problems related to CCE clusters in the Lite Cluster scenario, submit a service ticket to contact CCE technical support for troubleshooting.

Software Versions Required by BMSs

Table 1 BMS

Type

Card Type

RDMA Network Protocol

OS

Applicable Scope

Dependent Plug-in

NPU

ascend-snt9b

RoCE

  • OS: EulerOS 2.10 64-bit (recommended)
  • Kernel version: 4.19.90-vhulk2211.3.0.h1543.eulerosv2r10.aarch64
  • Architecture type: aarch64
  • Cluster type: CCE Standard
  • Cluster version: v1.23 (v1.23.5-r0 or later), v1.25, v1.28, or v1.31 (recommended)
  • Cluster scale: 50, 200, 1000, or 2000
  • Cluster network mode: container tunnel network or VPC
  • Cluster forwarding mode: iptables or ipvs
  • huawei-npu
  • volcano

For details about the plug-in version mapping, see Table 3.

RoCE

  • OS: Huawei Cloud EulerOS 2.0 64-bit
  • Kernel version: 5.10.0-60.18.0.50.r865_35.hce2.aarch64
  • Architecture type: aarch64
  • Cluster type: CCE Turbo
  • Cluster version: v1.23, v1.25, v1.28, or v1.31 (recommended)
  • Cluster scale: 50, 200, 1000, or 2000
  • Cluster network mode: ENI
  • Cluster forwarding mode: iptables or ipvs

ascend-snt9

RoCE

  • OS: EulerOS 2.8 64-bit
  • Kernel version: 4.19.36-vhulk1907.1.0.h619.eulerosv2r8.aarch64
  • Architecture type: aarch64
  • Cluster type: CCE Standard or Turbo
  • Cluster version: v1.23 (v1.23.5-r0 or later), v1.25, or v1.28 (recommended)
  • Cluster scale: 50, 200, 1,000, or 2,000
  • Cluster network mode: container tunnel network, VPC, or ENI
  • Cluster forwarding mode: iptables or ipvs

GPU

gp-ant8

RoCE

  • OS: EulerOS 2.10 64-bit
  • Kernel version: 4.18.0-147.5.2.15.h1109.eulerosv2r10.x86_64
  • Architecture type: x86
  • Cluster type: CCE Standard or Turbo
  • Cluster version: v1.23, v1.25, v1.28, or v1.31 (recommended)
  • Cluster scale: 50, 200, 1000, or 2000
  • Cluster network mode: container tunnel network or VPC

    Distributed training only supports container tunnel network.

  • Cluster forwarding mode: iptables or ipvs
  • gpu-beta
  • rdma-sriov-dev-plugin

For details about the plug-in version mapping, see Table 3.

gp-ant1

RoCE

  • OS: EulerOS 2.10 64-bit
  • 4.18.0-147.5.2.15.h1109.eulerosv2r10.x86_64
  • Architecture type: x86
  • Cluster type: CCE Standard or Turbo
  • Cluster version: v1.23, v1.25, v1.28, or v1.31 (recommended)
  • Cluster scale: 50, 200, 1000, or 2000
  • Cluster network mode: container tunnel network or VPC

    Distributed training only supports container tunnel network.

  • Cluster forwarding mode: iptables or ipvs

gp-vnt1

RoCE|IB

  • OS: EulerOS 2.9 64-bit (used only for p6 and p6s flavors in Shanghai 1)
  • Kernel version: 147.5.1.6.h1099.eulerosv2r9.x86_64
  • Architecture type: x86
  • Cluster type: CCE Standard
  • Cluster version: v1.23, v1.25, v1.28, or v1.31 (recommended)
  • Cluster scale: 50, 200, 1000, or 2000
  • Cluster network mode: container tunnel network or VPC

    Distributed training only supports container tunnel network.

  • Cluster forwarding mode: iptables or ipvs
  • OS: EulerOS 2.9 64-bit (recommended)
  • Kernel version: 4.18.0-147.5.1.6.h841.eulerosv2r9.x86_64
  • Architecture type: x86
  • Remote direct memory access (RDMA) is a direct memory access from the memory of one computer into that of another without involving either one's operating system.
  • RDMA over Converged Ethernet (RoCE) is a network protocol which allows RDMA over an Ethernet network.
  • InfiniBand (IB) is a computer networking communications standard used in high-performance computing. It is used for data interconnect both among and within computers.

Software Versions Required by ECSs

Table 2 ECS

Type

Card Type

OS

Applicable Scope

Dependent Plug-in

NPU

ascend-snt3p-300i

  • OS: Huawei Cloud EulerOS 2.0 64-bit
  • Architecture type: x86 or Arm
  • Cluster type: CCE Standard
  • Cluster version: v1.23 (v1.23.5-r0 or later), v1.25, v1.28, or v1.31 (recommended)
  • Cluster scale: 50, 200, 1,000, or 2,000
  • Cluster network mode: container tunnel network or VPC
  • Cluster forwarding mode: iptables or ipvs
  • huawei-npu
  • volcano

For details about the plug-in version mapping, see Table 3.

  • OS: EulerOS 2.9
  • Architecture type: x86
  • Cluster type: CCE Standard or Turbo
  • Cluster version: v1.23 (v1.23.5-r0 or later), v1.25, or v1.28 (recommended)
  • Cluster scale: 50, 200, 1000, or 2000
  • Cluster network mode: container tunnel network, VPC, or ENI
  • Cluster forwarding mode: iptables or ipvs

ascend-snt3

  • OS: EulerOS 2.5
  • Architecture type: x86
  • Cluster type: CCE Standard
  • Cluster version: v1.23 or v1.25
  • Cluster scale: 50, 200, 1000, or 2000
  • Cluster network mode: container tunnel network or VPC
  • Cluster forwarding mode: iptables or ipvs
  • OS: EulerOS 2.8
  • Architecture type: Arm
  • Cluster type: CCE Standard
  • Cluster version: v1.23 or v1.25
  • Cluster scale: 50, 200, 1000, or 2000
  • Cluster network mode: container tunnel network or VPC

    Cluster forwarding mode: iptables or ipvs

GPU

gp-vnt1

  • OS: EulerOS 2.9
  • Architecture type: x86
  • Cluster type: CCE Standard
  • Cluster version: v1.23, v1.25, v1.28, or v1.31 (recommended)
  • Cluster scale: 50, 200, 1000, or 2000
  • Cluster network mode: container tunnel network or VPC
  • Cluster forwarding mode: iptables or ipvs
  • gpu-beta
  • rdma-sriov-dev-plugin

For details about the plug-in version mapping, see Table 3.

gp-ant03

  • OS: EulerOS 2.9
  • Architecture type: x86
  • Cluster type: CCE Standard
  • Cluster version: v1.23, v1.25, v1.28, or v1.31 (recommended)
  • Cluster scale: 50, 200, 1000, or 2000
  • Cluster network mode: container tunnel network or VPC
  • Cluster forwarding mode: iptables or ipvs

gp-ant1-pcie40

  • OS: EulerOS 2.9
  • Architecture type: x86
  • Cluster type: CCE Standard
  • Cluster version: v1.23, v1.25, v1.28, or v1.31 (recommended)
  • Cluster scale: 50, 200, 1000, or 2000
  • Cluster network mode: container tunnel network or VPC
  • Cluster forwarding mode: iptables or ipvs

gp-tnt004

  • OS: EulerOS 2.9
  • Architecture type: x86
  • Cluster type: CCE Standard
  • Cluster version: v1.23, v1.25, v1.28, or v1.31 (recommended)
  • Cluster scale: 50, 200, 1000, or 2000
  • Cluster network mode: container tunnel network or VPC
  • Cluster forwarding mode: iptables or ipvs

Mapping Between Drivers, Plug-in Versions, and CCE Cluster Versions

Table 3 Mapping between driver and CCE cluster versions

Type

Driver

Driver Version

Matched CCE Cluster Version

Applicable Scope

Plug-in Function

npuDriver

npu-driver

7.1.0.9.220-23.0.6 (recommended)

7.1.0.7.220-23.0.5

7.1.0.5.220-23.0.3

None

NPU (snt9b)

Upgrades and rolls back NPU drivers.

gpuDriver

gpu-driver

515.65.01 (recommended)

510.47.03

470.182.03

470.57.02

None

GPU

Upgrades and rolls back GPU drivers. The plug-in depends on the gpu-beta version.

Table 4 Mapping between plug-in versions and CCE cluster versions

Plug-in

Plug-in Version

Matched CCE Cluster Version

Applicable Scope

Plug-in Function

gpu-beta

2.7.63 (recommended)

v1.(28|31).*

GPU

Allows containers to use GPU devices.

2.6.4

v1.28.*

2.0.48

v1.(23|25).*

huawei-npu

2.1.53 (recommended)

v1.(23|25|28|31).*

NPU

Allows containers to use Huawei NPU devices.

2.1.22

v1.(23|25|28).*

volcano

1.16.8 (recommended)

v1.(23|25|28|31).*

NPU

Kubernetes-based batch processing platform.

1.15.8

v1.(23|25|28).*

os-node-agent

7.0.0

None

None

OS plug-in, which is used for fault detection.

icagent

default

The matched CCE version is installed by default.

None

CCE basic component, which is used for logging and monitoring.