跳到主要内容

安装 S4000/S5000 3.3.5(Alinux)

本页适用于 S4000/S5000,Linux Driver Server 3.3.5,MUSA SDK 4.3.5。


目录

  1. 下载软件包

  2. 驱动安装

  3. MUSA Toolkit 安装指南

  4. muDNN 安装指南

  5. mccl 安装指南


1.下载软件包

软件包中应包含以下文件:

├── mccl_rc2.1.5.PH1.tar.gz
├── mudnn_rc3.1.5.PH1.tar.gz
├── musa_3.3.5-server_amd64.rpm
├── musa_toolkits_rc4.3.5.tar.gz

2.驱动安装

2.1 资源需求

  • GPU 卡:MTT S5000
  • 操作系统:Alibaba Cloud Linux 3 (Soaring Falcon)
  • 内核版本:5.10.134-13.al8.x86_64

2.2 配置 BIOS

  1. 将显示器连接到主板上,并断开电源重新启动计算机。在启动过程中,按下 delete 键进入 BIOS 设置界面。
  2. 在 BIOS 设置界面中,找到并开启 above 4G 选项。
  3. 查找并将 resize bar 开关置为 ON。大部分主板首次开机时此选项默认为 OFF,通常显示在 BIOS 菜单右上角。
  4. 关闭iommu选项。

2.3 安装 MT 驱动 rpm 包

2.3.1 环境依赖

  • 查看 PCI 是否识别到显卡

    lspci -nn | grep 1ed5
  • 安装依赖包(若已安装请忽略)

    dnf -y install dkms

2.3.2 卸载驱动

# 卸载 old musa rpm
sudo rpm -e musa

# 如果报错 mtgpu is inuse,直接重启即可
sudo rmmod mtgpu

# 查询是否仍有 mtgpu ko 残留
lsmod | grep mtgpu

2.3.3 安装驱动

# 安装 MT DDK 驱动,以 musa_3.0.0_amd64.deb 包为例
sudo rpm -ivh musa_3.3.5-server_amd64.rpm

# 强烈建议重启电脑来更换驱动
sudo reboot

# 特殊情况下希望不重启更换驱动,可以手动重新加载驱动
modprobe -rv mtgpu && modprobe mtgpu

2.3.4 设置用户组

非 Root 用户如需获取设备使用权限,可参考以下方式设置:

# 把普通用户账号,加进 render 和 video group
sudo usermod -aG render ${USER}
sudo usermod -aG video ${USER}
# 重新登入账户

2.4 验证安装 & 版本查询

2.4.1 mthreads-gmi

  • 通过 mthreads-gmi 校验 DDK 安装

    mthreads-gmi

2.4.2 查看版本信息

  • 查看 KMD 版本信息

    # 方式一
    dmesg | grep "Driver Version"
    # 方式二
    cat /var/log/dmesg | grep "Driver Version"

3. MUSA Toolkit 安装指南

本章主要描述 MUSA Toolkit 的详细安装步骤。

前提条件

  • 已安装兼容版本 MT GPU 驱动。

3.1 准备工作

  1. 检查 MT GPU 驱动版本

    mthreads-gmi
  2. 若未安装或未安装适配版本 MT GPU 驱动,请参考上一节驱动安装。

3.2 安装 MUSA Toolkit

MUSA Toolkit 支持安装到主机(host)或 docker 环境中。

  • 运行环境在 docker 上,建议在 docker 中安装 musa toolkit。

3.2.1 安装

以musa_toolkits_rc4.3.5.tar.gz 为例:

tar -zxvf musa_toolkits_rc4.3.5.tar.gz
cd musa_toolkits_install
bash ./install.sh -u # 先卸载环境已安装的 musa
bash ./install.sh # 非 root 用户需要 sudo 提权执行

安装完毕后,终端会输出如下提示:

Successfully installed MUSA in /usr/local/musa
Please export /usr/local/musa/bin to PATH in ~/.bashrc
export PATH=/usr/local/musa/bin:${PATH}
Please export /usr/local/musa/lib to LD_LIBRARY_PATH in ~/.bashrc
export LD_LIBRARY_PATH=/usr/local/musa/lib:${LD_LIBRARY_PATH}

3.2.2 设置环境变量

  1. 编辑用户终端配置文件:

    vim ~/.bashrc
  2. 在配置文件末尾追加如下命令:

    export PATH=/usr/local/musa/bin:$PATH
    export LD_LIBRARY_PATH=/usr/local/musa/lib:$LD_LIBRARY_PATH
  3. 配置文件会在下次启动终端时生效。如需立即生效:

    source ~/.bashrc

3.2.3 验证

验证 MUSA Toolkit 是否安装成功:

musa_version_query

输出示例:

musa_toolkits:
{
"version": "4.3.5",
"git branch": "release_musa_4.3.0",
"git tag": "No tag",
"commit id": "165aeaa9d4b70b3f2dd3a6bc813ceb11fc12ecef",
"commit date": "2026-01-09 09:14:56 +0800"
}
CUB:
{
"version": "1.12.1",
"git branch": "develop",
"commit id": "ba4f248dcfb5b79814b21bb2c96eeb6b4e6bd254",
"commit date": "2024-09-10 14:25:44 +0800"
}
mccl:
{
"version": "2.1.5",
"build archs": "31",
"git branch": "release_musa_4.3.0_s5000",
"git tag": "No tag",
"commit id": "b1ac16e756fa0d29ae46d81f9b3e69890d4bc33a",
"commit date": "2026-01-13 17:07:07 +0800"
}
mcc:
{
"version": "4.3.5",
"git branch": "release_musa_4.3.0",
"git tag": "No tag",
"commit id": "23a614460b10fbb6ac5be3673102c6d8524d969a",
"commit date": "2026-01-15 09:53:56 +0800"
}
mtjpeg:
{
"version": "1.0.2",
"git branch": "release_musa_4.3.0",
"git tag": "No tag",
"commit id": "799e676c97f6b7c5e08aaad51c2c1857d8f3e233",
"commit date": "2025-11-27 17:45:21 +0800"
}
mublasLt:
{
"version": "1.10.5",
"build archs": "22,31",
"git branch": "release_musa_4.3.0",
"git tag": "No tag",
"commit id": "aa6520e37c6778ed453f6f40f935455053767549",
"commit date": "2026-01-12 15:37:54 +0800"
}
mublas:
{
"version": "1.10.5",
"build archs": "22,31",
"git branch": "release_musa_4.3.0",
"git tag": "No tag",
"commit id": "6829928b617b3e0fca9209ce10d345a86e98c9f4",
"commit date": "2026-01-12 15:34:37 +0800"
}
mudnn:
{
"version": "3.1.5",
"git branch": "release_musa_4.3.0",
"git tag": "No tag",
"commit id": "38b420034939038b2beb7eaff0ad16d669f4e9b1",
"commit date": "2026-01-13 15:00:45 +0800"
}
mufft:
{
"version": "1.10.5",
"build archs": "22;31",
"git branch": "release_musa_4.3.0",
"git tag": "No tag",
"commit id": "1eb3164b9bc46417f07bd431811f937efe4b91d5",
"commit date": "2026-01-09 09:17:50 +0800"
}
muPP:
{
"version": "1.11.5",
"build archs": "22;31",
"git branch": "release_musa_4.3.0",
"git tag": "No tag",
"commit id": "bb449a16da0ac15a7dcb713e9005102dd39ab685",
"commit date": "2026-01-12 15:23:45 +0800"
}
MUPTI:
{
"version": "1.0.5",
"git branch": "release_musa_4.3.0",
"git tag": "No tag",
"commit id": "08c437ef32c2d8102d81dc99d660e4fe754aef49",
"commit date": "2026-01-12 12:09:42 +0800"
}
murand:
{
"version": "1.1.5",
"build archs": "22;31",
"git branch": "release_musa_4.3.0",
"git tag": "No tag",
"commit id": "0535a5cfd55774e3874cd53a25706b6c5772b76a",
"commit date": "2026-01-12 15:28:26 +0800"
}
musify:
{
"version": "1.1.5",
"git branch": "release_musa_4.3.0",
"git tag": "No tag",
"commit id": "ee363d2233e3933cbacc9d45a72e71b57cb3403a",
"commit date": "2026-01-12 11:28:12 +0800"
}
musolver:
{
"version": "1.4.5",
"build archs": "22,31",
"git branch": "release_musa_4.3.0",
"git tag": "No tag",
"commit id": "3bd7c64543ac306fa05996f6e5d77b1f585a174f",
"commit date": "2026-01-12 15:27:02 +0800"
}
musparse:
{
"version": "1.5.5",
"build archs": "22,31",
"git branch": "release_musa_4.3.0",
"git tag": "No tag",
"commit id": "2b4d06cd5a60052a094f53aa1d47bb945fb7248e",
"commit date": "2026-01-12 15:29:47 +0800"
}
Thrust:
{
"version": "1.12.1",
"git branch": "develop",
"commit id": "e1a1f8c1c587e20a4dd8429c26e8445466dec97f",
"commit date": "2024-09-14 08:57:33 +0800"
}
musa_runtime:
{
"version": "4.3.5",
"git branch": "release_musa_4.3.0",
"git tag": "No tag",
"commit id": "4f6c4df31cdee56b937b7b0bccfcc137deabd173",
"commit date": "2026-01-12 14:22:25 +0800"
}

检查子模块是否正确包含以及数学库的架构是否正确。

验证 MUSA Toolkit 能否正常调用:

musaInfo

输出示例::

compiler: mcc
--------------------------------------------------------------------------------
device# 0
Name: MTT S5000
Driver Version: 4.3
Runtime Version: 4.3
compute capability: 3.1
pciBusID: 0x3
pciDeviceID: 0x0
pciDomainID: 0x0
multiProcessorCount: 56
maxThreadsPerMultiProcessor: 3072
isMultiGpuBoard: 1
clockRate: 1500 Mhz
memoryClockRate: 20000 Mhz
memoryBusWidth: 640
totalGlobalMem: 79.92 GB
sharedMemPerMultiprocessor: 192.00 KB
totalConstMem: 8192
sharedMemPerBlock: 192.00 KB
canMapHostMemory: 1
regsPerBlock: 131072
warpSize: 32
l2CacheSize: 62914560
computeMode: 0
maxThreadsPerBlock: 1024
maxThreadsDim.x: 1024
maxThreadsDim.y: 1024
maxThreadsDim.z: 1024
maxGridSize.x: 2147483647
maxGridSize.y: 2147483647
maxGridSize.z: 2147483647
concurrentKernels: 1
cooperativeLaunch: 0
cooperativeMultiDeviceLaunch: 0
isIntegrated: 0
maxTexture1D: 32768
maxTexture2D.width: 32768
maxTexture2D.height: 32768
maxTexture3D.width: 16384
maxTexture3D.height: 16384
maxTexture3D.depth: 2048
peers: device#1 device#2 device#3 device#4 device#5 device#6 device#7
non-peers: device#0
memInfo.total: 79.92 GB
memInfo.free: 79.34 GB (99%)
--------------------------------------------------------------------------------
device# 1
Name: MTT S5000
Driver Version: 4.3
Runtime Version: 4.3
compute capability: 3.1
pciBusID: 0x24
pciDeviceID: 0x0
pciDomainID: 0x0
multiProcessorCount: 56
maxThreadsPerMultiProcessor: 3072
isMultiGpuBoard: 1
clockRate: 1500 Mhz
memoryClockRate: 20000 Mhz
memoryBusWidth: 640
totalGlobalMem: 79.92 GB
sharedMemPerMultiprocessor: 192.00 KB
totalConstMem: 8192
sharedMemPerBlock: 192.00 KB
canMapHostMemory: 1
regsPerBlock: 131072
warpSize: 32
l2CacheSize: 62914560
computeMode: 0
maxThreadsPerBlock: 1024
maxThreadsDim.x: 1024
maxThreadsDim.y: 1024
maxThreadsDim.z: 1024
maxGridSize.x: 2147483647
maxGridSize.y: 2147483647
maxGridSize.z: 2147483647
concurrentKernels: 1
cooperativeLaunch: 0
cooperativeMultiDeviceLaunch: 0
isIntegrated: 0
maxTexture1D: 32768
maxTexture2D.width: 32768
maxTexture2D.height: 32768
maxTexture3D.width: 16384
maxTexture3D.height: 16384
maxTexture3D.depth: 2048
peers: device#0 device#2 device#3 device#4 device#5 device#6 device#7
non-peers: device#1
memInfo.total: 79.92 GB
memInfo.free: 79.34 GB (99%)
--------------------------------------------------------------------------------
device# 2
Name: MTT S5000
Driver Version: 4.3
Runtime Version: 4.3
compute capability: 3.1
pciBusID: 0x29
pciDeviceID: 0x0
pciDomainID: 0x0
multiProcessorCount: 56
maxThreadsPerMultiProcessor: 3072
isMultiGpuBoard: 1
clockRate: 1500 Mhz
memoryClockRate: 20000 Mhz
memoryBusWidth: 640
totalGlobalMem: 79.92 GB
sharedMemPerMultiprocessor: 192.00 KB
totalConstMem: 8192
sharedMemPerBlock: 192.00 KB
canMapHostMemory: 1
regsPerBlock: 131072
warpSize: 32
l2CacheSize: 62914560
computeMode: 0
maxThreadsPerBlock: 1024
maxThreadsDim.x: 1024
maxThreadsDim.y: 1024
maxThreadsDim.z: 1024
maxGridSize.x: 2147483647
maxGridSize.y: 2147483647
maxGridSize.z: 2147483647
concurrentKernels: 1
cooperativeLaunch: 0
cooperativeMultiDeviceLaunch: 0
isIntegrated: 0
maxTexture1D: 32768
maxTexture2D.width: 32768
maxTexture2D.height: 32768
maxTexture3D.width: 16384
maxTexture3D.height: 16384
maxTexture3D.depth: 2048
peers: device#0 device#1 device#3 device#4 device#5 device#6 device#7
non-peers: device#2
memInfo.total: 79.92 GB
memInfo.free: 79.34 GB (99%)
--------------------------------------------------------------------------------
device# 3
Name: MTT S5000
Driver Version: 4.3
Runtime Version: 4.3
compute capability: 3.1
pciBusID: 0x75
pciDeviceID: 0x0
pciDomainID: 0x0
multiProcessorCount: 56
maxThreadsPerMultiProcessor: 3072
isMultiGpuBoard: 1
clockRate: 1500 Mhz
memoryClockRate: 20000 Mhz
memoryBusWidth: 640
totalGlobalMem: 79.92 GB
sharedMemPerMultiprocessor: 192.00 KB
totalConstMem: 8192
sharedMemPerBlock: 192.00 KB
canMapHostMemory: 1
regsPerBlock: 131072
warpSize: 32
l2CacheSize: 62914560
computeMode: 0
maxThreadsPerBlock: 1024
maxThreadsDim.x: 1024
maxThreadsDim.y: 1024
maxThreadsDim.z: 1024
maxGridSize.x: 2147483647
maxGridSize.y: 2147483647
maxGridSize.z: 2147483647
concurrentKernels: 1
cooperativeLaunch: 0
cooperativeMultiDeviceLaunch: 0
isIntegrated: 0
maxTexture1D: 32768
maxTexture2D.width: 32768
maxTexture2D.height: 32768
maxTexture3D.width: 16384
maxTexture3D.height: 16384
maxTexture3D.depth: 2048
peers: device#0 device#1 device#2 device#4 device#5 device#6 device#7
non-peers: device#3
memInfo.total: 79.92 GB
memInfo.free: 79.34 GB (99%)
--------------------------------------------------------------------------------
device# 4
Name: MTT S5000
Driver Version: 4.3
Runtime Version: 4.3
compute capability: 3.1
pciBusID: 0x95
pciDeviceID: 0x0
pciDomainID: 0x0
multiProcessorCount: 56
maxThreadsPerMultiProcessor: 3072
isMultiGpuBoard: 1
clockRate: 1500 Mhz
memoryClockRate: 20000 Mhz
memoryBusWidth: 640
totalGlobalMem: 79.92 GB
sharedMemPerMultiprocessor: 192.00 KB
totalConstMem: 8192
sharedMemPerBlock: 192.00 KB
canMapHostMemory: 1
regsPerBlock: 131072
warpSize: 32
l2CacheSize: 62914560
computeMode: 0
maxThreadsPerBlock: 1024
maxThreadsDim.x: 1024
maxThreadsDim.y: 1024
maxThreadsDim.z: 1024
maxGridSize.x: 2147483647
maxGridSize.y: 2147483647
maxGridSize.z: 2147483647
concurrentKernels: 1
cooperativeLaunch: 0
cooperativeMultiDeviceLaunch: 0
isIntegrated: 0
maxTexture1D: 32768
maxTexture2D.width: 32768
maxTexture2D.height: 32768
maxTexture3D.width: 16384
maxTexture3D.height: 16384
maxTexture3D.depth: 2048
peers: device#0 device#1 device#2 device#3 device#5 device#6 device#7
non-peers: device#4
memInfo.total: 79.92 GB
memInfo.free: 79.34 GB (99%)
--------------------------------------------------------------------------------
device# 5
Name: MTT S5000
Driver Version: 4.3
Runtime Version: 4.3
compute capability: 3.1
pciBusID: 0xab
pciDeviceID: 0x0
pciDomainID: 0x0
multiProcessorCount: 56
maxThreadsPerMultiProcessor: 3072
isMultiGpuBoard: 1
clockRate: 1500 Mhz
memoryClockRate: 20000 Mhz
memoryBusWidth: 640
totalGlobalMem: 79.92 GB
sharedMemPerMultiprocessor: 192.00 KB
totalConstMem: 8192
sharedMemPerBlock: 192.00 KB
canMapHostMemory: 1
regsPerBlock: 131072
warpSize: 32
l2CacheSize: 62914560
computeMode: 0
maxThreadsPerBlock: 1024
maxThreadsDim.x: 1024
maxThreadsDim.y: 1024
maxThreadsDim.z: 1024
maxGridSize.x: 2147483647
maxGridSize.y: 2147483647
maxGridSize.z: 2147483647
concurrentKernels: 1
cooperativeLaunch: 0
cooperativeMultiDeviceLaunch: 0
isIntegrated: 0
maxTexture1D: 32768
maxTexture2D.width: 32768
maxTexture2D.height: 32768
maxTexture3D.width: 16384
maxTexture3D.height: 16384
maxTexture3D.depth: 2048
peers: device#0 device#1 device#2 device#3 device#4 device#6 device#7
non-peers: device#5
memInfo.total: 79.92 GB
memInfo.free: 79.34 GB (99%)
--------------------------------------------------------------------------------
device# 6
Name: MTT S5000
Driver Version: 4.3
Runtime Version: 4.3
compute capability: 3.1
pciBusID: 0xb2
pciDeviceID: 0x0
pciDomainID: 0x0
multiProcessorCount: 56
maxThreadsPerMultiProcessor: 3072
isMultiGpuBoard: 1
clockRate: 1500 Mhz
memoryClockRate: 20000 Mhz
memoryBusWidth: 640
totalGlobalMem: 79.92 GB
sharedMemPerMultiprocessor: 192.00 KB
totalConstMem: 8192
sharedMemPerBlock: 192.00 KB
canMapHostMemory: 1
regsPerBlock: 131072
warpSize: 32
l2CacheSize: 62914560
computeMode: 0
maxThreadsPerBlock: 1024
maxThreadsDim.x: 1024
maxThreadsDim.y: 1024
maxThreadsDim.z: 1024
maxGridSize.x: 2147483647
maxGridSize.y: 2147483647
maxGridSize.z: 2147483647
concurrentKernels: 1
cooperativeLaunch: 0
cooperativeMultiDeviceLaunch: 0
isIntegrated: 0
maxTexture1D: 32768
maxTexture2D.width: 32768
maxTexture2D.height: 32768
maxTexture3D.width: 16384
maxTexture3D.height: 16384
maxTexture3D.depth: 2048
peers: device#0 device#1 device#2 device#3 device#4 device#5 device#7
non-peers: device#6
memInfo.total: 79.92 GB
memInfo.free: 79.34 GB (99%)
--------------------------------------------------------------------------------
device# 7
Name: MTT S5000
Driver Version: 4.3
Runtime Version: 4.3
compute capability: 3.1
pciBusID: 0xde
pciDeviceID: 0x0
pciDomainID: 0x0
multiProcessorCount: 56
maxThreadsPerMultiProcessor: 3072
isMultiGpuBoard: 1
clockRate: 1500 Mhz
memoryClockRate: 20000 Mhz
memoryBusWidth: 640
totalGlobalMem: 79.92 GB
sharedMemPerMultiprocessor: 192.00 KB
totalConstMem: 8192
sharedMemPerBlock: 192.00 KB
canMapHostMemory: 1
regsPerBlock: 131072
warpSize: 32
l2CacheSize: 62914560
computeMode: 0
maxThreadsPerBlock: 1024
maxThreadsDim.x: 1024
maxThreadsDim.y: 1024
maxThreadsDim.z: 1024
maxGridSize.x: 2147483647
maxGridSize.y: 2147483647
maxGridSize.z: 2147483647
concurrentKernels: 1
cooperativeLaunch: 0
cooperativeMultiDeviceLaunch: 0
isIntegrated: 0
maxTexture1D: 32768
maxTexture2D.width: 32768
maxTexture2D.height: 32768
maxTexture3D.width: 16384
maxTexture3D.height: 16384
maxTexture3D.depth: 2048
peers: device#0 device#1 device#2 device#3 device#4 device#5 device#6
non-peers: device#7
memInfo.total: 79.92 GB
memInfo.free: 79.34 GB (99%)

4. muDNN 安装指南

本章主要描述 muDNN 的详细安装步骤。

4.1 准备工作

  • 已安装兼容版本 MT GPU 驱动及 MUSA Toolkit。

4.2 安装 muDNN

4.2.1 安装

mudnn_rc3.1.5.PH1.tar.gz 为例:

tar -zxvf mudnn_rc3.1.5.PH1.tar.gz
cd mudnn && ./install_mudnn.sh -i #非 root 用户需要 sudo 提权执行
#卸载可通过如下命令
./install_mudnn.sh -u

4.2.2 设置环境变量

#若已设置则无需操作
export PATH=/usr/local/musa/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/musa/lib:$LD_LIBRARY_PATH

安装完成后,您可以查看 muDNN 的所有 API。

5. mccl 安装指南

本章主要描述 mccl 的详细安装步骤。

5.1 准备工作

  • 已安装兼容版本 MT GPU 驱动及 MUSA Toolkit以及muDNN。

5.2 安装 mccl

5.2.1 安装

mccl_rc2.1.5.PH1.tar.gz 为例:

tar -zxvf mccl_rc2.1.5.PH1.tar.gz
cd mccl && ./install.sh -i #非 root 用户需要 sudo 提权执行
#卸载可通过如下命令
./install.sh -u

5.2.2 设置环境变量

#若已设置则无需操作
export PATH=/usr/local/musa/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/musa/lib:$LD_LIBRARY_PATH

安装完成后,您可以查看 mccl 的所有 API。