Installation and Deployment

This section describes how to install and deploy GMEM.

Software and Hardware Requirements

• Kunpeng 920 CPU
• Ascend 910 processor
• openEuler 23.09

Environment Requirements

• The root permission is required for using and configuring GMEM.
• GMEM can be enabled or disabled only at the system level.
• The administrator must ensure that the GMEM configuration is safe and available.

Installing GMEM

• Prepare files.

  CANN Community Version History - Ascend Community (hiascend.com)

  Firmware and Driver - Ascend Community (hiascend.com)

  Source	Software Package
  openEuler 23.09	kernel-6.4.0-xxx.aarch64.rpm kernel-devel-6.4.0-xxx.aarch64.rpm libgmem-xxx.aarch64.rpm libgmem-devel-xxx.aarch64.rpm
  Ascend community	CANN package: Ascend-cann-toolkit-xxx-linux.aarch64.rpm NPU firmware and driver: Ascend-hdk-910-npu-driver-xxx.aarch64.rpm Ascend-hdk-910-npu-firmware-xxx.noarch.rpm
  Contact the maintainers of the GMEM community. @yang_yanchao email: yangyanchao6@huawei.com @LemmyHuang email: huangliming5@huawei.com	gmem-example-xxx.aarch64.rpm mindspore-xxx-linux_aarch64.whl

• Install the kernel.

  Ensure that GMEM compilation options are enabled (enabled by default) for the openEuler kernel.

  [root@localhost ~]# cat /boot/config-`uname -r` | grep CONFIG_GMEM
  CONFIG_GMEM=y
  CONFIG_GMEM_DEV=m
  
  [root@localhost ~]#  cat /boot/config-`uname -r` | grep CONFIG_REMOTE_PAGER
  CONFIG_REMOTE_PAGER=m
  CONFIG_REMOTE_PAGER_MASTER=m
  

  Add gmem=on to the boot options.

  [root@localhost gmem]# cat /proc/cmdline
  BOOT_IMAGE=/vmlinuz-xxx root=/dev/mapper/openeuler-root ... gmem=on
  

  Configure transparent_hugepage.

  echo always > /sys/kernel/mm/transparent_hugepage/enabled 
  

• Install the user-mode dynamic library libgmem.

  yum install libgmem libgmem-devel 
  

• Install the CANN framework.

  Install the matching CANN, including the toolkit, driver, and firmware. After the installation is complete, restart the system.

  rpm -ivh Ascend-cann-toolkit-xxx-linux.aarch64.rpm
  # Use the tool provided by libgmem to install the NPU driver.
  sh /usr/local/gmem/install_npu_driver.sh Ascend-hdk-910-npu-driver-xxx.aarch64.rpm
  rpm -ivh Ascend-hdk-910-npu-firmware-xxx.noarch.rpm
  

  Run the environment configuration script in the Ascend directory to configure environment variables.

  source /usr/local/Ascend/ascend-toolkit/set_env.sh
  

  Check whether the NPU is working properly.

  [root@localhost ~]# npu-smi info
  +-------------------------------------------------------------------------------------------+
  | npu-smi 22.0.4.1                 Version: 22.0.4.1                                        |
  +----------------------+---------------+----------------------------------------------------+
  | NPU   Name           | Health        | Power(W)    Temp(C)           Hugepages-Usage(page)|
  | Chip                 | Bus-Id        | AICore(%)   Memory-Usage(MB)  HBM-Usage(MB)        |
  +======================+===============+====================================================+
  | 0     910B           | OK            | 79.4        82                0    / 0             |
  | 0                    | 0000:81:00.0  | 0           1979 / 15039      0    / 32768         |
  +======================+===============+====================================================+
  

• Install the gmem-example software package.

  gmem-example updates the host driver, NPU driver, and NPU kernel. After the installation is complete, restart the system for the driver to take effect.

  rpm -ivh gmem-example-xxx.aarch64.rpm
  

• Install MindSpore.

  Obtain the correct MindSpore version and install it. After the installation, run the following command to check whether MindSpore functions are normal:

  python -c "import mindspore;mindspore.run_check()"
  MindSpore version:  x.x.x
  The result of multiplication calculation is correct, MindSpore has been installed on platform [Ascend] successfully!
  

Performing Training or Inference

After installation is complete, you can execute MindSpore-based training or inference directly without any adaptation.