Prepare the resource environment for acceleration

Environment requirements

To use TorchAcc for training acceleration, you must use a GPU-accelerated instance. The required versions and specifications are listed below.

• Version requirements

  Driver	Version
  CUDA driver	11.3 or later
  Nvidia driver	470 or later

• Specifications

  Instance type	Supported
  V100M16	Support
  V100M32	Supported
  GU50	Support
  GU100	Support
  GU108	Support
  A10M24	Support

  For more information about instance types, see Appendix: List of public resource specifications.

• Image requirements

  To use TorchAcc for training acceleration, you must use the specified TorchAcc test runtime image: registry.cn-hangzhou.aliyuncs.com/pai-dlc/pai-pytorch-training:torch-1.12-cuda11.3-py38-acc-230219.

  Note

  This image environment is available only in the China (Hangzhou) region.

Prepare the staging environment

Using a DSW environment

To test TorchAcc on the PAI platform, you can create a DSW instance and perform tests in a Jupyter Notebook as follows:

1. In the China (Hangzhou) region, create a resource and add a resource quota that meet the environment requirements. For more information, see Create a resource group and purchase general computing resources and General computing resource quotas.

2. Go to the workspace that is associated with the resource quota and create a developer instance. Configure the key parameters as follows. For more information, see Create and manage DSW instances.

   Parameter	Description
   Resource Type	Select Resource Quota.
   Resource Quota	Select the resource quota that you created in Step 1.
   Resource Specifications	Set CPU (Cores) to 30. Set Memory (GiB) to 180. Set Shared Memory (GiB) to 100. Set GPU (Cards) to 1.
   Image	On the Registry Address tab, set the registry address to registry.cn-hangzhou.aliyuncs.com/pai-dlc/pai-pytorch-training:torch-1.12-cuda11.3-py38-acc-230219. Note This image environment is available only in the China (Hangzhou) region.

Use your own ECS instance

To use your own ECS resources for TorchAcc training acceleration, you can prepare an ECS instance as follows:

1. In the China (Hangzhou) region, purchase an ECS instance that meets the environment requirements and install the required versions of the nvidia-smi and CUDA driver. For more information about how to purchase an instance, see Create an instance. The key parameters are as follows:

   • Set Instance to ecs.gn6v-c8g1.2xlarge.

   • For Image, select Public Image > Alibaba Cloud Linux > Alibaba Cloud Linux 3.2104 LST 64-bit. Select the Install GPU Driver check box and then select the following versions: CUDA Version 11.4.1 > Driver Version 470.161.03 > CUDNN Version 8.2.4.

   • For the System Disk, allocate at least 80 GiB of storage capacity.

2. Install Docker on the ECS instance. For more information, see Install and use Docker and Docker Compose.

3. Install the NVIDIA Container Toolkit. For more information, see Installing the NVIDIA Container Toolkit.

   Select the installation command for your operating system. The steps in this topic use yum or dnf. After the installation is complete, restart the Docker daemon.

4. You can run the following script to start the TorchAcc runtime image.

   DOCKER=registry.cn-hangzhou.aliyuncs.com/pai-dlc/pai-pytorch-training:torch-1.12-cuda11.3-py38-acc-230219
   name=TorchAcc_Tutorials
   
   set -x
   docker run \
       --name $name \
       --rm -it \
       --privileged \
       --ulimit memlock=-1:-1 \
       --gpus all \
       --shm-size 10G \
       -v /dev/shm:/dev/shm \
       --ipc host \
       --network host \
       --rm \
       --cap-add=CAP_SYS_ADMIN \
       -v /path/to/code:/workspace \
       -w /workspace \
       ${DOCKER} bash