样例执行

本节以单机8卡组网、通过rank table文件配置资源信息的方式为例，介绍如何运行代码示例中的样例代码。

1. 准备rank table文件。
   以 Atlas A2 训练系列产品 ，组网为单机8卡为例，rank_table.json配置示例如下，不同产品形态rank table文件的配置示例及详细参数说明可参见rank table文件配置资源信息。

   

   以下JSON文件仅为配置示例，使用时请根据实际组网信息进行配置，并删除注释。

   {
       "status":"completed",   // rank table可用标识，completed为可用
       "version": "1.0",
       "server_count": "1",   // 参与训练的AI Server数目
       "server_list": [
           {
               "server_id": "SERVER_ID_SV1",   // AI Server标识，String类型，请确保全局唯一
               "device": [                     // AI Server中的Device列表
                   {
                       "device_id": "0",
                       "device_ip": "192.168.1.8",
                       "rank_id": "0"
                   },
                   {
                       "device_id": "1",
                       "device_ip": "192.168.1.9",
                       "rank_id": "1"
                   },
                   {
                       "device_id": "2",
                       "device_ip": "192.168.1.10",
                       "rank_id": "2"
                   },
                   {
                       "device_id": "3",
                       "device_ip": "192.168.1.11",
                       "rank_id": "3"
                   },
                   {
                       "device_id": "4",
                       "device_ip": "192.168.1.12",
                       "rank_id": "4"
                   },
                   {
                       "device_id": "5",
                       "device_ip": "192.168.1.13",
                       "rank_id": "5"
                   },
                   {
                       "device_id": "6",
                       "device_ip": "192.168.1.14",
                       "rank_id": "6"
                   },
                   {
                       "device_id": "7",
                       "device_ip": "192.168.1.15",
                       "rank_id": "7"
                   }
               ]
           }
       ]
   }

2. 构造启动脚本。

   假设命名为hccl_start_8p.sh，示例如下：

   1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

   # 配置CANN软件环境变量（以root用户为例）： source /usr/local/Ascend/ascend-toolkit/set_env.sh # TF Adapter python库，其中${TFPLUGIN_INSTALL_PATH}为TF Adapter软件包安装路径。 export PYTHONPATH=${TFPLUGIN_INSTALL_PATH}:$PYTHONPATH export RANK_SIZE=8 export RANK_TABLE_FILE=/home/test/rank_table.json # rank table资源配置文件路径，请根据实际情况替换 export JOB_ID=10087 # 用户自定义，指定任务ID，可以包含大小写字母、数字、中划线或下划线 for((RANK_ID=0;RANK_ID<$((RANK_SIZE));RANK_ID++)); do export RANK_ID=$RANK_ID export ASCEND_DEVICE_ID=$RANK_ID # 执行脚本，脚本路径与名称请根据实际情况替换 nohup python3 /home/test/hccl_test.py & done

3. 执行启动脚本。

   1	bash hccl_start_8p.sh

   结果示例如下：

   1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27

   ... ... 'reduce_sum': array([[ 0, 0, 0, ..., 0, 0, 0], [ 0, 0, 0, ..., 0, 0, 0], [ 0, 0, 0, ..., 0, 0, 0], ..., [ 0, 0, 0, ..., 0, 0, 0], [ 0, 0, 0, ..., 0, 0, 0], [ 0, 0, 0, ..., 44, 44, 44]]), 'reduce_max': array([[4097, 4098, 4099, ..., 4222, 4223, 4224], [4225, 4226, 4227, ..., 4350, 4351, 4352], [4353, 4354, 4355, ..., 4478, 4479, 4480], ..., [4737, 4738, 4739, ..., 4862, 4863, 4864], [4865, 4866, 4867, ..., 4990, 4991, 4992], [4993, 4994, 4995, ..., 9, 9, 9]]), 'reduce_min': array([[0, 0, 0, ..., 0, 0, 0], [0, 0, 0, ..., 0, 0, 0], [0, 0, 0, ..., 0, 0, 0], ..., [0, 0, 0, ..., 0, 0, 0], [0, 0, 0, ..., 0, 0, 0], [0, 0, 0, ..., 2, 2, 2]]), 'reduce_prod': array([[ 0, 0, 0, ..., 0, 0, 0], [ 0, 0, 0, ..., 0, 0, 0], [ 0, 0, 0, ..., 0, 0, 0], ..., [ 0, 0, 0, ..., 0, 0, 0], [ 0, 0, 0, ..., 0, 0, 0], [ 0, 0, 0, ..., 362880, 362880, 362880]]), 'alltoallv_tensor': array([ 1, 2, 3, ..., 8246, 8247, 8248]), 'check_tensors': array([ 1, 2, 3, ..., 8246, 8247, 8248]) train success