从0制作自定义镜像用于创建训练作业(MindSpore+Ascend)
本案例介绍如何从0到1制作Ascend容器镜像,并使用该镜像在ModelArts平台上进行训练。镜像中使用的AI引擎是MindSpore,训练使用的资源是专属资源池的Ascend芯片。
场景描述
目标:构建安装如下软件的容器镜像,并在ModelArts平台上使用Ascend规格资源运行训练任务。
- ubuntu-18.04
- cann-6.3.RC2 (商用版本)
- python-3.7.13
- mindspore-2.1.1
- 本教程以cann-6.3.RC2.、mindspore-2.1.1为例介绍。
- 本示例仅用于示意Ascend容器镜像制作流程,且在匹配正确的Ascend驱动/固件版本的专属资源池上运行通过。
操作流程
使用自定义镜像创建训练作业时,需要您熟悉docker软件的使用,并具备一定的开发经验。详细步骤如下所示:
约束限制
- 由于案例中需要下载商用版CANN,因此本案例仅面向有下载权限的渠道用户,非渠道用户建议参考其他自定义镜像制作教程。
- Mindspore版本与CANN版本,CANN版本与Ascend驱动/固件版本均有严格的匹配关系,版本不匹配会导致训练失败。
前提条件
已注册华为账号并开通华为云,且在使用ModelArts前检查账号状态,账号不能处于欠费或冻结状态。
Step1 创建OBS桶和文件夹
在OBS服务中创建桶和文件夹,用于存放样例数据集以及训练代码。如下示例中,请创建命名为“test-modelarts”的桶,并创建如表1所示的文件夹。
请确保您使用的OBS与ModelArts在同一区域。
Step2 准备脚本文件并上传至OBS中
- 准备本案例所需训练脚本mindspore-verification.py文件和Ascend的启动脚本文件(共5个)。
- 训练脚本文件具体内容请参见训练mindspore-verification.py文件。
- Ascend的启动脚本文件包括以下5个,具体脚本内容请参见Ascend的启动脚本文件。
- run_ascend.py
- common.py
- rank_table.py
- manager.py
- fmk.py
mindspore-verification.py和run_ascend.py脚本文件在创建训练作业时的“启动命令”参数中调用,具体请参见启动命令。
run_ascend.py脚本运行时会调用common.py、rank_table.py、manager.py、fmk.py脚本。
- 上传训练脚本mindspore-verification.py文件至OBS桶的“obs://test-modelarts/ascend/demo-code/”文件夹下。
- 上传Ascend的启动脚本文件(共5个)至OBS桶的“obs://test-modelarts/ascend/demo-code/run_ascend/”文件夹下。
Step3 制作自定义镜像
此处介绍如何通过编写Dockerfile文件制作自定义镜像的操作步骤。
目标:构建安装好如下软件的容器镜像,并使用ModelArts训练服务运行。
- ubuntu-18.04
- cann-6.3.RC2(商用版本)
- python-3.7.13
- mindspore-2.1.1
Mindspore版本与CANN版本,CANN版本和Ascend驱动/固件版本均有严格的匹配关系,版本不匹配会导致训练失败。
本示例仅用于示意Ascend容器镜像制作流程,且在匹配正确的Ascend驱动/固件版本的专属资源池上运行通过。
- 准备一台Linux aarch64架构的主机,操作系统使用ubuntu-18.04。您可以准备相同规格的弹性云服务器ECS或者应用本地已有的主机进行自定义镜像的制作。
购买ECS服务器的具体操作请参考购买并登录Linux弹性云服务器。“CPU架构”选择“x86计算”,“镜像”选择“公共镜像”,推荐使用Ubuntu18.04的镜像。
- 安装Docker。
以Linux aarch64架构的操作系统为例,获取Docker安装包。您可以使用以下指令安装Docker。关于安装Docker的更多指导内容参见Docker官方文档。
curl -fsSL get.docker.com -o get-docker.sh sh get-docker.sh
如果docker images命令可以执行成功,表示Docker已安装,此步骤可跳过。
启动docker。systemctl start docker
- 确认Docker Engine版本。执行如下命令。
docker version | grep -A 1 Engine
命令回显如下。Engine: Version: 18.09.0
推荐使用大于等于该版本的Docker Engine来制作自定义镜像。
- 准备名为context的文件夹。
mkdir -p context
- 准备可用的pip源文件pip.conf。本示例使用华为开源镜像站提供的pip源,其pip.conf文件内容如下。
[global] index-url = https://repo.huaweicloud.com/repository/pypi/simple trusted-host = repo.huaweicloud.com timeout = 120
在华为开源镜像站https://mirrors.huaweicloud.com/home中,搜索pypi,可以查看pip.conf文件内容。
- 准备可用的apt源文件Ubuntu-Ports-bionic.list。本示例使用华为开源镜像站提供的apt源,执行如下命令获取apt源文件。
wget -O Ubuntu-Ports-bionic.list https://repo.huaweicloud.com/repository/conf/Ubuntu-Ports-bionic.list
在华为开源镜像站https://mirrors.huaweicloud.com/home中,搜索Ubuntu-Ports,可以查看获取apt源文件的命令。
- 下载CANN 6.3.RC2-linux aarch64与mindspore-2.1.1-cp37-cp37m-linux_aarch64.whl安装文件。
- 下载run文件“Ascend-cann-nnae_6.3.RC2_linux-aarch64.run”(下载链接)。
- 下载whl文件“mindspore-2.1.1-cp37-cp37m-linux_aarch64.whl”(下载链接)。
ModelArts当前仅支持CANN商用版本,不支持社区版。
- 下载Miniconda3安装文件。
使用地址https://repo.anaconda.com/miniconda/Miniconda3-py37_4.10.3-Linux-aarch64.sh,下载Miniconda3-py37-4.10.3安装文件(对应python 3.7.10)。
- 将上述pip源文件、*.run文件、*.whl文件、Miniconda3安装文件放置在context文件夹内,context文件夹内容如下。
context ├── Ascend-cann-nnae_6.3.RC2_linux-aarch64.run ├── mindspore-2.1.1-cp37-cp37m-linux_aarch64.whl ├── Miniconda3-py37_4.10.3-Linux-aarch64.sh ├── pip.conf └── Ubuntu-Ports-bionic.list
- 编写容器镜像Dockerfile文件。
在context文件夹内新建名为Dockerfile的空文件,并将下述内容写入其中。
# 容器镜像构建主机需要连通公网 FROM arm64v8/ubuntu:18.04 AS builder # 基础容器镜像的默认用户已经是 root # USER root # 安装 OS 依赖(使用华为开源镜像站) COPY Ubuntu-Ports-bionic.list /tmp RUN cp -a /etc/apt/sources.list /etc/apt/sources.list.bak && \ mv /tmp/Ubuntu-Ports-bionic.list /etc/apt/sources.list && \ echo > /etc/apt/apt.conf.d/00skip-verify-peer.conf "Acquire { https::Verify-Peer false }" && \ apt-get update && \ apt-get install -y \ # utils ca-certificates vim curl \ # CANN 6.3.RC2 gcc-7 g++ make cmake zlib1g zlib1g-dev openssl libsqlite3-dev libssl-dev libffi-dev unzip pciutils net-tools libblas-dev gfortran libblas3 && \ apt-get clean && \ mv /etc/apt/sources.list.bak /etc/apt/sources.list && \ # 修改 CANN 6.3.RC2 安装目录的父目录权限,使得 ma-user 可以写入 chmod o+w /usr/local RUN useradd -m -d /home/ma-user -s /bin/bash -g 100 -u 1000 ma-user # 设置容器镜像默认用户与工作目录 USER ma-user WORKDIR /home/ma-user # 使用华为开源镜像站提供的 pypi 配置 RUN mkdir -p /home/ma-user/.pip/ COPY --chown=ma-user:100 pip.conf /home/ma-user/.pip/pip.conf # 复制待安装文件到基础容器镜像中的 /tmp 目录 COPY --chown=ma-user:100 Miniconda3-py37_4.10.3-Linux-aarch64.sh /tmp # https://conda.io/projects/conda/en/latest/user-guide/install/linux.html#installing-on-linux # 安装 Miniconda3 到基础容器镜像的 /home/ma-user/miniconda3 目录中 RUN bash /tmp/Miniconda3-py37_4.10.3-Linux-aarch64.sh -b -p /home/ma-user/miniconda3 ENV PATH=$PATH:/home/ma-user/miniconda3/bin # 安装 CANN 6.3.RC2 Python Package 依赖 RUN pip install numpy~=1.14.3 decorator~=4.4.0 sympy~=1.4 cffi~=1.12.3 protobuf~=3.11.3 \ attrs pyyaml pathlib2 scipy requests psutil absl-py # 安装 CANN 6.3.RC2 至 /usr/local/Ascend 目录 COPY --chown=ma-user:100 Ascend-cann-nnae_6.3.RC2_linux-aarch64.run /tmp RUN chmod +x /tmp/Ascend-cann-nnae_6.3.RC2_linux-aarch64.run && \ /tmp/Ascend-cann-nnae_6.3.RC2_linux-aarch64.run --install --install-path=/usr/local/Ascend # 安装 MindSpore 2.1.1 COPY --chown=ma-user:100 mindspore-2.1.1-cp37-cp37m-linux_aarch64.whl /tmp RUN chmod +x /tmp/mindspore-2.1.1-cp37-cp37m-linux_aarch64.whl && \ pip install /tmp/mindspore-2.1.1-cp37-cp37m-linux_aarch64.whl # 构建最终容器镜像 FROM arm64v8/ubuntu:18.04 # 安装 OS 依赖(使用华为开源镜像站) COPY Ubuntu-Ports-bionic.list /tmp RUN cp -a /etc/apt/sources.list /etc/apt/sources.list.bak && \ mv /tmp/Ubuntu-Ports-bionic.list /etc/apt/sources.list && \ echo > /etc/apt/apt.conf.d/00skip-verify-peer.conf "Acquire { https::Verify-Peer false }" && \ apt-get update && \ apt-get install -y \ # utils ca-certificates vim curl \ # CANN 6.3.RC2 gcc-7 g++ make cmake zlib1g zlib1g-dev openssl libsqlite3-dev libssl-dev libffi-dev unzip pciutils net-tools libblas-dev gfortran libblas3 && \ apt-get clean && \ mv /etc/apt/sources.list.bak /etc/apt/sources.list RUN useradd -m -d /home/ma-user -s /bin/bash -g 100 -u 1000 ma-user # 从上述 builder stage 中复制目录到当前容器镜像的同名目录 COPY --chown=ma-user:100 --from=builder /home/ma-user/miniconda3 /home/ma-user/miniconda3 COPY --chown=ma-user:100 --from=builder /home/ma-user/Ascend /home/ma-user/Ascend COPY --chown=ma-user:100 --from=builder /home/ma-user/var /home/ma-user/var COPY --chown=ma-user:100 --from=builder /usr/local/Ascend /usr/local/Ascend # 设置容器镜像预置环境变量 # 请务必设置 CANN 相关环境变量 # 请务必设置 Ascend Driver 相关环境变量 # 请务必设置 PYTHONUNBUFFERED=1, 以免日志丢失 ENV PATH=$PATH:/usr/local/Ascend/nnae/latest/bin:/usr/local/Ascend/nnae/latest/compiler/ccec_compiler/bin:/home/ma-user/miniconda3/bin \ LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/Ascend/driver/lib64:/usr/local/Ascend/driver/lib64/common:/usr/local/Ascend/driver/lib64/driver:/usr/local/Ascend/nnae/latest/lib64:/usr/local/Ascend/nnae/latest/lib64/plugin/opskernel:/usr/local/Ascend/nnae/latest/lib64/plugin/nnengine \ PYTHONPATH=$PYTHONPATH:/usr/local/Ascend/nnae/latest/python/site-packages:/usr/local/Ascend/nnae/latest/opp/built-in/op_impl/ai_core/tbe \ ASCEND_AICPU_PATH=/usr/local/Ascend/nnae/latest \ ASCEND_OPP_PATH=/usr/local/Ascend/nnae/latest/opp \ ASCEND_HOME_PATH=/usr/local/Ascend/nnae/latest \ PYTHONUNBUFFERED=1 # 设置容器镜像默认用户与工作目录 USER ma-user WORKDIR /home/ma-user
关于Dockerfile文件编写的更多指导内容参见Docker官方文档。
- 确认已创建完成Dockerfile文件。此时context文件夹内容如下。
context ├── Ascend-cann-nnae_6.3.RC2_linux-aarch64.run ├── Dockerfile ├── mindspore-2.1.1-cp37-cp37m-linux_aarch64.whl ├── Miniconda3-py37_4.10.3-Linux-aarch64.sh ├── pip.conf └── Ubuntu-Ports-bionic.list
- 构建容器镜像。在Dockerfile文件所在的目录执行如下命令构建容器镜像。
1
docker build . -t mindspore:2.1.1-cann6.3.RC2
构建过程结束时出现如下构建日志说明镜像构建成功。Successfully tagged mindspore:2.1.1-cann6.3.RC2
- 将制作完成的镜像上传至SWR服务,具体参见Step4 上传镜像至SWR。
Step4 上传镜像至SWR
本章节介绍如何将制作好的镜像上传至SWR服务,方便后续在ModelArts上创建训练作业时调用。
- 登录容器镜像服务控制台,选择区域,要和ModelArts区域保持一致,否则无法选择到镜像。
- 单击右上角“创建组织”,输入组织名称完成组织创建。请自定义组织名称,本示例使用“deep-learning”,下面的命令中涉及到组织名称“deep-learning”也请替换为自定义的值。
- 单击右上角“登录指令”,获取登录访问指令,本文选择复制临时登录指令。
- 以root用户登录本地环境,输入复制的SWR临时登录指令。
- 上传镜像至容器镜像服务镜像仓库。
- 使用docker tag命令给上传镜像打标签。
#region和domain信息请替换为实际值,组织名称deep-learning也请替换为自定义的值。 sudo docker tag mindspore:2.1.1-cann6.3.RC2 swr.{region}.{domain}/deep-learning/mindspore:2.1.1-cann6.3.RC2 #以华为云北京四为例: sudo docker tag mindspore:2.1.1-cann6.3.RC2 swr.cn-north-4.myhuaweicloud.com/deep-learning/mindspore:2.1.1-cann6.3.RC2
- 使用docker push命令上传镜像。
#region和domain信息请替换为实际值,组织名称deep-learning也请替换为自定义的值。 sudo docker push swr.{region}.{domain}/deep-learning/mindspore:2.1.1-cann6.3.RC2 #以华为云北京四为例: sudo docker push swr.cn-north-4.myhuaweicloud.com/deep-learning/mindspore:2.1.1-cann6.3.RC2
- 使用docker tag命令给上传镜像打标签。
- 完成镜像上传后,在“容器镜像服务控制台>我的镜像”页面可查看已上传的自定义镜像。
“swr.cn-north-4.myhuaweicloud.com/deep-learning/mindspore:2.1.1-cann6.3.RC2”即为此自定义镜像的“SWR_URL”。
Step5 在ModelArts上创建Notebook并调试
- 将上传到SWR上的镜像注册到ModelArts的镜像管理中。
登录ModelArts管理控制台,在左侧导航栏中选择“镜像管理 ”,单击“注册镜像”,根据界面提示注册镜像。注册后的镜像可以用于创建Notebook。
- 在Notebook中使用自定义镜像创建Notebook并调试,调试成功后,保存镜像。
- 在Notebook中使用自定义镜像创建Notebook操作请参见基于自定义镜像创建Notebook实例。
- 保存Notebook镜像操作请参见保存Notebook镜像环境。
- 已有的镜像调试成功后,再使用ModelArts训练模块训练作业。
Step6 在ModelArts上创建训练作业
- 登录ModelArts管理控制台,在左侧导航栏中选择“训练管理 > 训练作业”,默认进入“训练作业”列表。
- 在“创建训练作业”页面,填写相关参数信息,然后单击“提交”。
- 创建方式:选择“自定义算法”
- 启动方式:选择“自定义”
- 镜像地址:“swr.cn-north-4.myhuaweicloud.com/deep-learning/mindspore:2.1.1-cann6.3.RC2”
- 代码目录:设置为OBS中存放启动脚本文件的目录,例如:“obs://test-modelarts/ascend/demo-code/”
- 启动命令:“python ${MA_JOB_DIR}/demo-code/run_ascend/run_ascend.py python ${MA_JOB_DIR}/demo-code/mindspore-verification.py”
- 资源池:选择专属资源池
- 类型:选择驱动/固件版本匹配的专属资源池Ascend规格。
- 作业日志路径:设置为OBS中存放训练日志的路径。例如:“obs://test-modelarts/ascend/log/”
- 在“规格确认”页面,确认训练作业的参数信息,确认无误后单击“提交”。
- 训练作业创建完成后,后台将自动完成容器镜像下载、代码目录下载、执行启动命令等动作。
训练作业一般需要运行一段时间,根据您的训练业务逻辑和选择的资源不同,训练时长将持续几十分钟到几小时不等。训练作业执行成功后,日志信息如图1所示。
训练mindspore-verification.py文件
mindspore-verification.py文件内容如下:
import os import numpy as np from mindspore import Tensor import mindspore.ops as ops import mindspore.context as context print('Ascend Envs') print('------') print('JOB_ID: ', os.environ['JOB_ID']) print('RANK_TABLE_FILE: ', os.environ['RANK_TABLE_FILE']) print('RANK_SIZE: ', os.environ['RANK_SIZE']) print('ASCEND_DEVICE_ID: ', os.environ['ASCEND_DEVICE_ID']) print('DEVICE_ID: ', os.environ['DEVICE_ID']) print('RANK_ID: ', os.environ['RANK_ID']) print('------') context.set_context(device_target="Ascend") x = Tensor(np.ones([1,3,3,4]).astype(np.float32)) y = Tensor(np.ones([1,3,3,4]).astype(np.float32)) print(ops.add(x, y))
Ascend的启动脚本文件
- run_ascend.py
import sys import os from common import RunAscendLog from common import RankTableEnv from rank_table import RankTable, RankTableTemplate1, RankTableTemplate2 from manager import FMKManager if __name__ == '__main__': log = RunAscendLog.setup_run_ascend_logger() if len(sys.argv) <= 1: log.error('there are not enough args') sys.exit(1) train_command = sys.argv[1:] log.info('training command') log.info(train_command) if os.environ.get(RankTableEnv.RANK_TABLE_FILE_V1) is not None: # new format rank table file rank_table_path = os.environ.get(RankTableEnv.RANK_TABLE_FILE_V1) RankTable.wait_for_available(rank_table_path) rank_table = RankTableTemplate1(rank_table_path) else: # old format rank table file rank_table_path_origin = RankTableEnv.get_rank_table_template2_file_path() RankTable.wait_for_available(rank_table_path_origin) rank_table = RankTableTemplate2(rank_table_path_origin) if rank_table.get_device_num() >= 1: log.info('set rank table %s env to %s' % (RankTableEnv.RANK_TABLE_FILE, rank_table.get_rank_table_path())) RankTableEnv.set_rank_table_env(rank_table.get_rank_table_path()) else: log.info('device num < 1, unset rank table %s env' % RankTableEnv.RANK_TABLE_FILE) RankTableEnv.unset_rank_table_env() instance = rank_table.get_current_instance() server = rank_table.get_server(instance.server_id) current_instance = RankTable.convert_server_to_instance(server) fmk_manager = FMKManager(current_instance) fmk_manager.run(rank_table.get_device_num(), train_command) return_code = fmk_manager.monitor() fmk_manager.destroy() sys.exit(return_code)
- common.py
import logging import os logo = 'Training' # Rank Table Constants class RankTableEnv: RANK_TABLE_FILE = 'RANK_TABLE_FILE' RANK_TABLE_FILE_V1 = 'RANK_TABLE_FILE_V_1_0' HCCL_CONNECT_TIMEOUT = 'HCCL_CONNECT_TIMEOUT' # jobstart_hccl.json is provided by the volcano controller of Cloud-Container-Engine(CCE) HCCL_JSON_FILE_NAME = 'jobstart_hccl.json' RANK_TABLE_FILE_DEFAULT_VALUE = '/user/config/%s' % HCCL_JSON_FILE_NAME @staticmethod def get_rank_table_template1_file_dir(): parent_dir = os.environ[ModelArts.MA_MOUNT_PATH_ENV] return os.path.join(parent_dir, 'rank_table') @staticmethod def get_rank_table_template2_file_path(): rank_table_file_path = os.environ.get(RankTableEnv.RANK_TABLE_FILE) if rank_table_file_path is None: return RankTableEnv.RANK_TABLE_FILE_DEFAULT_VALUE return os.path.join(os.path.normpath(rank_table_file_path), RankTableEnv.HCCL_JSON_FILE_NAME) @staticmethod def set_rank_table_env(path): os.environ[RankTableEnv.RANK_TABLE_FILE] = path @staticmethod def unset_rank_table_env(): del os.environ[RankTableEnv.RANK_TABLE_FILE] class ModelArts: MA_MOUNT_PATH_ENV = 'MA_MOUNT_PATH' MA_CURRENT_INSTANCE_NAME_ENV = 'MA_CURRENT_INSTANCE_NAME' MA_VJ_NAME = 'MA_VJ_NAME' MA_CURRENT_HOST_IP = 'MA_CURRENT_HOST_IP' CACHE_DIR = '/cache' TMP_LOG_DIR = '/tmp/log/' FMK_WORKSPACE = 'workspace' @staticmethod def get_current_instance_name(): return os.environ[ModelArts.MA_CURRENT_INSTANCE_NAME_ENV] @staticmethod def get_current_host_ip(): return os.environ.get(ModelArts.MA_CURRENT_HOST_IP) @staticmethod def get_job_id(): ma_vj_name = os.environ[ModelArts.MA_VJ_NAME] return ma_vj_name.replace('ma-job', 'modelarts-job', 1) @staticmethod def get_parent_working_dir(): if ModelArts.MA_MOUNT_PATH_ENV in os.environ: return os.path.join(os.environ.get(ModelArts.MA_MOUNT_PATH_ENV), ModelArts.FMK_WORKSPACE) return ModelArts.CACHE_DIR class RunAscendLog: @staticmethod def setup_run_ascend_logger(): name = logo formatter = logging.Formatter(fmt='[run ascend] %(asctime)s - %(levelname)s - %(message)s') handler = logging.StreamHandler() handler.setFormatter(formatter) logger = logging.getLogger(name) logger.setLevel(logging.INFO) logger.addHandler(handler) logger.propagate = False return logger @staticmethod def get_run_ascend_logger(): return logging.getLogger(logo)
- rank_table.py
import json import time import os from common import ModelArts from common import RunAscendLog from common import RankTableEnv log = RunAscendLog.get_run_ascend_logger() class Device: def __init__(self, device_id, device_ip, rank_id): self.device_id = device_id self.device_ip = device_ip self.rank_id = rank_id class Instance: def __init__(self, pod_name, server_id, devices): self.pod_name = pod_name self.server_id = server_id self.devices = self.parse_devices(devices) @staticmethod def parse_devices(devices): if devices is None: return [] device_object_list = [] for device in devices: device_object_list.append(Device(device['device_id'], device['device_ip'], '')) return device_object_list def set_devices(self, devices): self.devices = devices class Group: def __init__(self, group_name, device_count, instance_count, instance_list): self.group_name = group_name self.device_count = int(device_count) self.instance_count = int(instance_count) self.instance_list = self.parse_instance_list(instance_list) @staticmethod def parse_instance_list(instance_list): instance_object_list = [] for instance in instance_list: instance_object_list.append( Instance(instance['pod_name'], instance['server_id'], instance['devices'])) return instance_object_list class RankTable: STATUS_FIELD = 'status' COMPLETED_STATUS = 'completed' def __init__(self): self.rank_table_path = "" self.rank_table = {} @staticmethod def read_from_file(file_path): with open(file_path) as json_file: return json.load(json_file) @staticmethod def wait_for_available(rank_table_file, period=1): log.info('Wait for Rank table file at %s ready' % rank_table_file) complete_flag = False while not complete_flag: with open(rank_table_file) as json_file: data = json.load(json_file) if data[RankTable.STATUS_FIELD] == RankTable.COMPLETED_STATUS: log.info('Rank table file is ready for read') log.info('\n' + json.dumps(data, indent=4)) return True time.sleep(period) return False @staticmethod def convert_server_to_instance(server): device_list = [] for device in server['device']: device_list.append( Device(device_id=device['device_id'], device_ip=device['device_ip'], rank_id=device['rank_id'])) ins = Instance(pod_name='', server_id=server['server_id'], devices=[]) ins.set_devices(device_list) return ins def get_rank_table_path(self): return self.rank_table_path def get_server(self, server_id): for server in self.rank_table['server_list']: if server['server_id'] == server_id: log.info('Current server') log.info('\n' + json.dumps(server, indent=4)) return server log.error('server [%s] is not found' % server_id) return None class RankTableTemplate2(RankTable): def __init__(self, rank_table_template2_path): super().__init__() json_data = self.read_from_file(file_path=rank_table_template2_path) self.status = json_data[RankTableTemplate2.STATUS_FIELD] if self.status != RankTableTemplate2.COMPLETED_STATUS: return # sorted instance list by the index of instance # assert there is only one group json_data["group_list"][0]["instance_list"] = sorted(json_data["group_list"][0]["instance_list"], key=RankTableTemplate2.get_index) self.group_count = int(json_data['group_count']) self.group_list = self.parse_group_list(json_data['group_list']) self.rank_table_path, self.rank_table = self.convert_template2_to_template1_format_file() @staticmethod def parse_group_list(group_list): group_object_list = [] for group in group_list: group_object_list.append( Group(group['group_name'], group['device_count'], group['instance_count'], group['instance_list'])) return group_object_list @staticmethod def get_index(instance): # pod_name example: job94dc1dbf-job-bj4-yolov4-15 pod_name = instance["pod_name"] return int(pod_name[pod_name.rfind("-") + 1:]) def get_current_instance(self): """ get instance by pod name specially, return the first instance when the pod name is None :return: """ pod_name = ModelArts.get_current_instance_name() if pod_name is None: if len(self.group_list) > 0: if len(self.group_list[0].instance_list) > 0: return self.group_list[0].instance_list[0] return None for group in self.group_list: for instance in group.instance_list: if instance.pod_name == pod_name: return instance return None def convert_template2_to_template1_format_file(self): rank_table_template1_file = { 'status': 'completed', 'version': '1.0', 'server_count': '0', 'server_list': [] } logic_index = 0 server_map = {} # collect all devices in all groups for group in self.group_list: if group.device_count == 0: continue for instance in group.instance_list: if instance.server_id not in server_map: server_map[instance.server_id] = [] for device in instance.devices: template1_device = { 'device_id': device.device_id, 'device_ip': device.device_ip, 'rank_id': str(logic_index) } logic_index += 1 server_map[instance.server_id].append(template1_device) server_count = 0 for server_id in server_map: rank_table_template1_file['server_list'].append({ 'server_id': server_id, 'device': server_map[server_id] }) server_count += 1 rank_table_template1_file['server_count'] = str(server_count) log.info('Rank table file (Template1)') log.info('\n' + json.dumps(rank_table_template1_file, indent=4)) if not os.path.exists(RankTableEnv.get_rank_table_template1_file_dir()): os.makedirs(RankTableEnv.get_rank_table_template1_file_dir()) path = os.path.join(RankTableEnv.get_rank_table_template1_file_dir(), RankTableEnv.HCCL_JSON_FILE_NAME) with open(path, 'w') as f: f.write(json.dumps(rank_table_template1_file)) log.info('Rank table file (Template1) is generated at %s', path) return path, rank_table_template1_file def get_device_num(self): total_device_num = 0 for group in self.group_list: total_device_num += group.device_count return total_device_num class RankTableTemplate1(RankTable): def __init__(self, rank_table_template1_path): super().__init__() self.rank_table_path = rank_table_template1_path self.rank_table = self.read_from_file(file_path=rank_table_template1_path) def get_current_instance(self): current_server = None server_list = self.rank_table['server_list'] if len(server_list) == 1: current_server = server_list[0] elif len(server_list) > 1: host_ip = ModelArts.get_current_host_ip() if host_ip is not None: for server in server_list: if server['server_id'] == host_ip: current_server = server break else: current_server = server_list[0] if current_server is None: log.error('server is not found') return None return self.convert_server_to_instance(current_server) def get_device_num(self): server_list = self.rank_table['server_list'] device_num = 0 for server in server_list: device_num += len(server['device']) return device_num
- manager.py
import time import os import os.path import signal from common import RunAscendLog from fmk import FMK log = RunAscendLog.get_run_ascend_logger() class FMKManager: # max destroy time: ~20 (15 + 5) # ~ 15 (1 + 2 + 4 + 8) MAX_TEST_PROC_CNT = 4 def __init__(self, instance): self.instance = instance self.fmk = [] self.fmk_processes = [] self.get_sigterm = False self.max_test_proc_cnt = FMKManager.MAX_TEST_PROC_CNT # break the monitor and destroy processes when get terminate signal def term_handle(func): def receive_term(signum, stack): log.info('Received terminate signal %d, try to destroyed all processes' % signum) stack.f_locals['self'].get_sigterm = True def handle_func(self, *args, **kwargs): origin_handle = signal.getsignal(signal.SIGTERM) signal.signal(signal.SIGTERM, receive_term) res = func(self, *args, **kwargs) signal.signal(signal.SIGTERM, origin_handle) return res return handle_func def run(self, rank_size, command): for index, device in enumerate(self.instance.devices): fmk_instance = FMK(index, device) self.fmk.append(fmk_instance) self.fmk_processes.append(fmk_instance.run(rank_size, command)) @term_handle def monitor(self, period=1): # busy waiting for all fmk processes exit by zero # or there is one process exit by non-zero fmk_cnt = len(self.fmk_processes) zero_ret_cnt = 0 while zero_ret_cnt != fmk_cnt: zero_ret_cnt = 0 for index in range(fmk_cnt): fmk = self.fmk[index] fmk_process = self.fmk_processes[index] if fmk_process.poll() is not None: if fmk_process.returncode != 0: log.error('proc-rank-%s-device-%s (pid: %d) has exited with non-zero code: %d' % (fmk.rank_id, fmk.device_id, fmk_process.pid, fmk_process.returncode)) return fmk_process.returncode zero_ret_cnt += 1 if self.get_sigterm: break time.sleep(period) return 0 def destroy(self, base_period=1): log.info('Begin destroy training processes') self.send_sigterm_to_fmk_process() self.wait_fmk_process_end(base_period) log.info('End destroy training processes') def send_sigterm_to_fmk_process(self): # send SIGTERM to fmk processes (and process group) for r_index in range(len(self.fmk_processes) - 1, -1, -1): fmk = self.fmk[r_index] fmk_process = self.fmk_processes[r_index] if fmk_process.poll() is not None: log.info('proc-rank-%s-device-%s (pid: %d) has exited before receiving the term signal', fmk.rank_id, fmk.device_id, fmk_process.pid) del self.fmk_processes[r_index] del self.fmk[r_index] try: os.killpg(fmk_process.pid, signal.SIGTERM) except ProcessLookupError: pass def wait_fmk_process_end(self, base_period): test_cnt = 0 period = base_period while len(self.fmk_processes) > 0 and test_cnt < self.max_test_proc_cnt: for r_index in range(len(self.fmk_processes) - 1, -1, -1): fmk = self.fmk[r_index] fmk_process = self.fmk_processes[r_index] if fmk_process.poll() is not None: log.info('proc-rank-%s-device-%s (pid: %d) has exited', fmk.rank_id, fmk.device_id, fmk_process.pid) del self.fmk_processes[r_index] del self.fmk[r_index] if not self.fmk_processes: break time.sleep(period) period *= 2 test_cnt += 1 if len(self.fmk_processes) > 0: for r_index in range(len(self.fmk_processes) - 1, -1, -1): fmk = self.fmk[r_index] fmk_process = self.fmk_processes[r_index] if fmk_process.poll() is None: log.warn('proc-rank-%s-device-%s (pid: %d) has not exited within the max waiting time, ' 'send kill signal', fmk.rank_id, fmk.device_id, fmk_process.pid) os.killpg(fmk_process.pid, signal.SIGKILL)
- fmk.py
import os import subprocess import pathlib from contextlib import contextmanager from common import RunAscendLog from common import RankTableEnv from common import ModelArts log = RunAscendLog.get_run_ascend_logger() class FMK: def __init__(self, index, device): self.job_id = ModelArts.get_job_id() self.rank_id = device.rank_id self.device_id = str(index) def gen_env_for_fmk(self, rank_size): current_envs = os.environ.copy() current_envs['JOB_ID'] = self.job_id current_envs['ASCEND_DEVICE_ID'] = self.device_id current_envs['DEVICE_ID'] = self.device_id current_envs['RANK_ID'] = self.rank_id current_envs['RANK_SIZE'] = str(rank_size) FMK.set_env_if_not_exist(current_envs, RankTableEnv.HCCL_CONNECT_TIMEOUT, str(1800)) log_dir = FMK.get_log_dir() process_log_path = os.path.join(log_dir, self.job_id, 'ascend', 'process_log', 'rank_' + self.rank_id) FMK.set_env_if_not_exist(current_envs, 'ASCEND_PROCESS_LOG_PATH', process_log_path) pathlib.Path(current_envs['ASCEND_PROCESS_LOG_PATH']).mkdir(parents=True, exist_ok=True) return current_envs @contextmanager def switch_directory(self, directory): owd = os.getcwd() try: os.chdir(directory) yield directory finally: os.chdir(owd) def get_working_dir(self): fmk_workspace_prefix = ModelArts.get_parent_working_dir() return os.path.join(os.path.normpath(fmk_workspace_prefix), 'device%s' % self.device_id) @staticmethod def get_log_dir(): parent_path = os.getenv(ModelArts.MA_MOUNT_PATH_ENV) if parent_path: log_path = os.path.join(parent_path, 'log') if os.path.exists(log_path): return log_path return ModelArts.TMP_LOG_DIR @staticmethod def set_env_if_not_exist(envs, env_name, env_value): if env_name in os.environ: log.info('env already exists. env_name: %s, env_value: %s ' % (env_name, env_value)) return envs[env_name] = env_value def run(self, rank_size, command): envs = self.gen_env_for_fmk(rank_size) log.info('bootstrap proc-rank-%s-device-%s' % (self.rank_id, self.device_id)) log_dir = FMK.get_log_dir() if not os.path.exists(log_dir): os.makedirs(log_dir) log_file = '%s-proc-rank-%s-device-%s.txt' % (self.job_id, self.rank_id, self.device_id) log_file_path = os.path.join(log_dir, log_file) working_dir = self.get_working_dir() if not os.path.exists(working_dir): os.makedirs(working_dir) with self.switch_directory(working_dir): # os.setsid: change the process(forked) group id to itself training_proc = subprocess.Popen(command, env=envs, preexec_fn=os.setsid, stdout=subprocess.PIPE, stderr=subprocess.STDOUT) log.info('proc-rank-%s-device-%s (pid: %d)', self.rank_id, self.device_id, training_proc.pid) # https://docs.python.org/3/library/subprocess.html#subprocess.Popen.wait subprocess.Popen(['tee', log_file_path], stdin=training_proc.stdout) return training_proc