mirror of
https://github.com/tencentmusic/cube-studio.git
synced 2025-03-07 15:08:51 +08:00
343 lines
13 KiB
Python
343 lines
13 KiB
Python
|
||
import os,sys
|
||
base_dir = os.path.split(os.path.realpath(__file__))[0]
|
||
sys.path.append(base_dir)
|
||
|
||
import argparse
|
||
import datetime
|
||
import json
|
||
import time
|
||
import uuid
|
||
import os
|
||
import pysnooper
|
||
import os,sys
|
||
import re
|
||
import threading
|
||
import psutil
|
||
import copy
|
||
|
||
from kubernetes import client
|
||
|
||
# print(os.environ)
|
||
from job.pkgs.k8s.py_k8s import K8s
|
||
k8s_client = K8s()
|
||
|
||
KFJ_NAMESPACE = os.getenv('KFJ_NAMESPACE', '')
|
||
KFJ_TASK_ID = os.getenv('KFJ_TASK_ID', '')
|
||
KFJ_TASK_NAME = os.getenv('KFJ_TASK_NAME', '')
|
||
task_node_selectors = re.split(',|;|\n|\t', os.getenv('KFJ_TASK_NODE_SELECTOR', 'cpu=true,train=true'))
|
||
KFJ_TASK_NODE_SELECTOR = {}
|
||
for task_node_selector in task_node_selectors:
|
||
KFJ_TASK_NODE_SELECTOR[task_node_selector.split('=')[0]] = task_node_selector.split('=')[1]
|
||
|
||
KFJ_PIPELINE_ID = os.getenv('KFJ_PIPELINE_ID', '')
|
||
KFJ_RUN_ID = os.getenv('KFJ_RUN_ID', '')
|
||
KFJ_CREATOR = os.getenv('KFJ_CREATOR', '')
|
||
KFJ_RUNNER = os.getenv('KFJ_RUNNER','')
|
||
KFJ_PIPELINE_NAME = os.getenv('KFJ_PIPELINE_NAME', '')
|
||
KFJ_TASK_IMAGES = os.getenv('KFJ_TASK_IMAGES', '')
|
||
KFJ_TASK_VOLUME_MOUNT = os.getenv('KFJ_TASK_VOLUME_MOUNT', '')
|
||
KFJ_TASK_RESOURCE_CPU = os.getenv('KFJ_TASK_RESOURCE_CPU', '')
|
||
KFJ_TASK_RESOURCE_MEMORY = os.getenv('KFJ_TASK_RESOURCE_MEMORY', '')
|
||
NUM_WORKER = 3
|
||
INIT_FILE=''
|
||
crd_info={
|
||
"group": "kubeflow.org",
|
||
"version": "v1",
|
||
'kind': 'PyTorchJob',
|
||
"plural": "pytorchjobs",
|
||
"timeout": 60 * 60 * 24 * 2
|
||
}
|
||
|
||
|
||
k8s_volumes, k8s_volume_mounts = k8s_client.get_volume_mounts(KFJ_TASK_VOLUME_MOUNT,KFJ_CREATOR)
|
||
|
||
print(k8s_volumes)
|
||
print(k8s_volume_mounts)
|
||
|
||
GPU_TYPE= os.getenv('KFJ_GPU_TYPE', 'NVIDIA')
|
||
GPU_RESOURCE= os.getenv('KFJ_TASK_RESOURCE_GPU', '0')
|
||
print(GPU_TYPE,GPU_RESOURCE)
|
||
schedulerName = os.getenv('SCHEDULER', 'default-scheduler')
|
||
|
||
|
||
|
||
def default_job_name():
|
||
name = "pytorchjob-" + KFJ_PIPELINE_NAME.replace('_','-')+"-"+uuid.uuid4().hex[:4]
|
||
return name[0:54]
|
||
|
||
|
||
|
||
|
||
import subprocess
|
||
# @pysnooper.snoop()
|
||
def run_shell(shell):
|
||
print('begin run shell: %s'%shell,flush=True)
|
||
cmd = subprocess.Popen(shell, stdin=subprocess.PIPE, stderr=subprocess.PIPE,
|
||
stdout=subprocess.PIPE, universal_newlines=True, shell=True, bufsize=1)
|
||
# 实时输出
|
||
while True:
|
||
line = cmd.stdout.readline()
|
||
status = subprocess.Popen.poll(cmd)
|
||
if status:
|
||
print(status,line,end='', flush=True)
|
||
else:
|
||
print(line, end='', flush=True)
|
||
if status == 0: # 判断子进程是否结束
|
||
print('shell finish %s'%status,flush=True)
|
||
break
|
||
if status==-9 or status==-15 or status==143: # 外界触发kill
|
||
print('shell finish %s'%status,flush=True)
|
||
break
|
||
|
||
return cmd.returncode
|
||
|
||
|
||
|
||
|
||
|
||
|
||
# 监控指定名称的pytorchjob
|
||
def monitoring(crd_k8s,name,namespace):
|
||
time.sleep(10)
|
||
# 杀掉stern 进程
|
||
def get_pid(name):
|
||
'''
|
||
作用：根据进程名获取进程pid
|
||
'''
|
||
pids = psutil.process_iter()
|
||
print("[" + name + "]'s pid is:", flush=True)
|
||
back=[]
|
||
for pid in pids:
|
||
if name in pid.name():
|
||
print(pid.pid, flush=True)
|
||
back.append(pid.pid)
|
||
return back
|
||
check_time = datetime.datetime.now()
|
||
while(True):
|
||
pytorchjob = crd_k8s.get_one_crd(group=crd_info['group'],version=crd_info['version'],plural=crd_info['plural'],namespace=namespace,name=name)
|
||
if pytorchjob:
|
||
print('pytorchjob status %s'%pytorchjob['status'], flush=True)
|
||
else:
|
||
print('pytorchjob not exist', flush=True)
|
||
if pytorchjob and (pytorchjob['status']=="Succeeded" or pytorchjob['status']=="Failed"): # Created, Running, Restarting, Succeeded, or Failed
|
||
pids = get_pid("stern")
|
||
if pids:
|
||
for pid in pids:
|
||
pro = psutil.Process(int(pid))
|
||
pro.terminate()
|
||
print('kill process %s'%pid, flush=True)
|
||
break
|
||
if (datetime.datetime.now()-check_time).seconds>3600:
|
||
pids = get_pid("stern")
|
||
if pids:
|
||
for pid in pids:
|
||
pro = psutil.Process(int(pid))
|
||
pro.terminate()
|
||
print('kill process %s'%pid, flush=True)
|
||
check_time=datetime.datetime.now()
|
||
time.sleep(60)
|
||
|
||
|
||
|
||
# @pysnooper.snoop()
|
||
def make_pytorchjob(name,num_workers,image,working_dir,command):
|
||
# if type(command)==str:
|
||
# command=command.split(" ")
|
||
# command = [c for c in command if c]
|
||
pod_spec={
|
||
"replicas": 1,
|
||
"restartPolicy": "Never",
|
||
"template": {
|
||
"metadata": {
|
||
"labels": {
|
||
"pipeline-id": KFJ_PIPELINE_ID,
|
||
"pipeline-name": KFJ_PIPELINE_NAME,
|
||
"task-id": KFJ_TASK_ID,
|
||
"task-name": KFJ_TASK_NAME,
|
||
'rtx-user': KFJ_RUNNER,
|
||
"component": name,
|
||
"type": "pytorchjob",
|
||
"run-id": KFJ_RUN_ID,
|
||
}
|
||
},
|
||
"spec": {
|
||
"schedulerName": schedulerName,
|
||
"restartPolicy": "Never",
|
||
"volumes": k8s_volumes,
|
||
# "imagePullSecrets": [
|
||
# {
|
||
# "name": "hubsecret"
|
||
# }
|
||
# ],
|
||
"affinity": {
|
||
"nodeAffinity": {
|
||
"requiredDuringSchedulingIgnoredDuringExecution": {
|
||
"nodeSelectorTerms": [
|
||
{
|
||
"matchExpressions": [
|
||
{
|
||
"key": node_selector_key,
|
||
"operator": "In",
|
||
"values": [
|
||
KFJ_TASK_NODE_SELECTOR[node_selector_key]
|
||
]
|
||
} for node_selector_key in KFJ_TASK_NODE_SELECTOR
|
||
]
|
||
}
|
||
]
|
||
}
|
||
},
|
||
"podAntiAffinity": {
|
||
"preferredDuringSchedulingIgnoredDuringExecution": [
|
||
{
|
||
"weight": 5,
|
||
"podAffinityTerm": {
|
||
"topologyKey": "kubernetes.io/hostname",
|
||
"labelSelector": {
|
||
"matchLabels": {
|
||
"component": name,
|
||
"type": "pytorchjob"
|
||
}
|
||
}
|
||
}
|
||
}
|
||
]
|
||
}
|
||
},
|
||
"containers": [
|
||
{
|
||
"name": "pytorch",
|
||
"image": image if image else KFJ_TASK_IMAGES,
|
||
"imagePullPolicy": "Always",
|
||
"workingDir":working_dir,
|
||
"env":[
|
||
{
|
||
"name": "NCCL_DEBUG",
|
||
"value":"INFO"
|
||
},
|
||
{
|
||
"name": "NCCL_IB_DISABLE",
|
||
"value": "1"
|
||
},
|
||
# {
|
||
# "name": "NCCL_DEBUG_SUBSYS",
|
||
# "value": "ALL"
|
||
# },
|
||
{
|
||
"name": "NCCL_SOCKET_IFNAME",
|
||
"value": "eth0"
|
||
}
|
||
],
|
||
"command": ['bash','-c',command],
|
||
"volumeMounts": k8s_volume_mounts,
|
||
"resources": {
|
||
"requests": {
|
||
"cpu": KFJ_TASK_RESOURCE_CPU,
|
||
"memory": KFJ_TASK_RESOURCE_MEMORY,
|
||
},
|
||
"limits": {
|
||
"cpu": KFJ_TASK_RESOURCE_CPU,
|
||
"memory": KFJ_TASK_RESOURCE_MEMORY
|
||
}
|
||
}
|
||
}
|
||
]
|
||
}
|
||
}
|
||
}
|
||
|
||
|
||
if GPU_TYPE=='NVIDIA' and GPU_RESOURCE:
|
||
pod_spec['template']['spec']['containers'][0]['resources']['requests']['nvidia.com/gpu'] = GPU_RESOURCE.split(',')[0]
|
||
pod_spec['template']['spec']['containers'][0]['resources']['limits']['nvidia.com/gpu'] = GPU_RESOURCE.split(',')[0]
|
||
|
||
worker_pod_spec = copy.deepcopy(pod_spec)
|
||
worker_pod_spec['replicas']=int(num_workers)-1 # 因为master是其中一个worker
|
||
|
||
pytorch_deploy = {
|
||
"apiVersion": "kubeflow.org/v1",
|
||
"kind": "PyTorchJob",
|
||
"metadata": {
|
||
"namespace": KFJ_NAMESPACE,
|
||
"name": name,
|
||
"labels":{
|
||
"run-id":KFJ_RUN_ID,
|
||
"run-rtx":KFJ_RUNNER,
|
||
"pipeline-rtx": KFJ_CREATOR,
|
||
"pipeline-id": KFJ_PIPELINE_ID,
|
||
"pipeline-name": KFJ_PIPELINE_NAME,
|
||
"task-id": KFJ_TASK_ID,
|
||
"task-name": KFJ_TASK_NAME,
|
||
}
|
||
},
|
||
"spec": {
|
||
"backoffLimit":num_workers,
|
||
"cleanPodPolicy": "None",
|
||
"pytorchReplicaSpecs": {
|
||
"Master":pod_spec,
|
||
"Worker":worker_pod_spec
|
||
}
|
||
|
||
}
|
||
}
|
||
|
||
return pytorch_deploy
|
||
|
||
|
||
# @pysnooper.snoop()
|
||
def launch_pytorchjob(name, num_workers, image,working_dir, worker_command):
|
||
if KFJ_RUN_ID:
|
||
print('delete old pytorch, run-id %s'%KFJ_RUN_ID, flush=True)
|
||
k8s_client.delete_crd(group=crd_info['group'],version=crd_info['version'],plural=crd_info['plural'],namespace=KFJ_NAMESPACE,labels={"run-id":KFJ_RUN_ID})
|
||
time.sleep(10)
|
||
# 删除旧的pytorch
|
||
k8s_client.delete_crd(group=crd_info['group'], version=crd_info['version'], plural=crd_info['plural'],namespace=KFJ_NAMESPACE, name=name)
|
||
time.sleep(10)
|
||
# 创建新的pytorch
|
||
pytorchjob_json = make_pytorchjob(name=name,num_workers= num_workers,image = image,working_dir=working_dir,command=worker_command)
|
||
print('create new pytorch %s' % name, flush=True)
|
||
k8s_client.create_crd(group=crd_info['group'],version=crd_info['version'],plural=crd_info['plural'],namespace=KFJ_NAMESPACE,body=pytorchjob_json)
|
||
time.sleep(10)
|
||
|
||
print('begin start monitoring thread', flush=True)
|
||
# # 后台启动监控脚本
|
||
monitoring_thread = threading.Thread(target=monitoring,args=(k8s_client,name,KFJ_NAMESPACE))
|
||
monitoring_thread.start()
|
||
while True:
|
||
# 实时打印日志
|
||
line='>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>'
|
||
print('begin follow log\n%s'%line, flush=True)
|
||
command = '''stern %s --namespace %s --exclude-container init-pytorch --since 10s --template '{{.PodName}} {{.Message}} {{"\\n"}}' '''%(name,KFJ_NAMESPACE)
|
||
|
||
print(command, flush=True)
|
||
run_shell(command)
|
||
print('%s\nend follow log'%line, flush=True)
|
||
time.sleep(10)
|
||
|
||
pytorchjob = k8s_client.get_one_crd(group=crd_info['group'], version=crd_info['version'],plural=crd_info['plural'], namespace=KFJ_NAMESPACE, name=name)
|
||
if pytorchjob and (pytorchjob['status'] == "Succeeded" or pytorchjob['status'] == "Failed"):
|
||
break
|
||
|
||
pytorchjob = k8s_client.get_one_crd(group=crd_info['group'],version=crd_info['version'],plural=crd_info['plural'],namespace=KFJ_NAMESPACE,name=name)
|
||
print("pytorchJob %s finished, status %s"%(name, pytorchjob['status']))
|
||
|
||
if pytorchjob['status']!='Succeeded':
|
||
exit(1)
|
||
|
||
|
||
if __name__ == "__main__":
|
||
arg_parser = argparse.ArgumentParser("Pytorchjob launcher")
|
||
arg_parser.add_argument('--working_dir', type=str, help="运行job的工作目录", default='/mnt/')
|
||
arg_parser.add_argument('--command', type=str, help="运行job的命令", default='python3 mnist.py')
|
||
arg_parser.add_argument('--num_worker', type=int, help="分布式worker的数量", default=3)
|
||
arg_parser.add_argument('--image', type=str, help="运行job的镜像", default='')
|
||
|
||
args = arg_parser.parse_args()
|
||
print("{} args: {}".format(__file__, args))
|
||
|
||
|
||
launch_pytorchjob(name=default_job_name(),num_workers=args.num_worker,image=args.image,working_dir=args.working_dir,worker_command=args.command)
|
||
|
||
|