lightning/pytorch_lightning/trainer/connectors/slurm_connector.py

152 lines
5.3 KiB
Python

import os
import re
import signal
from subprocess import call
from pytorch_lightning import _logger as log
from pytorch_lightning.utilities.distributed import rank_zero_info
import torch.distributed as torch_distrib
import torch
class SLURMConnector:
def __init__(self, trainer):
self.trainer = trainer
def on_trainer_init(self, num_gpu_nodes):
self.configure_slurm_ddp(num_gpu_nodes)
def configure_slurm_ddp(self, num_gpu_nodes):
self.trainer.is_slurm_managing_tasks = False
# extract SLURM flag vars
# whenever we have the correct number of tasks, we let slurm manage processes
# otherwise we launch the required number of processes
if self.trainer.use_ddp:
self.trainer.num_requested_gpus = self.trainer.num_gpus * num_gpu_nodes
self.trainer.num_slurm_tasks = 0
try:
self.trainer.num_slurm_tasks = int(os.environ['SLURM_NTASKS'])
self.trainer.is_slurm_managing_tasks = self.trainer.num_slurm_tasks == self.trainer.num_requested_gpus
# in interactive mode we don't manage tasks
job_name = os.environ['SLURM_JOB_NAME']
if job_name == 'bash':
self.trainer.is_slurm_managing_tasks = False
except Exception:
# likely not on slurm, so set the slurm managed flag to false
self.trainer.is_slurm_managing_tasks = False
# used for tests only, set this flag to simulate slurm managing a task
try:
should_fake = int(os.environ['FAKE_SLURM_MANAGING_TASKS'])
if should_fake:
self.trainer.is_slurm_managing_tasks = True
except Exception:
pass
# notify user the that slurm is managing tasks
if self.trainer.is_slurm_managing_tasks:
rank_zero_info('Multi-processing is handled by Slurm.')
def resolve_root_node_address(self, root_node):
if '[' in root_node:
name, numbers = root_node.split('[', maxsplit=1)
number = numbers.split(',', maxsplit=1)[0]
if '-' in number:
number = number.split('-')[0]
number = re.sub('[^0-9]', '', number)
root_node = name + number
return root_node
def register_slurm_signal_handlers(self):
# see if we're using slurm (not interactive)
on_slurm = False
try:
job_name = os.environ['SLURM_JOB_NAME']
if job_name != 'bash':
on_slurm = True
except Exception:
pass
if on_slurm:
log.info('Set SLURM handle signals.')
signal.signal(signal.SIGUSR1, self.sig_handler)
signal.signal(signal.SIGTERM, self.term_handler)
def sig_handler(self, signum, frame): # pragma: no-cover
if self.trainer.is_global_zero:
# save weights
log.info('handling SIGUSR1')
self.trainer.hpc_save(self.trainer.weights_save_path, self.trainer.logger)
# find job id
job_id = os.environ['SLURM_JOB_ID']
cmd = ['scontrol', 'requeue', job_id]
# requeue job
log.info(f'requeing job {job_id}...')
result = call(cmd)
# print result text
if result == 0:
log.info(f'requeued exp {job_id}')
else:
log.warning('requeue failed...')
# close experiment to avoid issues
self.trainer.logger.close()
def term_handler(self, signum, frame):
# save
log.info("bypassing sigterm")
def connect_ddp(self, global_rank: int, world_size: int) -> None:
""""""
"""
Sets up environment variables necessary for pytorch distributed communications
based on slurm environment.
"""
# use slurm job id for the port number
# guarantees unique ports across jobs from same grid search
try:
# use the last 4 numbers in the job id as the id
default_port = os.environ["SLURM_JOB_ID"]
default_port = default_port[-4:]
# all ports should be in the 10k+ range
default_port = int(default_port) + 15000
except Exception:
default_port = 12910
# if user gave a port number, use that one instead
try:
default_port = os.environ["MASTER_PORT"]
except Exception:
os.environ["MASTER_PORT"] = str(default_port)
log.debug(f"MASTER_PORT: {os.environ['MASTER_PORT']}")
# figure out the root node addr
try:
root_node = os.environ["SLURM_NODELIST"].split(" ")[0]
except Exception:
root_node = "127.0.0.1"
root_node = self.trainer.slurm_connector.resolve_root_node_address(root_node)
os.environ["MASTER_ADDR"] = root_node
log.debug(f"MASTER_ADDR: {os.environ['MASTER_ADDR']}")
torch_backend = "nccl" if self.trainer.on_gpu else "gloo"
if not torch.distributed.is_initialized():
log.info(
f"initializing ddp (SLURM): GLOBAL_RANK: {global_rank}, MEMBER: {global_rank + 1}/{world_size}"
)
torch_distrib.init_process_group(
torch_backend, rank=global_rank, world_size=world_size
)