distrib_data_parallel.py

"""
Lightning supports model training on a cluster managed by SLURM in the following cases:

1. Training on a single cpu or single GPU.
2. Train on multiple GPUs on the same node using DataParallel or DistributedDataParallel
3. Training across multiple GPUs on multiple different nodes via DistributedDataParallel.

.. note:: A node means a machine with multiple GPUs

Running grid search on a cluster
--------------------------------

To use lightning to run a hyperparameter search (grid-search or random-search) on a cluster do 4 things:

(1). Define the parameters for the grid search

.. code-block:: python

    from test_tube import HyperOptArgumentParser

    # subclass of argparse
    parser = HyperOptArgumentParser(strategy='random_search')
    parser.add_argument('--learning_rate', default=0.002, type=float, help='the learning rate')

    # let's enable optimizing over the number of layers in the network
    parser.opt_list('--nb_layers', default=2, type=int, tunable=True, options=[2, 4, 8])

    hparams = parser.parse_args()

.. note:: You must set `Tunable=True` for that argument to be considered in the permutation set.
 Otherwise test-tube will use the default value. This flag is useful when you don't want
 to search over an argument and want to use the default instead.

(2). Define the cluster options in the
 `SlurmCluster object <https://williamfalcon.github.io/test-tube/hpc/SlurmCluster>`_ (over 5 nodes and 8 gpus)

.. code-block:: python

    from test_tube.hpc import SlurmCluster

    # hyperparameters is a test-tube hyper params object
    # see https://williamfalcon.github.io/test-tube/hyperparameter_optimization/HyperOptArgumentParser/
    hyperparams = args.parse()

    # init cluster
    cluster = SlurmCluster(
        hyperparam_optimizer=hyperparams,
        log_path='/path/to/log/results/to',
        python_cmd='python3'
    )

    # let the cluster know where to email for a change in job status (ie: complete, fail, etc...)
    cluster.notify_job_status(email='some@email.com', on_done=True, on_fail=True)

    # set the job options. In this instance, we'll run 20 different models
    # each with its own set of hyperparameters giving each one 1 GPU (ie: taking up 20 GPUs)
    cluster.per_experiment_nb_gpus = 8
    cluster.per_experiment_nb_nodes = 5

    # we'll request 10GB of memory per node
    cluster.memory_mb_per_node = 10000

    # set a walltime of 10 minues
    cluster.job_time = '10:00'


(3). Make a main function with your model and trainer. Each job will call this function with a particular
hparams configuration.::

    from pytorch_lightning import Trainer

    def train_fx(trial_hparams, cluster_manager, _):
        # hparams has a specific set of hyperparams

        my_model = MyLightningModel()

        # give the trainer the cluster object
        trainer = Trainer()
        trainer.fit(my_model)

    `

(4). Start the grid/random search::

    # run the models on the cluster
    cluster.optimize_parallel_cluster_gpu(
        train_fx,
        nb_trials=20,
        job_name='my_grid_search_exp_name',
        job_display_name='my_exp')

.. note:: `nb_trials` specifies how many of the possible permutations to use. If using `grid_search` it will use
 the depth first ordering. If using `random_search` it will use the first k shuffled options. FYI, random search
 has been shown to be just as good as any Bayesian optimization method when using a reasonable number of samples (60),
 see this `paper <http://www.jmlr.org/papers/volume13/bergstra12a/bergstra12a.pdf>`_  for more information.

Walltime auto-resubmit
----------------------

Lightning automatically resubmits jobs when they reach the walltime. Make sure to set the SIGUSR1 signal in
your SLURM script.::

    # 90 seconds before training ends
    #SBATCH --signal=SIGUSR1@90

When lightning receives the SIGUSR1 signal it will:
1. save a checkpoint with 'hpc_ckpt' in the name.
2. resubmit the job using the SLURM_JOB_ID

When the script starts again, Lightning will:
1. search for a 'hpc_ckpt' checkpoint.
2. restore the model, optimizers, schedulers, epoch, etc...

"""

import os
import re
from abc import ABC, abstractmethod
from typing import Union, List, Optional, Callable, Tuple
import subprocess
import sys
from time import sleep
import numpy as np
from os.path import abspath
from torch import distributed as dist
import queue

import torch
from pytorch_lightning import _logger as log
from pytorch_lightning.callbacks import ModelCheckpoint
from pytorch_lightning.loggers import LightningLoggerBase
from pytorch_lightning.utilities import NATIVE_AMP_AVALAIBLE
from pytorch_lightning.utilities.exceptions import MisconfigurationException
from pytorch_lightning.utilities.distributed import rank_zero_only, rank_zero_warn, rank_zero_info

try:
    from apex import amp
except ImportError:
    APEX_AVAILABLE = False
else:
    APEX_AVAILABLE = True

try:
    import horovod.torch as hvd
except (ModuleNotFoundError, ImportError):
    HOROVOD_AVAILABLE = False
else:
    HOROVOD_AVAILABLE = True


try:
    from hydra.utils import to_absolute_path
except ImportError:
    HYDRA_AVAILABLE = False
else:
    HYDRA_AVAILABLE = True


try:
    import torch_xla
    import torch_xla.core.xla_model as xm
    import torch_xla.distributed.xla_multiprocessing as xmp
except ImportError:
    XLA_AVAILABLE = False
else:
    XLA_AVAILABLE = True

pid = os.getpid()
rng1 = np.random.RandomState(pid)
RANDOM_PORTS = rng1.randint(10000, 19999, 100)


class TrainerDDPMixin(ABC):

    # this is just a summary on variables used in this abstract class,
    #  the proper values/initialisation should be done in child class
    on_gpu: bool
    num_gpu_nodes: int
    gpus: List[int]
    logger: Union[LightningLoggerBase, bool]
    data_parallel_device_ids: ...
    distributed_backend: Optional[str]
    amp_level: str
    use_tpu: bool
    default_root_dir: str
    progress_bar_callback: ...
    checkpoint_callback: ...
    num_processes: int
    num_nodes: int
    node_rank: int
    tpu_cores: int
    testing: bool

    @property
    @abstractmethod
    def is_global_zero(self) -> bool:
        """Warning: this is just empty shell for code implemented in other class."""

    @property
    @abstractmethod
    def num_gpus(self) -> int:
        """Warning: this is just empty shell for code implemented in other class."""

    @property
    @abstractmethod
    def use_amp(self) -> bool:
        """Warning: this is just empty shell for code implemented in other class."""

    @abstractmethod
    def copy_trainer_model_properties(self, *args):
        """Warning: this is just empty shell for code implemented in other class."""

    @abstractmethod
    def run_pretrain_routine(self, *args):
        """Warning: this is just empty shell for code implemented in other class."""

    @abstractmethod
    def init_optimizers(self, *args) -> Tuple[List, List, List]:
        """Warning: this is just empty shell for code implemented in other class."""

    @abstractmethod
    def reinit_scheduler_properties(self, *args):
        """Warning: this is just empty shell for code implemented in other class."""

    @abstractmethod
    def save_checkpoint(self, *args):
        """Warning: this is just empty shell for code implemented in other class."""

    @abstractmethod
    def setup(self, *args) -> None:
        """Warning: this is just empty shell for code implemented in other class."""

    @abstractmethod
    def is_function_implemented(self, *args) -> bool:
        """Warning: this is just empty shell for code implemented in other class."""

    def init_tpu(self):
        # turn off all the GPU stuff
        self.distributed_backend = None

        # enable tpu
        self.use_tpu = True

    def set_distributed_mode(self, distributed_backend):
        self.use_dp = False
        self.use_ddp = False
        self.use_ddp2 = False
        self.use_horovod = False
        self.single_gpu = False

        if distributed_backend is None:
            if self.has_horovodrun():
                self._set_horovod_backend()
            elif self.num_gpus == 0:
                if self.num_nodes > 1 or self.num_processes > 1:
                    self.use_ddp = True  # ddp_cpu
            elif self.num_gpus == 1:
                self.single_gpu = True
            elif self.num_gpus > 1:
                rank_zero_warn('You requested multiple GPUs but did not specify a backend, e.g.'
                               ' Trainer(distributed_backend=dp) (or ddp, ddp2).'
                               ' Setting distributed_backend=ddp_spawn for you.')
                self.distributed_backend = 'ddp_spawn'
                distributed_backend = 'ddp_spawn'

        if distributed_backend == "dp":
            # do nothing if num_gpus == 0
            if self.num_gpus == 1:
                self.single_gpu = True
                self.use_dp = True
            elif self.num_gpus > 1:
                self.use_dp = True

        elif distributed_backend in ['ddp', 'ddp_spawn']:
            if self.num_gpus == 0:
                if self.num_nodes > 1 or self.num_processes > 1:
                    self.use_ddp = True  # ddp_cpu
            elif self.num_gpus == 1:
                self.single_gpu = True
                self.use_ddp = True
            elif self.num_gpus > 1:
                self.use_ddp = True
                self.num_processes = self.num_gpus

        elif distributed_backend == "ddp2":
            # do nothing if num_gpus == 0
            if self.num_gpus >= 1:
                self.use_ddp2 = True
        elif distributed_backend == "ddp_cpu":
            if self.num_gpus > 0:
                rank_zero_warn('You requested one or more GPUs, but set the backend to `ddp_cpu`.'
                               ' Training will not use GPUs.')
            self.use_ddp = True
            self.data_parallel_device_ids = None
            self.on_gpu = False
        elif distributed_backend == 'horovod':
            self._set_horovod_backend()

        # throw error to force user ddp or ddp2 choice
        if self.num_nodes > 1 and not (self.use_ddp2 or self.use_ddp):
            raise MisconfigurationException(
                'DataParallel does not support num_nodes > 1. Switching to DistributedDataParallel for you. '
                'To silence this warning set distributed_backend=ddp or distributed_backend=ddp2'
            )

        rank_zero_info(f'GPU available: {torch.cuda.is_available()}, used: {self.on_gpu}')
        num_cores = self.tpu_cores if self.tpu_cores is not None else 0
        rank_zero_info(f'TPU available: {XLA_AVAILABLE}, using: {num_cores} TPU cores')

    def configure_slurm_ddp(self, num_gpu_nodes):
        self.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.use_ddp:
            self.num_requested_gpus = self.num_gpus * num_gpu_nodes
            self.num_slurm_tasks = 0
            try:
                self.num_slurm_tasks = int(os.environ['SLURM_NTASKS'])
                self.is_slurm_managing_tasks = self.num_slurm_tasks == self.num_requested_gpus

                # in interactive mode we don't manage tasks
                job_name = os.environ['SLURM_JOB_NAME']
                if job_name == 'bash':
                    self.is_slurm_managing_tasks = False

            except Exception:
                # likely not on slurm, so set the slurm managed flag to false
                self.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.is_slurm_managing_tasks = True
        except Exception:
            pass

        # notify user the that slurm is managing tasks
        if self.is_slurm_managing_tasks:
            rank_zero_info('Multi-processing is handled by Slurm.')

    def determine_local_rank(self):
        if self.is_slurm_managing_tasks:
            return int(os.environ['SLURM_LOCALID'])

        else:
            return int(os.environ.get('LOCAL_RANK', 0))

    def determine_ddp_node_rank(self):
        if self.is_slurm_managing_tasks:
            return int(os.environ['SLURM_NODEID'])

        # torchelastic uses the envvar GROUP_RANK, whereas other systems(?) use NODE_RANK.
        # otherwise use given node rank or default to node rank 0
        env_vars = ['NODE_RANK', 'GROUP_RANK']
        node_ids = [(k, os.environ.get(k, None)) for k in env_vars]
        node_ids = [(k, v) for k, v in node_ids if v is not None]
        if len(node_ids) == 0:
            return 0
        if len(node_ids) > 1:
            log.warning(f"Multiple environment variables ({node_ids}) defined for node rank. "
                        f"Using the first one.")
        k, rank = node_ids.pop()
        rank_zero_info(f"Using environment variable {k} for node rank ({rank}).")
        return int(rank)

    def set_nvidia_flags(self, is_slurm_managing_tasks, data_parallel_device_ids):
        if data_parallel_device_ids is None:
            return

        # set the correct cuda visible devices (using pci order)
        os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"

        # when slurm is managing the task it sets the visible devices
        if not is_slurm_managing_tasks and 'CUDA_VISIBLE_DEVICES' not in os.environ:
            if isinstance(data_parallel_device_ids, int):
                id_str = ','.join(str(x) for x in list(range(data_parallel_device_ids)))
                os.environ["CUDA_VISIBLE_DEVICES"] = id_str
            else:
                gpu_str = ','.join([str(x) for x in data_parallel_device_ids])
                os.environ["CUDA_VISIBLE_DEVICES"] = gpu_str

        # don't make this debug... this is good UX
        rank_zero_info(f'CUDA_VISIBLE_DEVICES: [{os.environ["CUDA_VISIBLE_DEVICES"]}]')

    def set_random_port(self, force=False):
        """
        When running DDP NOT managed by SLURM, the ports might collide
        """
        # pick a random port first
        assert self.num_nodes == 1, 'random port can only be called from single node training'
        global RANDOM_PORTS
        default_port = RANDOM_PORTS[-1]
        RANDOM_PORTS = RANDOM_PORTS[:-1]

        # when not forced, use the user port
        if not force:
            default_port = os.environ.get('MASTER_PORT', default_port)

        os.environ['MASTER_PORT'] = str(default_port)

    def spawn_ddp_children(self, model):
        port = os.environ['MASTER_PORT']

        master_address = '127.0.0.1' if 'MASTER_ADDR' not in os.environ else os.environ['MASTER_ADDR']
        os.environ['MASTER_PORT'] = f'{port}'
        os.environ['MASTER_ADDR'] = f'{master_address}'

        # allow the user to pass the node rank
        node_rank = '0'
        if 'NODE_RANK' in os.environ:
            node_rank = os.environ['NODE_RANK']
        if 'GROUP_RANK' in os.environ:
            node_rank = os.environ['GROUP_RANK']

        os.environ['NODE_RANK'] = node_rank
        os.environ['LOCAL_RANK'] = '0'

        # when user is using hydra find the absolute path
        path_lib = abspath if not HYDRA_AVAILABLE else to_absolute_path

        # pull out the commands used to run the script and resolve the abs file path
        command = sys.argv
        try:
            full_path = path_lib(command[0])
        except Exception as e:
            full_path = abspath(command[0])

        command[0] = full_path
        # use the same python interpreter and actually running
        command = [sys.executable] + command

        # since this script sets the visible devices we replace the gpus flag with a number
        num_gpus = os.environ['CUDA_VISIBLE_DEVICES'].split(',').__len__()

        if '--gpus' in command:
            gpu_flag_idx = command.index('--gpus')
            command[gpu_flag_idx + 1] = f'{num_gpus}'

        os.environ['WORLD_SIZE'] = f'{num_gpus * self.num_nodes}'

        self.interactive_ddp_procs = []
        for local_rank in range(1, self.num_processes):
            env_copy = os.environ.copy()
            env_copy['LOCAL_RANK'] = f'{local_rank}'

            # start process
            proc = subprocess.Popen(command, env=env_copy)
            self.interactive_ddp_procs.append(proc)

            # starting all processes at once can cause issues
            # with dataloaders delay between 1-10 seconds
            delay = np.random.uniform(1, 5, 1)[0]
            sleep(delay)

        local_rank = 0
        results = self.ddp_train(local_rank, q=None, model=model, is_master=True)
        del os.environ['WORLD_SIZE']

        return results

    def ddp_train(self, process_idx, q, model, is_master=False, proc_offset=0):
        """
        Entry point for ddp

        Args:
            process_idx:
            q:
            model:
            is_master:
            proc_offset:

        Returns:

        """
        # offset the process id if requested
        process_idx = process_idx + proc_offset

        # show progressbar only on progress_rank 0
        if (self.node_rank != 0 or process_idx != 0) and self.progress_bar_callback is not None:
            self.progress_bar_callback.disable()

        # determine which process we are and world size
        if self.use_ddp:
            self.local_rank = process_idx
            self.global_rank = self.node_rank * self.num_processes + process_idx
            self.world_size = self.num_nodes * self.num_processes

        elif self.use_ddp2:
            self.local_rank = self.node_rank
            self.global_rank = self.node_rank
            self.world_size = self.num_nodes

        # set warning rank
        rank_zero_only.rank = self.global_rank

        # set up server using proc 0's ip address
        # try to init for 20 times at max in case ports are taken
        # where to store ip_table
        model.trainer = self
        model.init_ddp_connection(self.global_rank, self.world_size, self.is_slurm_managing_tasks)

        # call setup after the ddp process has connected
        self.setup('fit')
        if self.is_function_implemented('setup', model):
            model.setup('fit')

        # on world_size=0 let everyone know training is starting
        if self.is_global_zero:
            log.info('-' * 100)
            log.info(f'distributed_backend={self.distributed_backend}')
            log.info(f'All DDP processes registered. Starting ddp with {self.world_size} processes')
            log.info('-' * 100)

        # CHOOSE OPTIMIZER
        # allow for lr schedulers as well
        self.optimizers, self.lr_schedulers, self.optimizer_frequencies = self.init_optimizers(model)

        # MODEL
        # copy model to each gpu
        if self.on_gpu:
            gpu_idx = process_idx
            if is_master:
                # source of truth is cuda for gpu idx
                gpus = os.environ['CUDA_VISIBLE_DEVICES'].split(',')
                gpu_idx = int(gpus[self.local_rank])

            self.root_gpu = gpu_idx
            torch.cuda.set_device(self.root_gpu)
            model.cuda(self.root_gpu)

        # set model properties before going into wrapper
        self.copy_trainer_model_properties(model)

        # AMP
        # run through amp wrapper before going to distributed DP
        # TODO: remove with dropping NVIDIA AMP support
        if self.use_amp and not NATIVE_AMP_AVALAIBLE:
            model, optimizers = model.configure_apex(amp, model, self.optimizers, self.amp_level)
            self.optimizers = optimizers
            self.reinit_scheduler_properties(self.optimizers, self.lr_schedulers)

        # DDP2 uses all GPUs on the machine
        if self.distributed_backend == 'ddp' or self.distributed_backend == 'ddp_spawn':
            device_ids = [self.root_gpu]
        elif self.use_ddp2:
            device_ids = self.data_parallel_device_ids
        else:  # includes ddp_cpu
            device_ids = None

        # allow user to configure ddp
        model = model.configure_ddp(model, device_ids)

        # continue training routine
        results = self.run_pretrain_routine(model)

        # persist info in ddp_spawn
        self.__transfer_ddp_spawn_state_on_fit_end(model, q, results)

        # clean up memory
        torch.cuda.empty_cache()

        if self.global_rank == 0 and self.distributed_backend not in ['ddp_spawn', 'ddp_cpu']:
            return results

    def __transfer_ddp_spawn_state_on_fit_end(self, model, q, results):
        if not self.distributed_backend in ['ddp_spawn', 'ddp_cpu']:
            return

        # track the best model path
        best_model_path = None
        if self.checkpoint_callback is not None:
            best_model_path = self.checkpoint_callback.best_model_path

        if self.global_rank == 0 and q is not None:
            rank_zero_warn('cleaning up ddp environment...')
            q.put(best_model_path)
            q.put(results)

            # save the last weights
            last_path = None
            if not self.testing:
                last_path = os.path.join(self.default_root_dir, '__temp_weight_ddp_end.ckpt')
                torch.save(model.state_dict(), last_path)
            q.put(last_path)

    def save_spawn_weights(self, model):
        """
        Dump a temporary checkpoint after ddp ends to get weights out of the process
        :param model:
        :return:
        """
        if self.is_global_zero:
            path = os.path.join(self.default_root_dir, '__temp_weight_ddp_end.ckpt')
            self.save_checkpoint(path)
            return path

    def load_spawn_weights(self, original_model):
        """
        Load the temp weights saved in the process
        To recover the trained model from the ddp process we load the saved weights
        :param model:
        :return:
        """

        loaded_model = original_model

        if self.is_global_zero:
            # load weights saved in ddp
            path = os.path.join(self.default_root_dir, '__temp_weight_ddp_end.ckpt')
            loaded_model = original_model.__class__.load_from_checkpoint(path)

            # copy loaded weights to old model
            original_model.load_state_dict(loaded_model.state_dict())

            # remove ddp weights
            os.remove(path)

        return loaded_model

    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 _set_horovod_backend(self):
        self.check_horovod()
        self.use_horovod = True

        # Initialize Horovod to get rank / size info
        hvd.init()
        if self.on_gpu:
            # Horovod assigns one local GPU per process
            self.root_gpu = hvd.local_rank()

    def check_horovod(self):
        """Raises a `MisconfigurationException` if the Trainer is not configured correctly for Horovod."""
        if not HOROVOD_AVAILABLE:
            raise MisconfigurationException(
                'Requested `distributed_backend="horovod"`, but Horovod is not installed.'
                'Install with \n $HOROVOD_WITH_PYTORCH=1 pip install horovod[pytorch]'
            )

        if self.num_gpus > 1 or self.num_nodes > 1:
            raise MisconfigurationException(
                'Horovod does not support setting num_nodes / num_gpus explicitly. Use '
                'horovodrun / mpirun to configure the number of processes.'
            )

    @staticmethod
    def has_horovodrun():
        """Returns True if running with `horovodrun` using Gloo or OpenMPI."""
        return 'OMPI_COMM_WORLD_RANK' in os.environ or 'HOROVOD_RANK' in os.environ