Shortcuts

Source code for torchrl.collectors.distributed.rpc

# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

r"""Generic distributed data-collector using torch.distributed.rpc backend."""
from __future__ import annotations

import collections
import os
import socket
import time
import warnings
from copy import copy, deepcopy
from typing import Callable, List, OrderedDict

from torchrl._utils import logger as torchrl_logger

from torchrl.collectors.distributed import DEFAULT_SLURM_CONF
from torchrl.collectors.distributed.default_configs import (
    DEFAULT_TENSORPIPE_OPTIONS,
    IDLE_TIMEOUT,
    TCP_PORT,
)
from torchrl.collectors.utils import _NON_NN_POLICY_WEIGHTS, split_trajectories
from torchrl.data.utils import CloudpickleWrapper

SUBMITIT_ERR = None
try:
    import submitit

    _has_submitit = True
except ModuleNotFoundError as err:
    _has_submitit = False
    SUBMITIT_ERR = err
import torch.cuda
from tensordict import TensorDict
from torch import nn

from torch.distributed import rpc
from torchrl._utils import _ProcessNoWarn, VERBOSE

from torchrl.collectors import MultiaSyncDataCollector
from torchrl.collectors.collectors import (
    DataCollectorBase,
    DEFAULT_EXPLORATION_TYPE,
    MultiSyncDataCollector,
    SyncDataCollector,
)
from torchrl.envs.common import EnvBase
from torchrl.envs.env_creator import EnvCreator


def _rpc_init_collection_node(
    rank,
    rank0_ip,
    tcp_port,
    world_size,
    visible_device,
    tensorpipe_options,
    verbose=VERBOSE,
):
    os.environ["MASTER_ADDR"] = str(rank0_ip)
    os.environ["MASTER_PORT"] = str(tcp_port)

    if isinstance(visible_device, list):
        pass
    elif isinstance(visible_device, (str, int, torch.device)):
        visible_device = [visible_device]
    elif visible_device is None:
        pass
    else:
        raise RuntimeError(f"unrecognised dtype {type(visible_device)}")

    options = rpc.TensorPipeRpcBackendOptions(
        devices=visible_device,
        **tensorpipe_options,
    )
    if verbose:
        torchrl_logger.info(
            f"init rpc with master addr: {os.environ['MASTER_ADDR']}:{os.environ['MASTER_PORT']}"
        )
    rpc.init_rpc(
        f"COLLECTOR_NODE_{rank}",
        rank=rank,
        backend=rpc.BackendType.TENSORPIPE,
        rpc_backend_options=options,
        world_size=world_size,
    )
    rpc.shutdown()


[docs]class RPCDataCollector(DataCollectorBase): """An RPC-based distributed data collector. Supports sync and async data collection. Args: create_env_fn (Callable or List[Callabled]): list of Callables, each returning an instance of :class:`~torchrl.envs.EnvBase`. policy (Callable): Policy to be executed in the environment. Must accept :class:`tensordict.tensordict.TensorDictBase` object as input. If ``None`` is provided, the policy used will be a :class:`~torchrl.collectors.RandomPolicy` instance with the environment ``action_spec``. Accepted policies are usually subclasses of :class:`~tensordict.nn.TensorDictModuleBase`. This is the recommended usage of the collector. Other callables are accepted too: If the policy is not a ``TensorDictModuleBase`` (e.g., a regular :class:`~torch.nn.Module` instances) it will be wrapped in a `nn.Module` first. Then, the collector will try to assess if these modules require wrapping in a :class:`~tensordict.nn.TensorDictModule` or not. - If the policy forward signature matches any of ``forward(self, tensordict)``, ``forward(self, td)`` or ``forward(self, <anything>: TensorDictBase)`` (or any typing with a single argument typed as a subclass of ``TensorDictBase``) then the policy won't be wrapped in a :class:`~tensordict.nn.TensorDictModule`. - In all other cases an attempt to wrap it will be undergone as such: ``TensorDictModule(policy, in_keys=env_obs_key, out_keys=env.action_keys)``. Keyword Args: frames_per_batch (int): A keyword-only argument representing the total number of elements in a batch. total_frames (int): A keyword-only argument representing the total number of frames returned by the collector during its lifespan. If the ``total_frames`` is not divisible by ``frames_per_batch``, an exception is raised. Endless collectors can be created by passing ``total_frames=-1``. Defaults to ``-1`` (endless collector). device (int, str or torch.device, optional): The generic device of the collector. The ``device`` args fills any non-specified device: if ``device`` is not ``None`` and any of ``storing_device``, ``policy_device`` or ``env_device`` is not specified, its value will be set to ``device``. Defaults to ``None`` (No default device). Lists of devices are supported. storing_device (int, str or torch.device, optional): The *remote* device on which the output :class:`~tensordict.TensorDict` will be stored. If ``device`` is passed and ``storing_device`` is ``None``, it will default to the value indicated by ``device``. For long trajectories, it may be necessary to store the data on a different device than the one where the policy and env are executed. Defaults to ``None`` (the output tensordict isn't on a specific device, leaf tensors sit on the device where they were created). Lists of devices are supported. env_device (int, str or torch.device, optional): The *remote* device on which the environment should be cast (or executed if that functionality is supported). If not specified and the env has a non-``None`` device, ``env_device`` will default to that value. If ``device`` is passed and ``env_device=None``, it will default to ``device``. If the value as such specified of ``env_device`` differs from ``policy_device`` and one of them is not ``None``, the data will be cast to ``env_device`` before being passed to the env (i.e., passing different devices to policy and env is supported). Defaults to ``None``. Lists of devices are supported. policy_device (int, str or torch.device, optional): The *remote* device on which the policy should be cast. If ``device`` is passed and ``policy_device=None``, it will default to ``device``. If the value as such specified of ``policy_device`` differs from ``env_device`` and one of them is not ``None``, the data will be cast to ``policy_device`` before being passed to the policy (i.e., passing different devices to policy and env is supported). Defaults to ``None``. Lists of devices are supported. max_frames_per_traj (int, optional): Maximum steps per trajectory. Note that a trajectory can span across multiple batches (unless ``reset_at_each_iter`` is set to ``True``, see below). Once a trajectory reaches ``n_steps``, the environment is reset. If the environment wraps multiple environments together, the number of steps is tracked for each environment independently. Negative values are allowed, in which case this argument is ignored. Defaults to ``None`` (i.e., no maximum number of steps). init_random_frames (int, optional): Number of frames for which the policy is ignored before it is called. This feature is mainly intended to be used in offline/model-based settings, where a batch of random trajectories can be used to initialize training. If provided, it will be rounded up to the closest multiple of frames_per_batch. Defaults to ``None`` (i.e. no random frames). reset_at_each_iter (bool, optional): Whether environments should be reset at the beginning of a batch collection. Defaults to ``False``. postproc (Callable, optional): A post-processing transform, such as a :class:`~torchrl.envs.Transform` or a :class:`~torchrl.data.postprocs.MultiStep` instance. Defaults to ``None``. split_trajs (bool, optional): Boolean indicating whether the resulting TensorDict should be split according to the trajectories. See :func:`~torchrl.collectors.utils.split_trajectories` for more information. Defaults to ``False``. exploration_type (ExplorationType, optional): interaction mode to be used when collecting data. Must be one of ``torchrl.envs.utils.ExplorationType.DETERMINISTIC``, ``torchrl.envs.utils.ExplorationType.RANDOM``, ``torchrl.envs.utils.ExplorationType.MODE`` or ``torchrl.envs.utils.ExplorationType.MEAN``. Defaults to ``torchrl.envs.utils.ExplorationType.RANDOM``. collector_class (type or str, optional): a collector class for the remote node. Can be :class:`~torchrl.collectors.SyncDataCollector`, :class:`~torchrl.collectors.MultiSyncDataCollector`, :class:`~torchrl.collectors.MultiaSyncDataCollector` or a derived class of these. The strings "single", "sync" and "async" correspond to respective class. Defaults to :class:`~torchrl.collectors.SyncDataCollector`. .. note:: Support for :class:`MultiSyncDataCollector` and :class:`MultiaSyncDataCollector` is experimental, and :class:`~torchrl.collectors.SyncDataCollector` should always be preferred. If multiple simultaneous environment need to be executed on a single node, consider using a :class:`~torchrl.envs.ParallelEnv` instance. collector_kwargs (dict or list, optional): a dictionary of parameters to be passed to the remote data-collector. If a list is provided, each element will correspond to an individual set of keyword arguments for the dedicated collector. num_workers_per_collector (int, optional): the number of copies of the env constructor that is to be used on the remote nodes. Defaults to 1 (a single env per collector). On a single worker node all the sub-workers will be executing the same environment. If different environments need to be executed, they should be dispatched across worker nodes, not subnodes. sync (bool, optional): if ``True``, the resulting tensordict is a stack of all the tensordicts collected on each node. If ``False`` (default), each tensordict results from a separate node in a "first-ready, first-served" fashion. slurm_kwargs (dict): a dictionary of parameters to be passed to the submitit executor. update_after_each_batch (bool, optional): if ``True``, the weights will be updated after each collection. For ``sync=True``, this means that all workers will see their weights updated. For ``sync=False``, only the worker from which the data has been gathered will be updated. Defaults to ``False``, ie. updates have to be executed manually through :meth:`~torchrl.collectors.distributed.DistributedDataCollector.update_policy_weights_`. max_weight_update_interval (int, optional): the maximum number of batches that can be collected before the policy weights of a worker is updated. For sync collections, this parameter is overwritten by ``update_after_each_batch``. For async collections, it may be that one worker has not seen its parameters being updated for a certain time even if ``update_after_each_batch`` is turned on. Defaults to -1 (no forced update). launcher (str, optional): how jobs should be launched. Can be one of "submitit" or "mp" for multiprocessing. The former can launch jobs across multiple nodes, whilst the latter will only launch jobs on a single machine. "submitit" requires the homonymous library to be installed. To find more about submitit, visit https://github.com/facebookincubator/submitit Defaults to "submitit". tcp_port (int, optional): the TCP port to be used. Defaults to 10003. visible_devices (list of Union[int, torch.device, str], optional): a list of the same length as the number of nodes containing the device used to pass data to main. tensorpipe_options (dict, optional): a dictionary of keyword argument to pass to :class:`torch.distributed.rpc.TensorPipeRpcBackendOption`. """ _VERBOSE = VERBOSE # for debugging def __init__( self, create_env_fn, policy, *, frames_per_batch: int, total_frames: int = -1, device: torch.device | List[torch.device] = None, storing_device: torch.device | List[torch.device] = None, env_device: torch.device | List[torch.device] = None, policy_device: torch.device | List[torch.device] = None, max_frames_per_traj: int = -1, init_random_frames: int = -1, reset_at_each_iter: bool = False, postproc: Callable | None = None, split_trajs: bool = False, exploration_type: "ExporationType" = DEFAULT_EXPLORATION_TYPE, # noqa collector_class=SyncDataCollector, collector_kwargs=None, num_workers_per_collector=1, sync=False, slurm_kwargs=None, update_after_each_batch=False, max_weight_update_interval=-1, launcher="submitit", tcp_port=None, visible_devices=None, tensorpipe_options=None, ): if collector_class == "async": collector_class = MultiaSyncDataCollector elif collector_class == "sync": collector_class = MultiSyncDataCollector elif collector_class == "single": collector_class = SyncDataCollector self.collector_class = collector_class self.env_constructors = create_env_fn self.policy = policy if isinstance(policy, nn.Module): policy_weights = TensorDict.from_module(policy) policy_weights = policy_weights.data.lock_() else: warnings.warn(_NON_NN_POLICY_WEIGHTS) policy_weights = TensorDict(lock=True) self.policy_weights = policy_weights self.num_workers = len(create_env_fn) self.frames_per_batch = frames_per_batch self.device = device self.storing_device = storing_device self.env_device = env_device self.policy_device = policy_device self.storing_device = storing_device # make private to avoid changes from users during collection self._sync = sync self.update_after_each_batch = update_after_each_batch self.max_weight_update_interval = max_weight_update_interval if self.update_after_each_batch and self.max_weight_update_interval > -1: raise RuntimeError( "Got conflicting udpate instructions: `update_after_each_batch` " "`max_weight_update_interval` are incompatible." ) self.launcher = launcher self._batches_since_weight_update = [0 for _ in range(self.num_workers)] if tcp_port is None: self.tcp_port = os.environ.get("TCP_PORT", TCP_PORT) else: self.tcp_port = str(tcp_port) self.visible_devices = visible_devices if self._sync: if self.frames_per_batch % self.num_workers != 0: raise RuntimeError( f"Cannot dispatch {self.frames_per_batch} frames across {self.num_workers}. " f"Consider using a number of frames per batch that is divisible by the number of workers." ) self._frames_per_batch_corrected = self.frames_per_batch // self.num_workers else: self._frames_per_batch_corrected = self.frames_per_batch self.num_workers_per_collector = num_workers_per_collector self.total_frames = total_frames self.slurm_kwargs = copy(DEFAULT_SLURM_CONF) if slurm_kwargs is not None: self.slurm_kwargs.update(slurm_kwargs) collector_kwargs = collector_kwargs if collector_kwargs is not None else {} self.collector_kwargs = ( deepcopy(collector_kwargs) if isinstance(collector_kwargs, (list, tuple)) else [copy(collector_kwargs) for _ in range(self.num_workers)] ) # update collector kwargs for i, collector_kwarg in enumerate(self.collector_kwargs): collector_kwarg["max_frames_per_traj"] = max_frames_per_traj collector_kwarg["init_random_frames"] = ( init_random_frames // self.num_workers ) if not self._sync and init_random_frames > 0: warnings.warn( "async distributed data collection with init_random_frames > 0 " "may have unforeseen consequences as we do not control that once " "non-random data is being collected all nodes are returning non-random data. " "If this is a feature that you feel should be fixed, please raise an issue on " "torchrl's repo." ) collector_kwarg["reset_at_each_iter"] = reset_at_each_iter collector_kwarg["exploration_type"] = exploration_type collector_kwarg["device"] = self.device[i] collector_kwarg["storing_device"] = self.storing_device[i] collector_kwarg["env_device"] = self.env_device[i] collector_kwarg["policy_device"] = self.policy_device[i] self.postproc = postproc self.split_trajs = split_trajs if tensorpipe_options is None: self.tensorpipe_options = copy(DEFAULT_TENSORPIPE_OPTIONS) else: self.tensorpipe_options = copy(DEFAULT_TENSORPIPE_OPTIONS).update( tensorpipe_options ) self._init() @property def device(self) -> List[torch.device]: return self._device @property def storing_device(self) -> List[torch.device]: return self._storing_device @property def env_device(self) -> List[torch.device]: return self._env_device @property def policy_device(self) -> List[torch.device]: return self._policy_device @device.setter def device(self, value): if isinstance(value, (tuple, list)): if len(value) != self.num_workers: raise RuntimeError( "The number of devices passed to the collector must match the number of workers." ) self._device = value else: self._device = [value] * self.num_workers @storing_device.setter def storing_device(self, value): if isinstance(value, (tuple, list)): if len(value) != self.num_workers: raise RuntimeError( "The number of devices passed to the collector must match the number of workers." ) self._storing_device = value else: self._storing_device = [value] * self.num_workers @env_device.setter def env_device(self, value): if isinstance(value, (tuple, list)): if len(value) != self.num_workers: raise RuntimeError( "The number of devices passed to the collector must match the number of workers." ) self._env_device = value else: self._env_device = [value] * self.num_workers @policy_device.setter def policy_device(self, value): if isinstance(value, (tuple, list)): if len(value) != self.num_workers: raise RuntimeError( "The number of devices passed to the collector must match the number of workers." ) self._policy_device = value else: self._policy_device = [value] * self.num_workers def _init_master_rpc( self, world_size, ): """Init RPC on main node.""" options = rpc.TensorPipeRpcBackendOptions(**self.tensorpipe_options) if torch.cuda.device_count(): if self.visible_devices: for i in range(self.num_workers): rank = i + 1 options.set_device_map( f"COLLECTOR_NODE_{rank}", {0: self.visible_devices[i]} ) if self._VERBOSE: torchrl_logger.info("init rpc") rpc.init_rpc( "TRAINER_NODE", rank=0, backend=rpc.BackendType.TENSORPIPE, rpc_backend_options=options, world_size=world_size, ) def _start_workers( self, world_size, env_constructors, collector_class, num_workers_per_collector, policy, frames_per_batch, total_frames, collector_kwargs, ): """Instantiate remote collectors.""" num_workers = world_size - 1 time_interval = 1.0 collector_infos = [] for i in range(num_workers): counter = 0 while True: counter += 1 time.sleep(time_interval) try: if self._VERBOSE: torchrl_logger.info( f"trying to connect to collector node {i + 1}" ) collector_info = rpc.get_worker_info(f"COLLECTOR_NODE_{i + 1}") break except RuntimeError as err: if counter * time_interval > self.tensorpipe_options["rpc_timeout"]: raise RuntimeError("Could not connect to remote node") from err continue collector_infos.append(collector_info) collector_rrefs = [] for i in range(num_workers): env_make = env_constructors[i] if not isinstance(env_make, (EnvBase, EnvCreator)): env_make = CloudpickleWrapper(env_make) if self._VERBOSE: torchrl_logger.info("Making collector in remote node") collector_rref = rpc.remote( collector_infos[i], collector_class, args=( [env_make] * num_workers_per_collector if collector_class is not SyncDataCollector else env_make, policy, ), kwargs={ "frames_per_batch": frames_per_batch, "total_frames": -1, "split_trajs": False, **collector_kwargs[i], }, ) collector_rrefs.append(collector_rref) futures = collections.deque(maxlen=self.num_workers) if not self._sync: for i in range(num_workers): if self._VERBOSE: torchrl_logger.info("Asking for the first batch") future = rpc.rpc_async( collector_infos[i], collector_class.next, args=(collector_rrefs[i],), ) futures.append((future, i)) self.futures = futures self.collector_rrefs = collector_rrefs self.collector_infos = collector_infos def _init_worker_rpc(self, executor, i): """Init RPC node if necessary.""" visible_device = ( self.visible_devices[i] if self.visible_devices is not None else None ) if self.launcher == "submitit": if not _has_submitit: raise ImportError("submitit not found.") from SUBMITIT_ERR job = executor.submit( _rpc_init_collection_node, i + 1, self.IPAddr, self.tcp_port, self.num_workers + 1, visible_device, self.tensorpipe_options, self._VERBOSE, ) if self._VERBOSE: torchrl_logger.info(f"job id {job.job_id}") # ID of your job return job elif self.launcher == "mp": job = _ProcessNoWarn( target=_rpc_init_collection_node, args=( i + 1, self.IPAddr, self.tcp_port, self.num_workers + 1, visible_device, self.tensorpipe_options, self._VERBOSE, ), ) job.start() return job elif self.launcher == "submitit_delayed": # job is already launched return None else: raise NotImplementedError(f"Unknown launcher {self.launcher}") def _init(self): self._shutdown = False if self.launcher == "submitit": executor = submitit.AutoExecutor(folder="log_test") executor.update_parameters(**self.slurm_kwargs) else: executor = None hostname = socket.gethostname() if self.launcher != "mp": IPAddr = socket.gethostbyname(hostname) else: IPAddr = "localhost" self.IPAddr = IPAddr os.environ["MASTER_ADDR"] = str(self.IPAddr) os.environ["MASTER_PORT"] = str(self.tcp_port) self.jobs = [] for i in range(self.num_workers): if self._VERBOSE: torchrl_logger.info(f"Submitting job {i}") job = self._init_worker_rpc( executor, i, ) self.jobs.append(job) self._init_master_rpc( self.num_workers + 1, ) self._start_workers( world_size=self.num_workers + 1, env_constructors=self.env_constructors, collector_class=self.collector_class, num_workers_per_collector=self.num_workers_per_collector, policy=self.policy, frames_per_batch=self._frames_per_batch_corrected, total_frames=self.total_frames, collector_kwargs=self.collector_kwargs, ) def iterator(self): self._collected_frames = 0 while self._collected_frames < self.total_frames: if self._sync: data = self._next_sync_rpc() else: data = self._next_async_rpc() if self.split_trajs: data = split_trajectories(data) if self.postproc is not None: data = self.postproc(data) yield data if self.max_weight_update_interval > -1 and not self._sync: for j in range(self.num_workers): if ( self._batches_since_weight_update[j] > self.max_weight_update_interval ): self.update_policy_weights_([j], wait=False) elif self.max_weight_update_interval > -1: ranks = [ 1 for j in range(self.num_workers) if self._batches_since_weight_update[j] > self.max_weight_update_interval ] self.update_policy_weights_(ranks, wait=True)
[docs] def update_policy_weights_(self, workers=None, wait=True) -> None: if workers is None: workers = list(range(self.num_workers)) futures = [] for i in workers: if self._VERBOSE: torchrl_logger.info(f"calling update on worker {i}") futures.append( rpc.rpc_async( self.collector_infos[i], self.collector_class.update_policy_weights_, args=(self.collector_rrefs[i], self.policy_weights.detach()), ) ) if wait: for i in workers: if self._VERBOSE: torchrl_logger.info(f"waiting for worker {i}") futures[i].wait() if self._VERBOSE: torchrl_logger.info("got it!")
def _next_async_rpc(self): if self._VERBOSE: torchrl_logger.info("next async") if not len(self.futures): raise StopIteration( f"The queue is empty, the collector has ran out of data after {self._collected_frames} collected frames." ) while True: future, i = self.futures.popleft() if future.done(): if self.update_after_each_batch: self.update_policy_weights_(workers=(i,), wait=False) if self._VERBOSE: torchrl_logger.info(f"future {i} is done") data = future.value() self._collected_frames += data.numel() if self._collected_frames < self.total_frames: future = rpc.rpc_async( self.collector_infos[i], self.collector_class.next, args=(self.collector_rrefs[i],), ) self.futures.append((future, i)) return data self.futures.append((future, i)) def _next_sync_rpc(self): if self._VERBOSE: torchrl_logger.info("next sync: futures") if self.update_after_each_batch: self.update_policy_weights_() for i in range(self.num_workers): future = rpc.rpc_async( self.collector_infos[i], self.collector_class.next, args=(self.collector_rrefs[i],), ) self.futures.append((future, i)) data = [] while len(self.futures): future, i = self.futures.popleft() # the order is NOT guaranteed: should we change that? if future.done(): data += [future.value()] if self._VERBOSE: torchrl_logger.info( f"got data from {i} // data has len {len(data)} / {self.num_workers}" ) else: self.futures.append((future, i)) data = torch.cat(data) traj_ids = data.get(("collector", "traj_ids"), None) if traj_ids is not None: for i in range(1, self.num_workers): traj_ids[i] += traj_ids[i - 1].max() data.set_(("collector", "traj_ids"), traj_ids) self._collected_frames += data.numel() return data def set_seed(self, seed: int, static_seed: bool = False) -> int: for worker in self.collector_infos: seed = rpc.rpc_sync(worker, self.collector_class.set_seed, args=(seed,)) def state_dict(self) -> OrderedDict: raise NotImplementedError def load_state_dict(self, state_dict: OrderedDict) -> None: raise NotImplementedError def shutdown(self): if not hasattr(self, "_shutdown"): warnings.warn("shutdown has no effect has `_init` has not been called yet.") return if self._shutdown: return if self._VERBOSE: torchrl_logger.info("shutting down") for future, i in self.futures: # clear the futures while future is not None and not future.done(): torchrl_logger.info(f"waiting for proc {i} to clear") future.wait() for i in range(self.num_workers): if self._VERBOSE: torchrl_logger.info(f"shutting down {i}") rpc.rpc_sync( self.collector_infos[i], self.collector_class.shutdown, args=(self.collector_rrefs[i],), timeout=int(IDLE_TIMEOUT), ) if self._VERBOSE: torchrl_logger.info("rpc shutdown") rpc.shutdown(timeout=int(IDLE_TIMEOUT)) if self.launcher == "mp": for job in self.jobs: job.join(int(IDLE_TIMEOUT)) elif self.launcher == "submitit": for job in self.jobs: _ = job.result() elif self.launcher == "submitit_delayed": pass else: raise NotImplementedError(f"Unknown launcher {self.launcher}") self._shutdown = True

Docs

Access comprehensive developer documentation for PyTorch

View Docs

Tutorials

Get in-depth tutorials for beginners and advanced developers

View Tutorials

Resources

Find development resources and get your questions answered

View Resources