Source code for torch.nn.parallel.distributed
import copy
import inspect
import itertools
import logging
import os
import sys
import warnings
import weakref
from contextlib import contextmanager
from dataclasses import dataclass, fields, is_dataclass
from enum import Enum, auto
from typing import Callable, Any, Type
import torch
import torch.distributed as dist
from torch.autograd import Function, Variable
from torch.distributed.algorithms.join import (
Join,
Joinable,
JoinHook,
)
from torch.utils._pytree import tree_flatten, tree_unflatten
RPC_AVAILABLE = False
if dist.is_available():
from torch.distributed.utils import (
_verify_param_shape_across_processes,
_sync_module_states,
_to_kwargs,
)
from torch.distributed.distributed_c10d import ReduceOp, _get_default_group
if torch.distributed.rpc.is_available():
RPC_AVAILABLE = True
from torch.distributed.rpc import RRef
from torch._utils import _get_device_index
from ..modules import Module
from ._replicated_tensor_ddp_utils import _ddp_with_replicated_tensor_enabled
from .scatter_gather import gather, scatter_kwargs # noqa: F401
__all__ = ["DistributedDataParallel"]
logger = logging.getLogger(__name__)
def _tree_flatten_with_rref(output):
output_is_rref = RPC_AVAILABLE and isinstance(output, RRef)
if output_is_rref:
output_tensor_list, treespec = tree_flatten(output.local_value())
else:
output_tensor_list, treespec = tree_flatten(output)
# Need to return flattened tensors, spec to re-pack them, as well
# as if the return type was actually an RRef to reconstruct.
return output_tensor_list, treespec, output_is_rref
def _tree_unflatten_with_rref(output, treespec, output_is_rref):
output = tree_unflatten(output, treespec)
if output_is_rref:
output = RRef(output)
return output
def _find_tensors(obj):
r"""
Recursively find all tensors contained in the specified object.
"""
if RPC_AVAILABLE and isinstance(obj, RRef):
# If the current node is the owner of the RRef, unwrap it and try to
# find Tensors.
# TODO: Expand to remote RRefs.
if obj.is_owner():
return _find_tensors(obj.local_value())
if isinstance(obj, torch.Tensor):
return [obj]
if isinstance(obj, (list, tuple)):
return itertools.chain(*map(_find_tensors, obj))
if isinstance(obj, dict):
return itertools.chain(*map(_find_tensors, obj.values()))
if is_dataclass(obj):
return itertools.chain(
*map(_find_tensors, (getattr(obj, f.name) for f in fields(obj)))
)
return []
def _dump_DDP_relevant_env_vars():
relevant_env_vars = [
"RANK",
"LOCAL_RANK",
"WORLD_SIZE",
"MASTER_PORT",
"MASTER_ADDR",
"CUDA_VISIBLE_DEVICES",
"GLOO_SOCKET_IFNAME",
"GLOO_DEVICE_TRANSPORT",
"NCCL_SOCKET_IFNAME",
"NCCL_BLOCKING_WAIT",
"NCCL_DEBUG",
"NCCL_DEBUG_SUBSYS",
"NCCL_IB_DISABLE",
# More NCCL env vars:
"NCCL_P2P_DISABLE",
"NCCL_P2P_LEVEL",
"NCCL_SHM_DISABLE",
"NCCL_SOCKET_NTHREADS",
"NCCL_NSOCKS_PERTHREAD",
"NCCL_BUFFSIZE",
"NCCL_NTHREADS",
"NCCL_RINGS",
"NCCL_MAX_NCHANNELS",
"NCCL_MIN_NCHANNELS",
"NCCL_CHECKS_DISABLE",
"NCCL_CHECK_POINTERS",
"NCCL_LAUNCH_MODE",
"NCCL_IB_HCA",
"NCCL_IB_TIMEOUT",
"NCCL_IB_RETRY_CNT",
"NCCL_IB_GID_INDEX",
"NCCL_IB_SL",
"NCCL_IB_TC",
"NCCL_IB_AR_THRESHOLD",
"NCCL_IB_CUDA_SUPPORT",
"NCCL_NET_GDR_LEVEL",
"NCCL_NET_GDR_READ",
"NCCL_SINGLE_RING_THRESHOLD",
"NCCL_LL_THRESHOLD",
"NCCL_TREE_THRESHOLD",
"NCCL_ALGO",
"NCCL_PROTO",
"NCCL_IGNORE_CPU_AFFINITY",
"NCCL_DEBUG_FILE",
"NCCL_COLLNET_ENABLE",
"NCCL_TOPO_FILE",
"NCCL_TOPO_DUMP_FILE",
"NCCL_ASYNC_ERROR_HANDLING",
]
formatted_output = ""
for var in relevant_env_vars:
value = os.environ[var] if var in os.environ else "N/A"
formatted_output += "env:%s=%s\n" % (var, value)
print(formatted_output)
class _BufferCommHookLocation(Enum):
PRE_FORWARD = auto()
POST_FORWARD = auto()
@dataclass
class _BufferCommHook:
buffer_comm_hook: Callable
buffer_comm_hook_state: Any
buffer_comm_hook_location: _BufferCommHookLocation
# Add a DDPSink to run various functions when backwards starts, such as
# queueing call back of out-most backward/graph task,
# this helps call back is fired after all gradients' calculation
# is completed.
class _DDPSink(Function):
@staticmethod
def forward(ctx, reducer, state_dict, *inputs):
# set_materialize_grads(False) will ensure that None gradients stay as
# None and are not filled with zeros.
ctx.set_materialize_grads(False)
ctx.reducer = reducer
ctx.state_dict = state_dict
ret = tuple(
inp.clone() if isinstance(inp, torch.Tensor) else inp
for inp in inputs
)
return ret
@staticmethod
def backward(ctx, *grad_outputs):
state_dict = ctx.state_dict
# Enqueue delay allreduce for static graph training on the first
# iteration.
if (
ctx.state_dict["static_graph"]
and ctx.state_dict["num_iterations"] == 1
):
Variable._execution_engine.queue_callback( # type: ignore[call-arg,misc]
ctx.reducer._delay_all_reduce
)
return (None, None, *grad_outputs)
class _DDPJoinHook(JoinHook):
def __init__(self, ddp, divide_by_initial_world_size):
"""
Sets config variables for internal usage.
"""
assert isinstance(ddp, DistributedDataParallel), (
"DDP join hook requires passing in a DistributedDataParallel "
"instance as the state"
)
assert ddp.logger is not None
ddp.logger._set_uneven_input_join()
self.ddp = ddp
self.ddp._divide_by_initial_world_size = divide_by_initial_world_size
super().__init__()
def main_hook(self):
"""
Shadows the DDP collective communication operations in the forward and
backward passes.
"""
ddp = self.ddp
# Buckets are rebuilt only once during a training period
ddp.reducer._rebuild_buckets()
# Schedule a broadcast if we are syncing module buffers in the
# forward pass
# TODO: make DDP uneven inputs context manager support buffer
# comm hook (https://github.com/pytorch/pytorch/issues/65436)
ddp._check_and_sync_module_buffers()
# Check if need to sync in the backward pass
work = ddp._check_global_requires_backward_grad_sync(
is_joined_rank=True
)
work.wait()
should_sync_backwards = work.result()[0].item() != 0
# Forward parameter sync is disabled in the next iteration if we
# are skipping gradient sync this iteration, so set
# `require_forward_param_sync` accordingly
ddp.require_forward_param_sync = should_sync_backwards
if not should_sync_backwards:
return
# Schedule one allreduce per gradient bucket to match the backward
# pass allreduce
ddp._match_all_reduce_for_bwd_pass()
# Check if we need to allreduce locally unused parameters
if ddp.find_unused_parameters:
ddp._match_unused_params_allreduce()
# Rebuilt parameters are pushed only once during a training period
ddp.reducer._push_all_rebuilt_params()
def post_hook(self, is_last_joiner: bool):
"""
Syncs the final model to ensure that the model is the same across all
processes.
"""
self.ddp._sync_final_model(is_last_joiner)
[docs]class DistributedDataParallel(Module, Joinable):
r"""Implements distributed data parallelism that is based on
``torch.distributed`` package at the module level.
This container provides data parallelism by synchronizing gradients
across each model replica. The devices to synchronize across are
specified by the input ``process_group``, which is the entire world
by default. Note that ``DistributedDataParallel`` does not chunk or
otherwise shard the input across participating GPUs; the user is
responsible for defining how to do so, for example through the use
of a :class:`DistributedSampler`.
See also: :ref:`distributed-basics` and :ref:`cuda-nn-ddp-instead`.
The same constraints on input as in :class:`torch.nn.DataParallel` apply.
Creation of this class requires that ``torch.distributed`` to be already
initialized, by calling :func:`torch.distributed.init_process_group`.
``DistributedDataParallel`` is proven to be significantly faster than
:class:`torch.nn.DataParallel` for single-node multi-GPU data
parallel training.
To use ``DistributedDataParallel`` on a host with N GPUs, you should spawn
up ``N`` processes, ensuring that each process exclusively works on a single
GPU from 0 to N-1. This can be done by either setting
``CUDA_VISIBLE_DEVICES`` for every process or by calling:
>>> # xdoctest: +SKIP("undefined variables")
>>> torch.cuda.set_device(i)
where i is from 0 to N-1. In each process, you should refer the following
to construct this module:
>>> # xdoctest: +SKIP("undefined variables")
>>> torch.distributed.init_process_group(
>>> backend='nccl', world_size=N, init_method='...'
>>> )
>>> model = DistributedDataParallel(model, device_ids=[i], output_device=i)
In order to spawn up multiple processes per node, you can use either
``torch.distributed.launch`` or ``torch.multiprocessing.spawn``.
.. note::
Please refer to `PyTorch Distributed Overview <https://pytorch.org/tutorials/beginner/dist_overview.html>`__
for a brief introduction to all features related to distributed training.
.. note::
``DistributedDataParallel`` can be used in conjunction with
:class:`torch.distributed.optim.ZeroRedundancyOptimizer` to reduce
per-rank optimizer states memory footprint. Please refer to
`ZeroRedundancyOptimizer recipe <https://pytorch.org/tutorials/recipes/zero_redundancy_optimizer.html>`__
for more details.
.. note:: ``nccl`` backend is currently the fastest and highly recommended
backend when using GPUs. This applies to both single-node and
multi-node distributed training.
.. note:: This module also supports mixed-precision distributed training.
This means that your model can have different types of parameters such
as mixed types of ``fp16`` and ``fp32``, the gradient reduction on these
mixed types of parameters will just work fine.
.. note:: If you use ``torch.save`` on one process to checkpoint the module,
and ``torch.load`` on some other processes to recover it, make sure that
``map_location`` is configured properly for every process. Without
``map_location``, ``torch.load`` would recover the module to devices
where the module was saved from.
.. note:: When a model is trained on ``M`` nodes with ``batch=N``, the
gradient will be ``M`` times smaller when compared to the same model
trained on a single node with ``batch=M*N`` if the loss is summed (NOT
averaged as usual) across instances in a batch (because the gradients
between different nodes are averaged). You should take this into
consideration when you want to obtain a mathematically equivalent
training process compared to the local training counterpart. But in most
cases, you can just treat a DistributedDataParallel wrapped model, a
DataParallel wrapped model and an ordinary model on a single GPU as the
same (E.g. using the same learning rate for equivalent batch size).
.. note::
Parameters are never broadcast between processes. The module performs
an all-reduce step on gradients and assumes that they will be modified
by the optimizer in all processes in the same way. Buffers
(e.g. BatchNorm stats) are broadcast from the module in process of rank
0, to all other replicas in the system in every iteration.
.. note::
If you are using DistributedDataParallel in conjunction with the
:ref:`distributed-rpc-framework`, you should always use
:meth:`torch.distributed.autograd.backward` to compute gradients and
:class:`torch.distributed.optim.DistributedOptimizer` for optimizing
parameters.
Example::
>>> # xdoctest: +SKIP("undefined variables")
>>> import torch.distributed.autograd as dist_autograd
>>> from torch.nn.parallel import DistributedDataParallel as DDP
>>> import torch
>>> from torch import optim
>>> from torch.distributed.optim import DistributedOptimizer
>>> import torch.distributed.rpc as rpc
>>> from torch.distributed.rpc import RRef
>>>
>>> t1 = torch.rand((3, 3), requires_grad=True)
>>> t2 = torch.rand((3, 3), requires_grad=True)
>>> rref = rpc.remote("worker1", torch.add, args=(t1, t2))
>>> ddp_model = DDP(my_model)
>>>
>>> # Setup optimizer
>>> optimizer_params = [rref]
>>> for param in ddp_model.parameters():
>>> optimizer_params.append(RRef(param))
>>>
>>> dist_optim = DistributedOptimizer(
>>> optim.SGD,
>>> optimizer_params,
>>> lr=0.05,
>>> )
>>>
>>> with dist_autograd.context() as context_id:
>>> pred = ddp_model(rref.to_here())
>>> loss = loss_func(pred, target)
>>> dist_autograd.backward(context_id, [loss])
>>> dist_optim.step(context_id)
.. note::
DistributedDataParallel currently offers limited support for gradient
checkpointing with :meth:`torch.utils.checkpoint`. DDP will work as
expected when there are no unused parameters in the model and each layer
is checkpointed at most once (make sure you are not passing
`find_unused_parameters=True` to DDP). We currently do not support the
case where a layer is checkpointed multiple times, or when there unused
parameters in the checkpointed model.
.. note::
To let a non-DDP model load a state dict from a DDP model,
:meth:`~torch.nn.modules.utils.consume_prefix_in_state_dict_if_present`
needs to be applied to strip the prefix "module." in the DDP state dict before loading.
.. warning::
Constructor, forward method, and differentiation of the output (or a
function of the output of this module) are distributed synchronization
points. Take that into account in case different processes might be
executing different code.
.. warning::
This module assumes all parameters are registered in the model by the
time it is created. No parameters should be added nor removed later.
Same applies to buffers.
.. warning::
This module assumes all parameters are registered in the model of each
distributed processes are in the same order. The module itself will
conduct gradient ``allreduce`` following the reverse order of the
registered parameters of the model. In other words, it is users'
responsibility to ensure that each distributed process has the exact
same model and thus the exact same parameter registration order.
.. warning::
This module allows parameters with non-rowmajor-contiguous strides.
For example, your model may contain some parameters whose
:class:`torch.memory_format` is ``torch.contiguous_format``
and others whose format is ``torch.channels_last``. However,
corresponding parameters in different processes must have the
same strides.
.. warning::
This module doesn't work with :func:`torch.autograd.grad` (i.e. it will
only work if gradients are to be accumulated in ``.grad`` attributes of
parameters).
.. warning::
If you plan on using this module with a ``nccl`` backend or a ``gloo``
backend (that uses Infiniband), together with a DataLoader that uses
multiple workers, please change the multiprocessing start method to
``forkserver`` (Python 3 only) or ``spawn``. Unfortunately
Gloo (that uses Infiniband) and NCCL2 are not fork safe, and you will
likely experience deadlocks if you don't change this setting.
.. warning::
You should never try to change your model's parameters after wrapping
up your model with ``DistributedDataParallel``. Because, when
wrapping up your model with ``DistributedDataParallel``, the constructor
of ``DistributedDataParallel`` will register the additional gradient
reduction functions on all the parameters of the model itself at the
time of construction. If you change the model's parameters afterwards,
gradient reduction functions no longer match the correct set of
parameters.
.. warning::
Using ``DistributedDataParallel`` in conjunction with the
:ref:`distributed-rpc-framework` is experimental and subject to change.
Args:
module (Module): module to be parallelized
device_ids (list of int or torch.device): CUDA devices.
1) For single-device modules, ``device_ids`` can
contain exactly one device id, which represents the only
CUDA device where the input module corresponding to this process resides.
Alternatively, ``device_ids`` can also be ``None``.
2) For multi-device modules and CPU modules,
``device_ids`` must be ``None``.
When ``device_ids`` is ``None`` for both cases,
both the input data for the forward pass and the actual module
must be placed on the correct device.
(default: ``None``)
output_device (int or torch.device): Device location of output for
single-device CUDA modules. For multi-device modules and
CPU modules, it must be ``None``, and the module itself
dictates the output location. (default: ``device_ids[0]``
for single-device modules)
broadcast_buffers (bool): Flag that enables syncing (broadcasting)
buffers of the module at beginning of the ``forward``
function. (default: ``True``)
process_group: The process group to be used for distributed data
all-reduction. If ``None``, the default process group, which
is created by :func:`torch.distributed.init_process_group`,
will be used. (default: ``None``)
bucket_cap_mb: ``DistributedDataParallel`` will bucket parameters into
multiple buckets so that gradient reduction of each
bucket can potentially overlap with backward computation.
:attr:`bucket_cap_mb` controls the bucket size in
MegaBytes (MB). (default: 25)
find_unused_parameters (bool): Traverse the autograd graph from all
tensors contained in the return value of the
wrapped module's ``forward`` function. Parameters
that don't receive gradients as part of this
graph are preemptively marked as being ready to
be reduced. In addition, parameters that may have
been used in the wrapped module's ``forward``
function but were not part of loss computation and
thus would also not receive gradients are
preemptively marked as ready to be reduced.
(default: ``False``)
check_reduction: This argument is deprecated.
gradient_as_bucket_view (bool): When set to ``True``, gradients will be views
pointing to different offsets of ``allreduce`` communication
buckets. This can reduce peak memory usage, where the
saved memory size will be equal to the total gradients
size. Moreover, it avoids the overhead of copying between
gradients and ``allreduce`` communication buckets. When
gradients are views, ``detach_()`` cannot be called on the
gradients. If hitting such errors, please fix it by
referring to the :meth:`~torch.optim.Optimizer.zero_grad`
function in ``torch/optim/optimizer.py`` as a solution.
Note that gradients will be views after first iteration, so
the peak memory saving should be checked after first iteration.
static_graph (bool): When set to ``True``, DDP knows the trained graph is
static. Static graph means 1) The set of used and unused
parameters will not change during the whole training loop; in
this case, it does not matter whether users set
``find_unused_parameters = True`` or not. 2) How the graph is trained
will not change during the whole training loop (meaning there is
no control flow depending on iterations).
When static_graph is set to be ``True``, DDP will support cases that
can not be supported in the past:
1) Reentrant backwards.
2) Activation checkpointing multiple times.
3) Activation checkpointing when model has unused parameters.
4) There are model parameters that are outside of forward function.
5) Potentially improve performance when there are unused parameters,
as DDP will not search graph in each iteration to detect unused
parameters when static_graph is set to be ``True``.
To check whether you can set static_graph to be ``True``, one way is to
check ddp logging data at the end of your previous model training,
if ``ddp_logging_data.get("can_set_static_graph") == True``, mostly you
can set ``static_graph = True`` as well.
Example::
>>> # xdoctest: +SKIP("undefined variables")
>>> model_DDP = torch.nn.parallel.DistributedDataParallel(model)
>>> # Training loop
>>> ...
>>> ddp_logging_data = model_DDP._get_ddp_logging_data()
>>> static_graph = ddp_logging_data.get("can_set_static_graph")
Attributes:
module (Module): the module to be parallelized.
Example::
>>> # xdoctest: +SKIP("undefined variables")
>>> torch.distributed.init_process_group(backend='nccl', world_size=4, init_method='...')
>>> net = torch.nn.parallel.DistributedDataParallel(model)
"""
# used to track whether the given thread is inside ddp forward for torchdynamo purposes
_active_ddp_module = None
def __init__(
self,
module,
device_ids=None,
output_device=None,
dim=0,
broadcast_buffers=True,
process_group=None,
bucket_cap_mb=25,
find_unused_parameters=False,
check_reduction=False,
gradient_as_bucket_view=False,
static_graph=False,
):
super().__init__()
Joinable.__init__(self)
self.logger = None
if hasattr(module, "_ddp_params_and_buffers_to_ignore"):
self.parameters_to_ignore = set(module._ddp_params_and_buffers_to_ignore)
else:
self.parameters_to_ignore = set()
self._module_parameters = [p for n, p in module.named_parameters() if n not in self.parameters_to_ignore]
if not any((p.requires_grad for p in self._module_parameters)):
self._log_and_throw(
RuntimeError,
"DistributedDataParallel is not needed when a module "
"doesn't have any parameter that requires a gradient.",
)
if device_ids is not None and len(device_ids) > 1:
self._log_and_throw(
ValueError,
"device_ids can only be None or contain a single element.",
)
self.is_multi_device_module = len({p.device for p in self._module_parameters}) > 1
distinct_device_types = {p.device.type for p in self._module_parameters if p.device is not None}
if len(distinct_device_types) != 1:
self._log_and_throw(
ValueError,
"DistributedDataParallel's input module must be on "
"the same type of devices, but input module parameters locate in {}.".format(
distinct_device_types
),
)
self.device_type = list(distinct_device_types)[0]
if (
device_ids is None
or len(device_ids) == 0 # For backward compatibility.
or self.device_type == "cpu"
or self.is_multi_device_module
):
if device_ids or output_device:
self._log_and_throw(
ValueError,
"DistributedDataParallel device_ids and output_device arguments "
"only work with single-device/multiple-device GPU modules or CPU modules, "
"but got device_ids {}, output_device {}, and module parameters {}.".format(
device_ids,
output_device,
{p.device for p in self._module_parameters},
),
)
self.device_ids = None
self.output_device = None
else:
self.device_ids = [_get_device_index(x, True) for x in device_ids]
if output_device is None:
output_device = device_ids[0]
self.output_device = _get_device_index(output_device, True)
if process_group is None:
self.process_group = _get_default_group()
else:
self.process_group = process_group
self.static_graph = False
self.dim = dim
self.module = module
self.device = list(self._module_parameters)[0].device
self.broadcast_buffers = broadcast_buffers
self.find_unused_parameters = find_unused_parameters
self.require_backward_grad_sync = True
self.require_forward_param_sync = True
self.gradient_as_bucket_view = gradient_as_bucket_view
self._use_replicated_tensor_module = (
_ddp_with_replicated_tensor_enabled()
)
self._build_replicated_tensor_module()
if check_reduction:
# This argument is no longer used since the reducer
# will ensure reduction completes even if some parameters
# do not receive gradients.
warnings.warn(
"The `check_reduction` argument in `DistributedDataParallel` "
"module is deprecated. Please avoid using it."
)
# Check that a module does not have Uninitialized parameters
for param in self._module_parameters:
if isinstance(param, torch.nn.parameter.UninitializedParameter):
self._log_and_throw(
RuntimeError,
"Modules with uninitialized parameters can't be used with `DistributedDataParallel`. "
"Run a dummy forward pass to correctly initialize the modules",
)
# used for intra-node param sync and inter-node sync as well
self.broadcast_bucket_size = int(250 * 1024 * 1024)
# reduction bucket size
self.bucket_bytes_cap = int(bucket_cap_mb * 1024 * 1024)
# Whether to perform input tensor CPU to GPU copies on a side-stream
self.use_side_stream_for_tensor_copies = (
os.environ.get("PYTORCH_DDP_USE_SIDE_STREAM", "1") == "1"
)
# Build parameters for reducer.
parameters, expect_sparse_gradient = self._build_params_for_reducer()
# Verify model equivalence.
_verify_param_shape_across_processes(self.process_group, parameters)
# Sync params and buffers. Ensures all DDP models start off at the same value.
_sync_module_states(
module=self.module,
process_group=self.process_group,
broadcast_bucket_size=self.broadcast_bucket_size,
src=0,
params_and_buffers_to_ignore=self.parameters_to_ignore,
)
# In debug mode, build a mapping of parameter index -> parameter.
param_to_name_mapping = self._build_debug_param_to_name_mapping(
parameters
)
# Builds reducer.
self._ddp_init_helper(
parameters,
expect_sparse_gradient,
param_to_name_mapping,
static_graph,
)
self._has_rebuilt_buckets = False
if static_graph:
self._set_static_graph()
self._setup_in_backward_optimizers()
def _setup_in_backward_optimizers(self):
# Check if user has used apply_optim_in_backward to overlap optimizer
# step + DDP backward. Current constraints:
# 1. Only allreduce is supported at the moment, no custom communication.
# 2. The reducer by default sets all grads for parameters DDP manages to
# None after they have been applied by the optimizer. There is no support
# for setting only some parameter grads to None, this must be done manually
# by user (and DDP_OVERLAPPED_OPTIM_SET_GRADS_TO_NONE=0 needs to be set.)
# If your use case requires some DDP managed parameters to run with
# an in-backward optimizer and some with a traditional optimizer, please
# ping https://github.com/pytorch/pytorch/issues/90052.
# NOTE: we use self._module_parameters instead of .parameters() since
# the former excludes ignored (non-DDP managed) parameters.
if any(
hasattr(p, '_in_backward_optimizers') for p in self._module_parameters
):
# Remove hooks that apply_optim_in_backward had registered because
# DDP customizes how optimizer is overlapped with backward due to
# the allreduce.
param_to_handle_map = dist.optim.apply_optimizer_in_backward.param_to_optim_hook_handle_map
for p in self._module_parameters:
for handle in param_to_handle_map.get(p, []):
handle.remove()
# Need a weakref to the reducer in order to run all_reduce.
reducer_weakref = weakref.ref(self.reducer)
# Note: importing in function, otherwise this will cause a circular
# import.
from torch.distributed.algorithms.ddp_comm_hooks.optimizer_overlap_hooks import (
_apply_optim_in_backward_hook
)
self.register_comm_hook(
(reducer_weakref, self.process_group),
_apply_optim_in_backward_hook(
gradient_is_bucket_view=self.gradient_as_bucket_view
),
)
# TODO (rohan-varma): this is a workaround that allows users to
# disable the default behavior of DDP managed parameters with
# optimizer runing in backwards having their gradients all set to None.
# Currently, it is an "all or nothing behavior" where DDP will set
# no grads to None or all of them, relaxing this behavior will be
# done dependent on use cases.
if os.getenv("DDP_OVERLAPPED_OPTIM_SET_GRADS_TO_NONE", "1") != "0":
warnings.warn(
"DDP + apply_optim_in_backward will currently set all "
"parameter gradients to None. If this is not the desired "
"behavior, please set env variable "
"DDP_OVERLAPPED_OPTIM_SET_GRADS_TO_NONE=0, and manually set"
"gradients to None/zero as desired."
)
self.reducer._set_grads_to_none() # type: ignore[attr-defined]
def _build_replicated_tensor_module(self):
if self._use_replicated_tensor_module:
# Create a module with ReplicatedTensor without copying tensors. Avoid
# registering '_replicated_tensor_module' as a submodule by directly
# adding to self.__dict__.
from ._replicated_tensor_ddp_interop import _replicate_module
self.__dict__["_replicated_tensor_module"] = _replicate_module(
self.module, self.process_group
)
def _log_and_throw(self, err_type, err_msg):
if self.logger is not None:
self.logger.set_error_and_log(f"{str(err_type)}: {err_msg}")
raise err_type(err_msg)
def _ddp_init_helper(
self,
parameters,
expect_sparse_gradient,
param_to_name_mapping,
static_graph,
):
"""
Initialization helper function that does the following:
(1) bucketing the parameters for reductions
(2) resetting the bucketing states
(3) registering the grad hooks
(4) Logging construction-time DDP logging data
(5) passing a handle of DDP to SyncBatchNorm Layer
"""
self.num_iterations = 0
# Notice, the parameters order is not in the order in which they are used,
# especially in models with control flow.
#
# Alongside parameters are not presented in the real execution order,
# if a certain model happens to also
# 1) have other collectives comm ops in its backward graph.
# 2) have unused parameter in subset ranks of the whole world.
# bucketing could insert ALL-REDUCE comm op too early on the rank with unused parameter,
# matching up with other collectives comm ops on other ranks unexpectedly.
#
# In order to handle this corner case, when the parameters are not in the real execution order,
# we don't do bucketing, thus only one ALL-REDUCE is inserted after all the gradients
# of the whole graph are computed.
#
# Notice, here we only disable bucketing for the first iteration.
# After the first iteration, it's OK to rebuild buckets,
# because "bucket rebuild" bucketizes parameters based on its real execution order in backward graph.
# Can remove this branching once #73732 is landed.
if static_graph is True or self.find_unused_parameters is False:
bucket_size_limits = [sys.maxsize]
else:
bucket_size_limits = [
dist._DEFAULT_FIRST_BUCKET_BYTES,
self.bucket_bytes_cap,
]
(
bucket_indices,
per_bucket_size_limits,
) = dist._compute_bucket_assignment_by_size(
parameters,
bucket_size_limits,
expect_sparse_gradient,
)
# Note: reverse list of buckets because we want to approximate the
# order in which their gradients are produced, and assume they
# are used in the forward pass in the order they are defined.
self.reducer = dist.Reducer(
parameters,
list(reversed(bucket_indices)),
list(reversed(per_bucket_size_limits)),
self.process_group,
expect_sparse_gradient,
# The bucket size limit is specified in the constructor.
# Additionally, we allow for a single small bucket for parameters
# that are defined first, such that their gradients don't spill into
# a much larger bucket, adding unnecessary latency after gradient
# computation finishes. Experiments showed 1MB is a reasonable value.
self.bucket_bytes_cap,
self.find_unused_parameters,
self.gradient_as_bucket_view,
param_to_name_mapping,
# User can set dist._DEFAULT_FIRST_BUCKET_BYTES to tune DDP first
# bucket.
dist._DEFAULT_FIRST_BUCKET_BYTES,
)
self.logger = dist.Logger(self.reducer)
# Set as a weak reference to avoid reference cycle between
# logger and reducer.
self.reducer.set_logger(self.logger)
has_sync_bn = False
for submodule in self.module.modules():
if isinstance(submodule, torch.nn.SyncBatchNorm):
has_sync_bn = True
break
# Set logging data that can be got during construction time.
self.logger.set_construction_data_and_log(
self.module.__class__.__name__,
[] if self.device_ids is None else self.device_ids,
-1 if self.output_device is None else self.output_device,
self.broadcast_buffers,
has_sync_bn,
static_graph,
)
# passing a handle to torch.nn.SyncBatchNorm layer
self._passing_sync_batchnorm_handle(self.module)
def __getstate__(self):
self._check_default_group()
attrs = copy.copy(self.__dict__)
del attrs["process_group"]
del attrs["reducer"]
del attrs["logger"]
if self._use_replicated_tensor_module:
del attrs["_replicated_tensor_module"]
return attrs
def __setstate__(self, state):
# If serializable, then the process group should be the default one
self.process_group = _get_default_group()
super().__setstate__(state)
self._build_replicated_tensor_module()
self.__dict__.setdefault("require_forward_param_sync", True)
self.__dict__.setdefault("require_backward_grad_sync", True)
parameters, expect_sparse_gradient = self._build_params_for_reducer()
# In debug mode, build a mapping of parameter index -> parameter.
param_to_name_mapping = self._build_debug_param_to_name_mapping(
parameters
)
# Builds reducer.
self._ddp_init_helper(
parameters,
expect_sparse_gradient,
param_to_name_mapping,
self.static_graph,
)
if self.static_graph:
self.reducer._set_static_graph()
assert self.logger is not None
self.logger._set_static_graph()
def _build_params_for_reducer(self):
# Build tuple of (module, parameter) for all parameters that require grads.
modules_and_parameters = [
(module, parameter)
for module_name, module in self.module.named_modules()
for parameter in [
param
# Note that we access module.named_parameters instead of
# parameters(module). parameters(module) is only needed in the
# single-process multi device case, where it accesses replicated
# parameters through _former_parameters.
for param_name, param in module.named_parameters(recurse=False)
if param.requires_grad
and f"{module_name}.{param_name}"
not in self.parameters_to_ignore
]
]
# Deduplicate any parameters that might be shared across child modules.
memo = set()
modules_and_parameters = [
# "p not in memo" is the deduplication check.
# "not memo.add(p)" is always True, and it's only there to cause "add(p)" if needed.
(m, p)
for m, p in modules_and_parameters
if p not in memo and not memo.add(p) # type: ignore[func-returns-value]
]
# Build list of parameters.
parameters = [parameter for _, parameter in modules_and_parameters]
# Checks if a module will produce a sparse gradient.
def produces_sparse_gradient(module):
if isinstance(module, (torch.nn.Embedding, torch.nn.EmbeddingBag)):
return module.sparse
return False
# Build list of booleans indicating whether or not to expect sparse
# gradients for the corresponding parameters.
expect_sparse_gradient = [
produces_sparse_gradient(module)
for module, _ in modules_and_parameters
]
self._assign_modules_buffers()
return parameters, expect_sparse_gradient
def _assign_modules_buffers(self):
"""
Assigns module buffers to self.modules_buffers which are then used to
broadcast across ranks when broadcast_buffers=True. Note that this
must be called every time buffers need to be synced because buffers can
be reassigned by user module,
see https://github.com/pytorch/pytorch/issues/63916.
"""
# Collect buffers for modules, filtering out buffers that should be ignored.
named_module_buffers = [
(buffer, buffer_name)
for buffer_name, buffer in self.module.named_buffers()
if buffer_name not in self.parameters_to_ignore
]
self.modules_buffers = [
buffer for (buffer, buffer_name) in named_module_buffers
]
# Dict[str, tensor] representing module buffers not ignored by DDP.
self.named_module_buffers = {
buffer_name: buffer
for (buffer, buffer_name) in named_module_buffers
}
def _build_debug_param_to_name_mapping(self, parameters):
if dist.get_debug_level() == dist.DebugLevel.OFF:
return {}
param_to_param_index = {
parameters[i]: i for i in range(len(parameters))
}
param_set = set(parameters)
param_index_to_param_fqn = {}
for module_name, module in self.module.named_modules():
for param_name, param in module.named_parameters(recurse=False):
fqn = f"{module_name}.{param_name}"
# Bypass ignored parameters since those are not reduced by DDP
# to begin with.
if fqn not in self.parameters_to_ignore and param.requires_grad:
if param not in param_set:
self._log_and_throw(
ValueError,
f"Param with name {fqn} found in module parameters, but not DDP parameters."
" This indicates a bug in DDP, please report an issue to PyTorch.",
)
param_index = param_to_param_index[param]
param_index_to_param_fqn[param_index] = fqn
# Ensure we covered all parameters
if len(param_set) != len(param_index_to_param_fqn):
self._log_and_throw(
ValueError,
(
"Expected param to name mapping to cover all parameters, but"
f" got conflicting lengths: {len(param_set)} vs "
f"{len(param_index_to_param_fqn)}. This indicates a bug in DDP"
", please report an issue to PyTorch."
),
)
return param_index_to_param_fqn
def _get_parameters(self, m, recurse=True):
"""
Returns a generator of module parameters
"""
def model_parameters(m):
ps = (
m._former_parameters.values()
if hasattr(m, "_former_parameters")
else m.parameters(recurse=False)
)
yield from ps
for m in m.modules() if recurse else [m]:
for p in model_parameters(m):
yield p
def _check_default_group(self):
pickle_not_supported = False
try:
if self.process_group != _get_default_group():
pickle_not_supported = True
except RuntimeError:
pickle_not_supported = True
if pickle_not_supported:
self._log_and_throw(
RuntimeError,
"DDP Pickling/Unpickling are only supported "
"when using DDP with the default process "
"group. That is, when you have called "
"init_process_group and have not passed "
"process_group argument to DDP constructor",
)
[docs] @contextmanager
def no_sync(self):
r"""
A context manager to disable gradient synchronizations across DDP
processes. Within this context, gradients will be accumulated on module
variables, which will later be synchronized in the first
forward-backward pass exiting the context.
Example::
>>> # xdoctest: +SKIP("undefined variables")
>>> ddp = torch.nn.parallel.DistributedDataParallel(model, pg)
>>> with ddp.no_sync():
>>> for input in inputs:
>>> ddp(input).backward() # no synchronization, accumulate grads
>>> ddp(another_input).backward() # synchronize grads
.. warning::
The forward pass should be included inside the context manager, or
else gradients will still be synchronized.
"""
old_require_backward_grad_sync = self.require_backward_grad_sync
self.require_backward_grad_sync = False
try:
yield
finally:
self.require_backward_grad_sync = old_require_backward_grad_sync
@classmethod
def _get_active_ddp_module(cls):
"""
TorchDynamo needs to know whether DDP is currently active, and access the DDP module in order to cooperatively optimize it.
"""
return cls._active_ddp_module
# note, this ctxmgr function is marked 'skip' in torchdynamo, so dynamo only kicks in
# for the 'module_to_run' underneath
# see torch._dynamo/eval_frame.py TorchPatcher.patch for more details
@contextmanager
def _inside_ddp_forward(self):
DistributedDataParallel._active_ddp_module = self
try:
yield
except Exception:
raise
finally:
DistributedDataParallel._active_ddp_module = None
def _run_ddp_forward(self, *inputs, **kwargs):
module_to_run = (
self._replicated_tensor_module
if self._use_replicated_tensor_module
else self.module
)
if self.device_ids:
inputs, kwargs = _to_kwargs(
inputs,
kwargs,
self.device_ids[0],
self.use_side_stream_for_tensor_copies,
)
with self._inside_ddp_forward():
return module_to_run(*inputs[0], **kwargs[0]) # type: ignore[index]
else:
with self._inside_ddp_forward():
return module_to_run(*inputs, **kwargs)
def forward(self, *inputs, **kwargs):
with torch.autograd.profiler.record_function(
"DistributedDataParallel.forward"
):
if torch.is_grad_enabled() and self.require_backward_grad_sync:
assert self.logger is not None
self.logger.set_runtime_stats_and_log()
self.num_iterations += 1
self.reducer.prepare_for_forward()
# Notify the join context that this process has not joined, if
# needed
work = Join.notify_join_context(self)
if work:
self.reducer._set_forward_pass_work_handle(
work, self._divide_by_initial_world_size # type: ignore[arg-type]
)
# Calling _rebuild_buckets before forward compuation,
# It may allocate new buckets before deallocating old buckets
# inside _rebuild_buckets. To save peak memory usage,
# call _rebuild_buckets before the peak memory usage increases
# during forward computation.
# This should be called only once during whole training period.
if torch.is_grad_enabled() and self.reducer._rebuild_buckets():
logger.info(
"Reducer buckets have been rebuilt in this iteration."
)
self._has_rebuilt_buckets = True
# sync params according to location (before/after forward) user
# specified as part of hook, if hook was specified.
if self._check_sync_bufs_pre_fwd():
self._sync_buffers()
if self._join_config.enable:
# Notify joined ranks whether they should sync in backwards pass or not.
self._check_global_requires_backward_grad_sync(
is_joined_rank=False
)
output = self._run_ddp_forward(*inputs, **kwargs)
# sync params according to location (before/after forward) user
# specified as part of hook, if hook was specified.
if self._check_sync_bufs_post_fwd():
self._sync_buffers()
if torch.is_grad_enabled() and self.require_backward_grad_sync:
self.require_forward_param_sync = True
# We'll return the output object verbatim since it is a freeform
# object. We need to find any tensors in this object, though,
# because we need to figure out which parameters were used during
# this forward pass, to ensure we short circuit reduction for any
# unused parameters. Only if `find_unused_parameters` is set.
if self.find_unused_parameters and not self.static_graph:
# Do not need to populate this for static graph.
self.reducer.prepare_for_backward(
list(_find_tensors(output))
)
else:
self.reducer.prepare_for_backward([])
else:
self.require_forward_param_sync = False
# TODO: DDPSink is currently enabled for unused parameter detection and
# static graph training for first iteration.
if (self.find_unused_parameters and not self.static_graph) or (
self.static_graph and self.num_iterations == 1
):
state_dict = {
"static_graph": self.static_graph,
"num_iterations": self.num_iterations,
}
(
output_tensor_list,
treespec,
output_is_rref,
) = _tree_flatten_with_rref(output)
output_placeholders = [None for _ in range(len(output_tensor_list))]
# Do not touch tensors that have no grad_fn, which can cause issues
# such as https://github.com/pytorch/pytorch/issues/60733
for i, output in enumerate(output_tensor_list):
if torch.is_tensor(output) and output.grad_fn is None:
output_placeholders[i] = output
# When find_unused_parameters=True, makes tensors which require grad
# run through the DDPSink backward pass. When not all outputs are
# used in loss, this makes those corresponding tensors receive
# undefined gradient which the reducer then handles to ensure
# param.grad field is not touched and we don't error out.
passthrough_tensor_list = _DDPSink.apply(
self.reducer,
state_dict,
*output_tensor_list,
)
for i in range(len(output_placeholders)):
if output_placeholders[i] is None:
output_placeholders[i] = passthrough_tensor_list[i]
# Reconstruct output data structure.
output = _tree_unflatten_with_rref(
output_placeholders, treespec, output_is_rref
)
return output
def scatter(self, inputs, kwargs, device_ids):
return scatter_kwargs(inputs, kwargs, device_ids, dim=self.dim)
def to_kwargs(self, inputs, kwargs, device_id):
# Kept for BC
return _to_kwargs(
inputs, kwargs, device_id, self.use_side_stream_for_tensor_copies
)
def gather(self, outputs, output_device):
return gather(outputs, output_device, dim=self.dim)
def train(self, mode=True):
super().train(mode)
if self._use_replicated_tensor_module:
self._replicated_tensor_module.train(mode) # type: ignore[union-attr]
return self
# When running in join mode, schedules an allreduce to notify joined ranks
# of whether backwards pass synchronization will run this iteration or not.
def _check_global_requires_backward_grad_sync(self, is_joined_rank):
if not is_joined_rank and self.require_backward_grad_sync:
requires_sync_tensor = torch.ones(1, device=self.device)
else:
requires_sync_tensor = torch.zeros(1, device=self.device)
work = dist.all_reduce(
requires_sync_tensor, group=self.process_group, async_op=True
)
return work
# When running in join mode, checks and performs sync of module buffers if
# the models have buffers that should be synchronized in the forward pass.
def _check_and_sync_module_buffers(self):
if self._check_sync_bufs_pre_fwd():
authoritative_rank = self._find_common_rank(
self._distributed_rank, False
)
self._sync_module_buffers(authoritative_rank)
# When running in join model, agrees upon a common rank and broadcast model
# parameters to all other ranks.
def _sync_final_model(self, is_last_joiner):
# Agree upon the process that will be the authoritative model copy.
# The current rank is a candidate for being the authoritative copy if
# is_last_joiner=True. We break ties via picking the larger rank.
self._authoritative_rank = self._find_common_rank(
self._distributed_rank, is_last_joiner
)
_sync_module_states(
module=self.module,
process_group=self.process_group,
broadcast_bucket_size=self.broadcast_bucket_size,
src=self._authoritative_rank,
params_and_buffers_to_ignore=self.parameters_to_ignore,
)
# Schedule comm ops to match those scheduled in the reducer's backward
# pass.
def _match_all_reduce_for_bwd_pass(self):
comm_work = []
# Schedule comm in the same order as Reducer schedules them, i.e.
# the order of the buckets. Retrieving the bucket order from the reducer
# ensures that we keep the same order in join mode, such as when bucket
# order is rebuilt dynamically.
# Returns grad_buckets in order, but real tensors are substituted with
# zero tensors of the same shape.
grad_buckets = self.reducer._get_zeros_like_grad_buckets()
for grad_bucket in grad_buckets:
# Joined processes contribute zero gradient. In the case that
# divide_by_initial_world_size=True, we divide grads by the static
# world size, if not, the dividing factor is reduced by the number
# of joined processes.
work = self.reducer._run_comm_hook(grad_bucket)
comm_work.append(work)
for work in comm_work:
work.wait()
# Allreduces the used parameter mapping across ranks.
def _match_unused_params_allreduce(self):
locally_used_param_map = self.reducer._get_local_used_map()
self.process_group.allreduce(locally_used_param_map)
[docs] def join(
self,
divide_by_initial_world_size: bool = True,
enable: bool = True,
throw_on_early_termination: bool = False,
):
r"""
A context manager to be used in conjunction with an instance of
:class:`torch.nn.parallel.DistributedDataParallel` to be
able to train with uneven inputs across participating processes.
This context manager will keep track of already-joined DDP processes,
and "shadow" the forward and backward passes by inserting collective
communication operations to match with the ones created by non-joined
DDP processes. This will ensure each collective call has a corresponding
call by already-joined DDP processes, preventing hangs or errors that
would otherwise happen when training with uneven inputs across
processes. Alternatively, if the flag ``throw_on_early_termination`` is
specified to be ``True``, all trainers will throw an error once one rank
runs out of inputs, allowing these errors to be caught and handled
according to application logic.
Once all DDP processes have joined, the context manager will broadcast
the model corresponding to the last joined process to all processes to
ensure the model is the same across all processes
(which is guaranteed by DDP).
To use this to enable training with uneven inputs across processes,
simply wrap this context manager around your training loop. No further
modifications to the model or data loading is required.
.. warning::
If the model or training loop this context manager is wrapped around
has additional distributed collective operations, such as
``SyncBatchNorm`` in the model's forward pass, then the flag
``throw_on_early_termination`` must be enabled. This is because this
context manager is not aware of non-DDP collective communication.
This flag will cause all ranks to throw when any one rank
exhausts inputs, allowing these errors to be caught and recovered
from across all ranks.
Args:
divide_by_initial_world_size (bool): If ``True``, will divide
gradients by the initial ``world_size`` DDP training was launched
with. If ``False``, will compute the effective world size
(number of ranks that have not depleted their inputs yet) and
divide gradients by that during allreduce. Set
``divide_by_initial_world_size=True`` to ensure every input
sample including the uneven inputs have equal weight in terms of
how much they contribute to the global gradient. This is
achieved by always dividing the gradient by the initial
``world_size`` even when we encounter uneven inputs. If you set
this to ``False``, we divide the gradient by the remaining
number of nodes. This ensures parity with training on a smaller
``world_size`` although it also means the uneven inputs would
contribute more towards the global gradient. Typically, you
would want to set this to ``True`` for cases where the last few
inputs of your training job are uneven. In extreme cases, where
there is a large discrepancy in the number of inputs, setting
this to ``False`` might provide better results.
enable (bool): Whether to enable uneven input detection or not. Pass
in ``enable=False`` to disable in cases where you know that
inputs are even across participating processes. Default is
``True``.
throw_on_early_termination (bool): Whether to throw an error
or continue training when at least one rank has exhausted
inputs. If ``True``, will throw upon the first rank reaching end
of data. If ``False``, will continue training with a smaller
effective world size until all ranks are joined. Note that if
this flag is specified, then the flag
``divide_by_initial_world_size`` would be ignored. Default
is ``False``.
Example::
>>> # xdoctest: +SKIP("Distributed")
>>> import torch
>>> import torch.distributed as dist
>>> import os
>>> import torch.multiprocessing as mp
>>> import torch.nn as nn
>>> # On each spawned worker
>>> def worker(rank):
>>> dist.init_process_group("nccl", rank=rank, world_size=2)
>>> torch.cuda.set_device(rank)
>>> model = nn.Linear(1, 1, bias=False).to(rank)
>>> model = torch.nn.parallel.DistributedDataParallel(
>>> model, device_ids=[rank], output_device=rank
>>> )
>>> # Rank 1 gets one more input than rank 0.
>>> inputs = [torch.tensor([1]).float() for _ in range(10 + rank)]
>>> with model.join():
>>> for _ in range(5):
>>> for inp in inputs:
>>> loss = model(inp).sum()
>>> loss.backward()
>>> # Without the join() API, the below synchronization will hang
>>> # blocking for rank 1's allreduce to complete.
>>> torch.cuda.synchronize(device=rank)
"""
return Join(
[self],
enable,
throw_on_early_termination,
divide_by_initial_world_size=divide_by_initial_world_size,
)
[docs] def join_hook(
self,
**kwargs,
):
r"""
Returns the DDP join hook, which enables training on uneven inputs by
shadowing the collective communications in the forward and backward
passes.
Arguments:
kwargs (dict): a :class:`dict` containing any keyword arguments
to modify the behavior of the join hook at run time; all
:class:`Joinable` instances sharing the same join context
manager are forwarded the same value for ``kwargs``.
The hook supports the following keyword arguments:
divide_by_initial_world_size (bool, optional):
If ``True``, then gradients are divided by the initial world
size that DDP was launched with.
If ``False``, then gradients are divided by the effective world
size (i.e. the number of non-joined processes), meaning that
the uneven inputs contribute more toward the global gradient.
Typically, this should be set to ``True`` if the degree of
unevenness is small but can be set to ``False`` in extreme
cases for possibly better results.
Default is ``True``.
"""
divide_by_initial_world_size = kwargs.get(
"divide_by_initial_world_size", True
)
return _DDPJoinHook(
self, divide_by_initial_world_size=divide_by_initial_world_size
)
@property
def join_device(self):
return self.device
@property
def join_process_group(self):
return self.process_group
def _register_buffer_comm_hook(
self,
state,
hook: Callable,
comm_hook_location=_BufferCommHookLocation.POST_FORWARD,
):
r"""
Allows custom registration of hooks that define how buffer are
synchronized across ranks. The hook takes in an optional state
and is passed in a Dict[str, Tensor] corresponding to buffer names
and the buffers, and can run arbitrary reductions on buffers as
opposed to DDP's default broadcast from rank 0. This is useful for
example if a counter needs to be summed or averaged across ranks
every iteration.
Args:
state (Any): Optional state that is passed to the hook.
hook (Callable): Callable with the following signature:
``hook(state: object, bucket: dist.GradBucket) -> torch.futures.Future[torch.Tensor]``
comm_hook_location (_BufferCommHookLocation): Enum value indicating
where to run the hook.
_BufferCommHookLocation.PRE_FORWARD means that the
hook will run _before_ the forward pass, and
_BufferCommHookLocation.POST_FORWARD means that the
hook will run _after_ the forward pass.
NOTE: To maximize performance, users can return a
List[torch.futures.Future] from their hook, and DDP will
install and await these hooks appropriately at the end of
the backward pass. This will ensure all buffers are
synchronized by the end of the backward pass. If this
setting is used, it is recommended to pass
comm_hook_location=_BufferCommHookLocation.POST_FORWARD,
which will trigger the hook after the forward pass.
If _BufferCommHookLocation.PRE_FORWARD is used, users must
ensure appropriate synchronization when manipulating GPU
buffers in the forward pass.
"""
assert callable(hook)
self.buffer_hook = _BufferCommHook(
buffer_comm_hook=hook,
buffer_comm_hook_state=state,
buffer_comm_hook_location=comm_hook_location,
)
[docs] def register_comm_hook(self, state: object, hook: Callable):
r"""
Registers a communication hook which is an enhancement that provides a
flexible hook to users where they can specify how DDP aggregates gradients
across multiple workers.
This hook would be very useful for researchers to try out new ideas. For
example, this hook can be used to implement several algorithms like GossipGrad
and gradient compression which involve different communication strategies for
parameter syncs while running Distributed DataParallel training.
Args:
state (object): Passed to the hook to maintain any state information during the training process.
Examples include error feedback in gradient compression,
peers to communicate with next in GossipGrad, etc.
It is locally stored by each worker
and shared by all the gradient tensors on the worker.
hook (Callable): Callable with the following signature:
``hook(state: object, bucket: dist.GradBucket) -> torch.futures.Future[torch.Tensor]``:
This function is called once the bucket is ready. The
hook can perform whatever processing is needed and return
a Future indicating completion of any async work (ex: allreduce).
If the hook doesn't perform any communication, it still
must return a completed Future. The Future should hold the
new value of grad bucket's tensors. Once a bucket is ready,
c10d reducer would call this hook and use the tensors returned
by the Future and copy grads to individual parameters.
Note that the future's return type must be a single tensor.
We also provide an API called ``get_future`` to retrieve a
Future associated with the completion of ``c10d.ProcessGroup.Work``.
``get_future`` is currently supported for NCCL and also supported for most
operations on GLOO and MPI, except for peer to peer operations (send/recv).
.. warning ::
Grad bucket's tensors will not be predivided by world_size. User is responsible
to divide by the world_size in case of operations like allreduce.
.. warning ::
DDP communication hook can only be registered once and should be registered
before calling backward.
.. warning ::
The Future object that hook returns should contain a single tensor
that has the same shape with the tensors inside grad bucket.
.. warning ::
``get_future`` API supports NCCL, and partially GLOO and MPI backends (no support
for peer-to-peer operations like send/recv) and will return a ``torch.futures.Future``.
Example::
Below is an example of a noop hook that returns the same tensor.
>>> # xdoctest: +SKIP('undefined name')
>>> def noop(state: object, bucket: dist.GradBucket) -> torch.futures.Future[torch.Tensor]:
>>> fut = torch.futures.Future()
>>> fut.set_result(bucket.buffer())
>>> return fut
>>> ddp.register_comm_hook(state=None, hook=noop)
Example::
Below is an example of a Parallel SGD algorithm where gradients are encoded before
allreduce, and then decoded after allreduce.
>>> # xdoctest: +SKIP('undefined name')
>>> def encode_and_decode(state: object, bucket: dist.GradBucket) -> torch.futures.Future[torch.Tensor]:
>>> encoded_tensor = encode(bucket.buffer()) # encode gradients
>>> fut = torch.distributed.all_reduce(encoded_tensor).get_future()
>>> # Define the then callback to decode.
>>> def decode(fut):
>>> decoded_tensor = decode(fut.value()[0]) # decode gradients
>>> return decoded_tensor
>>> return fut.then(decode)
>>> ddp.register_comm_hook(state=None, hook=encode_and_decode)
"""
self._check_comm_hook(hook)
assert self.logger is not None
self.logger._set_comm_hook_name(hook.__qualname__)
dist._register_comm_hook(self.reducer, state, hook)
def _register_builtin_comm_hook(self, comm_hook_type):
r"""
Registers a built-in communication hook that specifies how DDP
aggregates gradients across multiple workers.
The built-in hooks aim to provide efficient C++ implementations for certain hooks,
which might not be as efficient if implemented in Python using a Python communication hook.
Args:
comm_hook_type (dist.BuiltinCommHookType): type of communication hook, such as ALLREDUCE, FP16_COMPRESS, etc.
.. warning ::
DDP communication hook can only be registered once and should be registered
before calling backward.
Example::
Below is an example of a FP16 compression where gradients are
compressed into 16-bit floating-point numbers before allreduce, and
then decompressed after allreduce.
>>> # xdoctest: +SKIP('undefined name')
>>> ddp._register_builtin_comm_hook(dist.BuiltinCommHookType.FP16_COMPRESS)
"""
assert self.logger is not None
self.logger._set_comm_hook_name(str(comm_hook_type))
dist._register_builtin_comm_hook(self.reducer, comm_hook_type)
def _register_fused_optim(
self, optim: Type, *args, optim_params=None, **kwargs
):
r"""
Registers an optimizer with DDP such that the optimization for a
parameter will run immediately when that parameter's gradient is
finished with reduction, instead of waiting for all parameters'
gradients to finish reduction. This can result in a training speedup
depending on your workload since the optimizer can run while gradient
reduction for other parameters are still ongoing. In addition, this has
the potential to reduce peak memory consumption during training, as it
only needs to load the per-parameter optimizer states of a single
parameter at a time, instead of loading all per-parameter optimizer
states at once.
Args:
optim (Type): a ``torch.optim.Optimizer`` class to be registered
as a fused optimizer.
*args (Sequence[Any]): Arguments to forward to `optim`.
optim_params (Optional[Iterable[torch.Tensor]]): Set of parameters
to optimize, similar to `params` argument of traditional `torch.optim`
Optimizers. If this is omitted, all DDP model parameters will be
optimized.
**kwargs: (Dict[str, Any]): Keyword arguments to forward to `optim`.
.. warning ::
_register_fused_optim should only be called once on a DDP instance,
and registering multiple fused optimizers for the same DDP model
is not currently supported. Please ping
https://github.com/pytorch/pytorch/issues/71595 if this is necessary
for your use case.
.. warning ::
_register_fused_optim and register_comm_hook currently do not
compose together, meaning that custom DDP communication hooks are
not supported with overlapped optimizers. Please ping
https://github.com/pytorch/pytorch/issues/71595 if this is necessary
for your use case.
.. warning ::
Gradient accumulation and DDP `no_sync` are currently not supported
with overlapped optimizer. Please ping
https://github.com/pytorch/pytorch/issues/71595 if this is necessary
for your use case.
Example::
>>> # xdoctest: +SKIP("No rendezvous handler")
>>> torch.distributed.init_process_group(backend='nccl', world_size=4, init_method='...')
>>> net = torch.nn.parallel.DistributedDataParallel(model, pg)
>>> lr = 1e-2
>>> betas = (0.9, 0.99)
>>> eps = 1e-6
>>> net._register_fused_optim(torch.optim.Adam, lr, betas=betas, eps=eps)
>>> # Example with subset of parameters
>>> params_to_opt = [list(net.parameters())[0]]
>>> net._register_fused_optim(
... torch.optim.Adam, lr, optim_params=params_to_opt, betas=betas, eps=eps
... )
"""
# Note: importing in function, otherwise this will cause a circular
# import as optimizer_overlap module needs to import DistributedDataParallel.
from torch.distributed.algorithms._optimizer_overlap import (
_as_overlapped_optim,
)
overlapped_optim = _as_overlapped_optim(
optim, optim_params, *args, **kwargs
)
try:
overlapped_optim.register_ddp(self)
except NotImplementedError as e:
raise RuntimeError(
f"{optim} does not support overlapped DDP. Please file an issue to PyTorch or the respective owner of {optim}."
) from e
def _distributed_broadcast_coalesced(
self, tensors, buffer_size, authoritative_rank=0
):
dist._broadcast_coalesced(
self.process_group, tensors, buffer_size, authoritative_rank
)
def _check_sync_bufs_post_fwd(self):
return (
self.will_sync_module_buffers()
and hasattr(self, "buffer_hook")
and self.buffer_hook.buffer_comm_hook_location
== _BufferCommHookLocation.POST_FORWARD
)
def _check_sync_bufs_pre_fwd(self):
return self.will_sync_module_buffers() and (
not hasattr(self, "buffer_hook")
or self.buffer_hook.buffer_comm_hook_location
== _BufferCommHookLocation.PRE_FORWARD
)
def will_sync_module_buffers(self):
return (
self.require_forward_param_sync
and self.broadcast_buffers
and len(self.modules_buffers) > 0
)
def _find_common_rank(self, input_rank, rank_cond):
# -1 indicates that this rank is not under consideration to be the
# common_rank
rank_to_use = torch.tensor(
[input_rank if rank_cond else -1],
device=self.device,
)
dist.all_reduce(rank_to_use, op=ReduceOp.MAX, group=self.process_group)
if rank_to_use.item() == -1:
self._log_and_throw(
ValueError,
"BUG! Expected rank_cond to be true for at least one process."
" This indicates a bug in PyTorch, please report an issue.",
)
return rank_to_use.item()
def _sync_buffers(self):
with torch.no_grad():
# module buffer sync
# Synchronize buffers across processes.
# If we are running DDP with the join manager, we have to agree
# upon a rank to sync module buffers from, since rank 0 may
# already have been joined and have stale module buffers.
if self._join_config.enable:
authoritative_rank = self._find_common_rank(
self._distributed_rank, True
)
else:
# The process with rank 0 is considered the authoritative copy.
authoritative_rank = 0
# Update self.modules_buffers incase any buffers were
# reassigned.
self._assign_modules_buffers()
self._sync_module_buffers(authoritative_rank)
def _sync_module_buffers(self, authoritative_rank):
if not hasattr(self, "buffer_hook"):
self._default_broadcast_coalesced(
authoritative_rank=authoritative_rank
)
else:
hook = self.buffer_hook.buffer_comm_hook
state = self.buffer_hook.buffer_comm_hook_state
futs = hook(state, self.named_module_buffers)
if futs is not None:
self.reducer._install_post_backward_futures(futs)
def _default_broadcast_coalesced(
self, bufs=None, bucket_size=None, authoritative_rank=0
):
"""
Broadcasts buffers from rank 0 to rest of workers. If bufs, bucket_size
are None, default values self.modules_buffers and
self.broadcast_bucket_size are used instead.
"""
if bufs is None:
bufs = self.modules_buffers
if bucket_size is None:
bucket_size = self.broadcast_bucket_size
self._distributed_broadcast_coalesced(
bufs, bucket_size, authoritative_rank
)
def _passing_sync_batchnorm_handle(self, module):
for layer in module.modules():
if isinstance(layer, torch.nn.modules.SyncBatchNorm):
if self.device_type == "cpu":
self._log_and_throw(
ValueError,
"SyncBatchNorm layers only work with GPU modules",
)
def _check_comm_hook(self, hook):
if not callable(hook):
self._log_and_throw(
TypeError, "Communication hook must be callable."
)
sig = inspect.signature(hook)
if (
sig.parameters["bucket"].annotation != inspect._empty
and sig.parameters["bucket"].annotation != dist.GradBucket
):
self._log_and_throw(
ValueError,
"Communication hook: bucket annotation should be dist.GradBucket.",
)
if (
sig.return_annotation != inspect._empty
and sig.return_annotation != torch.futures.Future[torch.Tensor]
):
self._log_and_throw(
ValueError,
"Communication hook: return annotation should be torch.futures.Future[torch.Tensor].",
)
if hook.__name__ in [
"bf16_compress_hook",
"bf16_compress_wrapper_hook",
] and (
(torch.version.cuda is None and torch.version.hip is None)
or (
torch.version.cuda is not None
and int(torch.version.cuda.split(".")[0]) < 11
)
or not dist.is_available()
or not dist.is_nccl_available()
or torch.cuda.nccl.version() < (2, 10)
):
self._log_and_throw(
TypeError,
"BF16 all reduce communication hook required CUDA 11+ and NCCL 2.10+.",
)
@property
def _distributed_rank(self):
return dist.get_rank(self.process_group)
@staticmethod
def _set_params_and_buffers_to_ignore_for_model(
module, params_and_buffers_to_ignore
):
"""
Sets parameters and buffers to be ignored by DDP. Expected format for
parameters is the fully qualified name: {module_name}.{param_name}, and
similarly, {module_name}.{buffer_name} for buffers. For example:
params_to_ignore = []
# NB: model here is vanilla PyTorch module, not yet wrapped with DDP.
for module_name, module in model.named_modules():
for param_name, param in module.named_parameters(recurse=False):
if should_ignore(param):
# Create expected format
fqn = f"{module_name}.{param_name}"
params_to_ignore.append(fqn)
torch.nn.parallel.DistributedDataParallel._set_params_and_buffers_to_ignore_for_model(
model,
params_to_ignore
)
"""
# This is a workaround to set parameters and buffers DDP should ignore
# during synchronization. It will be removed when the API is finalized
# as part of addressing https://github.com/pytorch/pytorch/issues/43690.
module._ddp_params_and_buffers_to_ignore = params_and_buffers_to_ignore
for name, param in module.named_parameters():
if name in params_and_buffers_to_ignore:
param._ddp_ignored = True
for name, buffer in module.named_buffers():
if name in params_and_buffers_to_ignore:
buffer._ddp_ignored = True
def _get_ddp_logging_data(self):
r"""
This interface can be called after DistributedDataParallel() is
constructed. It returns a dictionary of logging data. It could help
for debugging and analysis. The loggind data includes DistributedDataParallel
constructor input parameters, some internal states of DistributedDataParallel
and performance metrics. Simply print the dictorinary and see what
these metrics are.
This is a prototype interface and subject to change in the future.
"""
assert self.logger is not None
ddp_logging_data = self.logger._get_ddp_logging_data()
return {**ddp_logging_data.strs_map, **ddp_logging_data.ints_map}
def _set_ddp_runtime_logging_sample_rate(self, sample_rate):
r"""
This interface allows users to set sample_rate of collecting
runtime stats. The runtime stats will be recorded for the
first 10 iterations, after 10 iterations runtime stats will be
recorded once every "sample_rate" training iterations. In
default, runtime stats are recorded for the first 10 iterations,
after 10 iterations runtime stats are recorded once every
"kDDPRuntimeLoggingSampleRate=100" training iterations.
This is a prototype interface and subject to change in the future.
"""
if sample_rate < 1:
self._log_and_throw(
ValueError,
"DDP runtime logging sample rate should be equal or greater than 1",
)
self.reducer._set_ddp_runtime_logging_sample_rate(sample_rate)
def _set_static_graph(self):
"""
It is recommended to set static graph in the DDP constructor, which will
call this private API internally.
"""
# If self.static_graph has been set, no need to set it again
if self.static_graph:
warnings.warn(
"You've set static_graph to be True, no need to set it again."
)
return
self.static_graph = True
self.reducer._set_static_graph()
assert self.logger is not None
self.logger._set_static_graph()
if self.find_unused_parameters:
warnings.warn(
"You passed find_unused_parameters=true to DistributedDataParallel, "
"`_set_static_graph` will detect unused parameters automatically, so "
"you do not need to set find_unused_parameters=true, just be sure these "
"unused parameters will not change during training loop while calling "
"`_set_static_graph`."
)
