DistributedDataParallel(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)¶
Implements distributed data parallelism that is based on
torch.distributedpackage at the module level.
This container parallelizes the application of the given module by splitting the input across the specified devices by chunking in the batch dimension. The module is replicated on each machine and each device, and each such replica handles a portion of the input. During the backwards pass, gradients from each node are averaged.
The batch size should be larger than the number of GPUs used locally.
See also: Basics and Use nn.parallel.DistributedDataParallel instead of multiprocessing or nn.DataParallel. The same constraints on input as in
Creation of this class requires that
torch.distributedto be already initialized, by calling
DistributedDataParallelis proven to be significantly faster than
torch.nn.DataParallelfor single-node multi-GPU data parallel training.
DistributedDataParallelon a host with N GPUs, you should spawn up
Nprocesses, ensuring that each process exclusively works on a single GPU from 0 to N-1. This can be done by either setting
CUDA_VISIBLE_DEVICESfor every process or by calling:
where i is from 0 to N-1. In each process, you should refer the following to construct this module:
>>> 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
Please refer to PyTorch Distributed Overview for a brief introduction to all features related to distributed training.
DistributedDataParallelcan be used in conjunction with
torch.distributed.optim.ZeroRedundancyOptimizerto reduce per-rank optimizer states memory footprint. Please refer to ZeroRedundancyOptimizer recipe for more details.
ncclbackend is currently the fastest and highly recommended backend when using GPUs. This applies to both single-node and multi-node distributed training.
This module also supports mixed-precision distributed training. This means that your model can have different types of parameters such as mixed types of
fp32, the gradient reduction on these mixed types of parameters will just work fine.
If you use
torch.saveon one process to checkpoint the module, and
torch.loadon some other processes to recover it, make sure that
map_locationis configured properly for every process. Without
torch.loadwould recover the module to devices where the module was saved from.
When a model is trained on
batch=N, the gradient will be
Mtimes smaller when compared to the same model trained on a single node with
batch=M*Nif 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).
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.
If you are using DistributedDataParallel in conjunction with the Distributed RPC Framework, you should always use
torch.distributed.autograd.backward()to compute gradients and
torch.distributed.optim.DistributedOptimizerfor optimizing parameters.
>>> import torch.distributed.autograd as dist_autograd >>> from torch.nn.parallel import DistributedDataParallel as DDP >>> from torch import optim >>> from torch.distributed.optim import DistributedOptimizer >>> 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, loss) >>> dist_autograd.backward(context_id, loss) >>> dist_optim.step()
To let a non-DDP model load a state dict from a DDP model,
consume_prefix_in_state_dict_if_present()needs to be applied to strip the prefix “module.” in the DDP state dict before loading.
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.
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.
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
allreducefollowing 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.
This module allows parameters with non-rowmajor-contiguous strides. For example, your model may contain some parameters whose
torch.contiguous_formatand others whose format is
torch.channels_last. However, corresponding parameters in different processes must have the same strides.
This module doesn’t work with
torch.autograd.grad()(i.e. it will only work if gradients are to be accumulated in
.gradattributes of parameters).
If you plan on using this module with a
ncclbackend or a
gloobackend (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.
Forward and backward hooks defined on
moduleand its submodules won’t be invoked anymore, unless the hooks are initialized in the
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
DistributedDataParallelwill 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 redunction functions no longer match the correct set of parameters.
DistributedDataParallelin conjunction with the Distributed RPC Framework is experimental and subject to change.
module (Module) – module to be parallelized
device_ids (list of python:int or torch.device) –
CUDA devices. 1) For single-device modules,
device_idscan contain exactly one device id, which represents the only CUDA device where the input module corresponding to this process resides. Alternatively,
device_idscan also be
None. 2) For multi-device modules and CPU modules,
Nonefor both cases, both the input data for the forward pass and the actual module must be placed on the correct device. (default:
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_idsfor single-device modules)
broadcast_buffers (bool) – Flag that enables syncing (broadcasting) buffers of the module at beginning of the
process_group – The process group to be used for distributed data all-reduction. If
None, the default process group, which is created by
torch.distributed.init_process_group(), will be used. (default:
DistributedDataParallelwill bucket parameters into multiple buckets so that gradient reduction of each bucket can potentially overlap with backward computation.
bucket_cap_mbcontrols 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
forwardfunction. 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
forwardfunction but were not part of loss computation and thus would also not receive gradients are preemptively marked as ready to be reduced. (default:
check_reduction – This argument is deprecated.
gradient_as_bucket_view (bool) – When set to
True, gradients will be views pointing to different offsets of
allreducecommunication 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
allreducecommunication buckets. When gradients are views,
detach_()cannot be called on the gradients. If hitting such errors, please fix it by referring to the
torch/optim/optimizer.pyas a solution.
~DistributedDataParallel.module (Module) – the module to be parallelized.
>>> torch.distributed.init_process_group(backend='nccl', world_size=4, init_method='...') >>> net = torch.nn.parallel.DistributedDataParallel(model, pg)
join(divide_by_initial_world_size=True, enable=True, throw_on_early_termination=False)¶
A context manager to be used in conjunction with an instance of
torch.nn.parallel.DistributedDataParallelto 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_terminationis 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.
If the model or training loop this context manager is wrapped around has additional distributed collective operations, such as
SyncBatchNormin the model’s forward pass, then the flag
throw_on_early_terminationmust 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.
divide_by_initial_world_size (bool) – If
True, will divide gradients by the initial
world_sizeDDP 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=Trueto 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_sizeeven 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_sizealthough it also means the uneven inputs would contribute more towards the global gradient. Typically, you would want to set this to
Truefor 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
Falsemight provide better results.
enable (bool) – Whether to enable uneven input detection or not. Pass in
enable=Falseto disable in cases where you know that inputs are even across participating processes. Default is
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_sizewould be ignored. Default is
>>> 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().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)
Returns the DDP join hook, which enables training on uneven inputs by shadowing the collective communications in the forward and backward passes.
kwargs (dict) – a
dictcontaining any keyword arguments to modify the behavior of the join hook at run time; all
Joinableinstances sharing the same join context manager are forwarded the same value for
- The hook supports the following keyword arguments:
- divide_by_initial_world_size (bool, optional):
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
Trueif the degree of unevenness is small but can be set to
Falsein extreme cases for possibly better results. Default is
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.
>>> 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
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.
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_futureto retrieve a Future associated with the completion of
get_futureis currently supported for NCCL and also supported for most operations on GLOO and MPI, except for peer to peer operations (send/recv).
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.
DDP communication hook can only be registered once and should be registered before calling backward.
The Future object that hook returns should contain a single tensor that has the same shape with the tensors inside grad bucket.
get_futureAPI supports NCCL, and partially GLOO and MPI backends (no support for peer-to-peer operations like send/recv) and will return a
Below is an example of a noop hook that returns the same tensor.
>>> 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)
Below is an example of a Parallel SGD algorithm where gradients are encoded before allreduce, and then decoded after allreduce.
>>> 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()) # decode gradients >>> return decoded_tensor >>> return fut.then(decode)
>>> ddp.register_comm_hook(state=None, hook=encode_and_decode)