PyTorch Distributed Overview¶
Author: Shen Li
This is the overview page for the
torch.distributed package. The goal of
this page is to categorize documents into different topics and briefly
describe each of them. If this is your first time building distributed training
applications using PyTorch, it is recommended to use this document to navigate
to the technology that can best serve your use case.
As of PyTorch v1.6.0, features in
torch.distributed can be categorized into
three main components:
Distributed Data-Parallel Training (DDP) is a widely adopted single-program multiple-data training paradigm. With DDP, the model is replicated on every process, and every model replica will be fed with a different set of input data samples. DDP takes care of gradient communication to keep model replicas synchronized and overlaps it with the gradient computations to speed up training.
RPC-Based Distributed Training (RPC) supports general training structures that cannot fit into data-parallel training such as distributed pipeline parallelism, parameter server paradigm, and combinations of DDP with other training paradigms. It helps manage remote object lifetime and extends the autograd engine beyond machine boundaries.
Collective Communication (c10d) library supports sending tensors across processes within a group. It offers both collective communication APIs (e.g., all_reduce and all_gather) and P2P communication APIs (e.g., send and isend). DDP and RPC (ProcessGroup Backend) are built on c10d, where the former uses collective communications and the latter uses P2P communications. Usually, developers do not need to directly use this raw communication API, as the DDP and RPC APIs can serve many distributed training scenarios. However, there are use cases where this API is still helpful. One example would be distributed parameter averaging, where applications would like to compute the average values of all model parameters after the backward pass instead of using DDP to communicate gradients. This can decouple communications from computations and allow finer-grain control over what to communicate, but on the other hand, it also gives up the performance optimizations offered by DDP. The Writing Distributed Applications with PyTorch shows examples of using c10d communication APIs.
Data Parallel Training¶
PyTorch provides several options for data-parallel training. For applications that gradually grow from simple to complex and from prototype to production, the common development trajectory would be:
Use single-device training if the data and model can fit in one GPU, and training speed is not a concern.
Use single-machine multi-GPU DataParallel to make use of multiple GPUs on a single machine to speed up training with minimal code changes.
Use single-machine multi-GPU DistributedDataParallel, if you would like to further speed up training and are willing to write a little more code to set it up.
Use torch.distributed.elastic to launch distributed training if errors (e.g., out-of-memory) are expected or if resources can join and leave dynamically during training.
Data-parallel training also works with Automatic Mixed Precision (AMP).
package enables single-machine multi-GPU parallelism with the lowest coding
hurdle. It only requires a one-line change to the application code. The tutorial
Optional: Data Parallelism
shows an example. Although
DataParallel is very easy to
use, it usually does not offer the best performance because it replicates the
model in every forward pass, and its single-process multi-thread parallelism
naturally suffers from
GIL contention. To get
better performance, consider using
Compared to DataParallel, DistributedDataParallel requires one more step to set up, i.e., calling init_process_group. DDP uses multi-process parallelism, and hence there is no GIL contention across model replicas. Moreover, the model is broadcast at DDP construction time instead of in every forward pass, which also helps to speed up training. DDP is shipped with several performance optimization technologies. For a more in-depth explanation, refer to this paper (VLDB’20).
DDP materials are listed below:
DDP notes offer a starter example and some brief descriptions of its design and implementation. If this is your first time using DDP, start from this document.
Getting Started with Distributed Data Parallel explains some common problems with DDP training, including unbalanced workload, checkpointing, and multi-device models. Note that, DDP can be easily combined with single-machine multi-device model parallelism which is described in the Single-Machine Model Parallel Best Practices tutorial.
The Launching and configuring distributed data parallel applications document shows how to use the DDP launching script.
The Shard Optimizer States With ZeroRedundancyOptimizer recipe demonstrates how ZeroRedundancyOptimizer helps to reduce optimizer memory footprint.
The Distributed Training with Uneven Inputs Using the Join Context Manager tutorial walks through using the generic join context for distributed training with uneven inputs.
With the growth of the application complexity and scale, failure recovery
becomes a requirement. Sometimes it is inevitable to hit errors
like out-of-memory (OOM) when using DDP, but DDP itself cannot recover from those errors,
and it is not possible to handle them using a standard
This is because DDP requires all processes to operate in a closely synchronized manner
AllReduce communications launched in different processes must match.
If one of the processes in the group
throws an exception, it is likely to lead to desynchronization (mismatched
AllReduce operations) which would then cause a crash or hang.
adds fault tolerance and the ability to make use of a dynamic pool of machines (elasticity).
RPC-Based Distributed Training¶
Many training paradigms do not fit into data parallelism, e.g., parameter server paradigm, distributed pipeline parallelism, reinforcement learning applications with multiple observers or agents, etc. torch.distributed.rpc aims at supporting general distributed training scenarios.
torch.distributed.rpc has four main pillars:
RPC supports running a given function on a remote worker.
Distributed Optimizer automatically reaches out to all participating workers to update parameters using gradients computed by the distributed autograd engine.
RPC Tutorials are listed below:
The Getting Started with Distributed RPC Framework tutorial first uses a simple Reinforcement Learning (RL) example to demonstrate RPC and RRef. Then, it applies a basic distributed model parallelism to an RNN example to show how to use distributed autograd and distributed optimizer.
The Implementing a Parameter Server Using Distributed RPC Framework tutorial borrows the spirit of HogWild! training and applies it to an asynchronous parameter server (PS) training application.
The Distributed Pipeline Parallelism Using RPC tutorial extends the single-machine pipeline parallel example (presented in Single-Machine Model Parallel Best Practices) to a distributed environment and shows how to implement it using RPC.
The Implementing Batch RPC Processing Using Asynchronous Executions tutorial demonstrates how to implement RPC batch processing using the @rpc.functions.async_execution decorator, which can help speed up inference and training. It uses RL and PS examples similar to those in the above tutorials 1 and 2.
The Combining Distributed DataParallel with Distributed RPC Framework tutorial demonstrates how to combine DDP with RPC to train a model using distributed data parallelism combined with distributed model parallelism.