Distributed communication package - torch.distributed

Note

Please refer to PyTorch Distributed Overview for a brief introduction to all features related to distributed training.

Backends

torch.distributed supports three built-in backends, each with different capabilities. The table below shows which functions are available for use with CPU / CUDA tensors. MPI supports CUDA only if the implementation used to build PyTorch supports it.

Backend	gloo		mpi		nccl
Device	CPU	GPU	CPU	GPU	CPU	GPU
send	✓	✘	✓	?	✘	✘
recv	✓	✘	✓	?	✘	✘
broadcast	✓	✓	✓	?	✘	✓
all_reduce	✓	✓	✓	?	✘	✓
reduce	✓	✘	✓	?	✘	✓
all_gather	✓	✘	✓	?	✘	✓
gather	✓	✘	✓	?	✘	✘
scatter	✓	✘	✓	?	✘	✘
reduce_scatter	✘	✘	✘	✘	✘	✓
all_to_all	✘	✘	✓	?	✘	✘
barrier	✓	✘	✓	?	✘	✓

Backends that come with PyTorch

PyTorch distributed package supports Linux (stable), MacOS (stable), and Windows (prototype). By default for Linux, the Gloo and NCCL backends are built and included in PyTorch distributed (NCCL only when building with CUDA). MPI is an optional backend that can only be included if you build PyTorch from source. (e.g.building PyTorch on a host that has MPI installed.)

Warning

As of PyTorch v1.7, Windows support for the distributed package only covers collective communications with Gloo backend, FileStore, and DistributedDataParallel. Therefore, the init_method argument in init_process_group() must point to a file. This works for both local and shared file systems:

• Local file system, init_method="file:///d:/tmp/some_file"

• Shared file system, init_method="file://////{machine_name}/{share_folder_name}/some_file"

Similarly, if you directly pass in a store argument, it must be a FileStore instance.

Which backend to use?

In the past, we were often asked: “which backend should I use?”.

• Rule of thumb

  • Use the NCCL backend for distributed GPU training

  • Use the Gloo backend for distributed CPU training.

• GPU hosts with InfiniBand interconnect

  • Use NCCL, since it’s the only backend that currently supports InfiniBand and GPUDirect.

• GPU hosts with Ethernet interconnect

  • Use NCCL, since it currently provides the best distributed GPU training performance, especially for multiprocess single-node or multi-node distributed training. If you encounter any problem with NCCL, use Gloo as the fallback option. (Note that Gloo currently runs slower than NCCL for GPUs.)

• CPU hosts with InfiniBand interconnect

  • If your InfiniBand has enabled IP over IB, use Gloo, otherwise, use MPI instead. We are planning on adding InfiniBand support for Gloo in the upcoming releases.

• CPU hosts with Ethernet interconnect

  • Use Gloo, unless you have specific reasons to use MPI.

Common environment variables

Choosing the network interface to use

By default, both the NCCL and Gloo backends will try to find the right network interface to use. If the automatically detected interface is not correct, you can override it using the following environment variables (applicable to the respective backend):

• NCCL_SOCKET_IFNAME, for example export NCCL_SOCKET_IFNAME=eth0

• GLOO_SOCKET_IFNAME, for example export GLOO_SOCKET_IFNAME=eth0

If you’re using the Gloo backend, you can specify multiple interfaces by separating them by a comma, like this: export GLOO_SOCKET_IFNAME=eth0,eth1,eth2,eth3. The backend will dispatch operations in a round-robin fashion across these interfaces. It is imperative that all processes specify the same number of interfaces in this variable.

Other NCCL environment variables

NCCL has also provided a number of environment variables for fine-tuning purposes.

Commonly used ones include the following for debugging purposes:

• export NCCL_DEBUG=INFO

• export NCCL_DEBUG_SUBSYS=ALL

For the full list of NCCL environment variables, please refer to NVIDIA NCCL’s official documentation

Basics

The torch.distributed package provides PyTorch support and communication primitives for multiprocess parallelism across several computation nodes running on one or more machines. The class torch.nn.parallel.DistributedDataParallel() builds on this functionality to provide synchronous distributed training as a wrapper around any PyTorch model. This differs from the kinds of parallelism provided by Multiprocessing package - torch.multiprocessing and torch.nn.DataParallel() in that it supports multiple network-connected machines and in that the user must explicitly launch a separate copy of the main training script for each process.

In the single-machine synchronous case, torch.distributed or the torch.nn.parallel.DistributedDataParallel() wrapper may still have advantages over other approaches to data-parallelism, including torch.nn.DataParallel():

• Each process maintains its own optimizer and performs a complete optimization step with each iteration. While this may appear redundant, since the gradients have already been gathered together and averaged across processes and are thus the same for every process, this means that no parameter broadcast step is needed, reducing time spent transferring tensors between nodes.

• Each process contains an independent Python interpreter, eliminating the extra interpreter overhead and “GIL-thrashing” that comes from driving several execution threads, model replicas, or GPUs from a single Python process. This is especially important for models that make heavy use of the Python runtime, including models with recurrent layers or many small components.

Initialization

The package needs to be initialized using the torch.distributed.init_process_group() function before calling any other methods. This blocks until all processes have joined.

torch.distributed.is_available()

Returns True if the distributed package is available. Otherwise, torch.distributed does not expose any other APIs. Currently, torch.distributed is available on Linux, MacOS and Windows. Set USE_DISTRIBUTED=1 to enable it when building PyTorch from source. Currently, the default value is USE_DISTRIBUTED=1 for Linux and Windows, USE_DISTRIBUTED=0 for MacOS.

---

Currently three initialization methods are supported:

TCP initialization

There are two ways to initialize using TCP, both requiring a network address reachable from all processes and a desired world_size. The first way requires specifying an address that belongs to the rank 0 process. This initialization method requires that all processes have manually specified ranks.

Note that multicast address is not supported anymore in the latest distributed package. group_name is deprecated as well.

import torch.distributed as dist

# Use address of one of the machines
dist.init_process_group(backend, init_method='tcp://10.1.1.20:23456',
                        rank=args.rank, world_size=4)

Shared file-system initialization

Another initialization method makes use of a file system that is shared and visible from all machines in a group, along with a desired world_size. The URL should start with file:// and contain a path to a non-existent file (in an existing directory) on a shared file system. File-system initialization will automatically create that file if it doesn’t exist, but will not delete the file. Therefore, it is your responsibility to make sure that the file is cleaned up before the next init_process_group() call on the same file path/name.

Note that automatic rank assignment is not supported anymore in the latest distributed package and group_name is deprecated as well.

Warning

This method assumes that the file system supports locking using fcntl - most local systems and NFS support it.

Warning

This method will always create the file and try its best to clean up and remove the file at the end of the program. In other words, each initialization with the file init method will need a brand new empty file in order for the initialization to succeed. If the same file used by the previous initialization (which happens not to get cleaned up) is used again, this is unexpected behavior and can often cause deadlocks and failures. Therefore, even though this method will try its best to clean up the file, if the auto-delete happens to be unsuccessful, it is your responsibility to ensure that the file is removed at the end of the training to prevent the same file to be reused again during the next time. This is especially important if you plan to call init_process_group() multiple times on the same file name. In other words, if the file is not removed/cleaned up and you call init_process_group() again on that file, failures are expected. The rule of thumb here is that, make sure that the file is non-existent or empty every time init_process_group() is called.

import torch.distributed as dist

# rank should always be specified
dist.init_process_group(backend, init_method='file:///mnt/nfs/sharedfile',
                        world_size=4, rank=args.rank)

Environment variable initialization

This method will read the configuration from environment variables, allowing one to fully customize how the information is obtained. The variables to be set are:

• MASTER_PORT - required; has to be a free port on machine with rank 0

• MASTER_ADDR - required (except for rank 0); address of rank 0 node

• WORLD_SIZE - required; can be set either here, or in a call to init function

• RANK - required; can be set either here, or in a call to init function

The machine with rank 0 will be used to set up all connections.

This is the default method, meaning that init_method does not have to be specified (or can be env://).

Distributed Key-Value Store

The distributed package comes with a distributed key-value store, which can be used to share information between processes in the group as well as to initialize the distributed pacakge in torch.distributed.init_process_group() (by explicitly creating the store as an alternative to specifying init_method.) There are 3 choices for Key-Value Stores: TCPStore, FileStore, and HashStore.

Groups

By default collectives operate on the default group (also called the world) and require all processes to enter the distributed function call. However, some workloads can benefit from more fine-grained communication. This is where distributed groups come into play. new_group() function can be used to create new groups, with arbitrary subsets of all processes. It returns an opaque group handle that can be given as a group argument to all collectives (collectives are distributed functions to exchange information in certain well-known programming patterns).

Point-to-point communication

isend() and irecv() return distributed request objects when used. In general, the type of this object is unspecified as they should never be created manually, but they are guaranteed to support two methods:

• is_completed() - returns True if the operation has finished

• wait() - will block the process until the operation is finished. is_completed() is guaranteed to return True once it returns.

Synchronous and asynchronous collective operations

Every collective operation function supports the following two kinds of operations, depending on the setting of the async_op flag passed into the collective:

Synchronous operation - the default mode, when async_op is set to False. When the function returns, it is guaranteed that the collective operation is performed. In the case of CUDA operations, it is not guaranteed that the CUDA operation is completed, since CUDA operations are asynchronous. For CPU collectives, any further function calls utilizing the output of the collective call will behave as expected. For CUDA collectives, function calls utilizing the output on the same CUDA stream will behave as expected. Users must take care of synchronization under the scenario of running under different streams. For details on CUDA semantics such as stream synchronization, see CUDA Semantics. See the below script to see examples of differences in these semantics for CPU and CUDA operations.

Asynchronous operation - when async_op is set to True. The collective operation function returns a distributed request object. In general, you don’t need to create it manually and it is guaranteed to support two methods:

• is_completed() - in the case of CPU collectives, returns True if completed. In the case of CUDA operations, returns True if the operation has been successfully enqueued onto a CUDA stream and the output can be utilized on the default stream without further synchronization.

• wait() - in the case of CPU collectives, will block the process until the operation is completed. In the case of CUDA collectives, will block until the operation has been successfully enqueued onto a CUDA stream and the output can be utilized on the default stream without further synchronization.

Example

The following code can serve as a reference regarding semantics for CUDA operations when using distributed collectives. It shows the explicit need to synchronize when using collective outputs on different CUDA streams:

# Code runs on each rank.
dist.init_process_group("nccl", rank=rank, world_size=2)
output = torch.tensor([rank]).cuda(rank)
s = torch.cuda.Stream()
handle = dist.all_reduce(output, async_op=True)
# Wait ensures the operation is enqueued, but not necessarily complete.
handle.wait()
# Using result on non-default stream.
with torch.cuda.stream(s):
    s.wait_stream(torch.cuda.default_stream())
    output.add_(100)
if rank == 0:
    # if the explicit call to wait_stream was omitted, the output below will be
    # non-deterministically 1 or 101, depending on whether the allreduce overwrote
    # the value after the add completed.
    print(output)

Collective functions

class torch.distributed.reduce_op

Deprecated enum-like class for reduction operations: SUM, PRODUCT, MIN, and MAX.

ReduceOp is recommended to use instead.

Autograd-enabled communication primitives

If you want to use collective communication functions supporting autograd you can find an implementation of those in the torch.distributed.nn.* module.

Functions here are synchronous and will be inserted in the autograd graph, so you need to ensure that all the processes that participated in the collective operation will do the backward pass for the backward communication to effectively happen and don’t cause a deadlock.

Please notice that currently the only backend where all the functions are guaranteed to work is gloo. .. autofunction:: torch.distributed.nn.broadcast .. autofunction:: torch.distributed.nn.gather .. autofunction:: torch.distributed.nn.scatter .. autofunction:: torch.distributed.nn.reduce .. autofunction:: torch.distributed.nn.all_gather .. autofunction:: torch.distributed.nn.all_to_all .. autofunction:: torch.distributed.nn.all_reduce

Multi-GPU collective functions

If you have more than one GPU on each node, when using the NCCL and Gloo backend, broadcast_multigpu() all_reduce_multigpu() reduce_multigpu() all_gather_multigpu() and reduce_scatter_multigpu() support distributed collective operations among multiple GPUs within each node. These functions can potentially improve the overall distributed training performance and be easily used by passing a list of tensors. Each Tensor in the passed tensor list needs to be on a separate GPU device of the host where the function is called. Note that the length of the tensor list needs to be identical among all the distributed processes. Also note that currently the multi-GPU collective functions are only supported by the NCCL backend.

For example, if the system we use for distributed training has 2 nodes, each of which has 8 GPUs. On each of the 16 GPUs, there is a tensor that we would like to all-reduce. The following code can serve as a reference:

Code running on Node 0

import torch
import torch.distributed as dist

dist.init_process_group(backend="nccl",
                        init_method="file:///distributed_test",
                        world_size=2,
                        rank=0)
tensor_list = []
for dev_idx in range(torch.cuda.device_count()):
    tensor_list.append(torch.FloatTensor([1]).cuda(dev_idx))

dist.all_reduce_multigpu(tensor_list)

Code running on Node 1

import torch
import torch.distributed as dist

dist.init_process_group(backend="nccl",
                        init_method="file:///distributed_test",
                        world_size=2,
                        rank=1)
tensor_list = []
for dev_idx in range(torch.cuda.device_count()):
    tensor_list.append(torch.FloatTensor([1]).cuda(dev_idx))

dist.all_reduce_multigpu(tensor_list)

After the call, all 16 tensors on the two nodes will have the all-reduced value of 16

Third-party backends

Besides the GLOO/MPI/NCCL backends, PyTorch distributed supports third-party backends through a run-time register mechanism. For references on how to develop a third-party backend through C++ Extension, please refer to Tutorials - Custom C++ and CUDA Extensions and test/cpp_extensions/cpp_c10d_extension.cpp. The capability of third-party backends are decided by their own implementations.

The new backend derives from c10d.ProcessGroup and registers the backend name and the instantiating interface through torch.distributed.Backend.register_backend() when imported.

When manually importing this backend and invoking torch.distributed.init_process_group() with the corresponding backend name, the torch.distributed package runs on the new backend.

Warning

The support of third-party backend is experimental and subject to change.

Launch utility

The torch.distributed package also provides a launch utility in torch.distributed.launch. This helper utility can be used to launch multiple processes per node for distributed training.

torch.distributed.launch is a module that spawns up multiple distributed training processes on each of the training nodes.

The utility can be used for single-node distributed training, in which one or more processes per node will be spawned. The utility can be used for either CPU training or GPU training. If the utility is used for GPU training, each distributed process will be operating on a single GPU. This can achieve well-improved single-node training performance. It can also be used in multi-node distributed training, by spawning up multiple processes on each node for well-improved multi-node distributed training performance as well. This will especially be benefitial for systems with multiple Infiniband interfaces that have direct-GPU support, since all of them can be utilized for aggregated communication bandwidth.

In both cases of single-node distributed training or multi-node distributed training, this utility will launch the given number of processes per node (--nproc_per_node). If used for GPU training, this number needs to be less or equal to the number of GPUs on the current system (nproc_per_node), and each process will be operating on a single GPU from GPU 0 to GPU (nproc_per_node - 1).

How to use this module:

1. Single-Node multi-process distributed training

>>> python -m torch.distributed.launch --nproc_per_node=NUM_GPUS_YOU_HAVE
           YOUR_TRAINING_SCRIPT.py (--arg1 --arg2 --arg3 and all other
           arguments of your training script)

2. Multi-Node multi-process distributed training: (e.g. two nodes)

Node 1: (IP: 192.168.1.1, and has a free port: 1234)

>>> python -m torch.distributed.launch --nproc_per_node=NUM_GPUS_YOU_HAVE
           --nnodes=2 --node_rank=0 --master_addr="192.168.1.1"
           --master_port=1234 YOUR_TRAINING_SCRIPT.py (--arg1 --arg2 --arg3
           and all other arguments of your training script)

Node 2:

>>> python -m torch.distributed.launch --nproc_per_node=NUM_GPUS_YOU_HAVE
           --nnodes=2 --node_rank=1 --master_addr="192.168.1.1"
           --master_port=1234 YOUR_TRAINING_SCRIPT.py (--arg1 --arg2 --arg3
           and all other arguments of your training script)

3. To look up what optional arguments this module offers:

>>> python -m torch.distributed.launch --help

Important Notices:

1. This utility and multi-process distributed (single-node or multi-node) GPU training currently only achieves the best performance using the NCCL distributed backend. Thus NCCL backend is the recommended backend to use for GPU training.

2. In your training program, you must parse the command-line argument: --local_rank=LOCAL_PROCESS_RANK, which will be provided by this module. If your training program uses GPUs, you should ensure that your code only runs on the GPU device of LOCAL_PROCESS_RANK. This can be done by:

Parsing the local_rank argument

>>> import argparse
>>> parser = argparse.ArgumentParser()
>>> parser.add_argument("--local_rank", type=int)
>>> args = parser.parse_args()

Set your device to local rank using either

>>> torch.cuda.set_device(args.local_rank)  # before your code runs

or

>>> with torch.cuda.device(args.local_rank):
>>>    # your code to run

3. In your training program, you are supposed to call the following function at the beginning to start the distributed backend. You need to make sure that the init_method uses env://, which is the only supported init_method by this module.

torch.distributed.init_process_group(backend='YOUR BACKEND',
                                     init_method='env://')

4. In your training program, you can either use regular distributed functions or use torch.nn.parallel.DistributedDataParallel() module. If your training program uses GPUs for training and you would like to use torch.nn.parallel.DistributedDataParallel() module, here is how to configure it.

model = torch.nn.parallel.DistributedDataParallel(model,
                                                  device_ids=[args.local_rank],
                                                  output_device=args.local_rank)

Please ensure that device_ids argument is set to be the only GPU device id that your code will be operating on. This is generally the local rank of the process. In other words, the device_ids needs to be [args.local_rank], and output_device needs to be args.local_rank in order to use this utility

5. Another way to pass local_rank to the subprocesses via environment variable LOCAL_RANK. This behavior is enabled when you launch the script with --use_env=True. You must adjust the subprocess example above to replace args.local_rank with os.environ['LOCAL_RANK']; the launcher will not pass --local_rank when you specify this flag.

Warning

local_rank is NOT globally unique: it is only unique per process on a machine. Thus, don’t use it to decide if you should, e.g., write to a networked filesystem. See https://github.com/pytorch/pytorch/issues/12042 for an example of how things can go wrong if you don’t do this correctly.

Spawn utility

The Multiprocessing package - torch.multiprocessing package also provides a spawn function in torch.multiprocessing.spawn(). This helper function can be used to spawn multiple processes. It works by passing in the function that you want to run and spawns N processes to run it. This can be used for multiprocess distributed training as well.

For references on how to use it, please refer to PyTorch example - ImageNet implementation

Note that this function requires Python 3.4 or higher.