• Docs >
  • Distributed communication package - torch.distributed
Shortcuts

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.)

Note

As of PyTorch v1.8, Windows supports all collective communications backend but NCCL, If the init_method argument of init_process_group() points to a file it must adhere to the following schema:

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

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

Same as on Linux platform, you can enable TcpStore by setting environment variables, MASTER_ADDR and MASTER_PORT.

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

Debugging - in case of NCCL failure, you can set NCCL_DEBUG=INFO to print an explicit warning message as well as basic NCCL initialization information.

You may also use NCCL_DEBUG_SUBSYS to get more details about a specific aspect of NCCL. For example, NCCL_DEBUG_SUBSYS=COLL would print logs of collective calls, which may be helpful when debugging hangs, especially those caused by collective type or message size mismatch. In case of topology detection failure, it would be helpful to set NCCL_DEBUG_SUBSYS=GRAPH to inspect the detailed detection result and save as reference if further help from NCCL team is needed.

Performance tuning - NCCL performs automatic tuning based on its topology detection to save users’ tuning effort. On some socket-based systems, users may still try tuning NCCL_SOCKET_NTHREADS and NCCL_NSOCKS_PERTHREAD to increase socket network bandwidth. These two environment variables have been pre-tuned by NCCL for some cloud providers, such as AWS or GCP.

For a 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() or torch.distributed.device_mesh.init_device_mesh() function before calling any other methods. Both block until all processes have joined.

torch.distributed.is_available()[source]

Return 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.

Return type

bool

torch.distributed.init_process_group(backend=None, init_method=None, timeout=None, world_size=-1, rank=-1, store=None, group_name='', pg_options=None, device_id=None)[source]

Initialize the default distributed process group.

This will also initialize the distributed package.

There are 2 main ways to initialize a process group:
  1. Specify store, rank, and world_size explicitly.

  2. Specify init_method (a URL string) which indicates where/how to discover peers. Optionally specify rank and world_size, or encode all required parameters in the URL and omit them.

If neither is specified, init_method is assumed to be “env://”.

Parameters
  • backend (str or Backend, optional) – The backend to use. Depending on build-time configurations, valid values include mpi, gloo, nccl, and ucc. If the backend is not provided, then both a gloo and nccl backend will be created, see notes below for how multiple backends are managed. This field can be given as a lowercase string (e.g., "gloo"), which can also be accessed via Backend attributes (e.g., Backend.GLOO). If using multiple processes per machine with nccl backend, each process must have exclusive access to every GPU it uses, as sharing GPUs between processes can result in deadlocks. ucc backend is experimental.

  • init_method (str, optional) – URL specifying how to initialize the process group. Default is “env://” if no init_method or store is specified. Mutually exclusive with store.

  • world_size (int, optional) – Number of processes participating in the job. Required if store is specified.

  • rank (int, optional) – Rank of the current process (it should be a number between 0 and world_size-1). Required if store is specified.

  • store (Store, optional) – Key/value store accessible to all workers, used to exchange connection/address information. Mutually exclusive with init_method.

  • timeout (timedelta, optional) – Timeout for operations executed against the process group. Default value is 10 minutes for NCCL and 30 minutes for other backends. This is the duration after which collectives will be aborted asynchronously and the process will crash. This is done since CUDA execution is async and it is no longer safe to continue executing user code since failed async NCCL operations might result in subsequent CUDA operations running on corrupted data. When TORCH_NCCL_BLOCKING_WAIT is set, the process will block and wait for this timeout.

  • group_name (str, optional, deprecated) – Group name. This argument is ignored

  • pg_options (ProcessGroupOptions, optional) – process group options specifying what additional options need to be passed in during the construction of specific process groups. As of now, the only options we support is ProcessGroupNCCL.Options for the nccl backend, is_high_priority_stream can be specified so that the nccl backend can pick up high priority cuda streams when there’re compute kernels waiting.

  • device_id (torch.device, optional) – a single, specific device to “bind” this process to, allowing for backend-specific optimizations. Currently this has two effects, only under NCCL: the communicator is immediately formed (calling ncclCommInit* immediately rather than the normal lazy call) and sub-groups will use ncclCommSplit when possible to avoid unnecessary overhead of group creation. If you want to know NCCL initialization error early, you can also use this field.

Note

To enable backend == Backend.MPI, PyTorch needs to be built from source on a system that supports MPI.

Note

Support for multiple backends is experimental. Currently when no backend is specified, both gloo and nccl backends will be created. The gloo backend will be used for collectives with CPU tensors and the nccl backend will be used for collectives with CUDA tensors. A custom backend can be specified by passing in a string with format “<device_type>:<backend_name>,<device_type>:<backend_name>”, e.g. “cpu:gloo,cuda:custom_backend”.

torch.distributed.device_mesh.init_device_mesh(device_type, mesh_shape, *, mesh_dim_names=None)[source]

Initializes a DeviceMesh based on device_type, mesh_shape, and mesh_dim_names parameters.

This creates a DeviceMesh with an n-dimensional array layout, where n is the length of mesh_shape. If mesh_dim_names is provided, each dimension is labeled as mesh_dim_names[i].

Note

init_device_mesh follows SPMD programming model, meaning the same PyTorch Python program runs on all processes/ranks in the cluster. Ensure mesh_shape (the dimensions of the nD array describing device layout) is identical across all ranks. Inconsistent mesh_shape may lead to hanging.

Note

If no process group is found, init_device_mesh will initialize distributed process group/groups required for distributed communications behind the scene.

Parameters
  • device_type (str) – The device type of the mesh. Currently supports: “cpu”, “cuda/cuda-like”.

  • mesh_shape (Tuple[int]) – A tuple defining the dimensions of the multi-dimensional array describing the layout of devices.

  • mesh_dim_names (Tuple[str], optional) – A tuple of mesh dimension names to assign to each dimension of the multi-dimensional array describing the layout of devices. Its length must match the length of mesh_shape. Each string in mesh_dim_names must be unique.

Returns

A DeviceMesh object representing the device layout.

Return type

DeviceMesh

Example::
>>> from torch.distributed.device_mesh import init_device_mesh
>>>
>>> mesh_1d = init_device_mesh("cuda", mesh_shape=(8,))
>>> mesh_2d = init_device_mesh("cuda", mesh_shape=(2, 8), mesh_dim_names=("dp", "tp"))
torch.distributed.is_initialized()[source]

Check if the default process group has been initialized.

Return type

bool

torch.distributed.is_mpi_available()[source]

Check if the MPI backend is available.

Return type

bool

torch.distributed.is_nccl_available()[source]

Check if the NCCL backend is available.

Return type

bool

torch.distributed.is_gloo_available()[source]

Check if the Gloo backend is available.

Return type

bool

torch.distributed.is_torchelastic_launched()[source]

Check whether this process was launched with torch.distributed.elastic (aka torchelastic).

The existence of TORCHELASTIC_RUN_ID environment variable is used as a proxy to determine whether the current process was launched with torchelastic. This is a reasonable proxy since TORCHELASTIC_RUN_ID maps to the rendezvous id which is always a non-null value indicating the job id for peer discovery purposes..

Return type

bool


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://).

Post-Initialization

Once torch.distributed.init_process_group() was run, the following functions can be used. To check whether the process group has already been initialized use torch.distributed.is_initialized().

class torch.distributed.Backend(name)[source]

An enum-like class for backends.

Available backends: GLOO, NCCL, UCC, MPI, and other registered backends.

The values of this class are lowercase strings, e.g., "gloo". They can be accessed as attributes, e.g., Backend.NCCL.

This class can be directly called to parse the string, e.g., Backend(backend_str) will check if backend_str is valid, and return the parsed lowercase string if so. It also accepts uppercase strings, e.g., Backend("GLOO") returns "gloo".

Note

The entry Backend.UNDEFINED is present but only used as initial value of some fields. Users should neither use it directly nor assume its existence.

classmethod register_backend(name, func, extended_api=False, devices=None)[source]

Register a new backend with the given name and instantiating function.

This class method is used by 3rd party ProcessGroup extension to register new backends.

Parameters
  • name (str) – Backend name of the ProcessGroup extension. It should match the one in init_process_group().

  • func (function) – Function handler that instantiates the backend. The function should be implemented in the backend extension and takes four arguments, including store, rank, world_size, and timeout.

  • extended_api (bool, optional) – Whether the backend supports extended argument structure. Default: False. If set to True, the backend will get an instance of c10d::DistributedBackendOptions, and a process group options object as defined by the backend implementation.

  • device (str or list of str, optional) – device type this backend supports, e.g. “cpu”, “cuda”, etc. If None, assuming both “cpu” and “cuda”

Note

This support of 3rd party backend is experimental and subject to change.

torch.distributed.get_backend(group=None)[source]

Return the backend of the given process group.

Parameters

group (ProcessGroup, optional) – The process group to work on. The default is the general main process group. If another specific group is specified, the calling process must be part of group.

Returns

The backend of the given process group as a lower case string.

Return type

Backend

torch.distributed.get_rank(group=None)[source]

Return the rank of the current process in the provided group, default otherwise.

Rank is a unique identifier assigned to each process within a distributed process group. They are always consecutive integers ranging from 0 to world_size.

Parameters

group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

Returns

The rank of the process group -1, if not part of the group

Return type

int

torch.distributed.get_world_size(group=None)[source]

Return the number of processes in the current process group.

Parameters

group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

Returns

The world size of the process group -1, if not part of the group

Return type

int


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 package 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.

class torch.distributed.Store

Base class for all store implementations, such as the 3 provided by PyTorch distributed: (TCPStore, FileStore, and HashStore).

class torch.distributed.TCPStore

A TCP-based distributed key-value store implementation. The server store holds the data, while the client stores can connect to the server store over TCP and perform actions such as set() to insert a key-value pair, get() to retrieve a key-value pair, etc. There should always be one server store initialized because the client store(s) will wait for the server to establish a connection.

Parameters
  • host_name (str) – The hostname or IP Address the server store should run on.

  • port (int) – The port on which the server store should listen for incoming requests.

  • world_size (int, optional) – The total number of store users (number of clients + 1 for the server). Default is None (None indicates a non-fixed number of store users).

  • is_master (bool, optional) – True when initializing the server store and False for client stores. Default is False.

  • timeout (timedelta, optional) – Timeout used by the store during initialization and for methods such as get() and wait(). Default is timedelta(seconds=300)

  • wait_for_workers (bool, optional) – Whether to wait for all the workers to connect with the server store. This is only applicable when world_size is a fixed value. Default is True.

  • multi_tenant (bool, optional) – If True, all TCPStore instances in the current process with the same host/port will use the same underlying TCPServer. Default is False.

  • master_listen_fd (int, optional) – If specified, the underlying TCPServer will listen on this file descriptor, which must be a socket already bound to port. Useful to avoid port assignment races in some scenarios. Default is None (meaning the server creates a new socket and attempts to bind it to port).

Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> # Run on process 1 (server)
>>> server_store = dist.TCPStore("127.0.0.1", 1234, 2, True, timedelta(seconds=30))
>>> # Run on process 2 (client)
>>> client_store = dist.TCPStore("127.0.0.1", 1234, 2, False)
>>> # Use any of the store methods from either the client or server after initialization
>>> server_store.set("first_key", "first_value")
>>> client_store.get("first_key")
class torch.distributed.HashStore

A thread-safe store implementation based on an underlying hashmap. This store can be used within the same process (for example, by other threads), but cannot be used across processes.

Example::
>>> import torch.distributed as dist
>>> store = dist.HashStore()
>>> # store can be used from other threads
>>> # Use any of the store methods after initialization
>>> store.set("first_key", "first_value")
class torch.distributed.FileStore

A store implementation that uses a file to store the underlying key-value pairs.

Parameters
  • file_name (str) – path of the file in which to store the key-value pairs

  • world_size (int, optional) – The total number of processes using the store. Default is -1 (a negative value indicates a non-fixed number of store users).

Example::
>>> import torch.distributed as dist
>>> store1 = dist.FileStore("/tmp/filestore", 2)
>>> store2 = dist.FileStore("/tmp/filestore", 2)
>>> # Use any of the store methods from either the client or server after initialization
>>> store1.set("first_key", "first_value")
>>> store2.get("first_key")
class torch.distributed.PrefixStore

A wrapper around any of the 3 key-value stores (TCPStore, FileStore, and HashStore) that adds a prefix to each key inserted to the store.

Parameters
  • prefix (str) – The prefix string that is prepended to each key before being inserted into the store.

  • store (torch.distributed.store) – A store object that forms the underlying key-value store.

torch.distributed.Store.set(self: torch._C._distributed_c10d.Store, arg0: str, arg1: str) None

Inserts the key-value pair into the store based on the supplied key and value. If key already exists in the store, it will overwrite the old value with the new supplied value.

Parameters
  • key (str) – The key to be added to the store.

  • value (str) – The value associated with key to be added to the store.

Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> store = dist.TCPStore("127.0.0.1", 0, 1, True, timedelta(seconds=30))
>>> store.set("first_key", "first_value")
>>> # Should return "first_value"
>>> store.get("first_key")
torch.distributed.Store.get(self: torch._C._distributed_c10d.Store, arg0: str) bytes

Retrieves the value associated with the given key in the store. If key is not present in the store, the function will wait for timeout, which is defined when initializing the store, before throwing an exception.

Parameters

key (str) – The function will return the value associated with this key.

Returns

Value associated with key if key is in the store.

Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> store = dist.TCPStore("127.0.0.1", 0, 1, True, timedelta(seconds=30))
>>> store.set("first_key", "first_value")
>>> # Should return "first_value"
>>> store.get("first_key")
torch.distributed.Store.add(self: torch._C._distributed_c10d.Store, arg0: str, arg1: int) int

The first call to add for a given key creates a counter associated with key in the store, initialized to amount. Subsequent calls to add with the same key increment the counter by the specified amount. Calling add() with a key that has already been set in the store by set() will result in an exception.

Parameters
  • key (str) – The key in the store whose counter will be incremented.

  • amount (int) – The quantity by which the counter will be incremented.

Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> # Using TCPStore as an example, other store types can also be used
>>> store = dist.TCPStore("127.0.0.1", 0, 1, True, timedelta(seconds=30))
>>> store.add("first_key", 1)
>>> store.add("first_key", 6)
>>> # Should return 7
>>> store.get("first_key")
torch.distributed.Store.compare_set(self: torch._C._distributed_c10d.Store, arg0: str, arg1: str, arg2: str) bytes

Inserts the key-value pair into the store based on the supplied key and performs comparison between expected_value and desired_value before inserting. desired_value will only be set if expected_value for the key already exists in the store or if expected_value is an empty string.

Parameters
  • key (str) – The key to be checked in the store.

  • expected_value (str) – The value associated with key to be checked before insertion.

  • desired_value (str) – The value associated with key to be added to the store.

Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> store = dist.TCPStore("127.0.0.1", 0, 1, True, timedelta(seconds=30))
>>> store.set("key", "first_value")
>>> store.compare_set("key", "first_value", "second_value")
>>> # Should return "second_value"
>>> store.get("key")
torch.distributed.Store.wait(*args, **kwargs)

Overloaded function.

  1. wait(self: torch._C._distributed_c10d.Store, arg0: list[str]) -> None

Waits for each key in keys to be added to the store. If not all keys are set before the timeout (set during store initialization), then wait will throw an exception.

Parameters

keys (list) – List of keys on which to wait until they are set in the store.

Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> # Using TCPStore as an example, other store types can also be used
>>> store = dist.TCPStore("127.0.0.1", 0, 1, True, timedelta(seconds=30))
>>> # This will throw an exception after 30 seconds
>>> store.wait(["bad_key"])
  1. wait(self: torch._C._distributed_c10d.Store, arg0: list[str], arg1: datetime.timedelta) -> None

Waits for each key in keys to be added to the store, and throws an exception if the keys have not been set by the supplied timeout.

Parameters
  • keys (list) – List of keys on which to wait until they are set in the store.

  • timeout (timedelta) – Time to wait for the keys to be added before throwing an exception.

Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> # Using TCPStore as an example, other store types can also be used
>>> store = dist.TCPStore("127.0.0.1", 0, 1, True, timedelta(seconds=30))
>>> # This will throw an exception after 10 seconds
>>> store.wait(["bad_key"], timedelta(seconds=10))
torch.distributed.Store.num_keys(self: torch._C._distributed_c10d.Store) int

Returns the number of keys set in the store. Note that this number will typically be one greater than the number of keys added by set() and add() since one key is used to coordinate all the workers using the store.

Warning

When used with the TCPStore, num_keys returns the number of keys written to the underlying file. If the store is destructed and another store is created with the same file, the original keys will be retained.

Returns

The number of keys present in the store.

Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> # Using TCPStore as an example, other store types can also be used
>>> store = dist.TCPStore("127.0.0.1", 0, 1, True, timedelta(seconds=30))
>>> store.set("first_key", "first_value")
>>> # This should return 2
>>> store.num_keys()
torch.distributed.Store.delete_key(self: torch._C._distributed_c10d.Store, arg0: str) bool

Deletes the key-value pair associated with key from the store. Returns true if the key was successfully deleted, and false if it was not.

Warning

The delete_key API is only supported by the TCPStore and HashStore. Using this API with the FileStore will result in an exception.

Parameters

key (str) – The key to be deleted from the store

Returns

True if key was deleted, otherwise False.

Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> # Using TCPStore as an example, HashStore can also be used
>>> store = dist.TCPStore("127.0.0.1", 0, 1, True, timedelta(seconds=30))
>>> store.set("first_key")
>>> # This should return true
>>> store.delete_key("first_key")
>>> # This should return false
>>> store.delete_key("bad_key")
torch.distributed.Store.set_timeout(self: torch._C._distributed_c10d.Store, arg0: datetime.timedelta) None

Sets the store’s default timeout. This timeout is used during initialization and in wait() and get().

Parameters

timeout (timedelta) – timeout to be set in the store.

Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> # Using TCPStore as an example, other store types can also be used
>>> store = dist.TCPStore("127.0.0.1", 0, 1, True, timedelta(seconds=30))
>>> store.set_timeout(timedelta(seconds=10))
>>> # This will throw an exception after 10 seconds
>>> store.wait(["bad_key"])

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).

torch.distributed.new_group(ranks=None, timeout=None, backend=None, pg_options=None, use_local_synchronization=False)[source]

Create a new distributed group.

This function requires that all processes in the main group (i.e. all processes that are part of the distributed job) enter this function, even if they are not going to be members of the group. Additionally, groups should be created in the same order in all processes.

Warning

Using multiple process groups with the NCCL backend concurrently is not safe and the user should perform explicit synchronization in their application to ensure only one process group is used at a time. This means collectives from one process group should have completed execution on the device (not just enqueued since CUDA execution is async) before collectives from another process group are enqueued. See Using multiple NCCL communicators concurrently for more details.

Parameters
  • ranks (list[int]) – List of ranks of group members. If None, will be set to all ranks. Default is None.

  • timeout (timedelta, optional) – see init_process_group for details and default value.

  • backend (str or Backend, optional) – The backend to use. Depending on build-time configurations, valid values are gloo and nccl. By default uses the same backend as the global group. This field should be given as a lowercase string (e.g., "gloo"), which can also be accessed via Backend attributes (e.g., Backend.GLOO). If None is passed in, the backend corresponding to the default process group will be used. Default is None.

  • pg_options (ProcessGroupOptions, optional) – process group options specifying what additional options need to be passed in during the construction of specific process groups. i.e. for the nccl backend, is_high_priority_stream can be specified so that process group can pick up high priority cuda streams.

  • use_local_synchronization (bool, optional) – perform a group-local barrier at the end of the process group creation. This is different in that non-member ranks don’t need to call into API and don’t join the barrier.

Returns

A handle of distributed group that can be given to collective calls or None if the rank is not part of ranks.

N.B. use_local_synchronization doesn’t work with MPI.

N.B. While use_local_synchronization=True can be significantly faster with larger clusters and small process groups, care must be taken since it changes cluster behavior as non-member ranks don’t join the group barrier().

N.B. use_local_synchronization=True can lead to deadlocks when each rank creates multiple overlaping process groups. To avoid that, make sure all ranks follow the same global creation order.

torch.distributed.get_group_rank(group, global_rank)[source]

Translate a global rank into a group rank.

global_rank must be part of group otherwise this raises RuntimeError.

Parameters
  • group (ProcessGroup) – ProcessGroup to find the relative rank.

  • global_rank (int) – Global rank to query.

Returns

Group rank of global_rank relative to group

Return type

int

N.B. calling this function on the default process group returns identity

torch.distributed.get_global_rank(group, group_rank)[source]

Translate a group rank into a global rank.

group_rank must be part of group otherwise this raises RuntimeError.

Parameters
  • group (ProcessGroup) – ProcessGroup to find the global rank from.

  • group_rank (int) – Group rank to query.

Returns

Global rank of group_rank relative to group

Return type

int

N.B. calling this function on the default process group returns identity

torch.distributed.get_process_group_ranks(group)[source]

Get all ranks associated with group.

Parameters

group (ProcessGroup) – ProcessGroup to get all ranks from.

Returns

List of global ranks ordered by group rank.

Return type

List[int]

DeviceMesh

DeviceMesh is a higher level abstraction that manages process groups (or NCCL communicators). It allows user to easily create inter node and intra node process groups without worrying about how to set up the ranks correctly for different sub process groups, and it helps manage those distributed process group easily. init_device_mesh() function can be used to create new DeviceMesh, with a mesh shape describing the device topology.

class torch.distributed.device_mesh.DeviceMesh(device_type, mesh, *, mesh_dim_names=None, _init_backend=True)[source]

DeviceMesh represents a mesh of devices, where layout of devices could be represented as a n-d dimension array, and each value of the n-d dimensional array is the global id of the default process group ranks.

DeviceMesh could be used to describe the layout of devices across the cluster, and serves as a proxy for communication among the device lists within the cluster.

DeviceMesh can be used as a context manager.

Note

DeviceMesh follows SPMD programming model, which means the same PyTorch Python program is running on all processes/ranks in the cluster. Therefore, users need to make sure the mesh array (which describes the layout of devices) should be identical across all ranks. Inconsistent mesh will lead to silent hang.

Parameters
  • device_type (str) – The device type of the mesh. Currently supports: “cpu”, “cuda/cuda-like”.

  • mesh (ndarray) – A multi-dimensional array or an integer tensor describing the layout of devices, where the IDs are global IDs of the default process group.

Returns

A DeviceMesh object representing the device layout.

Return type

DeviceMesh

The following program runs on each process/rank in an SPMD manner. In this example, we have 2 hosts with 4 GPUs each. A reduction over the first dimension of mesh will reduce across columns (0, 4), .. and (3, 7), a reduction over the second dimension of mesh reduces across rows (0, 1, 2, 3) and (4, 5, 6, 7).

Example::
>>> from torch.distributed.device_mesh import DeviceMesh
>>>
>>> # Initialize device mesh as (2, 4) to represent the topology
>>> # of cross-host(dim 0), and within-host (dim 1).
>>> mesh = DeviceMesh(device_type="cuda", mesh=[[0, 1, 2, 3],[4, 5, 6, 7]])

Point-to-point communication

torch.distributed.send(tensor, dst, group=None, tag=0)[source]

Send a tensor synchronously.

Parameters
  • tensor (Tensor) – Tensor to send.

  • dst (int) – Destination rank on global process group (regardless of group argument). Destination rank should not be the same as the rank of the current process.

  • group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

  • tag (int, optional) – Tag to match send with remote recv

torch.distributed.recv(tensor, src=None, group=None, tag=0)[source]

Receives a tensor synchronously.

Parameters
  • tensor (Tensor) – Tensor to fill with received data.

  • src (int, optional) – Source rank on global process group (regardless of group argument). Will receive from any process if unspecified.

  • group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

  • tag (int, optional) – Tag to match recv with remote send

Returns

Sender rank -1, if not part of the group

Return type

int

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.

torch.distributed.isend(tensor, dst, group=None, tag=0)[source]

Send a tensor asynchronously.

Warning

Modifying tensor before the request completes causes undefined behavior.

Warning

tag is not supported with the NCCL backend.

Parameters
  • tensor (Tensor) – Tensor to send.

  • dst (int) – Destination rank on global process group (regardless of group argument)

  • group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

  • tag (int, optional) – Tag to match send with remote recv

Returns

A distributed request object. None, if not part of the group

Return type

Optional[Work]

torch.distributed.irecv(tensor, src=None, group=None, tag=0)[source]

Receives a tensor asynchronously.

Warning

tag is not supported with the NCCL backend.

Parameters
  • tensor (Tensor) – Tensor to fill with received data.

  • src (int, optional) – Source rank on global process group (regardless of group argument). Will receive from any process if unspecified.

  • group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

  • tag (int, optional) – Tag to match recv with remote send

Returns

A distributed request object. None, if not part of the group

Return type

Optional[Work]

torch.distributed.batch_isend_irecv(p2p_op_list)[source]

Send or Receive a batch of tensors asynchronously and return a list of requests.

Process each of the operations in p2p_op_list and return the corresponding requests. NCCL, Gloo, and UCC backend are currently supported.

Parameters

p2p_op_list – A list of point-to-point operations(type of each operator is torch.distributed.P2POp). The order of the isend/irecv in the list matters and it needs to match with corresponding isend/irecv on the remote end.

Returns

A list of distributed request objects returned by calling the corresponding op in the op_list.

Examples

>>> send_tensor = torch.arange(2, dtype=torch.float32) + 2 * rank
>>> recv_tensor = torch.randn(2, dtype=torch.float32)
>>> send_op = dist.P2POp(dist.isend, send_tensor, (rank + 1)%world_size)
>>> recv_op = dist.P2POp(dist.irecv, recv_tensor, (rank - 1 + world_size)%world_size)
>>> reqs = batch_isend_irecv([send_op, recv_op])
>>> for req in reqs:
>>>     req.wait()
>>> recv_tensor
tensor([2, 3])     # Rank 0
tensor([0, 1])     # Rank 1

Note

Note that when this API is used with the NCCL PG backend, users must set the current GPU device with torch.cuda.set_device, otherwise it will lead to unexpected hang issues.

In addition, if this API is the first collective call in the group passed to dist.P2POp, all ranks of the group must participate in this API call; otherwise, the behavior is undefined. If this API call is not the first collective call in the group, batched P2P operations involving only a subset of ranks of the group are allowed.

class torch.distributed.P2POp(op, tensor, peer, group=None, tag=0)[source]

A class to build point-to-point operations for batch_isend_irecv.

This class builds the type of P2P operation, communication buffer, peer rank, Process Group, and tag. Instances of this class will be passed to batch_isend_irecv for point-to-point communications.

Parameters
  • op (Callable) – A function to send data to or receive data from a peer process. The type of op is either torch.distributed.isend or torch.distributed.irecv.

  • tensor (Tensor) – Tensor to send or receive.

  • peer (int) – Destination or source rank.

  • group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

  • tag (int, optional) – Tag to match send with recv.

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.

  • get_future() - returns torch._C.Future object. Supported for NCCL, also supported for most operations on GLOO and MPI, except for peer to peer operations. Note: as we continue adopting Futures and merging APIs, get_future() call might become redundant.

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

torch.distributed.broadcast(tensor, src, group=None, async_op=False)[source]

Broadcasts the tensor to the whole group.

tensor must have the same number of elements in all processes participating in the collective.

Parameters
  • tensor (Tensor) – Data to be sent if src is the rank of current process, and tensor to be used to save received data otherwise.

  • src (int) – Source rank on global process group (regardless of group argument).

  • group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

  • async_op (bool, optional) – Whether this op should be an async op

Returns

Async work handle, if async_op is set to True. None, if not async_op or if not part of the group

torch.distributed.broadcast_object_list(object_list, src=0, group=None, device=None)[source]

Broadcasts picklable objects in object_list to the whole group.

Similar to broadcast(), but Python objects can be passed in. Note that all objects in object_list must be picklable in order to be broadcasted.

Parameters
  • object_list (List[Any]) – List of input objects to broadcast. Each object must be picklable. Only objects on the src rank will be broadcast, but each rank must provide lists of equal sizes.

  • src (int) – Source rank from which to broadcast object_list. Source rank is based on global process group (regardless of group argument)

  • group – (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. Default is None.

  • device (torch.device, optional) – If not None, the objects are serialized and converted to tensors which are moved to the device before broadcasting. Default is None.

Returns

None. If rank is part of the group, object_list will contain the broadcasted objects from src rank.

Note

For NCCL-based process groups, internal tensor representations of objects must be moved to the GPU device before communication takes place. In this case, the device used is given by torch.cuda.current_device() and it is the user’s responsibility to ensure that this is set so that each rank has an individual GPU, via torch.cuda.set_device().

Note

Note that this API differs slightly from the all_gather() collective since it does not provide an async_op handle and thus will be a blocking call.

Warning

broadcast_object_list() uses pickle module implicitly, which is known to be insecure. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling. Only call this function with data you trust.

Warning

Calling broadcast_object_list() with GPU tensors is not well supported and inefficient as it incurs GPU -> CPU transfer since tensors would be pickled. Please consider using broadcast() instead.

Example::
>>> # Note: Process group initialization omitted on each rank.
>>> import torch.distributed as dist
>>> if dist.get_rank() == 0:
>>>     # Assumes world_size of 3.
>>>     objects = ["foo", 12, {1: 2}] # any picklable object
>>> else:
>>>     objects = [None, None, None]
>>> # Assumes backend is not NCCL
>>> device = torch.device("cpu")
>>> dist.broadcast_object_list(objects, src=0, device=device)
>>> objects
['foo', 12, {1: 2}]
torch.distributed.all_reduce(tensor, op=<RedOpType.SUM: 0>, group=None, async_op=False)[source]

Reduces the tensor data across all machines in a way that all get the final result.

After the call tensor is going to be bitwise identical in all processes.

Complex tensors are supported.

Parameters
  • tensor (Tensor) – Input and output of the collective. The function operates in-place.

  • op (optional) – One of the values from torch.distributed.ReduceOp enum. Specifies an operation used for element-wise reductions.

  • group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

  • async_op (bool, optional) – Whether this op should be an async op

Returns

Async work handle, if async_op is set to True. None, if not async_op or if not part of the group

Examples

>>> # All tensors below are of torch.int64 type.
>>> # We have 2 process groups, 2 ranks.
>>> device = torch.device(f'cuda:{rank}')
>>> tensor = torch.arange(2, dtype=torch.int64, device=device) + 1 + 2 * rank
>>> tensor
tensor([1, 2], device='cuda:0') # Rank 0
tensor([3, 4], device='cuda:1') # Rank 1
>>> dist.all_reduce(tensor, op=ReduceOp.SUM)
>>> tensor
tensor([4, 6], device='cuda:0') # Rank 0
tensor([4, 6], device='cuda:1') # Rank 1
>>> # All tensors below are of torch.cfloat type.
>>> # We have 2 process groups, 2 ranks.
>>> tensor = torch.tensor([1+1j, 2+2j], dtype=torch.cfloat, device=device) + 2 * rank * (1+1j)
>>> tensor
tensor([1.+1.j, 2.+2.j], device='cuda:0') # Rank 0
tensor([3.+3.j, 4.+4.j], device='cuda:1') # Rank 1
>>> dist.all_reduce(tensor, op=ReduceOp.SUM)
>>> tensor
tensor([4.+4.j, 6.+6.j], device='cuda:0') # Rank 0
tensor([4.+4.j, 6.+6.j], device='cuda:1') # Rank 1
torch.distributed.reduce(tensor, dst, op=<RedOpType.SUM: 0>, group=None, async_op=False)[source]

Reduces the tensor data across all machines.

Only the process with rank dst is going to receive the final result.

Parameters
  • tensor (Tensor) – Input and output of the collective. The function operates in-place.

  • dst (int) – Destination rank on global process group (regardless of group argument)

  • op (optional) – One of the values from torch.distributed.ReduceOp enum. Specifies an operation used for element-wise reductions.

  • group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

  • async_op (bool, optional) – Whether this op should be an async op

Returns

Async work handle, if async_op is set to True. None, if not async_op or if not part of the group

torch.distributed.all_gather(tensor_list, tensor, group=None, async_op=False)[source]

Gathers tensors from the whole group in a list.

Complex tensors are supported.

Parameters
  • tensor_list (list[Tensor]) – Output list. It should contain correctly-sized tensors to be used for output of the collective.

  • tensor (Tensor) – Tensor to be broadcast from current process.

  • group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

  • async_op (bool, optional) – Whether this op should be an async op

Returns

Async work handle, if async_op is set to True. None, if not async_op or if not part of the group

Examples

>>> # All tensors below are of torch.int64 dtype.
>>> # We have 2 process groups, 2 ranks.
>>> device = torch.device(f'cuda:{rank}')
>>> tensor_list = [torch.zeros(2, dtype=torch.int64, device=device) for _ in range(2)]
>>> tensor_list
[tensor([0, 0], device='cuda:0'), tensor([0, 0], device='cuda:0')] # Rank 0
[tensor([0, 0], device='cuda:0'), tensor([0, 0], device='cuda:1')] # Rank 1
>>> tensor = torch.arange(2, dtype=torch.int64, device=device) + 1 + 2 * rank
>>> tensor
tensor([1, 2], device='cuda:0') # Rank 0
tensor([3, 4], device='cuda:1') # Rank 1
>>> dist.all_gather(tensor_list, tensor)
>>> tensor_list
[tensor([1, 2], device='cuda:0'), tensor([3, 4], device='cuda:0')] # Rank 0
[tensor([1, 2], device='cuda:1'), tensor([3, 4], device='cuda:1')] # Rank 1
>>> # All tensors below are of torch.cfloat dtype.
>>> # We have 2 process groups, 2 ranks.
>>> tensor_list = [torch.zeros(2, dtype=torch.cfloat, device=device) for _ in range(2)]
>>> tensor_list
[tensor([0.+0.j, 0.+0.j], device='cuda:0'), tensor([0.+0.j, 0.+0.j], device='cuda:0')] # Rank 0
[tensor([0.+0.j, 0.+0.j], device='cuda:1'), tensor([0.+0.j, 0.+0.j], device='cuda:1')] # Rank 1
>>> tensor = torch.tensor([1+1j, 2+2j], dtype=torch.cfloat, device=device) + 2 * rank * (1+1j)
>>> tensor
tensor([1.+1.j, 2.+2.j], device='cuda:0') # Rank 0
tensor([3.+3.j, 4.+4.j], device='cuda:1') # Rank 1
>>> dist.all_gather(tensor_list, tensor)
>>> tensor_list
[tensor([1.+1.j, 2.+2.j], device='cuda:0'), tensor([3.+3.j, 4.+4.j], device='cuda:0')] # Rank 0
[tensor([1.+1.j, 2.+2.j], device='cuda:1'), tensor([3.+3.j, 4.+4.j], device='cuda:1')] # Rank 1
torch.distributed.all_gather_into_tensor(output_tensor, input_tensor, group=None, async_op=False)[source]

Gather tensors from all ranks and put them in a single output tensor.

Parameters
  • output_tensor (Tensor) – Output tensor to accommodate tensor elements from all ranks. It must be correctly sized to have one of the following forms: (i) a concatenation of all the input tensors along the primary dimension; for definition of “concatenation”, see torch.cat(); (ii) a stack of all the input tensors along the primary dimension; for definition of “stack”, see torch.stack(). Examples below may better explain the supported output forms.

  • input_tensor (Tensor) – Tensor to be gathered from current rank. Different from the all_gather API, the input tensors in this API must have the same size across all ranks.

  • group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

  • async_op (bool, optional) – Whether this op should be an async op

Returns

Async work handle, if async_op is set to True. None, if not async_op or if not part of the group

Examples

>>> # All tensors below are of torch.int64 dtype and on CUDA devices.
>>> # We have two ranks.
>>> device = torch.device(f'cuda:{rank}')
>>> tensor_in = torch.arange(2, dtype=torch.int64, device=device) + 1 + 2 * rank
>>> tensor_in
tensor([1, 2], device='cuda:0') # Rank 0
tensor([3, 4], device='cuda:1') # Rank 1
>>> # Output in concatenation form
>>> tensor_out = torch.zeros(world_size * 2, dtype=torch.int64, device=device)
>>> dist.all_gather_into_tensor(tensor_out, tensor_in)
>>> tensor_out
tensor([1, 2, 3, 4], device='cuda:0') # Rank 0
tensor([1, 2, 3, 4], device='cuda:1') # Rank 1
>>> # Output in stack form
>>> tensor_out2 = torch.zeros(world_size, 2, dtype=torch.int64, device=device)
>>> dist.all_gather_into_tensor(tensor_out2, tensor_in)
>>> tensor_out2
tensor([[1, 2],
        [3, 4]], device='cuda:0') # Rank 0
tensor([[1, 2],
        [3, 4]], device='cuda:1') # Rank 1

Warning

The Gloo backend does not support this API.

torch.distributed.all_gather_object(object_list, obj, group=None)[source]

Gathers picklable objects from the whole group into a list.

Similar to all_gather(), but Python objects can be passed in. Note that the object must be picklable in order to be gathered.

Parameters
  • object_list (list[Any]) – Output list. It should be correctly sized as the size of the group for this collective and will contain the output.

  • obj (Any) – Pickable Python object to be broadcast from current process.

  • group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used. Default is None.

Returns

None. If the calling rank is part of this group, the output of the collective will be populated into the input object_list. If the calling rank is not part of the group, the passed in object_list will be unmodified.

Note

Note that this API differs slightly from the all_gather() collective since it does not provide an async_op handle and thus will be a blocking call.

Note

For NCCL-based processed groups, internal tensor representations of objects must be moved to the GPU device before communication takes place. In this case, the device used is given by torch.cuda.current_device() and it is the user’s responsiblity to ensure that this is set so that each rank has an individual GPU, via torch.cuda.set_device().

Warning

all_gather_object() uses pickle module implicitly, which is known to be insecure. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling. Only call this function with data you trust.

Warning

Calling all_gather_object() with GPU tensors is not well supported and inefficient as it incurs GPU -> CPU transfer since tensors would be pickled. Please consider using all_gather() instead.

Example::
>>> # Note: Process group initialization omitted on each rank.
>>> import torch.distributed as dist
>>> # Assumes world_size of 3.
>>> gather_objects = ["foo", 12, {1: 2}] # any picklable object
>>> output = [None for _ in gather_objects]
>>> dist.all_gather_object(output, gather_objects[dist.get_rank()])
>>> output
['foo', 12, {1: 2}]
torch.distributed.gather(tensor, gather_list=None, dst=0, group=None, async_op=False)[source]

Gathers a list of tensors in a single process.

Parameters
  • tensor (Tensor) – Input tensor.

  • gather_list (list[Tensor], optional) – List of appropriately-sized tensors to use for gathered data (default is None, must be specified on the destination rank)

  • dst (int, optional) – Destination rank on global process group (regardless of group argument). (default is 0)

  • group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

  • async_op (bool, optional) – Whether this op should be an async op

Returns

Async work handle, if async_op is set to True. None, if not async_op or if not part of the group

torch.distributed.gather_object(obj, object_gather_list=None, dst=0, group=None)[source]

Gathers picklable objects from the whole group in a single process.

Similar to gather(), but Python objects can be passed in. Note that the object must be picklable in order to be gathered.

Parameters
  • obj (Any) – Input object. Must be picklable.

  • object_gather_list (list[Any]) – Output list. On the dst rank, it should be correctly sized as the size of the group for this collective and will contain the output. Must be None on non-dst ranks. (default is None)

  • dst (int, optional) – Destination rank on global process group (regardless of group argument). (default is 0)

  • group – (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. Default is None.

Returns

None. On the dst rank, object_gather_list will contain the output of the collective.

Note

Note that this API differs slightly from the gather collective since it does not provide an async_op handle and thus will be a blocking call.

Note

For NCCL-based processed groups, internal tensor representations of objects must be moved to the GPU device before communication takes place. In this case, the device used is given by torch.cuda.current_device() and it is the user’s responsiblity to ensure that this is set so that each rank has an individual GPU, via torch.cuda.set_device().

Warning

gather_object() uses pickle module implicitly, which is known to be insecure. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling. Only call this function with data you trust.

Warning

Calling gather_object() with GPU tensors is not well supported and inefficient as it incurs GPU -> CPU transfer since tensors would be pickled. Please consider using gather() instead.

Example::
>>> # Note: Process group initialization omitted on each rank.
>>> import torch.distributed as dist
>>> # Assumes world_size of 3.
>>> gather_objects = ["foo", 12, {1: 2}] # any picklable object
>>> output = [None for _ in gather_objects]
>>> dist.gather_object(
...     gather_objects[dist.get_rank()],
...     output if dist.get_rank() == 0 else None,
...     dst=0
... )
>>> # On rank 0
>>> output
['foo', 12, {1: 2}]
torch.distributed.scatter(tensor, scatter_list=None, src=0, group=None, async_op=False)[source]

Scatters a list of tensors to all processes in a group.

Each process will receive exactly one tensor and store its data in the tensor argument.

Complex tensors are supported.

Parameters
  • tensor (Tensor) – Output tensor.

  • scatter_list (list[Tensor]) – List of tensors to scatter (default is None, must be specified on the source rank)

  • src (int) – Source rank on global process group (regardless of group argument). Default is 0

  • group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

  • async_op (bool, optional) – Whether this op should be an async op

Returns

Async work handle, if async_op is set to True. None, if not async_op or if not part of the group

Note

Note that all Tensors in scatter_list must have the same size.

Example::
>>> # Note: Process group initialization omitted on each rank.
>>> import torch.distributed as dist
>>> tensor_size = 2
>>> t_ones = torch.ones(tensor_size)
>>> t_fives = torch.ones(tensor_size) * 5
>>> output_tensor = torch.zeros(tensor_size)
>>> if dist.get_rank() == 0:
>>>     # Assumes world_size of 2.
>>>     # Only tensors, all of which must be the same size.
>>>     scatter_list = [t_ones, t_fives]
>>> else:
>>>     scatter_list = None
>>> dist.scatter(output_tensor, scatter_list, src=0)
>>> # Rank i gets scatter_list[i]. For example, on rank 1:
>>> output_tensor
tensor([5., 5.])
torch.distributed.scatter_object_list(scatter_object_output_list, scatter_object_input_list, src=0, group=None)[source]

Scatters picklable objects in scatter_object_input_list to the whole group.

Similar to scatter(), but Python objects can be passed in. On each rank, the scattered object will be stored as the first element of scatter_object_output_list. Note that all objects in scatter_object_input_list must be picklable in order to be scattered.

Parameters
  • scatter_object_output_list (List[Any]) – Non-empty list whose first element will store the object scattered to this rank.

  • scatter_object_input_list (List[Any]) – List of input objects to scatter. Each object must be picklable. Only objects on the src rank will be scattered, and the argument can be None for non-src ranks.

  • src (int) – Source rank from which to scatter scatter_object_input_list. Source rank is based on global process group (regardless of group argument).

  • group – (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. Default is None.

Returns

None. If rank is part of the group, scatter_object_output_list will have its first element set to the scattered object for this rank.

Note

Note that this API differs slightly from the scatter collective since it does not provide an async_op handle and thus will be a blocking call.

Warning

scatter_object_list() uses pickle module implicitly, which is known to be insecure. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling. Only call this function with data you trust.

Warning

Calling scatter_object_list() with GPU tensors is not well supported and inefficient as it incurs GPU -> CPU transfer since tensors would be pickled. Please consider using scatter() instead.

Example::
>>> # Note: Process group initialization omitted on each rank.
>>> import torch.distributed as dist
>>> if dist.get_rank() == 0:
>>>     # Assumes world_size of 3.
>>>     objects = ["foo", 12, {1: 2}] # any picklable object
>>> else:
>>>     # Can be any list on non-src ranks, elements are not used.
>>>     objects = [None, None, None]
>>> output_list = [None]
>>> dist.scatter_object_list(output_list, objects, src=0)
>>> # Rank i gets objects[i]. For example, on rank 2:
>>> output_list
[{1: 2}]
torch.distributed.reduce_scatter(output, input_list, op=<RedOpType.SUM: 0>, group=None, async_op=False)[source]

Reduces, then scatters a list of tensors to all processes in a group.

Parameters
  • output (Tensor) – Output tensor.

  • input_list (list[Tensor]) – List of tensors to reduce and scatter.

  • op (optional) – One of the values from torch.distributed.ReduceOp enum. Specifies an operation used for element-wise reductions.

  • group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

  • async_op (bool, optional) – Whether this op should be an async op.

Returns

Async work handle, if async_op is set to True. None, if not async_op or if not part of the group.

torch.distributed.reduce_scatter_tensor(output, input, op=<RedOpType.SUM: 0>, group=None, async_op=False)[source]

Reduces, then scatters a tensor to all ranks in a group.

Parameters
  • output (Tensor) – Output tensor. It should have the same size across all ranks.

  • input (Tensor) – Input tensor to be reduced and scattered. Its size should be output tensor size times the world size. The input tensor can have one of the following shapes: (i) a concatenation of the output tensors along the primary dimension, or (ii) a stack of the output tensors along the primary dimension. For definition of “concatenation”, see torch.cat(). For definition of “stack”, see torch.stack().

  • group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

  • async_op (bool, optional) – Whether this op should be an async op.

Returns

Async work handle, if async_op is set to True. None, if not async_op or if not part of the group.

Examples

>>> # All tensors below are of torch.int64 dtype and on CUDA devices.
>>> # We have two ranks.
>>> device = torch.device(f'cuda:{rank}')
>>> tensor_out = torch.zeros(2, dtype=torch.int64, device=device)
>>> # Input in concatenation form
>>> tensor_in = torch.arange(world_size * 2, dtype=torch.int64, device=device)
>>> tensor_in
tensor([0, 1, 2, 3], device='cuda:0') # Rank 0
tensor([0, 1, 2, 3], device='cuda:1') # Rank 1
>>> dist.reduce_scatter_tensor(tensor_out, tensor_in)
>>> tensor_out
tensor([0, 2], device='cuda:0') # Rank 0
tensor([4, 6], device='cuda:1') # Rank 1
>>> # Input in stack form
>>> tensor_in = torch.reshape(tensor_in, (world_size, 2))
>>> tensor_in
tensor([[0, 1],
        [2, 3]], device='cuda:0') # Rank 0
tensor([[0, 1],
        [2, 3]], device='cuda:1') # Rank 1
>>> dist.reduce_scatter_tensor(tensor_out, tensor_in)
>>> tensor_out
tensor([0, 2], device='cuda:0') # Rank 0
tensor([4, 6], device='cuda:1') # Rank 1

Warning

The Gloo backend does not support this API.

torch.distributed.all_to_all_single(output, input, output_split_sizes=None, input_split_sizes=None, group=None, async_op=False)[source]

Split input tensor and then scatter the split list to all processes in a group.

Later the received tensors are concatenated from all the processes in the group and returned as a single output tensor.

Complex tensors are supported.

Parameters
  • output (Tensor) – Gathered concatenated output tensor.

  • input (Tensor) – Input tensor to scatter.

  • output_split_sizes – (list[Int], optional): Output split sizes for dim 0 if specified None or empty, dim 0 of output tensor must divide equally by world_size.

  • input_split_sizes – (list[Int], optional): Input split sizes for dim 0 if specified None or empty, dim 0 of input tensor must divide equally by world_size.

  • group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

  • async_op (bool, optional) – Whether this op should be an async op.

Returns

Async work handle, if async_op is set to True. None, if not async_op or if not part of the group.

Warning

all_to_all_single is experimental and subject to change.

Examples

>>> input = torch.arange(4) + rank * 4
>>> input
tensor([0, 1, 2, 3])     # Rank 0
tensor([4, 5, 6, 7])     # Rank 1
tensor([8, 9, 10, 11])   # Rank 2
tensor([12, 13, 14, 15]) # Rank 3
>>> output = torch.empty([4], dtype=torch.int64)
>>> dist.all_to_all_single(output, input)
>>> output
tensor([0, 4, 8, 12])    # Rank 0
tensor([1, 5, 9, 13])    # Rank 1
tensor([2, 6, 10, 14])   # Rank 2
tensor([3, 7, 11, 15])   # Rank 3
>>> # Essentially, it is similar to following operation:
>>> scatter_list = list(input.chunk(world_size))
>>> gather_list  = list(output.chunk(world_size))
>>> for i in range(world_size):
>>>     dist.scatter(gather_list[i], scatter_list if i == rank else [], src = i)
>>> # Another example with uneven split
>>> input
tensor([0, 1, 2, 3, 4, 5])                                       # Rank 0
tensor([10, 11, 12, 13, 14, 15, 16, 17, 18])                     # Rank 1
tensor([20, 21, 22, 23, 24])                                     # Rank 2
tensor([30, 31, 32, 33, 34, 35, 36])                             # Rank 3
>>> input_splits
[2, 2, 1, 1]                                                     # Rank 0
[3, 2, 2, 2]                                                     # Rank 1
[2, 1, 1, 1]                                                     # Rank 2
[2, 2, 2, 1]                                                     # Rank 3
>>> output_splits
[2, 3, 2, 2]                                                     # Rank 0
[2, 2, 1, 2]                                                     # Rank 1
[1, 2, 1, 2]                                                     # Rank 2
[1, 2, 1, 1]                                                     # Rank 3
>>> output = ...
>>> dist.all_to_all_single(output, input, output_splits, input_splits)
>>> output
tensor([ 0,  1, 10, 11, 12, 20, 21, 30, 31])                     # Rank 0
tensor([ 2,  3, 13, 14, 22, 32, 33])                             # Rank 1
tensor([ 4, 15, 16, 23, 34, 35])                                 # Rank 2
tensor([ 5, 17, 18, 24, 36])                                     # Rank 3
>>> # Another example with tensors of torch.cfloat type.
>>> input = torch.tensor([1+1j, 2+2j, 3+3j, 4+4j], dtype=torch.cfloat) + 4 * rank * (1+1j)
>>> input
tensor([1+1j, 2+2j, 3+3j, 4+4j])                                # Rank 0
tensor([5+5j, 6+6j, 7+7j, 8+8j])                                # Rank 1
tensor([9+9j, 10+10j, 11+11j, 12+12j])                          # Rank 2
tensor([13+13j, 14+14j, 15+15j, 16+16j])                        # Rank 3
>>> output = torch.empty([4], dtype=torch.int64)
>>> dist.all_to_all_single(output, input)
>>> output
tensor([1+1j, 5+5j, 9+9j, 13+13j])                              # Rank 0
tensor([2+2j, 6+6j, 10+10j, 14+14j])                            # Rank 1
tensor([3+3j, 7+7j, 11+11j, 15+15j])                            # Rank 2
tensor([4+4j, 8+8j, 12+12j, 16+16j])                            # Rank 3
torch.distributed.all_to_all(output_tensor_list, input_tensor_list, group=None, async_op=False)[source]

Scatters list of input tensors to all processes in a group and return gathered list of tensors in output list.

Complex tensors are supported.

Parameters
  • output_tensor_list (list[Tensor]) – List of tensors to be gathered one per rank.

  • input_tensor_list (list[Tensor]) – List of tensors to scatter one per rank.

  • group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

  • async_op (bool, optional) – Whether this op should be an async op.

Returns

Async work handle, if async_op is set to True. None, if not async_op or if not part of the group.

Warning

all_to_all is experimental and subject to change.

Examples

>>> input = torch.arange(4) + rank * 4
>>> input = list(input.chunk(4))
>>> input
[tensor([0]), tensor([1]), tensor([2]), tensor([3])]     # Rank 0
[tensor([4]), tensor([5]), tensor([6]), tensor([7])]     # Rank 1
[tensor([8]), tensor([9]), tensor([10]), tensor([11])]   # Rank 2
[tensor([12]), tensor([13]), tensor([14]), tensor([15])] # Rank 3
>>> output = list(torch.empty([4], dtype=torch.int64).chunk(4))
>>> dist.all_to_all(output, input)
>>> output
[tensor([0]), tensor([4]), tensor([8]), tensor([12])]    # Rank 0
[tensor([1]), tensor([5]), tensor([9]), tensor([13])]    # Rank 1
[tensor([2]), tensor([6]), tensor([10]), tensor([14])]   # Rank 2
[tensor([3]), tensor([7]), tensor([11]), tensor([15])]   # Rank 3
>>> # Essentially, it is similar to following operation:
>>> scatter_list = input
>>> gather_list  = output
>>> for i in range(world_size):
>>>     dist.scatter(gather_list[i], scatter_list if i == rank else [], src=i)
>>> input
tensor([0, 1, 2, 3, 4, 5])                                       # Rank 0
tensor([10, 11, 12, 13, 14, 15, 16, 17, 18])                     # Rank 1
tensor([20, 21, 22, 23, 24])                                     # Rank 2
tensor([30, 31, 32, 33, 34, 35, 36])                             # Rank 3
>>> input_splits
[2, 2, 1, 1]                                                     # Rank 0
[3, 2, 2, 2]                                                     # Rank 1
[2, 1, 1, 1]                                                     # Rank 2
[2, 2, 2, 1]                                                     # Rank 3
>>> output_splits
[2, 3, 2, 2]                                                     # Rank 0
[2, 2, 1, 2]                                                     # Rank 1
[1, 2, 1, 2]                                                     # Rank 2
[1, 2, 1, 1]                                                     # Rank 3
>>> input = list(input.split(input_splits))
>>> input
[tensor([0, 1]), tensor([2, 3]), tensor([4]), tensor([5])]                   # Rank 0
[tensor([10, 11, 12]), tensor([13, 14]), tensor([15, 16]), tensor([17, 18])] # Rank 1
[tensor([20, 21]), tensor([22]), tensor([23]), tensor([24])]                 # Rank 2
[tensor([30, 31]), tensor([32, 33]), tensor([34, 35]), tensor([36])]         # Rank 3
>>> output = ...
>>> dist.all_to_all(output, input)
>>> output
[tensor([0, 1]), tensor([10, 11, 12]), tensor([20, 21]), tensor([30, 31])]   # Rank 0
[tensor([2, 3]), tensor([13, 14]), tensor([22]), tensor([32, 33])]           # Rank 1
[tensor([4]), tensor([15, 16]), tensor([23]), tensor([34, 35])]              # Rank 2
[tensor([5]), tensor([17, 18]), tensor([24]), tensor([36])]                  # Rank 3
>>> # Another example with tensors of torch.cfloat type.
>>> input = torch.tensor([1+1j, 2+2j, 3+3j, 4+4j], dtype=torch.cfloat) + 4 * rank * (1+1j)
>>> input = list(input.chunk(4))
>>> input
[tensor([1+1j]), tensor([2+2j]), tensor([3+3j]), tensor([4+4j])]            # Rank 0
[tensor([5+5j]), tensor([6+6j]), tensor([7+7j]), tensor([8+8j])]            # Rank 1
[tensor([9+9j]), tensor([10+10j]), tensor([11+11j]), tensor([12+12j])]      # Rank 2
[tensor([13+13j]), tensor([14+14j]), tensor([15+15j]), tensor([16+16j])]    # Rank 3
>>> output = list(torch.empty([4], dtype=torch.int64).chunk(4))
>>> dist.all_to_all(output, input)
>>> output
[tensor([1+1j]), tensor([5+5j]), tensor([9+9j]), tensor([13+13j])]          # Rank 0
[tensor([2+2j]), tensor([6+6j]), tensor([10+10j]), tensor([14+14j])]        # Rank 1
[tensor([3+3j]), tensor([7+7j]), tensor([11+11j]), tensor([15+15j])]        # Rank 2
[tensor([4+4j]), tensor([8+8j]), tensor([12+12j]), tensor([16+16j])]        # Rank 3
torch.distributed.barrier(group=None, async_op=False, device_ids=None)[source]

Synchronize all processes.

This collective blocks processes until the whole group enters this function, if async_op is False, or if async work handle is called on wait().

Parameters
  • group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

  • async_op (bool, optional) – Whether this op should be an async op

  • device_ids ([int], optional) – List of device/GPU ids.

Returns

Async work handle, if async_op is set to True. None, if not async_op or if not part of the group

Note

ProcessGroupNCCL now relies on stream synchronization instead of device synchronization to block the CPU. Thus, please do not assume that barrier() would perform a device synchronization.

torch.distributed.monitored_barrier(group=None, timeout=None, wait_all_ranks=False)[source]

Synchronize processes similar to torch.distributed.barrier, but consider a configurable timeout.

It is able to report ranks that did not pass this barrier within the provided timeout. Specifically, for non-zero ranks, will block until a send/recv is processed from rank 0. Rank 0 will block until all send /recv from other ranks are processed, and will report failures for ranks that failed to respond in time. Note that if one rank does not reach the monitored_barrier (for example due to a hang), all other ranks would fail in monitored_barrier.

This collective will block all processes/ranks in the group, until the whole group exits the function successfully, making it useful for debugging and synchronizing. However, it can have a performance impact and should only be used for debugging or scenarios that require full synchronization points on the host-side. For debugging purposes, this barrier can be inserted before the application’s collective calls to check if any ranks are desynchronized.

Note

Note that this collective is only supported with the GLOO backend.

Parameters
  • group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.

  • timeout (datetime.timedelta, optional) – Timeout for monitored_barrier. If None, the default process group timeout will be used.

  • wait_all_ranks (bool, optional) – Whether to collect all failed ranks or not. By default, this is False and monitored_barrier on rank 0 will throw on the first failed rank it encounters in order to fail fast. By setting wait_all_ranks=True monitored_barrier will collect all failed ranks and throw an error containing information about all failed ranks.

Returns

None.

Example::
>>> # Note: Process group initialization omitted on each rank.
>>> import torch.distributed as dist
>>> if dist.get_rank() != 1:
>>>     dist.monitored_barrier() # Raises exception indicating that
>>> # rank 1 did not call into monitored_barrier.
>>> # Example with wait_all_ranks=True
>>> if dist.get_rank() == 0:
>>>     dist.monitored_barrier(wait_all_ranks=True) # Raises exception
>>> # indicating that ranks 1, 2, ... world_size - 1 did not call into
>>> # monitored_barrier.
class torch.distributed.Work

A Work object represents the handle to a pending asynchronous operation in PyTorch’s distributed package. It is returned by non-blocking collective operations, such as dist.all_reduce(tensor, async_op=True).

class torch.distributed.ReduceOp

An enum-like class for available reduction operations: SUM, PRODUCT, MIN, MAX, BAND, BOR, BXOR, and PREMUL_SUM.

BAND, BOR, and BXOR reductions are not available when using the NCCL backend.

AVG divides values by the world size before summing across ranks. AVG is only available with the NCCL backend, and only for NCCL versions 2.10 or later.

PREMUL_SUM multiplies inputs by a given scalar locally before reduction. PREMUL_SUM is only available with the NCCL backend, and only available for NCCL versions 2.11 or later. Users are supposed to use torch.distributed._make_nccl_premul_sum.

Additionally, MAX, MIN and PRODUCT are not supported for complex tensors.

The values of this class can be accessed as attributes, e.g., ReduceOp.SUM. They are used in specifying strategies for reduction collectives, e.g., reduce().

This class does not support __members__ property.

class torch.distributed.reduce_op

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

ReduceOp is recommended to use instead.

Profiling Collective Communication

Note that you can use torch.profiler (recommended, only available after 1.8.1) or torch.autograd.profiler to profile collective communication and point-to-point communication APIs mentioned here. All out-of-the-box backends (gloo, nccl, mpi) are supported and collective communication usage will be rendered as expected in profiling output/traces. Profiling your code is the same as any regular torch operator:

import torch
import torch.distributed as dist
with torch.profiler():
    tensor = torch.randn(20, 10)
    dist.all_reduce(tensor)

Please refer to the profiler documentation for a full overview of profiler features.

Multi-GPU collective functions

Warning

The multi-GPU functions (which stand for multiple GPUs per CPU thread) are deprecated. As of today, PyTorch Distributed’s preferred programming model is one device per thread, as exemplified by the APIs in this document. If you are a backend developer and want to support multiple devices per thread, please contact PyTorch Distributed’s maintainers.

Third-party backends

Besides the builtin 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.

Module torch.distributed.launch.

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

Warning

This module is going to be deprecated in favor of torchrun.

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 beneficial 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)
  1. 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)
  1. 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. It is strongly recommended that init_method=env://. Other init methods (e.g. tcp://) may work, but env:// is the one that is officially supported 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.

Debugging torch.distributed applications

Debugging distributed applications can be challenging due to hard to understand hangs, crashes, or inconsistent behavior across ranks. torch.distributed provides a suite of tools to help debug training applications in a self-serve fashion:

Python Breakpoint

It is extremely convenient to use python’s debugger in a distributed environment, but because it does not work out of the box many people do not use it at all. PyTorch offers a customized wrapper around pdb that streamlines the process.

torch.distributed.breakpoint makes this process easy. Internally, it customizes pdb’s breakpoint behavior in two ways but otherwise behaves as normal pdb. 1. Attaches the debugger only on one rank (specified by the user). 2. Ensures all other ranks stop, by using a torch.distributed.barrier() that will release once the debugged rank issues a continue 3. Reroutes stdin from the child process such that it connects to your terminal.

To use it, simply issue torch.distributed.breakpoint(rank) on all ranks, using the same value for rank in each case.

Monitored Barrier

As of v1.10, torch.distributed.monitored_barrier() exists as an alternative to torch.distributed.barrier() which fails with helpful information about which rank may be faulty when crashing, i.e. not all ranks calling into torch.distributed.monitored_barrier() within the provided timeout. torch.distributed.monitored_barrier() implements a host-side barrier using send/recv communication primitives in a process similar to acknowledgements, allowing rank 0 to report which rank(s) failed to acknowledge the barrier in time. As an example, consider the following function where rank 1 fails to call into torch.distributed.monitored_barrier() (in practice this could be due to an application bug or hang in a previous collective):

import os
from datetime import timedelta

import torch
import torch.distributed as dist
import torch.multiprocessing as mp


def worker(rank):
    dist.init_process_group("nccl", rank=rank, world_size=2)
    # monitored barrier requires gloo process group to perform host-side sync.
    group_gloo = dist.new_group(backend="gloo")
    if rank not in [1]:
        dist.monitored_barrier(group=group_gloo, timeout=timedelta(seconds=2))


if __name__ == "__main__":
    os.environ["MASTER_ADDR"] = "localhost"
    os.environ["MASTER_PORT"] = "29501"
    mp.spawn(worker, nprocs=2, args=())

The following error message is produced on rank 0, allowing the user to determine which rank(s) may be faulty and investigate further:

RuntimeError: Rank 1 failed to pass monitoredBarrier in 2000 ms
 Original exception:
[gloo/transport/tcp/pair.cc:598] Connection closed by peer [2401:db00:eef0:1100:3560:0:1c05:25d]:8594

TORCH_DISTRIBUTED_DEBUG

With TORCH_CPP_LOG_LEVEL=INFO, the environment variable TORCH_DISTRIBUTED_DEBUG can be used to trigger additional useful logging and collective synchronization checks to ensure all ranks are synchronized appropriately. TORCH_DISTRIBUTED_DEBUG can be set to either OFF (default), INFO, or DETAIL depending on the debugging level required. Please note that the most verbose option, DETAIL may impact the application performance and thus should only be used when debugging issues.

Setting TORCH_DISTRIBUTED_DEBUG=INFO will result in additional debug logging when models trained with torch.nn.parallel.DistributedDataParallel() are initialized, and TORCH_DISTRIBUTED_DEBUG=DETAIL will additionally log runtime performance statistics a select number of iterations. These runtime statistics include data such as forward time, backward time, gradient communication time, etc. As an example, given the following application:

import os

import torch
import torch.distributed as dist
import torch.multiprocessing as mp


class TwoLinLayerNet(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.a = torch.nn.Linear(10, 10, bias=False)
        self.b = torch.nn.Linear(10, 1, bias=False)

    def forward(self, x):
        a = self.a(x)
        b = self.b(x)
        return (a, b)


def worker(rank):
    dist.init_process_group("nccl", rank=rank, world_size=2)
    torch.cuda.set_device(rank)
    print("init model")
    model = TwoLinLayerNet().cuda()
    print("init ddp")
    ddp_model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[rank])

    inp = torch.randn(10, 10).cuda()
    print("train")

    for _ in range(20):
        output = ddp_model(inp)
        loss = output[0] + output[1]
        loss.sum().backward()


if __name__ == "__main__":
    os.environ["MASTER_ADDR"] = "localhost"
    os.environ["MASTER_PORT"] = "29501"
    os.environ["TORCH_CPP_LOG_LEVEL"]="INFO"
    os.environ[
        "TORCH_DISTRIBUTED_DEBUG"
    ] = "DETAIL"  # set to DETAIL for runtime logging.
    mp.spawn(worker, nprocs=2, args=())

The following logs are rendered at initialization time:

I0607 16:10:35.739390 515217 logger.cpp:173] [Rank 0]: DDP Initialized with:
broadcast_buffers: 1
bucket_cap_bytes: 26214400
find_unused_parameters: 0
gradient_as_bucket_view: 0
is_multi_device_module: 0
iteration: 0
num_parameter_tensors: 2
output_device: 0
rank: 0
total_parameter_size_bytes: 440
world_size: 2
backend_name: nccl
bucket_sizes: 440
cuda_visible_devices: N/A
device_ids: 0
dtypes: float
master_addr: localhost
master_port: 29501
module_name: TwoLinLayerNet
nccl_async_error_handling: N/A
nccl_blocking_wait: N/A
nccl_debug: WARN
nccl_ib_timeout: N/A
nccl_nthreads: N/A
nccl_socket_ifname: N/A
torch_distributed_debug: INFO

The following logs are rendered during runtime (when TORCH_DISTRIBUTED_DEBUG=DETAIL is set):

I0607 16:18:58.085681 544067 logger.cpp:344] [Rank 1 / 2] Training TwoLinLayerNet unused_parameter_size=0
 Avg forward compute time: 40838608
 Avg backward compute time: 5983335
Avg backward comm. time: 4326421
 Avg backward comm/comp overlap time: 4207652
I0607 16:18:58.085693 544066 logger.cpp:344] [Rank 0 / 2] Training TwoLinLayerNet unused_parameter_size=0
 Avg forward compute time: 42850427
 Avg backward compute time: 3885553
Avg backward comm. time: 2357981
 Avg backward comm/comp overlap time: 2234674

In addition, TORCH_DISTRIBUTED_DEBUG=INFO enhances crash logging in torch.nn.parallel.DistributedDataParallel() due to unused parameters in the model. Currently, find_unused_parameters=True must be passed into torch.nn.parallel.DistributedDataParallel() initialization if there are parameters that may be unused in the forward pass, and as of v1.10, all model outputs are required to be used in loss computation as torch.nn.parallel.DistributedDataParallel() does not support unused parameters in the backwards pass. These constraints are challenging especially for larger models, thus when crashing with an error, torch.nn.parallel.DistributedDataParallel() will log the fully qualified name of all parameters that went unused. For example, in the above application, if we modify loss to be instead computed as loss = output[1], then TwoLinLayerNet.a does not receive a gradient in the backwards pass, and thus results in DDP failing. On a crash, the user is passed information about parameters which went unused, which may be challenging to manually find for large models:

RuntimeError: Expected to have finished reduction in the prior iteration before starting a new one. This error indicates that your module has parameters that were not used in producing loss. You can enable unused parameter detection by passing
 the keyword argument `find_unused_parameters=True` to `torch.nn.parallel.DistributedDataParallel`, and by
making sure all `forward` function outputs participate in calculating loss.
If you already have done the above, then the distributed data parallel module wasn't able to locate the output tensors in the return value of your module's `forward` function. Please include the loss function and the structure of the return va
lue of `forward` of your module when reporting this issue (e.g. list, dict, iterable).
Parameters which did not receive grad for rank 0: a.weight
Parameter indices which did not receive grad for rank 0: 0

Setting TORCH_DISTRIBUTED_DEBUG=DETAIL will trigger additional consistency and synchronization checks on every collective call issued by the user either directly or indirectly (such as DDP allreduce). This is done by creating a wrapper process group that wraps all process groups returned by torch.distributed.init_process_group() and torch.distributed.new_group() APIs. As a result, these APIs will return a wrapper process group that can be used exactly like a regular process group, but performs consistency checks before dispatching the collective to an underlying process group. Currently, these checks include a torch.distributed.monitored_barrier(), which ensures all ranks complete their outstanding collective calls and reports ranks which are stuck. Next, the collective itself is checked for consistency by ensuring all collective functions match and are called with consistent tensor shapes. If this is not the case, a detailed error report is included when the application crashes, rather than a hang or uninformative error message. As an example, consider the following function which has mismatched input shapes into torch.distributed.all_reduce():

import torch
import torch.distributed as dist
import torch.multiprocessing as mp


def worker(rank):
    dist.init_process_group("nccl", rank=rank, world_size=2)
    torch.cuda.set_device(rank)
    tensor = torch.randn(10 if rank == 0 else 20).cuda()
    dist.all_reduce(tensor)
    torch.cuda.synchronize(device=rank)


if __name__ == "__main__":
    os.environ["MASTER_ADDR"] = "localhost"
    os.environ["MASTER_PORT"] = "29501"
    os.environ["TORCH_CPP_LOG_LEVEL"]="INFO"
    os.environ["TORCH_DISTRIBUTED_DEBUG"] = "DETAIL"
    mp.spawn(worker, nprocs=2, args=())

With the NCCL backend, such an application would likely result in a hang which can be challenging to root-cause in nontrivial scenarios. If the user enables TORCH_DISTRIBUTED_DEBUG=DETAIL and reruns the application, the following error message reveals the root cause:

work = default_pg.allreduce([tensor], opts)
RuntimeError: Error when verifying shape tensors for collective ALLREDUCE on rank 0. This likely indicates that input shapes into the collective are mismatched across ranks. Got shapes:  10
[ torch.LongTensor{1} ]

Note

For fine-grained control of the debug level during runtime the functions torch.distributed.set_debug_level(), torch.distributed.set_debug_level_from_env(), and torch.distributed.get_debug_level() can also be used.

In addition, TORCH_DISTRIBUTED_DEBUG=DETAIL can be used in conjunction with TORCH_SHOW_CPP_STACKTRACES=1 to log the entire callstack when a collective desynchronization is detected. These collective desynchronization checks will work for all applications that use c10d collective calls backed by process groups created with the torch.distributed.init_process_group() and torch.distributed.new_group() APIs.

Logging

In addition to explicit debugging support via torch.distributed.monitored_barrier() and TORCH_DISTRIBUTED_DEBUG, the underlying C++ library of torch.distributed also outputs log messages at various levels. These messages can be helpful to understand the execution state of a distributed training job and to troubleshoot problems such as network connection failures. The following matrix shows how the log level can be adjusted via the combination of TORCH_CPP_LOG_LEVEL and TORCH_DISTRIBUTED_DEBUG environment variables.

TORCH_CPP_LOG_LEVEL

TORCH_DISTRIBUTED_DEBUG

Effective Log Level

ERROR

ignored

Error

WARNING

ignored

Warning

INFO

ignored

Info

INFO

INFO

Debug

INFO

DETAIL

Trace (a.k.a. All)

Distributed components raise custom Exception types derived from RuntimeError:

  • torch.distributed.DistError: This is the base type of all distributed exceptions.

  • torch.distributed.DistBackendError: This exception is thrown when a backend-specific error occurs. For example, if the NCCL backend is used and the user attempts to use a GPU that is not available to the NCCL library.

  • torch.distributed.DistNetworkError: This exception is thrown when networking libraries encounter errors (ex: Connection reset by peer)

  • torch.distributed.DistStoreError: This exception is thrown when the Store encounters an error (ex: TCPStore timeout)

class torch.distributed.DistError

Exception raised when an error occurs in the distributed library

class torch.distributed.DistBackendError

Exception raised when a backend error occurs in distributed

class torch.distributed.DistNetworkError

Exception raised when a network error occurs in distributed

class torch.distributed.DistStoreError

Exception raised when an error occurs in the distributed store

If you are running single node training, it may be convenient to interactively breakpoint your script. We offer a way to conveniently breakpoint a single rank:

torch.distributed.breakpoint(rank=0)[source]

Set a breakpoint, but only on a single rank. All other ranks will wait for you to be done with the breakpoint before continuing.

Parameters

rank (int) – Which rank to break on. Default: 0

Docs

Access comprehensive developer documentation for PyTorch

View Docs

Tutorials

Get in-depth tutorials for beginners and advanced developers

View Tutorials

Resources

Find development resources and get your questions answered

View Resources