Getting Started with DeviceMesh — PyTorch Tutorials 2.7.0+cu126 documentation (original) (raw)

recipes/distributed_device_mesh

Run in Google Colab

Colab

Download Notebook

Notebook

View on GitHub

GitHub

Created On: Jan 24, 2024 | Last Updated: Feb 24, 2025 | Last Verified: Nov 05, 2024

Author: Iris Zhang, Wanchao Liang

Note

edit View and edit this tutorial in github.

Prerequisites:

Setting up distributed communicators, i.e. NVIDIA Collective Communication Library (NCCL) communicators, for distributed training can pose a significant challenge. For workloads where users need to compose different parallelisms, users would need to manually set up and manage NCCL communicators (for example, ProcessGroup) for each parallelism solution. This process could be complicated and susceptible to errors.DeviceMesh can simplify this process, making it more manageable and less prone to errors.

What is DeviceMesh

DeviceMesh is a higher level abstraction that manages ProcessGroup. It allows users to effortlessly create inter-node and intra-node process groups without worrying about how to set up ranks correctly for different sub process groups. Users can also easily manage the underlying process_groups/devices for multi-dimensional parallelism via DeviceMesh.

PyTorch DeviceMesh

Why DeviceMesh is Useful

DeviceMesh is useful when working with multi-dimensional parallelism (i.e. 3-D parallel) where parallelism composability is required. For example, when your parallelism solutions require both communication across hosts and within each host. The image above shows that we can create a 2D mesh that connects the devices within each host, and connects each device with its counterpart on the other hosts in a homogeneous setup.

Without DeviceMesh, users would need to manually set up NCCL communicators, cuda devices on each process before applying any parallelism, which could be quite complicated. The following code snippet illustrates a hybrid sharding 2-D Parallel pattern setup without DeviceMesh. First, we need to manually calculate the shard group and replicate group. Then, we need to assign the correct shard and replicate group to each rank.

import os

import torch import torch.distributed as dist

Understand world topology

rank = int(os.environ["RANK"]) world_size = int(os.environ["WORLD_SIZE"]) print(f"Running example on {rank=} in a world with {world_size=}")

Create process groups to manage 2-D like parallel pattern

dist.init_process_group("nccl") torch.cuda.set_device(rank)

Create shard groups (e.g. (0, 1, 2, 3), (4, 5, 6, 7))

and assign the correct shard group to each rank

num_node_devices = torch.cuda.device_count() shard_rank_lists = list(range(0, num_node_devices // 2)), list(range(num_node_devices // 2, num_node_devices)) shard_groups = ( dist.new_group(shard_rank_lists[0]), dist.new_group(shard_rank_lists[1]), ) current_shard_group = ( shard_groups[0] if rank in shard_rank_lists[0] else shard_groups[1] )

Create replicate groups (for example, (0, 4), (1, 5), (2, 6), (3, 7))

and assign the correct replicate group to each rank

current_replicate_group = None shard_factor = len(shard_rank_lists[0]) for i in range(num_node_devices // 2): replicate_group_ranks = list(range(i, num_node_devices, shard_factor)) replicate_group = dist.new_group(replicate_group_ranks) if rank in replicate_group_ranks: current_replicate_group = replicate_group

To run the above code snippet, we can leverage PyTorch Elastic. Let’s create a file named 2d_setup.py. Then, run the following torch elastic/torchrun command.

torchrun --nproc_per_node=8 --rdzv_id=100 --rdzv_endpoint=localhost:29400 2d_setup.py

Note

For simplicity of demonstration, we are simulating 2D parallel using only one node. Note that this code snippet can also be used when running on multi hosts setup.

With the help of init_device_mesh(), we can accomplish the above 2D setup in just two lines, and we can still access the underlying ProcessGroup if needed.

from torch.distributed.device_mesh import init_device_mesh mesh_2d = init_device_mesh("cuda", (2, 4), mesh_dim_names=("replicate", "shard"))

Users can access the underlying process group thru get_group API.

replicate_group = mesh_2d.get_group(mesh_dim="replicate") shard_group = mesh_2d.get_group(mesh_dim="shard")

Let’s create a file named 2d_setup_with_device_mesh.py. Then, run the following torch elastic/torchrun command.

torchrun --nproc_per_node=8 2d_setup_with_device_mesh.py

How to use DeviceMesh with HSDP

Hybrid Sharding Data Parallel(HSDP) is 2D strategy to perform FSDP within a host and DDP across hosts.

Let’s see an example of how DeviceMesh can assist with applying HSDP to your model with a simple setup. With DeviceMesh, users would not need to manually create and manage shard group and replicate group.

import torch import torch.nn as nn

from torch.distributed.device_mesh import init_device_mesh from torch.distributed.fsdp import FullyShardedDataParallel as FSDP, ShardingStrategy

class ToyModel(nn.Module): def init(self): super(ToyModel, self).init() self.net1 = nn.Linear(10, 10) self.relu = nn.ReLU() self.net2 = nn.Linear(10, 5)

def forward(self, x):
    return self.net2(self.relu(self.net1(x)))

HSDP: MeshShape(2, 4)

mesh_2d = init_device_mesh("cuda", (2, 4)) model = FSDP( ToyModel(), device_mesh=mesh_2d, sharding_strategy=ShardingStrategy.HYBRID_SHARD )

Let’s create a file named hsdp.py. Then, run the following torch elastic/torchrun command.

torchrun --nproc_per_node=8 hsdp.py

How to use DeviceMesh for your custom parallel solutions

When working with large scale training, you might have more complex custom parallel training composition. For example, you may need to slice out sub-meshes for different parallelism solutions. DeviceMesh allows users to slice child mesh from the parent mesh and re-use the NCCL communicators already created when the parent mesh is initialized.

from torch.distributed.device_mesh import init_device_mesh mesh_3d = init_device_mesh("cuda", (2, 2, 2), mesh_dim_names=("replicate", "shard", "tp"))

Users can slice child meshes from the parent mesh.

hsdp_mesh = mesh_3d["replicate", "shard"] tp_mesh = mesh_3d["tp"]

Users can access the underlying process group thru get_group API.

replicate_group = hsdp_mesh["replicate"].get_group() shard_group = hsdp_mesh["shard"].get_group() tp_group = tp_mesh.get_group()