Author: Shen Li
DistributedDataParallel
(DDP) implements data parallelism at the module level. It uses communication
collectives in the torch.distributed
package to synchronize gradients, parameters, and buffers. Parallelism is
available both within a process and across processes. Within a process, DDP
replicates the input module to devices specified in device_ids, scatters
inputs along the batch dimension accordingly, and gathers outputs to the
output_device, which is similar to
DataParallel.
Across processes, DDP inserts necessary parameter synchronizations in forward
passes and gradient synchronizations in backward passes. It is up to users to
map processes to available resources, as long as processes do not share GPU
devices. The recommended (usually fastest) approach is to create a process for
every module replica, i.e., no module replication within a process. The code in
this tutorial runs on an 8-GPU server, but it can be easily generalized to
other environments.
To create DDP modules, first set up process groups properly. More details can be found in WRITING DISTRIBUTED APPLICATIONS WITH PYTORCH.
import os
import tempfile
import torch
import torch.distributed as dist
import torch.nn as nn
import torch.optim as optim
import torch.multiprocessing as mp
from torch.nn.parallel import DistributedDataParallel as DDP
def setup(rank, world_size):
os.environ['MASTER_ADDR'] = 'localhost'
os.environ['MASTER_PORT'] = '12355'
# initialize the process group
dist.init_process_group("gloo", rank=rank, world_size=world_size)
# Explicitly setting seed to make sure that models created in two processes
# start from same random weights and biases.
torch.manual_seed(42)
def cleanup():
dist.destroy_process_group()Now, let's create a toy module, wrap it with DDP, and feed it with some dummy input data. Please note, if training starts from random parameters, you might want to make sure that all DDP processes use the same initial values. Otherwise, global gradient synchronizes will not make sense.
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)))
def demo_basic(rank, world_size):
setup(rank, world_size)
# setup devices for this process, rank 1 uses GPUs [0, 1, 2, 3] and
# rank 2 uses GPUs [4, 5, 6, 7].
n = torch.cuda.device_count() // world_size
device_ids = list(range(rank * n, (rank + 1) * n))
# create model and move it to device_ids[0]
model = ToyModel().to(device_ids[0])
# output_device defaults to device_ids[0]
ddp_model = DDP(model, device_ids=device_ids)
loss_fn = nn.MSELoss()
optimizer = optim.SGD(ddp_model.parameters(), lr=0.001)
optimizer.zero_grad()
outputs = ddp_model(torch.randn(20, 10))
labels = torch.randn(20, 5).to(device_ids[0])
loss_fn(outputs, labels).backward()
optimizer.step()
cleanup()
def run_demo(demo_fn, world_size):
mp.spawn(demo_fn,
args=(world_size,),
nprocs=world_size,
join=True)As you can see, DDP wraps lower level distributed communication details, and provides a clean API as if it is a local model. For basic use cases, DDP only requires a few more LoCs to set up the process group. When applying DDP to more advanced use cases, there are some caveats that require cautions.
In DDP, constructor, forward method, and differentiation of the outputs are
distributed synchronization points. Different processes are expected to reach
synchronization points in the same order and enter each synchronization point
at roughly the same time. Otherwise, fast processes might arrive early and
timeout on waiting for stragglers. Hence, users are responsible for balancing
workloads distributions across processes. Sometimes, skewed processing speeds
are inevitable due to, e.g., network delays, resource contentions,
unpredictable workload spikes. To avoid timeouts in these situations, make
sure that you pass a sufficiently large timeout value when calling
init_process_group.
It's common to use torch.save and torch.load to checkpoint modules
during training and recover from checkpoints. See
SAVING AND LOADING MODELS
for more details. When using DDP, one optimization is to save the model in
only one process and then load it to all processes, reducing write overhead.
This is correct because all processes start from the same parameters and
gradients are synchronized in backward passes, and hence optimizers should keep
setting parameters to same values. If you use this optimization, make sure all
processes do not start loading before the saving is finished. Besides, when
loading the module, you need to provide an appropriate map_location
argument to prevent a process to step into others' devices. If map_location
is missing, torch.load will first load the module to CPU and then copy each
parameter to where it was saved, which would result in all processes on the
same machine using the same set of devices.
def demo_checkpoint(rank, world_size):
setup(rank, world_size)
# setup devices for this process, rank 1 uses GPUs [0, 1, 2, 3] and
# rank 2 uses GPUs [4, 5, 6, 7].
n = torch.cuda.device_count() // world_size
device_ids = list(range(rank * n, (rank + 1) * n))
model = ToyModel().to(device_ids[0])
# output_device defaults to device_ids[0]
ddp_model = DDP(model, device_ids=device_ids)
loss_fn = nn.MSELoss()
optimizer = optim.SGD(ddp_model.parameters(), lr=0.001)
CHECKPOINT_PATH = tempfile.gettempdir() + "/model.checkpoint"
if rank == 0:
# All processes should see same parameters as they all start from same
# random parameters and gradients are synchronized in backward passes.
# Therefore, saving it in one process is sufficient.
torch.save(ddp_model.state_dict(), CHECKPOINT_PATH)
# Use a barrier() to make sure that process 1 loads the model after process
# 0 saves it.
dist.barrier()
# configure map_location properly
rank0_devices = [x - rank * len(device_ids) for x in device_ids]
device_pairs = zip(rank0_devices, device_ids)
map_location = {'cuda:%d' % x: 'cuda:%d' % y for x, y in device_pairs}
ddp_model.load_state_dict(
torch.load(CHECKPOINT_PATH, map_location=map_location))
optimizer.zero_grad()
outputs = ddp_model(torch.randn(20, 10))
labels = torch.randn(20, 5).to(device_ids[0])
loss_fn = nn.MSELoss()
loss_fn(outputs, labels).backward()
optimizer.step()
# Use a barrier() to make sure that all processes have finished reading the
# checkpoint
dist.barrier()
if rank == 0:
os.remove(CHECKPOINT_PATH)
cleanup()DDP also works with multi-GPU models, but replications within a process are not supported. You need to create one process per module replica, which usually leads to better performance compared to multiple replicas per process. DDP wrapping multi-GPU models is especially helpful when training large models with a huge amount of data. When using this feature, the multi-GPU model needs to be carefully implemented to avoid hard-coded devices, because different model replicas will be placed to different devices.
class ToyMpModel(nn.Module):
def __init__(self, dev0, dev1):
super(ToyMpModel, self).__init__()
self.dev0 = dev0
self.dev1 = dev1
self.net1 = torch.nn.Linear(10, 10).to(dev0)
self.relu = torch.nn.ReLU()
self.net2 = torch.nn.Linear(10, 5).to(dev1)
def forward(self, x):
x = x.to(self.dev0)
x = self.relu(self.net1(x))
x = x.to(self.dev1)
return self.net2(x)When passing a multi-GPU model to DDP, device_ids and output_device
must NOT be set. Input and output data will be placed in proper devices by
either the application or the model forward() method.
def demo_model_parallel(rank, world_size):
setup(rank, world_size)
# setup mp_model and devices for this process
dev0 = rank * 2
dev1 = rank * 2 + 1
mp_model = ToyMpModel(dev0, dev1)
ddp_mp_model = DDP(mp_model)
loss_fn = nn.MSELoss()
optimizer = optim.SGD(ddp_mp_model.parameters(), lr=0.001)
optimizer.zero_grad()
# outputs will be on dev1
outputs = ddp_mp_model(torch.randn(20, 10))
labels = torch.randn(20, 5).to(dev1)
loss_fn(outputs, labels).backward()
optimizer.step()
cleanup()
if __name__ == "__main__":
run_demo(demo_basic, 2)
run_demo(demo_checkpoint, 2)
if torch.cuda.device_count() >= 8:
run_demo(demo_model_parallel, 4)