Multi-GPU training

Lightning supports multiple ways of doing distributed training.

Preparing your code

To train on CPU/GPU/TPU without changing your code, we need to build a few good habits :)

Delete .cuda() or .to() calls

Delete any calls to .cuda() or .to(device).

# before lightning
def forward(self, x):
    x = x.cuda(0)
    x_hat = layer_1(x)

# after lightning
def forward(self, x):
    x_hat = layer_1(x)

Init tensors using type_as and register_buffer

When you need to create a new tensor, use type_as. This will make your code scale to any arbitrary number of GPUs or TPUs with Lightning.

# before lightning
def forward(self, x):
    z = torch.Tensor(2, 3)
    z = z.cuda(0)

# with lightning
def forward(self, x):
    z = torch.Tensor(2, 3)
    z = z.type_as(x)

The LightningModule knows what device it is on. You can access the reference via self.device. Sometimes it is necessary to store tensors as module attributes. However, if they are not parameters they will remain on the CPU even if the module gets moved to a new device. To prevent that and remain device agnostic, register the tensor as a buffer in your modules’s __init__ method with register_buffer().

class LitModel(LightningModule):

    def __init__(self):
        self.register_buffer("sigma", torch.eye(3))
        # you can now access self.sigma anywhere in your module

Remove samplers

In PyTorch, you must use torch.nn.DistributedSampler for multi-node or TPU training. The sampler makes sure each GPU sees the appropriate part of your data.

# without lightning
def train_dataloader(self):
    dataset = MNIST(...)
    sampler = None

    if self.on_tpu:
        sampler = DistributedSampler(dataset)

    return DataLoader(dataset, sampler=sampler)

Lightning adds the correct samplers when needed, so no need to explicitly add samplers.

# with lightning
def train_dataloader(self):
    dataset = MNIST(...)
    return DataLoader(dataset)


You can disable this behavior with Trainer(replace_sampler_ddp=False)


For iterable datasets, we don’t do this automatically.

Make models pickleable

It’s very likely your code is already pickleable, in that case no change in necessary. However, if you run a distributed model and get the following error:

File "/net/software/local/python/3.6.5/lib/python3.6/multiprocessing/", line 47,
in _launch reduction.dump(process_obj, fp)
File "/net/software/local/python/3.6.5/lib/python3.6/multiprocessing/", line 60, in dump
ForkingPickler(file, protocol).dump(obj)
_pickle.PicklingError: Can't pickle <function <lambda> at 0x2b599e088ae8>:
attribute lookup <lambda> on __main__ failed

This means something in your model definition, transforms, optimizer, dataloader or callbacks cannot be pickled, and the following code will fail:

import pickle

This is a limitation of using multiple processes for distributed training within PyTorch. To fix this issue, find your piece of code that cannot be pickled. The end of the stacktrace is usually helpful. ie: in the stacktrace example here, there seems to be a lambda function somewhere in the code which cannot be pickled.

File "/net/software/local/python/3.6.5/lib/python3.6/multiprocessing/", line 47,
in _launch reduction.dump(process_obj, fp)
File "/net/software/local/python/3.6.5/lib/python3.6/multiprocessing/", line 60, in dump
ForkingPickler(file, protocol).dump(obj)
_pickle.PicklingError: Can't pickle [THIS IS THE THING TO FIND AND DELETE]:
attribute lookup <lambda> on __main__ failed

Select GPU devices

You can select the GPU devices using ranges, a list of indices or a string containing a comma separated list of GPU ids:

# DEFAULT (int) specifies how many GPUs to use per node

# Above is equivalent to

# Specify which GPUs to use (don't use when running on cluster)
Trainer(gpus=[0, 1])

# Equivalent using a string
Trainer(gpus='0, 1')

# To use all available GPUs put -1 or '-1'
# equivalent to list(range(torch.cuda.device_count()))

The table below lists examples of possible input formats and how they are interpreted by Lightning. Note in particular the difference between gpus=0, gpus=[0] and gpus=”0”.















[0, 1, 2]

first 3 GPUs



[0, 1, 2, …]

all available GPUs





[1, 3]


[1, 3]

GPUs 1 and 3









“1, 3”


[1, 3]

GPUs 1 and 3



[0, 1, 2, …]

all available GPUs


When specifying number of gpus as an integer gpus=k, setting the trainer flag auto_select_gpus=True will automatically help you find k gpus that are not occupied by other processes. This is especially useful when GPUs are configured to be in “exclusive mode”, such that only one process at a time can access them.

Remove CUDA flags

CUDA flags make certain GPUs visible to your script. Lightning sets these for you automatically, there’s NO NEED to do this yourself.

# lightning will set according to what you give the trainer
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "0"

However, when using a cluster, Lightning will NOT set these flags (and you should not either). SLURM will set these for you. For more details see the SLURM cluster guide.

Distributed modes

Lightning allows multiple ways of training

  • Data Parallel (distributed_backend=’dp’) (multiple-gpus, 1 machine)

  • DistributedDataParallel (distributed_backend=’ddp’) (multiple-gpus across many machines (python script based)).

  • DistributedDataParallel (distributed_backend=’ddp_spawn’) (multiple-gpus across many machines (spawn based)).

  • DistributedDataParallel 2 (distributed_backend=’ddp2’) (DP in a machine, DDP across machines).

  • Horovod (distributed_backend=’horovod’) (multi-machine, multi-gpu, configured at runtime)

  • TPUs (tpu_cores=8|x) (tpu or TPU pod)


If you request multiple GPUs or nodes without setting a mode, DDP will be automatically used.

For a deeper understanding of what Lightning is doing, feel free to read this guide.

Data Parallel

DataParallel (DP) splits a batch across k GPUs. That is, if you have a batch of 32 and use DP with 2 gpus, each GPU will process 16 samples, after which the root node will aggregate the results.


DP use is discouraged by PyTorch and Lightning. Use DDP which is more stable and at least 3x faster

# train on 2 GPUs (using DP mode)
trainer = Trainer(gpus=2, distributed_backend='dp')

Distributed Data Parallel

DistributedDataParallel (DDP) works as follows:

  1. Each GPU across each node gets its own process.

  2. Each GPU gets visibility into a subset of the overall dataset. It will only ever see that subset.

  3. Each process inits the model.


Make sure to set the random seed before the instantiation of a Trainer() so that each model initializes with the same weights.

  1. Each process performs a full forward and backward pass in parallel.

  2. The gradients are synced and averaged across all processes.

  3. Each process updates its optimizer.

# train on 8 GPUs (same machine (ie: node))
trainer = Trainer(gpus=8, distributed_backend='ddp')

# train on 32 GPUs (4 nodes)
trainer = Trainer(gpus=8, distributed_backend='ddp', num_nodes=4)

This Lightning implementation of DDP calls your script under the hood multiple times with the correct environment variables:

# example for 3 GPUs DDP
MASTER_ADDR=localhost MASTER_PORT=random() WORLD_SIZE=3 NODE_RANK=0 LOCAL_RANK=0 python --gpus 3 --etc
MASTER_ADDR=localhost MASTER_PORT=random() WORLD_SIZE=3 NODE_RANK=1 LOCAL_RANK=0 python --gpus 3 --etc
MASTER_ADDR=localhost MASTER_PORT=random() WORLD_SIZE=3 NODE_RANK=2 LOCAL_RANK=0 python --gpus 3 --etc

We use DDP this way because ddp_spawn has a few limitations (due to Python and PyTorch):

  1. Since .spawn() trains the model in subprocesses, the model on the main process does not get updated.

  2. Dataloader(num_workers=N), where N is large, bottlenecks training with DDP… ie: it will be VERY slow or won’t work at all. This is a PyTorch limitation.

  3. Forces everything to be picklable.

There are cases in which it is NOT possible to use DDP. Examples are:

  • Jupyter Notebook, Google COLAB, Kaggle, etc.

  • You have a nested script without a root package

  • Your script needs to invoke both .fit and .test, or one of them multiple times

In these situations you should use dp or ddp_spawn instead.

Distributed Data Parallel 2

In certain cases, it’s advantageous to use all batches on the same machine instead of a subset. For instance, you might want to compute a NCE loss where it pays to have more negative samples.

In this case, we can use DDP2 which behaves like DP in a machine and DDP across nodes. DDP2 does the following:

  1. Copies a subset of the data to each node.

  2. Inits a model on each node.

  3. Runs a forward and backward pass using DP.

  4. Syncs gradients across nodes.

  5. Applies the optimizer updates.

# train on 32 GPUs (4 nodes)
trainer = Trainer(gpus=8, distributed_backend='ddp2', num_nodes=4)

Distributed Data Parallel Spawn

ddp_spawn is exactly like ddp except that it uses .spawn to start the training processes.


It is STRONGLY recommended to use DDP for speed and performance.

mp.spawn(self.ddp_train, nprocs=self.num_processes, args=(model, ))

If your script does not support being called from the command line (ie: it is nested without a root project module) you can use the following method:

# train on 8 GPUs (same machine (ie: node))
trainer = Trainer(gpus=8, distributed_backend='ddp')

We STRONGLY discourage this use because it has limitations (due to Python and PyTorch):

  1. The model you pass in will not update. Please save a checkpoint and restore from there.

  2. Set Dataloader(num_workers=0) or it will bottleneck training.

ddp is MUCH faster than ddp_spawn. We recommend you

  1. Install a top-level module for your project using

#!/usr/bin/env python

from setuptools import setup, find_packages

      description='Describe Your Cool Project',
  1. Setup your project like so:

  1. Install as a root-level package

cd /project
pip install -e .

You can then call your scripts anywhere

cd /project/src
python --distributed_backend 'ddp' --gpus 8


Horovod allows the same training script to be used for single-GPU, multi-GPU, and multi-node training.

Like Distributed Data Parallel, every process in Horovod operates on a single GPU with a fixed subset of the data. Gradients are averaged across all GPUs in parallel during the backward pass, then synchronously applied before beginning the next step.

The number of worker processes is configured by a driver application (horovodrun or mpirun). In the training script, Horovod will detect the number of workers from the environment, and automatically scale the learning rate to compensate for the increased total batch size.

Horovod can be configured in the training script to run with any number of GPUs / processes as follows:

# train Horovod on GPU (number of GPUs / machines provided on command-line)
trainer = Trainer(distributed_backend='horovod', gpus=1)

# train Horovod on CPU (number of processes / machines provided on command-line)
trainer = Trainer(distributed_backend='horovod')

When starting the training job, the driver application will then be used to specify the total number of worker processes:

# run training with 4 GPUs on a single machine
horovodrun -np 4 python

# run training with 8 GPUs on two machines (4 GPUs each)
horovodrun -np 8 -H hostname1:4,hostname2:4 python

See the official Horovod documentation for details on installation and performance tuning.

DP/DDP2 caveats

In DP and DDP2 each GPU within a machine sees a portion of a batch. DP and ddp2 roughly do the following:

def distributed_forward(batch, model):
    batch = torch.Tensor(32, 8)
    gpu_0_batch = batch[:8]
    gpu_1_batch = batch[8:16]
    gpu_2_batch = batch[16:24]
    gpu_3_batch = batch[24:]

    y_0 = model_copy_gpu_0(gpu_0_batch)
    y_1 = model_copy_gpu_1(gpu_1_batch)
    y_2 = model_copy_gpu_2(gpu_2_batch)
    y_3 = model_copy_gpu_3(gpu_3_batch)

    return [y_0, y_1, y_2, y_3]

So, when Lightning calls any of the training_step, validation_step, test_step you will only be operating on one of those pieces.

# the batch here is a portion of the FULL batch
def training_step(self, batch, batch_idx):
    y_0 = batch

For most metrics, this doesn’t really matter. However, if you want to add something to your computational graph (like softmax) using all batch parts you can use the training_step_end step.

def training_step_end(self, outputs):
    # only use when  on dp
    outputs =, dim=1)
    softmax = softmax(outputs, dim=1)
    out = softmax.mean()
    return out

In pseudocode, the full sequence is:

# get data
batch = next(dataloader)

# copy model and data to each gpu
batch_splits = split_batch(batch, num_gpus)
models = copy_model_to_gpus(model)

# in parallel, operate on each batch chunk
all_results = []
for gpu_num in gpus:
    batch_split = batch_splits[gpu_num]
    gpu_model = models[gpu_num]
    out = gpu_model(batch_split)

# use the full batch for something like softmax
full out = model.training_step_end(all_results)

To illustrate why this is needed, let’s look at DataParallel

def training_step(self, batch, batch_idx):
    x, y = batch
    y_hat = self(batch)

    # on dp or ddp2 if we did softmax now it would be wrong
    # because batch is actually a piece of the full batch
    return y_hat

def training_step_end(self, batch_parts_outputs):
    # batch_parts_outputs has outputs of each part of the batch

    # do softmax here
    outputs =, dim=1)
    softmax = softmax(outputs, dim=1)
    out = softmax.mean()

    return out

If training_step_end is defined it will be called regardless of TPU, DP, DDP, etc… which means it will behave the same regardless of the backend.

Validation and test step have the same option when using DP.

def validation_step_end(self, batch_parts_outputs):

def test_step_end(self, batch_parts_outputs):

Distributed and 16-bit precision

Due to an issue with Apex and DataParallel (PyTorch and NVIDIA issue), Lightning does not allow 16-bit and DP training. We tried to get this to work, but it’s an issue on their end.

Below are the possible configurations we support.


1+ GPUs









Trainer(gpus=1, use_amp=True)



Trainer(gpus=k, distributed_backend=’dp’)



Trainer(gpus=k, distributed_backend=’ddp’)




Trainer(gpus=k, distributed_backend=’ddp’, use_amp=True)

Implement Your Own Distributed (DDP) training

If you need your own way to init PyTorch DDP you can override pytorch_lightning.core.LightningModule.().

If you also need to use your own DDP implementation, override: pytorch_lightning.core.LightningModule.configure_ddp().

Batch size

When using distributed training make sure to modify your learning rate according to your effective batch size.

Let’s say you have a batch size of 7 in your dataloader.

class LitModel(LightningModule):

    def train_dataloader(self):
        return Dataset(..., batch_size=7)

In (DDP, Horovod) your effective batch size will be 7 * gpus * num_nodes.

# effective batch size = 7 * 8
Trainer(gpus=8, distributed_backend='ddp|horovod')

# effective batch size = 7 * 8 * 10
Trainer(gpus=8, num_nodes=10, distributed_backend='ddp|horovod')

In DDP2, your effective batch size will be 7 * num_nodes. The reason is that the full batch is visible to all GPUs on the node when using DDP2.

# effective batch size = 7
Trainer(gpus=8, distributed_backend='ddp2')

# effective batch size = 7 * 10
Trainer(gpus=8, num_nodes=10, distributed_backend='ddp2')


Huge batch sizes are actually really bad for convergence. Check out: Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour


Lightning supports the use of PytorchElastic to enable fault-tolerent and elastic distributed job scheduling. To use it, specify the ‘ddp’ or ‘ddp2’ backend and the number of gpus you want to use in the trainer.

Trainer(gpus=8, distributed_backend='ddp')

Following the PytorchElastic Quickstart documentation, you then need to start a single-node etcd server on one of the hosts:

etcd --enable-v2
     --advertise-client-urls PUBLIC_HOSTNAME:2379

And then launch the elastic job with:

python -m torchelastic.distributed.launch
        --rdzv_endpoint=ETCD_HOST:ETCD_PORT (--arg1 ... train script args...)

See the official PytorchElastic documentation for details on installation and more use cases.

Jupyter Notebooks

Unfortunately any ddp_ is not supported in jupyter notebooks. Please use dp for multiple GPUs. This is a known Jupyter issue. If you feel like taking a stab at adding this support, feel free to submit a PR!

Pickle Errors

Multi-GPU training sometimes requires your model to be pickled. If you run into an issue with pickling try the following to figure out the issue

import pickle

model = YourModel()

However, if you use ddp the pickling requirement is not there and you should be fine. If you use ddp_spawn the pickling requirement remains. This is a limitation of Python.

Read the Docs v: stable
On Read the Docs
Project Home

Free document hosting provided by Read the Docs.