Shortcuts

GPU training (Intermediate)

Audience: Users looking to train across machines or experiment with different scaling techniques.


Distributed Training strategies

Lightning supports multiple ways of doing distributed training.


  • Data Parallel (strategy='dp') (multiple-gpus, 1 machine)

  • DistributedDataParallel (multiple-gpus across many machines)
    • Regular (strategy='ddp')

    • Spawn (strategy='ddp_spawn')

    • Notebook/Fork (strategy='ddp_notebook')

  • Horovod (strategy='horovod') (multi-machine, multi-gpu, configured at runtime)

  • Bagua (strategy='bagua') (multiple-gpus across many machines with advanced training algorithms)

Note

If you request multiple GPUs or nodes without setting a mode, DDP Spawn 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.

Warning

DP use is discouraged by PyTorch and Lightning. State is not maintained on the replicas created by the DataParallel wrapper and you may see errors or misbehavior if you assign state to the module in the forward() or *_step() methods. For the same reason we cannot fully support Manual Optimization with DP. Use DDP which is more stable and at least 3x faster.

Warning

DP only supports scattering and gathering primitive collections of tensors like lists, dicts, etc. Therefore transfer_batch_to_device() and on_after_batch_transfer() do not apply in this mode and if you have overridden any of them, an exception will be raised.

# train on 2 GPUs (using DP mode)
trainer = Trainer(accelerator="gpu", devices=2, strategy="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.

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

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

  6. Each process updates its optimizer.

# train on 8 GPUs (same machine (ie: node))
trainer = Trainer(accelerator="gpu", devices=8, strategy="ddp")

# train on 32 GPUs (4 nodes)
trainer = Trainer(accelerator="gpu", devices=8, strategy="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 my_file.py --accelerator 'gpu' --devices 3 --etc
MASTER_ADDR=localhost MASTER_PORT=random() WORLD_SIZE=3 NODE_RANK=1 LOCAL_RANK=0 python my_file.py --accelerator 'gpu' --devices 3 --etc
MASTER_ADDR=localhost MASTER_PORT=random() WORLD_SIZE=3 NODE_RANK=2 LOCAL_RANK=0 python my_file.py --accelerator 'gpu' --devices 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

In these situations you should use ddp_notebook or dp instead.

Distributed Data Parallel Spawn

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

Warning

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(accelerator="gpu", devices=8, strategy="ddp_spawn")

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 setup.py

# setup.py
#!/usr/bin/env python

from setuptools import setup, find_packages

setup(
    name="src",
    version="0.0.1",
    description="Describe Your Cool Project",
    author="",
    author_email="",
    url="https://github.com/YourSeed",  # REPLACE WITH YOUR OWN GITHUB PROJECT LINK
    install_requires=["pytorch-lightning"],
    packages=find_packages(),
)
  1. Setup your project like so:

/project
    /src
        some_file.py
        /or_a_folder
    setup.py
  1. Install as a root-level package

cd /project
pip install -e .

You can then call your scripts anywhere

cd /project/src
python some_file.py --accelerator 'gpu' --devices 8 --strategy 'ddp'

Distributed Data Parallel in Notebooks

DDP Notebook/Fork is an alternative to Spawn that can be used in interactive Python and Jupyter notebooks, Google Colab, Kaggle notebooks, and so on: The Trainer enables it by default when such environments are detected.

# train on 8 GPUs in a Jupyter notebook
trainer = Trainer(accelerator="gpu", devices=8)

# can be set explicitly
trainer = Trainer(accelerator="gpu", devices=8, strategy="ddp_notebook")

# can also be used in non-interactive environments
trainer = Trainer(accelerator="gpu", devices=8, strategy="ddp_fork")

Data Parallel (strategy="dp") is the only other strategy supported in interactive environments but is slower, is discouraged by PyTorch and has other limitations. Among the native distributed strategies, regular DDP (strategy="ddp") is still recommended as the go-to strategy over Spawn and Fork/Notebook for its speed and stability but it can only be used with scripts.

Comparison of DDP variants and tradeoffs

DDP variants and their tradeoffs

DDP

DDP Spawn

DDP Notebook/Fork

Works in Jupyter notebooks / IPython environments

No

No

Yes

Supports multi-node

Yes

Yes

Yes

Supported platforms

Linux, Mac, Win

Linux, Mac, Win

Linux, Mac

Requires all objects to be picklable

No

Yes

No

Is the guard if __name__=="__main__" required?

Yes

Yes

No

Limitations in the main process

None

None

GPU operations such as moving tensors to the GPU or calling torch.cuda functions before invoking Trainer.fit is not allowed.

Process creation time

Slow

Slow

Fast

Horovod

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(strategy="horovod", accelerator="gpu", devices=1)

# train Horovod on CPU (number of processes / machines provided on command-line)
trainer = Trainer(strategy="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 train.py

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

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

Bagua

Bagua is a deep learning training acceleration framework which supports multiple advanced distributed training algorithms including:

  • Gradient AllReduce for centralized synchronous communication, where gradients are averaged among all workers.

  • Decentralized SGD for decentralized synchronous communication, where each worker exchanges data with one or a few specific workers.

  • ByteGrad and QAdam for low precision communication, where data is compressed into low precision before communication.

  • Asynchronous Model Average for asynchronous communication, where workers are not required to be synchronized in the same iteration in a lock-step style.

By default, Bagua uses Gradient AllReduce algorithm, which is also the algorithm implemented in Distributed Data Parallel and Horovod, but Bagua can usually produce a higher training throughput due to its backend written in Rust.

# train on 4 GPUs (using Bagua mode)
trainer = Trainer(strategy="bagua", accelerator="gpu", devices=4)

By specifying the algorithm in the BaguaStrategy, you can select more advanced training algorithms featured by Bagua:

# train on 4 GPUs, using Bagua Gradient AllReduce algorithm
trainer = Trainer(
    strategy=BaguaStrategy(algorithm="gradient_allreduce"),
    accelerator="gpu",
    devices=4,
)

# train on 4 GPUs, using Bagua ByteGrad algorithm
trainer = Trainer(
    strategy=BaguaStrategy(algorithm="bytegrad"),
    accelerator="gpu",
    devices=4,
)

# train on 4 GPUs, using Bagua Decentralized SGD
trainer = Trainer(
    strategy=BaguaStrategy(algorithm="decentralized"),
    accelerator="gpu",
    devices=4,
)

# train on 4 GPUs, using Bagua Low Precision Decentralized SGD
trainer = Trainer(
    strategy=BaguaStrategy(algorithm="low_precision_decentralized"),
    accelerator="gpu",
    devices=4,
)

# train on 4 GPUs, using Asynchronous Model Average algorithm, with a synchronization interval of 100ms
trainer = Trainer(
    strategy=BaguaStrategy(algorithm="async", sync_interval_ms=100),
    accelerator="gpu",
    devices=4,
)

To use QAdam, we need to initialize QAdamOptimizer first:

from pytorch_lightning.strategies import BaguaStrategy
from bagua.torch_api.algorithms.q_adam import QAdamOptimizer


class MyModel(pl.LightningModule):
    ...

    def configure_optimizers(self):
        # initialize QAdam Optimizer
        return QAdamOptimizer(self.parameters(), lr=0.05, warmup_steps=100)


model = MyModel()
trainer = Trainer(
    accelerator="gpu",
    devices=4,
    strategy=BaguaStrategy(algorithm="qadam"),
)
trainer.fit(model)

Bagua relies on its own launcher to schedule jobs. Below, find examples using bagua.distributed.launch which follows torch.distributed.launch API:

# start training with 8 GPUs on a single node
python -m bagua.distributed.launch --nproc_per_node=8 train.py

If the ssh service is available with passwordless login on each node, you can launch the distributed job on a single node with baguarun which has a similar syntax as mpirun. When staring the job, baguarun will automatically spawn new processes on each of your training node provided by --host_list option and each node in it is described as an ip address followed by a ssh port.

# Run on node1 (or node2) to start training on two nodes (node1 and node2), 8 GPUs per node
baguarun --host_list hostname1:ssh_port1,hostname2:ssh_port2 --nproc_per_node=8 --master_port=port1 train.py

Note

You can also start training in the same way as Distributed Data Parallel. However, system optimizations like Bagua-Net and Performance autotuning can only be enabled through bagua launcher. It is worth noting that with Bagua-Net, Distributed Data Parallel can also achieve better performance without modifying the training script.

See Bagua Tutorials for more details on installation and advanced features.

DP caveats

In DP each GPU within a machine sees a portion of a batch. It does roughly 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 using all batch parts you can use the training_step_end step.

def training_step_end(self, outputs):
    # only use when  on dp
    outputs = torch.cat(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)
    all_results.append(out)

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

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, step_output):
    ...


def test_step_end(self, step_output):
    ...

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 GPU

1+ GPUs

DP

DDP

16-bit

command

Y

Trainer(accelerator=”gpu”, devices=1)

Y

Y

Trainer(accelerator=”gpu”, devices=1, precision=16)

Y

Y

Trainer(accelerator=”gpu”, devices=k, strategy=’dp’)

Y

Y

Trainer(accelerator=”gpu”, devices=k, strategy=’ddp’)

Y

Y

Y

Trainer(accelerator=”gpu”, devices=k, strategy=’ddp’, precision=16)

Implement Your Own Distributed (DDP) training

If you need your own way to init PyTorch DDP you can override pytorch_lightning.strategies.ddp.DDPStrategy.setup_distributed().

If you also need to use your own DDP implementation, override pytorch_lightning.strategies.ddp.DDPStrategy.configure_ddp().


Torch Distributed Elastic

Lightning supports the use of Torch Distributed Elastic to enable fault-tolerant and elastic distributed job scheduling. To use it, specify the ‘ddp’ backend and the number of GPUs you want to use in the trainer.

Trainer(accelerator="gpu", devices=8, strategy="ddp")

To launch a fault-tolerant job, run the following on all nodes.

python -m torch.distributed.run
        --nnodes=NUM_NODES
        --nproc_per_node=TRAINERS_PER_NODE
        --rdzv_id=JOB_ID
        --rdzv_backend=c10d
        --rdzv_endpoint=HOST_NODE_ADDR
        YOUR_LIGHTNING_TRAINING_SCRIPT.py (--arg1 ... train script args...)

To launch an elastic job, run the following on at least MIN_SIZE nodes and at most MAX_SIZE nodes.

python -m torch.distributed.run
        --nnodes=MIN_SIZE:MAX_SIZE
        --nproc_per_node=TRAINERS_PER_NODE
        --rdzv_id=JOB_ID
        --rdzv_backend=c10d
        --rdzv_endpoint=HOST_NODE_ADDR
        YOUR_LIGHTNING_TRAINING_SCRIPT.py (--arg1 ... train script args...)

See the official Torch Distributed Elastic documentation for details on installation and more use cases.

Optimize multi-machine communication

By default, Lightning will select the nccl backend over gloo when running on GPUs. Find more information about PyTorch’s supported backends here.

Lightning allows explicitly specifying the backend via the process_group_backend constructor argument on the relevant Strategy classes. By default, Lightning will select the appropriate process group backend based on the hardware used.

from pytorch_lightning.strategies import DDPStrategy

# Explicitly specify the process group backend if you choose to
ddp = DDPStrategy(process_group_backend="nccl")

# Configure the strategy on the Trainer
trainer = Trainer(strategy=ddp, accelerator="gpu", devices=8)

© Copyright Copyright (c) 2018-2022, Lightning AI et al... Revision 92fe1887.

Built with Sphinx using a theme provided by Read the Docs.
Read the Docs v: stable
Versions
latest
stable
1.8.3post1
1.8.3.post0
1.8.3
1.8.2
1.8.1
1.8.0.post1
1.8.0
1.7.7
1.7.6
1.7.5
1.7.4
1.7.3
1.7.2
1.7.1
1.7.0
1.6.5
1.6.4
1.6.3
1.6.2
1.6.1
1.6.0
1.5.10
1.5.9
1.5.8
1.5.7
1.5.6
1.5.5
1.5.4
1.5.3
1.5.2
1.5.1
1.5.0
1.4.9
1.4.8
1.4.7
1.4.6
1.4.5
1.4.4
1.4.3
1.4.2
1.4.1
1.4.0
1.3.8
1.3.7
1.3.6
1.3.5
1.3.4
1.3.3
1.3.2
1.3.1
1.3.0
1.2.10
1.2.8
1.2.7
1.2.6
1.2.5
1.2.4
1.2.3
1.2.2
1.2.1
1.2.0
1.1.8
1.1.7
1.1.6
1.1.5
1.1.4
1.1.3
1.1.2
1.1.1
1.1.0
1.0.8
1.0.7
1.0.6
1.0.5
1.0.4
1.0.3
1.0.2
1.0.1
1.0.0
0.10.0
0.9.0
0.8.5
0.8.4
0.8.3
0.8.2
0.8.1
0.8.0
0.7.6
0.7.5
0.7.4
0.7.3
0.7.2
0.7.1
0.7.0
0.6.0
0.5.3
0.4.9
Downloads
html
On Read the Docs
Project Home
Builds

Free document hosting provided by Read the Docs.