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.

  • Regular (strategy='ddp')

  • Spawn (strategy='ddp_spawn')

  • Notebook/Fork (strategy='ddp_notebook')

Note

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

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


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=0 LOCAL_RANK=1 python my_file.py --accelerator 'gpu' --devices 3 --etc
MASTER_ADDR=localhost MASTER_PORT=random() WORLD_SIZE=3 NODE_RANK=0 LOCAL_RANK=2 python my_file.py --accelerator 'gpu' --devices 3 --etc

Using DDP this way has a few disadvantages over torch.multiprocessing.spawn():

  1. All processes (including the main process) participate in training and have the updated state of the model and Trainer state.

  2. No multiprocessing pickle errors

  3. Easily scales to multi-node training


It is NOT possible to use DDP in interactive environments like Jupyter Notebook, Google COLAB, Kaggle, etc. In these situations you should use ddp_notebook.


Distributed Data Parallel Spawn

Warning

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

The ddp_spawn strategy is similar to ddp except that it uses torch.multiprocessing.spawn() to start the training processes. Use this for debugging only, or if you are converting a code base to Lightning that relies on spawn.

# 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. After .fit(), only the model’s weights get restored to the main process, but no other state of the Trainer.

  2. Does not support multi-node training.

  3. It is generally slower than 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")

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

Limitations in the main process

None

The state of objects is not up-to-date after returning to the main process (Trainer.fit() etc). Only the model parameters get transferred over.

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


TorchRun (TorchElastic)

Lightning supports the use of TorchRun (previously known as TorchElastic) to enable fault-tolerant and elastic distributed job scheduling. To use it, specify the DDP strategy and the number of GPUs you want to use in the Trainer.

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

Then simply launch your script with the torchrun command.


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 lightning.pytorch.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)