Advanced GPU Optimized Training¶

When training large models, fitting larger batch sizes, or trying to increase throughput using multi-GPU compute, Lightning provides advanced optimized distributed training plugins to support these cases and offer substantial improvements in memory usage.

Note that some of the extreme memory saving configurations will affect the speed of training. This Speed/Memory trade-off in most cases can be adjusted.

Some of these memory-efficient plugins rely on offloading onto other forms of memory, such as CPU RAM or NVMe. This means you can even see memory benefits on a single GPU, using a plugin such as DeepSpeed ZeRO Stage 3 Offload.

Choosing an Advanced Distributed GPU Plugin¶

If you would like to stick with PyTorch DDP, see DDP Optimizations.

Unlike PyTorch’s DistributedDataParallel (DDP) where the maximum trainable model size and batch size do not change with respect to the number of GPUs, memory-optimized plugins can accommodate bigger models and larger batches as more GPUs are used. This means as you scale up the number of GPUs, you can reach the number of model parameters you’d like to train.

Pre-training vs Fine-tuning¶

When fine-tuning, we often use a magnitude less data compared to pre-training a model. This is important when choosing a distributed plugin as usually for pre-training, where we are compute-bound. This means we cannot sacrifice throughput as much as if we were fine-tuning, because in fine-tuning the data requirement is smaller.

Overall:

When fine-tuning a model, use advanced memory efficient plugins such as DeepSpeed ZeRO Stage 3 or DeepSpeed ZeRO Stage 3 Offload, allowing you to fine-tune larger models if you are limited on compute
When pre-training a model, use simpler optimizations such Sharded Training, DeepSpeed ZeRO Stage 2 or Fully Sharded Training, scaling the number of GPUs to reach larger parameter sizes
For both fine-tuning and pre-training, use DeepSpeed Activation Checkpointing or FairScale Activation Checkpointing as the throughput degradation is not significant

For example when using 128 GPUs, you can pre-train large 10 to 20 Billion parameter models using DeepSpeed ZeRO Stage 2 without having to take a performance hit with more advanced optimized multi-gpu plugins.

But for fine-tuning a model, you can reach 10 to 20 Billion parameter models using DeepSpeed ZeRO Stage 3 Offload on a single GPU. This does come with a significant throughput hit, which needs to be weighed accordingly.

When Shouldn’t I use an Optimized Distributed Plugin?¶

Sharding techniques help when model sizes are fairly large; roughly 500M+ parameters is where we’ve seen benefits. However, in cases where your model is small (ResNet50 of around 80M Parameters) it may be best to stick to ordinary distributed training, unless you are using unusually large batch sizes or inputs.

Sharded Training¶

Lightning integration of optimizer sharded training provided by FairScale. The technique can be found within DeepSpeed ZeRO and ZeRO-2, however the implementation is built from the ground up to be pytorch compatible and standalone. Sharded Training allows you to maintain GPU scaling efficiency, whilst reducing memory overhead drastically. In short, expect near-normal linear scaling (if your network allows), and significantly reduced memory usage when training large models.

Sharded Training still utilizes Data Parallel Training under the hood, except optimizer states and gradients are sharded across GPUs. This means the memory overhead per GPU is lower, as each GPU only has to maintain a partition of your optimizer state and gradients.

The benefits vary by model and parameter sizes, but we’ve recorded up to a 63% memory reduction per GPU allowing us to double our model sizes. Because of efficient communication, these benefits in multi-GPU setups are almost free and throughput scales well with multi-node setups.

It is highly recommended to use Sharded Training in multi-GPU environments where memory is limited, or where training larger models are beneficial (500M+ parameter models). A technical note: as batch size scales, storing activations for the backwards pass becomes the bottleneck in training. As a result, sharding optimizer state and gradients becomes less impactful. Use FairScale Activation Checkpointing to see even more benefit at the cost of some throughput.

To use Sharded Training, you need to first install FairScale using the command below.

pip install fairscale

# train using Sharded DDP
trainer = Trainer(plugins="ddp_sharded")

Sharded Training can work across all DDP variants by adding the additional --plugins ddp_sharded flag.

Internally we re-initialize your optimizers and shard them across your machines and processes. We handle all communication using PyTorch distributed, so no code changes are required.

Fully Sharded Training¶

Warning

Fully Sharded Training is in beta and the API is subject to change. Please create an issue if you run into any issues.

Fully Sharded shards optimizer state, gradients and parameters across data parallel workers. This allows you to fit much larger models onto multiple GPUs into memory.

Fully Sharded Training alleviates the need to worry about balancing layers onto specific devices using some form of pipe parallelism, and optimizes for distributed communication with minimal effort.

Shard Parameters to Reach 10+ Billion Parameters¶

To reach larger parameter sizes and be memory efficient, we have to shard parameters. There are various ways to enable this.

Note

Currently Fully Sharded Training relies on the user to wrap the model with Fully Sharded within the LightningModule. This means you must create a single model that is treated as a torch.nn.Module within the LightningModule. This is a limitation of Fully Sharded Training that will be resolved in the future.

Wrap the Model¶

To activate parameter sharding, you must wrap your model using provided wrap or auto_wrap functions as described below. Internally in Lightning, we enable a context manager around the configure_sharded_model function to make sure the wrap and auto_wrap parameters are passed correctly.

When not using Fully Sharded these wrap functions are a no-op. This means once the changes have been made, there is no need to remove the changes for other plugins.

This is a requirement for really large models and also saves on instantiation time as modules are sharded instantly, rather than after the entire model is created in memory.

auto_wrap will recursively wrap torch.nn.Modules within the LightningModule with nested Fully Sharded Wrappers, signalling that we’d like to partition these modules across data parallel devices, discarding the full weights when not required (information here).

auto_wrap can have varying level of success based on the complexity of your model. Auto Wrap does not support models with shared parameters.

wrap will simply wrap the module with a Fully Sharded Parallel class with the correct parameters from the Lightning context manager.

Below is an example of using both wrap and auto_wrap to create your model.

import torch
import torch.nn as nn
import pytorch_lightning as pl
from pytorch_lightning import Trainer
from fairscale.nn import checkpoint_wrapper, auto_wrap, wrap


class MyModel(pl.LightningModule):
    ...

    def configure_sharded_model(self):
        # Created within sharded model context, modules are instantly sharded across processes
        # as soon as they are wrapped with ``wrap`` or ``auto_wrap``

        # Wraps the layer in a Fully Sharded Wrapper automatically
        linear_layer = wrap(nn.Linear(32, 32))

        # Wraps the module recursively
        # based on a minimum number of parameters (default 100M parameters)
        block = auto_wrap(nn.Sequential(nn.Linear(32, 32), nn.ReLU()))

        # For best memory efficiency,
        # add fairscale activation checkpointing
        final_block = auto_wrap(checkpoint_wrapper(nn.Sequential(nn.Linear(32, 32), nn.ReLU())))
        self.model = nn.Sequential(linear_layer, nn.ReLU(), block, final_block)

    def configure_optimizers(self):
        return torch.optim.AdamW(self.model.parameters())


model = MyModel()
trainer = Trainer(gpus=4, plugins="fsdp", precision=16)
trainer.fit(model)

trainer.test()
trainer.predict()

FairScale Activation Checkpointing¶

Activation checkpointing frees activations from memory as soon as they are not needed during the forward pass. They are then re-computed for the backwards pass as needed.

FairScales’ checkpointing wrapper also handles batch norm layers correctly unlike the PyTorch implementation, ensuring stats are tracked correctly due to the multiple forward passes.

This saves memory when training larger models however requires wrapping modules you’d like to use activation checkpointing on. See here for more information.

from pytorch_lightning import Trainer
from fairscale.nn import checkpoint_wrapper


class MyModel(pl.LightningModule):
    def __init__(self):
        # Wrap layers using checkpoint_wrapper
        self.block = checkpoint_wrapper(nn.Sequential(nn.Linear(32, 32), nn.ReLU()))

DeepSpeed¶

Note

The DeepSpeed plugin is in beta and the API is subject to change. Please create an issue if you run into any issues.

DeepSpeed is a deep learning training optimization library, providing the means to train massive billion parameter models at scale. Using the DeepSpeed plugin, we were able to train model sizes of 10 Billion parameters and above, with a lot of useful information in this benchmark and the DeepSpeed docs. DeepSpeed also offers lower level training optimizations, and efficient optimizers such as 1-bit Adam. We recommend using DeepSpeed in environments where speed and memory optimizations are important (such as training large billion parameter models).

Below is a summary of all the configurations of DeepSpeed.

DeepSpeed ZeRO Stage 2 - Shard optimizer states and gradients, remains at parity with DDP with memory improvement
DeepSpeed ZeRO Stage 2 Offload - Offload optimizer states and gradients to CPU. Increases communication, but significant memory improvement
DeepSpeed ZeRO Stage 3 - Shard optimizer states, gradients, (Optional) activations and parameters. Increases communication volume, but even more memory improvement
DeepSpeed ZeRO Stage 3 Offload - Offload optimizer states, gradients, (Optional) activations and parameters to CPU. Increases communication, but even more signficant memory improvement.
DeepSpeed Activation Checkpointing - Free activations after forward pass. Increases computation, but provides memory improvement for all stages.

To use DeepSpeed, you first need to install DeepSpeed using the commands below.

pip install deepspeed

If you run into an issue with the install or later in training, ensure that the CUDA version of the pytorch you’ve installed matches your locally installed CUDA (you can see which one has been recognized by running nvcc --version).

Note

DeepSpeed currently only supports single optimizer, single scheduler within the training loop.

DeepSpeed ZeRO Stage 2¶

By default, we enable DeepSpeed ZeRO Stage 2, which partitions your optimizer states (Stage 1) and your gradients (Stage 2) across your GPUs to reduce memory. In most cases, this is more efficient or at parity with DDP, primarily due to the optimized custom communications written by the DeepSpeed team. As a result, benefits can also be seen on a single GPU. Do note that the default bucket sizes allocate around 3.6GB of VRAM to use during distributed communications, which can be tweaked when instantiating the plugin described in a few sections below.

Note

To use ZeRO, you must use precision=16.

from pytorch_lightning import Trainer

model = MyModel()
trainer = Trainer(gpus=4, plugins="deepspeed_stage_2", precision=16)
trainer.fit(model)

python train.py --plugins deepspeed_stage_2 --precision 16 --gpus 4

DeepSpeed ZeRO Stage 2 Offload¶

Below we show an example of running ZeRO-Offload. ZeRO-Offload leverages the host CPU to offload optimizer memory/computation, reducing the overall memory consumption.

Note

To use ZeRO-Offload, you must use precision=16.

from pytorch_lightning import Trainer
from pytorch_lightning.plugins import DeepSpeedPlugin

model = MyModel()
trainer = Trainer(gpus=4, plugins="deepspeed_stage_2_offload", precision=16)
trainer.fit(model)

This can also be done via the command line using a Pytorch Lightning script:

python train.py --plugins deepspeed_stage_2_offload --precision 16 --gpus 4

You can also modify the ZeRO-Offload parameters via the plugin as below.

from pytorch_lightning import Trainer
from pytorch_lightning.plugins import DeepSpeedPlugin

model = MyModel()
trainer = Trainer(
    gpus=4, plugins=DeepSpeedPlugin(cpu_offload=True, allgather_bucket_size=5e8, reduce_bucket_size=5e8), precision=16
)
trainer.fit(model)

Note

We suggest tuning the allgather_bucket_size parameter and reduce_bucket_size parameter to find optimum parameters based on your model size. These control how large a buffer we limit the model to using when reducing gradients/gathering updated parameters. Smaller values will result in less memory, but tradeoff with speed.

DeepSpeed allocates a reduce buffer size multiplied by 4.5x so take that into consideration when tweaking the parameters.

The plugin sets a reasonable default of 2e8, which should work for most low VRAM GPUs (less than 7GB), allocating roughly 3.6GB of VRAM as buffer. Higher VRAM GPUs should aim for values around 5e8.

For even more speed benefit, DeepSpeed offers an optimized CPU version of ADAM called DeepSpeedCPUAdam to run the offloaded computation, which is faster than the standard PyTorch implementation.

import pytorch_lightning
from pytorch_lightning import Trainer
from pytorch_lightning.plugins import DeepSpeedPlugin
from deepspeed.ops.adam import DeepSpeedCPUAdam


class MyModel(pl.LightningModule):
    ...

    def configure_optimizers(self):
        # DeepSpeedCPUAdam provides 5x to 7x speedup over torch.optim.adam(w)
        return DeepSpeedCPUAdam(self.parameters())


model = MyModel()
trainer = Trainer(gpus=4, plugins="deepspeed_stage_2_offload", precision=16)
trainer.fit(model)

DeepSpeed ZeRO Stage 3¶

DeepSpeed ZeRO Stage 3 shards the optimizer states, gradients and the model parameters (also optionally activations). Sharding model parameters and activations comes with an increase in distributed communication, however allows you to scale your models massively from one GPU to multiple GPUs. The DeepSpeed team report the ability to fine-tune models with over 40B parameters on a single GPU and over 2 Trillion parameters on 512 GPUs. For more information we suggest checking the DeepSpeed ZeRO-3 Offload documentation.

We’ve ran benchmarks for all these features and given a simple example of how all these features work in Lightning, which you can see at minGPT.

To reach the highest memory efficiency or model size, you must:

Use the DeepSpeed Plugin with the stage 3 parameter
Use CPU Offloading to offload weights to CPU, plus have a reasonable amount of CPU RAM to offload onto
Use DeepSpeed Activation Checkpointing to shard activations

Below we describe how to enable all of these to see benefit. With all these improvements we reached 45 Billion parameters training a GPT model on 8 GPUs with ~1TB of CPU RAM available.

Also please have a look at our DeepSpeed ZeRO Stage 3 Tips which contains a lot of helpful information when configuring your own models.

from pytorch_lightning import Trainer
from pytorch_lightning.plugins import DeepSpeedPlugin
from deepspeed.ops.adam import FusedAdam


class MyModel(pl.LightningModule):
    ...

    def configure_optimizers(self):
        return FusedAdam(self.parameters())


model = MyModel()
trainer = Trainer(gpus=4, plugins="deepspeed_stage_3", precision=16)
trainer.fit(model)

trainer.test()
trainer.predict()

Shard Model Instantly to Reduce Initialization Time/Memory¶

When instantiating really large models, it is sometimes necessary to shard the model layers instantly.

This is the case if layers may not fit on one single machines CPU or GPU memory, but would fit once sharded across multiple machines. We expose a hook that layers initialized within the hook will be sharded instantly on a per layer basis, allowing you to instantly shard models.

This reduces the time taken to initialize very large models, as well as ensure we do not run out of memory when instantiating larger models. For more information you can refer to the DeepSpeed docs for Constructing Massive Models.

import torch.nn as nn
from pytorch_lightning import Trainer
from pytorch_lightning.plugins import DeepSpeedPlugin
from deepspeed.ops.adam import FusedAdam


class MyModel(pl.LightningModule):
    ...

    def configure_sharded_model(self):
        # Created within sharded model context, modules are instantly sharded across processes
        # as soon as they are made.
        self.block = nn.Sequential(nn.Linear(32, 32), nn.ReLU())

    def configure_optimizers(self):
        return FusedAdam(self.parameters())


model = MyModel()
trainer = Trainer(gpus=4, plugins="deepspeed_stage_3", precision=16)
trainer.fit(model)

trainer.test()
trainer.predict()

DeepSpeed ZeRO Stage 3 Offload¶

DeepSpeed ZeRO Stage 3 Offloads optimizer state, gradients to the host CPU to reduce memory usage as ZeRO Stage 2 does, however additionally allows you to offload the parameters as well for even more memory saving.

from pytorch_lightning import Trainer
from pytorch_lightning.plugins import DeepSpeedPlugin

# Enable CPU Offloading
model = MyModel()
trainer = Trainer(gpus=4, plugins="deepspeed_stage_3_offload", precision=16)
trainer.fit(model)

# Enable CPU Offloading, and offload parameters to CPU
model = MyModel()
trainer = Trainer(
    gpus=4,
    plugins=DeepSpeedPlugin(
        stage=3,
        offload_optimizer=True,
        offload_parameters=True,
    ),
    precision=16,
)
trainer.fit(model)

DeepSpeed Infinity (NVMe Offloading)¶

Additionally, DeepSpeed supports offloading to NVMe drives for even larger models, utilizing the large memory space found in NVMes. DeepSpeed reports the ability to fine-tune 1 Trillion+ parameters using NVMe Offloading on one 8 GPU machine. Below shows how to enable this, assuming the NVMe drive is mounted in a directory called /local_nvme.

from pytorch_lightning import Trainer
from pytorch_lightning.plugins import DeepSpeedPlugin

# Enable CPU Offloading
model = MyModel()
trainer = Trainer(gpus=4, plugins="deepspeed_stage_3_offload", precision=16)
trainer.fit(model)

# Enable CPU Offloading, and offload parameters to CPU
model = MyModel()
trainer = Trainer(
    gpus=4,
    plugins=DeepSpeedPlugin(
        stage=3,
        offload_optimizer=True,
        offload_parameters=True,
        remote_device="nvme",
        offload_params_device="nvme",
        offload_optimizer_device="nvme",
        nvme_path="/local_nvme",
    ),
    precision=16,
)
trainer.fit(model)

When offloading to NVMe you may notice that the speed is slow. There are parameters that need to be tuned based on the drives that you are using. Running the aio_bench_perf_sweep.py script can help you to find optimum parameters. See the issue for more information on how to parse the information.

DeepSpeed Activation Checkpointing¶

Activation checkpointing frees activations from memory as soon as they are not needed during the forward pass. They are then re-computed for the backwards pass as needed.

This saves memory when training larger models however requires using a checkpoint function to run the module as shown below.

from pytorch_lightning import Trainer
from pytorch_lightning.plugins import DeepSpeedPlugin
import deepspeed


class MyModel(pl.LightningModule):
    ...

    def configure_sharded_model(self):
        self.block = nn.Sequential(nn.Linear(32, 32), nn.ReLU())

    def forward(self, x):
        # Use the DeepSpeed checkpointing function instead of calling the module directly
        output = deepspeed.checkpointing.checkpoint(self.block, x)
        return output


model = MyModel()


trainer = Trainer(gpus=4, plugins="deepspeed_stage_3_offload", precision=16)

# Enable CPU Activation Checkpointing
trainer = Trainer(
    gpus=4,
    plugins=DeepSpeedPlugin(
        stage=3,
        cpu_offload=True,  # Enable CPU Offloading
        cpu_checkpointing=True,  # (Optional) offload activations to CPU
    ),
    precision=16,
)
trainer.fit(model)

DeepSpeed ZeRO Stage 3 Tips¶

Here is some helpful information when setting up DeepSpeed ZeRO Stage 3 with Lightning.

If you’re using Adam or AdamW, ensure to use FusedAdam or DeepSpeedCPUAdam (for CPU Offloading) rather than the default torch optimizers as they come with large speed benefits
Treat your GPU/CPU memory as one large pool. In some cases, you may not want to offload certain things (like activations) to provide even more space to offload model parameters
When offloading to the CPU, make sure to bump up the batch size as GPU memory will be freed
We also support sharded checkpointing. By passing save_full_weights=False to the DeepSpeedPlugin, we’ll save shards of the model which allows you to save extremely large models. However to load the model and run test/validation/predict you must use the Trainer object.

Custom DeepSpeed Config¶

In some cases you may want to define your own DeepSpeed Config, to access all parameters defined. We’ve exposed most of the important parameters, however, there may be debugging parameters to enable. Also, DeepSpeed allows the use of custom DeepSpeed optimizers and schedulers defined within a config file that is supported.

Note

All plugin default parameters will be ignored when a config object is passed. All compatible arguments can be seen in the DeepSpeed docs.

from pytorch_lightning import Trainer
from pytorch_lightning.plugins import DeepSpeedPlugin

deepspeed_config = {
    "zero_allow_untested_optimizer": True,
    "optimizer": {
        "type": "OneBitAdam",
        "params": {
            "lr": 3e-5,
            "betas": [0.998, 0.999],
            "eps": 1e-5,
            "weight_decay": 1e-9,
            "cuda_aware": True,
        },
    },
    "scheduler": {
        "type": "WarmupLR",
        "params": {
            "last_batch_iteration": -1,
            "warmup_min_lr": 0,
            "warmup_max_lr": 3e-5,
            "warmup_num_steps": 100,
        },
    },
    "zero_optimization": {
        "stage": 2,  # Enable Stage 2 ZeRO (Optimizer/Gradient state partitioning)
        "cpu_offload": True,  # Enable Offloading optimizer state/calculation to the host CPU
        "contiguous_gradients": True,  # Reduce gradient fragmentation.
        "overlap_comm": True,  # Overlap reduce/backward operation of gradients for speed.
        "allgather_bucket_size": 2e8,  # Number of elements to all gather at once.
        "reduce_bucket_size": 2e8,  # Number of elements we reduce/allreduce at once.
    },
}

model = MyModel()
trainer = Trainer(gpus=4, plugins=DeepSpeedPlugin(deepspeed_config), precision=16)
trainer.fit(model)

We support taking the config as a json formatted file:

from pytorch_lightning import Trainer
from pytorch_lightning.plugins import DeepSpeedPlugin

model = MyModel()
trainer = Trainer(gpus=4, plugins=DeepSpeedPlugin("/path/to/deepspeed_config.json"), precision=16)
trainer.fit(model)

You can use also use an environment variable via your PyTorch Lightning script:

PL_DEEPSPEED_CONFIG_PATH=/path/to/deepspeed_config.json python train.py --plugins deepspeed

DDP Optimizations¶

Gradients as Bucket View¶

Enabling gradient_as_bucket_view=True in the DDPPlugin will make gradients views point to different offsets of the allreduce communication buckets. See DistributedDataParallel for more information.

This can reduce peak memory usage and throughput as saved memory will be equal to the total gradient memory + removes the need to copy gradients to the allreduce communication buckets.

Note

When gradient_as_bucket_view=True you cannot call detach_() on gradients. If hitting such errors, please fix it by referring to the zero_grad() function in torch/optim/optimizer.py as a solution (source).

from pytorch_lightning import Trainer
from pytorch_lightning.plugins import DDPPlugin

model = MyModel()
trainer = Trainer(gpus=4, plugins=DDPPlugin(gradient_as_bucket_view=True))
trainer.fit(model)

DDP Communication Hooks¶

DDP Communication hooks is an interface to control how gradients are communicated across workers, overriding the standard allreduce in DistributedDataParallel. This allows you to enable performance improving communication hooks when using multiple nodes.

Note

DDP communication hooks needs pytorch version at least 1.8.0

Enable FP16 Compress Hook for multi-node throughput improvement:

from pytorch_lightning import Trainer
from pytorch_lightning.plugins import DDPPlugin
from torch.distributed.algorithms.ddp_comm_hooks import (
    default_hooks as default,
    powerSGD_hook as powerSGD,
)

model = MyModel()
trainer = Trainer(gpus=4, plugins=DDPPlugin(ddp_comm_hook=default.fp16_compress_hook))
trainer.fit(model)

Enable PowerSGD for multi-node throughput improvement:

Note

PowerSGD typically requires extra memory of the same size as the model’s gradients to enable error feedback, which can compensate for biased compressed communication and improve accuracy (source).

from pytorch_lightning import Trainer
from pytorch_lightning.plugins import DDPPlugin
from torch.distributed.algorithms.ddp_comm_hooks import powerSGD_hook as powerSGD

model = MyModel()
trainer = Trainer(
    gpus=4,
    plugins=DDPPlugin(
        ddp_comm_state=powerSGD.PowerSGDState(
            process_group=None,
            matrix_approximation_rank=1,
            start_powerSGD_iter=5000,
        ),
        ddp_comm_hook=powerSGD.powerSGD_hook,
    ),
)
trainer.fit(model)

Combine hooks for accumulated benefit:

Note

DDP communication wrappers needs pytorch version at least 1.9.0

from pytorch_lightning import Trainer
from pytorch_lightning.plugins import DDPPlugin
from torch.distributed.algorithms.ddp_comm_hooks import (
    default_hooks as default,
    powerSGD_hook as powerSGD,
)

model = MyModel()
trainer = Trainer(
    gpus=4,
    plugins=DDPPlugin(
        ddp_comm_state=powerSGD.PowerSGDState(
            process_group=None,
            matrix_approximation_rank=1,
            start_powerSGD_iter=5000,
        ),
        ddp_comm_hook=powerSGD.powerSGD_hook,
        ddp_comm_wrapper=default.fp16_compress_wrapper,
    ),
)
trainer.fit(model)