DeepSpeed is a deep learning optimization library that makes distributed training easy, efficient, and effective.
10x Larger Models
5x Faster Training
Minimal Code Change
DeepSpeed can train DL models with over a hundred billion parameters on current generation of GPU clusters, while achieving over 5x in system performance compared to the state-of-art. Early adopters of DeepSpeed have already produced a language model (LM) with over 17B parameters called Turing-NLG, establishing a new SOTA in the LM category.
Section | Description |
---|---|
Why DeepSpeed? | DeepSpeed overview |
Getting Started | DeepSpeed first steps |
Further Reading | DeepSpeed features, tutorials, etc. |
Contributing | Instructions for contributing to DeepSpeed |
Publications | DeepSpeed publications |
Training advanced deep learning models is challenging. Beyond model design, model scientists also need to set up the state-of-the-art training techniques such as distributed training, mixed precision, gradient accumulation, and checkpointing. Yet still, scientists may not achieve the desired system performance and convergence rate. Large model sizes are even more challenging: a large model easily runs out of memory with pure data parallelism and it is difficult to use model parallelism. DeepSpeed addresses these challenges to accelerate model development and training.
The DeepSpeed API is a lightweight wrapper on PyTorch. This means that you can use everything you love in PyTorch and without learning a new platform. In addition, DeepSpeed manages all of the boilerplate state-of-the-art training techniques, such as distributed training, mixed precision, gradient accumulation, and checkpoints so that you can focus on your model development. Most importantly, you can leverage the distinctive efficiency and effectiveness benefit of DeepSpeed to boost speed and scale with just a few lines of code changes to your PyTorch models.
DeepSpeed achieves high performance and fast convergence through a combination of efficiency optimizations on compute/communication/memory/IO and effectiveness optimizations on advanced hyperparameter tuning and optimizers. For example:
-
DeepSpeed trains BERT-large to parity in 14 hours using 64 GPUs (4 DGX-2 boxes) and in 3.7 hours using 256 GPUs (16 DGX-2 boxes).
BERT-large Training Times
Devices Source Training Time (hours) 64 TPUs Google 96 64 V100 GPUs DeepSpeed 14 256 V100 GPUs NVIDIA 3.9 256 V100 GPUs DeepSpeed 3.7 BERT Tutorial: Coming Soon
-
DeepSpeed trains GPT2 (1.5 billion parameters) 3.75x faster than state-of-art, NVIDIA Megatron on Azure GPUs.
Read more: GPT tutorial
DeepSpeed provides memory-efficient data parallelism and enables training models without model parallelism. For example, DeepSpeed can train models with up to 6 billion parameters on NVIDIA V100 GPUs with 32GB of device memory. In comparison, existing frameworks (e.g., PyTorch's Distributed Data Parallel) run out of memory with 1.5 billion parameter models.
DeepSpeed reduces the training memory footprint through a novel solution called Zero Redundancy Optimizer (ZeRO). Unlike basic data parallelism where memory states are replicated across data-parallel processes, ZeRO partitions model states to save significant memory. The current implementation (stage 1 of ZeRO) reduces memory by up to 4x relative to the state-of-art. You can read more about ZeRO in our paper.
With this impressive memory reduction, early adopters of DeepSpeed have already produced a language model (LM) with over 17B parameters called Turing-NLG, establishing a new SOTA in the LM category.
DeepSpeed supports efficient data parallelism, model parallelism, and their combination. ZeRO boosts the scaling capability and efficiency further.
-
DeepSpeed provides system support to run models up to 100 billion parameters, 10x larger than the state-of-art (8 billion NVIDIA GPT, 11 billion Google T5).
-
DeepSpeed can run large models more efficiently, up to 6x faster for models with various sizes spanning 1.5B to 100B. More specifically, the data parallelism powered by ZeRO is complementary and can be combined with different types of model parallelism. It allows DeepSpeed to fit models using lower degree of model parallelism and higher batch size, offering significant performance gains compared to using model parallelism alone.
Read more: technical report, and GPT tutorial.
The figure depicts system throughput improvements of DeepSpeed (combining ZeRO-powered data parallelism with model parallelism of NVIDIA Megatron-LM) over using Megatron-LM alone.
DeepSpeed supports advanced hyperparameter tuning and large batch size optimizers such as LAMB. These improve the effectiveness of model training and reduce the number of samples required to convergence to desired accuracy.
Read more: Tuning tutorial,
Only a few lines of code changes are needed to enable a PyTorch model to use DeepSpeed and ZeRO. Compared to current model parallelism libraries, DeepSpeed does not require a code redesign or model refactoring. It also does not put limitations on model dimensions (such as number of attention heads, hidden sizes, and others), batch size, or any other training parameters. For models of up to six billion parameters, you can use ZeRO-powered data parallelism conveniently without requiring model parallelism, while in contrast, standard data parallelism will run out of memory for models with more than 1.3 billion parameters. In addition, DeepSpeed conveniently supports flexible combination of ZeRO-powered data parallelism with custom model parallelisms, such as tensor slicing of NVIDIA's Megatron-LM.
Below we provide a brief feature list, see our detailed feature overview for descriptions and usage.
- Distributed Training with Mixed Precision
- 16-bit mixed precision
- Single-GPU/Multi-GPU/Multi-Node
- Model Parallelism
- Support for Custom Model Parallelism
- Integration with Megatron-LM
- Memory and Bandwidth Optimizations
- The Zero Redundancy Optimizer (ZeRO)
- Constant Buffer Optimization (CBO)
- Smart Gradient Accumulation
- Training Features
- Simplified training API
- Gradient Clipping
- Automatic loss scaling with mixed precision
- Training Optimizers
- Fused Adam optimizer and arbitrary
torch.optim.Optimizer
- Memory bandwidth optimized FP16 Optimizer
- Large Batch Training with LAMB Optimizer
- Memory efficient Training with ZeRO Optimizer
- Fused Adam optimizer and arbitrary
- Training Agnostic Checkpointing
- Advanced Parameter Search
- Learning Rate Range Test
- 1Cycle Learning Rate Schedule
- Simplified Data Loader
- Performance Analysis and Debugging
- Please see our Azure tutorial to get started with DeepSpeed on Azure!
- If you're not on Azure, we recommend using our docker image via
docker pull deepspeed/deepspeed:latest
which contains a pre-installed version of DeepSpeed and all the necessary dependencies. - If you want to install DeepSpeed manually, we provide an install script
install.sh
to help install on a local machine or across an entire cluster.
DeepSpeed model training is accomplished using the DeepSpeed engine. The engine
can wrap any arbitrary model of type torch.nn.module
and has a minimal set of APIs
for training and checkpointing the model. Please see the tutorials for detailed
examples.
To initialize the DeepSpeed engine:
model_engine, optimizer, _, _ = deepspeed.initialize(args=cmd_args,
model=model,
model_parameters=params)
deepspeed.inialize
ensures that all of the necessary setup required for
distributed data parallel or mixed precision training are done
appropriately under the hood. In addition to wrapping the model, DeepSpeed can
construct and manage the training optimizer, data loader, and the learning rate
scheduler based on the parameters passed to deepspeed.initialze
and the
DeepSpeed configuration file.
Once the DeepSpeed engine has been initialized, it can be used to train the
model using three simple APIs for forward propagation (()
), backward
propagation (backward
), and weight updates (step
).
for step, batch in enumerate(data_loader):
#forward() method
loss = model_engine(batch)
#runs backpropagation
model_engine.backward(loss)
#weight update
model_engine.step()
Under the hood, DeepSpeed automatically performs the necessary operations required for distributed data parallel training, in mixed precision, with a pre-defined learning rate schedule:
-
Gradient Averaging: in distributed data parallel training,
backward
ensures that gradients are averaged across data parallel processes after training on antrain_batch_size
. -
Loss Scaling: in FP16/mixed precision training, the DeepSpeed engine automatically handles scaling the loss to avoid precision loss in the gradients.
-
Learning Rate Schedule: if using DeepSpeed's learning rate schedule, then DeepSpeed automatically handles any updates to the learning rate when
step
is executed.
Saving and loading the training state is handled via the save_checkpoint
and
load_checkpoint
API in DeepSpeed which takes two arguments to uniquely
identify a checkpoint:
ckpt_dir
: the directory where checkpoints will be saved.ckpt_id
: an identifier that uniquely identifies a checkpoint in the directory. In the following code snippet, we use the loss value as the checkpoint identifier.
#load checkpoint
_, client_sd = model_engine.load_checkpoint(args.load_dir, args.ckpt_id)
step = client_sd['step']
#advance data loader to ckpt step
dataloader_to_step(data_loader, step + 1)
for step, batch in enumerate(data_loader):
#forward() method
loss = model_engine(batch)
#runs backpropagation
model_engine.backward(loss)
#weight update
model_engine.step()
#save checkpoint
if step % args.save_interval:
client_sd['step'] = step
ckpt_id = loss.item()
model_engine.save_checkpoint(args.save_dir, ckpt_id, client_sd = client_sd)
DeepSpeed can automatically save and restore the model, optimizer, and the
learning rate scheduler states while hiding away these details from the user.
However, the user may want to save other data in addition to these that are
unique to a given model training. To support these items, save_checkpoint
accepts a client state dictionary client_sd
for saving. These items can be
retrieved from load_checkpoint
as a return argument. In the example above,
the step
value is stored as part of the client_sd
.
DeepSpeed features can be enabled, disabled, or configured using a config JSON
file that should be specified as args.deepspeed_config
. Available configs are at
deepspeed/pt/deepspeed_constants.py.
A sample config file is shown below. For a full set of features see core API
doc.
{
"train_batch_size": 8,
"gradient_accumulation_steps": 1,
"optimizer": {
"type": "Adam",
"params": {
"lr": 0.00015
}
},
"fp16": {
"enabled": true
},
"zero_optimization": true
}
When training across multiple nodes we have found it useful to support
propagating user-defined environment variables. By default DeepSpeed will
propagate all NCCL and PYTHON related environment variables that are set. If
you would like to propagate additional variables you can specify them in a
dot-file named .deepspeed_env
that contains a new-line separated list of
VAR=VAL
entries. The DeepSpeed launcher will look in the local path you are
executing from and also in your home directory (~/
).
As a concrete example, some clusters require special NCCL variables to set
prior to training. The user can simply add these variables to a
.deepspeed_env
file in their home directory that looks like this:
NCCL_IB_DISABLE=1
NCCL_SOCKET_IFNAME=eth0
DeepSpeed will then make sure that these environment variables are set when launching each process on every node across their training job.
DeepSpeed installs the entry point deepspeed
to launch distributed training.
We illustrate an example usage of DeepSpeed with the following assumptions:
- You have already integrated DeepSpeed into your model
client_entry.py
is the entry script for your modelclient args
is theargparse
command line argumentsds_config.json
is the configuration file for DeepSpeed
DeepSpeed configures multi-node compute resources with hostfiles that are compatible with OpenMPI and Horovod. A hostfile is a list of hostnames (or SSH aliases), which are machines accessible via passwordless SSH, and slot counts, which specify the number of GPUs available on the system. For example,
worker-1 slots=4
worker-2 slots=4
specifies that two machines named worker-1 and worker-2 each have four GPUs to use for training.
Hostfiles are specified with the --hostfile
command line option. If no hostfile is
specified, DeepSpeed searches for /job/hostfile
. If no hostfile is specified or found,
DeepSpeed queries the number of GPUs on the local machine to discover the number of local
slots available.
The following command launches a PyTorch training job across all available nodes and GPUs
specified in myhostfile
:
deepspeed <client_entry.py> <client args> \
--deepspeed --deepspeed_config ds_config.json --hostfile=myhostfile
Alternatively, DeepSpeed allows you to restrict distributed training of your model to a
subset of the available nodes and GPUs. This feature is enabled through two command line
arguments: --num_nodes
and --num_gpus
. For example, distributed training can be
restricted to use only two nodes with the following command:
deepspeed --num_nodes=2 \
<client_entry.py> <client args> \
--deepspeed --deepspeed_config ds_config.json
You can instead include or exclude specific resources using the --include
and
--exclude
flags. For example, to use all available resources except GPU 0 on node
worker-2 and GPUs 0 and 1 on worker-3:
deepspeed --exclude="worker-2:0@worker-3:0,1" \
<client_entry.py> <client args> \
--deepspeed --deepspeed_config ds_config.json
Similarly, you can use only GPUs 0 and 1 on worker-2:
deepspeed --include="worker-2:0,1" \
<client_entry.py> <client args> \
--deepspeed --deepspeed_config ds_config.json
As described above, DeepSpeed provides its own parallel launcher to help launch
multi-node/multi-gpu training jobs. If you prefer to launch your training job
using MPI (e.g., mpirun), we provide support for this. It should be noted that
DeepSpeed will still use the torch distributed NCCL backend and not the MPI
backend. To launch your training job with mpirun + DeepSpeed you simply pass us
an additional flag --deepspeed_mpi
. DeepSpeed will then use
mpi4py to discover the MPI environment (e.g.,
rank, world size) and properly initialize torch distributed for training. In this
case you will explicitly invoke python
to launch your model script instead of using
the deepspeed
launcher, here is an example:
mpirun <mpi-args> python \
<client_entry.py> <client args> \
--deepspeed_mpi --deepspeed --deepspeed_config ds_config.json
If you want to use this feature of DeepSpeed, please ensure that mpi4py is
installed via pip install mpi4py
.
In the case that we are only running on a single node (with one or more GPUs)
DeepSpeed does not require a hostfile as described above. If a hostfile is
not detected or passed in then DeepSpeed will query the number of GPUs on the
local machine to discover the number of slots available. The --include
and
--exclude
arguments work as normal, but the user should specify 'localhost'
as the hostname.
Article | Description |
---|---|
DeepSpeed Features | DeepSpeed features |
DeepSpeed JSON Configuration | Configuring DeepSpeed |
API Documentation | Generated DeepSpeed API documentation |
CIFAR-10 Tutorial | Getting started with CIFAR-10 and DeepSpeed |
Megatron-LM Tutorial | Train GPT2 with DeepSpeed and Megatron-LM |
Learning Rate Range Test Tutorial | Faster training with large learning rates |
1Cycle Tutorial | SOTA learning schedule in DeepSpeed |
DeepSpeed welcomes your contributions! Please see our contributing guide for more details on formatting, testing, etc.
This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.
When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.
This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.
- Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, Yuxiong He. (2019) ZeRO: Memory Optimization Towards Training A Trillion Parameter Models. ArXiv:1910.02054