February 6, 2020

2404 words 12 mins read



DeepSpeed is a deep learning optimization library that makes distributed training easy, efficient, and effective.

repo name microsoft/DeepSpeed
repo link https://github.com/microsoft/DeepSpeed
language Python
size (curr.) 551 kB
stars (curr.) 2064
created 2020-01-23
license MIT License

Build Status License MIT

DeepSpeed is a deep learning optimization library that makes distributed training easy, efficient, and effective.

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.

Table of Contents

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

Why DeepSpeed?

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.

Distributed, Effective, and Efficient Training with Ease

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

Memory efficiency

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.


Fast convergence for effectiveness

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,

Good Usability

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.

Getting Started


  • 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.

Writing DeepSpeed Models

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,

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

    #weight update

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 an train_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.

Model Checkpointing

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

    #weight update

    #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 Configuration

DeepSpeed features can be enabled, disabled, or configured using a config JSON file that should be specified as args.deepspeed_config. 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

Multi-Node Environment Variables

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:


DeepSpeed will then make sure that these environment variables are set when launching each process on every node across their training job.

Launching DeepSpeed Training

DeepSpeed installs the entry point deepspeed to launch distributed training. We illustrate an example usage of DeepSpeed with the following assumptions:

  1. You have already integrated DeepSpeed into your model
  2. client_entry.py is the entry script for your model
  3. client args is the argparse command line arguments
  4. ds_config.json is the configuration file for DeepSpeed

Resource Configuration (multi-node)

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

MPI Compatibility

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.

Resource Configuration (single-node)

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.

Further Reading

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.

Contributor License Agreement

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.

Code of Conduct

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.


  1. Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, Yuxiong He. (2019) ZeRO: Memory Optimization Towards Training A Trillion Parameter Models. ArXiv:1910.02054
comments powered by Disqus