May 17, 2020

713 words 4 mins read



A PyTorch Extension: Tools for easy mixed precision and distributed training in Pytorch

repo name NVIDIA/apex
repo link
language Python
size (curr.) 14098 kB
stars (curr.) 3940
created 2018-04-23
license BSD 3-Clause “New” or “Revised” License


This repository holds NVIDIA-maintained utilities to streamline mixed precision and distributed training in Pytorch. Some of the code here will be included in upstream Pytorch eventually. The intention of Apex is to make up-to-date utilities available to users as quickly as possible.

Full API Documentation:

GTC 2019 and Pytorch DevCon 2019 Slides


1. Amp: Automatic Mixed Precision

apex.amp is a tool to enable mixed precision training by changing only 3 lines of your script. Users can easily experiment with different pure and mixed precision training modes by supplying different flags to amp.initialize.

Webinar introducing Amp (The flag cast_batchnorm has been renamed to keep_batchnorm_fp32).

API Documentation

Comprehensive Imagenet example

DCGAN example coming soon…

Moving to the new Amp API (for users of the deprecated “Amp” and “FP16_Optimizer” APIs)

2. Distributed Training

apex.parallel.DistributedDataParallel is a module wrapper, similar to torch.nn.parallel.DistributedDataParallel. It enables convenient multiprocess distributed training, optimized for NVIDIA’s NCCL communication library.

API Documentation

Python Source


The Imagenet example shows use of apex.parallel.DistributedDataParallel along with apex.amp.

Synchronized Batch Normalization

apex.parallel.SyncBatchNorm extends torch.nn.modules.batchnorm._BatchNorm to support synchronized BN. It allreduces stats across processes during multiprocess (DistributedDataParallel) training. Synchronous BN has been used in cases where only a small local minibatch can fit on each GPU. Allreduced stats increase the effective batch size for the BN layer to the global batch size across all processes (which, technically, is the correct formulation). Synchronous BN has been observed to improve converged accuracy in some of our research models.


To properly save and load your amp training, we introduce the amp.state_dict(), which contains all loss_scalers and their corresponding unskipped steps, as well as amp.load_state_dict() to restore these attributes.

In order to get bitwise accuracy, we recommend the following workflow:

# Initialization
opt_level = 'O1'
model, optimizer = amp.initialize(model, optimizer, opt_level=opt_level)

# Train your model
with amp.scale_loss(loss, optimizer) as scaled_loss:

# Save checkpoint
checkpoint = {
    'model': model.state_dict(),
    'optimizer': optimizer.state_dict(),
    'amp': amp.state_dict()
}, '')

# Restore
model = ...
optimizer = ...
checkpoint = torch.load('')

model, optimizer = amp.initialize(model, optimizer, opt_level=opt_level)

# Continue training

Note that we recommend restoring the model using the same opt_level. Also note that we recommend calling the load_state_dict methods after amp.initialize.


Python 3

CUDA 9 or newer

PyTorch 0.4 or newer. The CUDA and C++ extensions require pytorch 1.0 or newer.

We recommend the latest stable release, obtainable from We also test against the latest master branch, obtainable from

It’s often convenient to use Apex in Docker containers. Compatible options include:

  • NVIDIA Pytorch containers from NGC, which come with Apex preinstalled. To use the latest Amp API, you may need to pip uninstall apex then reinstall Apex using the Quick Start commands below.
  • official Pytorch -devel Dockerfiles, e.g. docker pull pytorch/pytorch:nightly-devel-cuda10.0-cudnn7, in which you can install Apex using the Quick Start commands.

See the Docker example folder for details.

Quick Start


For performance and full functionality, we recommend installing Apex with CUDA and C++ extensions via

$ git clone
$ cd apex
$ pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

Apex also supports a Python-only build (required with Pytorch 0.4) via

$ pip install -v --no-cache-dir ./

A Python-only build omits:

  • Fused kernels required to use apex.optimizers.FusedAdam.
  • Fused kernels required to use apex.normalization.FusedLayerNorm.
  • Fused kernels that improve the performance and numerical stability of apex.parallel.SyncBatchNorm.
  • Fused kernels that improve the performance of apex.parallel.DistributedDataParallel and apex.amp. DistributedDataParallel, amp, and SyncBatchNorm will still be usable, but they may be slower.

To enable PyProf support, you need to install the packages required by PyProf. To do so, add the “–pyprof” option at installation time:

$ pip install -v --no-cache-dir --global-option="--pyprof" --global-option="--cpp_ext" --global-option="--cuda_ext" ./

Windows support

Windows support is experimental, and Linux is recommended. pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" . may work if you were able to build Pytorch from source on your system. pip install -v --no-cache-dir . (without CUDA/C++ extensions) is more likely to work. If you installed Pytorch in a Conda environment, make sure to install Apex in that same environment.

comments powered by Disqus