Skip to main content

Neural network inference engine that delivers GPU-class performance for sparsified models on CPUs

Project description

tool icon  DeepSparse Engine

Sparsity-aware neural network inference engine for GPU-class performance on CPUs

Documentation Main GitHub release Contributor Covenant

A CPU runtime that takes advantage of sparsity within neural networks to reduce compute. Read more about sparsification here.

Neural Magic's DeepSparse Engine is able to integrate into popular deep learning libraries (e.g., Hugging Face, Ultralytics) allowing you to leverage DeepSparse for loading and deploying sparse models with ONNX. ONNX gives the flexibility to serve your model in a framework-agnostic environment. Support includes PyTorch, TensorFlow, Keras, and many other frameworks.



The DeepSparse Engine is tested on Python 3.6-3.10, ONNX 1.5.0-1.12.0, ONNX opset version 11+, and manylinux compliant. Using a virtual environment is highly recommended. Install the engine using the following command:

pip install deepsparse

🔌 DeepSparse Server

The DeepSparse Server allows you to serve models and pipelines from the terminal. The server runs on top of the popular FastAPI web framework and Uvicorn web server. Install the server using the following command:

pip install deepsparse[server]

Single Model

Once installed, the following example CLI command is available for running inference with a single BERT model:

deepsparse.server \
    task question_answering \
    --model_path "zoo:nlp/question_answering/bert-base/pytorch/huggingface/squad/12layer_pruned80_quant-none-vnni"

To look up arguments run: deepsparse.server --help.

Multiple Models

To serve multiple models in your deployment you can easily build a config.yaml. In the example below, we define two BERT models in our configuration for the question answering task:

num_cores: 1
num_workers: 1
    - task: question_answering
      route: /predict/question_answering/base
      model: zoo:nlp/question_answering/bert-base/pytorch/huggingface/squad/base-none
      batch_size: 1
    - task: question_answering
      route: /predict/question_answering/pruned_quant
      model: zoo:nlp/question_answering/bert-base/pytorch/huggingface/squad/12layer_pruned80_quant-none-vnni
      batch_size: 1

Finally, after your config.yaml file is built, run the server with the config file path as an argument:

deepsparse.server config config.yaml

Getting Started with the DeepSparse Server for more info.

📜 DeepSparse Benchmark

The benchmark tool is available on your CLI to run expressive model benchmarks on the DeepSparse Engine with minimal parameters.

Run deepsparse.benchmark -h to look up arguments:

deepsparse.benchmark [-h] [-b BATCH_SIZE] [-shapes INPUT_SHAPES]
                          [-ncores NUM_CORES] [-s {async,sync}] [-t TIME]
                          [-nstreams NUM_STREAMS] [-pin {none,core,numa}]
                          [-q] [-x EXPORT_PATH]

Getting Started with CLI Benchmarking includes examples of select inference scenarios:

  • Synchronous (Single-stream) Scenario
  • Asynchronous (Multi-stream) Scenario

👩‍💻 NLP Inference Example

from deepsparse import Pipeline

# SparseZoo model stub or path to ONNX file
model_path = "zoo:nlp/question_answering/bert-base/pytorch/huggingface/squad/12layer_pruned80_quant-none-vnni"

qa_pipeline = Pipeline.create(

my_name = qa_pipeline(question="What's my name?", context="My name is Snorlax")

NLP Tutorials:

Tasks Supported:

🦉 SparseZoo ONNX vs. Custom ONNX Models

DeepSparse can accept ONNX models from two sources:

  • SparseZoo ONNX: our open-source collection of sparse models available for download. SparseZoo hosts inference-optimized models, trained on repeatable sparsification recipes using state-of-the-art techniques from SparseML.

  • Custom ONNX: your own ONNX model, can be dense or sparse. Plug in your model to compare performance with other solutions.

> wget
Saving to: ‘mobilenetv2-7.onnx’

Custom ONNX Benchmark example:

from deepsparse import compile_model
from deepsparse.utils import generate_random_inputs
onnx_filepath = "mobilenetv2-7.onnx"
batch_size = 16

# Generate random sample input
inputs = generate_random_inputs(onnx_filepath, batch_size)

# Compile and run
engine = compile_model(onnx_filepath, batch_size)
outputs =

The GitHub repository includes package APIs along with examples to quickly get started benchmarking and inferencing sparse models.

Scheduling Single-Stream, Multi-Stream, and Elastic Inference

The DeepSparse Engine offers up to three types of inferences based on your use case. Read more details here: Inference Types.

1 ⚡ Single-stream scheduling: the latency/synchronous scenario, requests execute serially. [default]

single stream diagram

Use Case: It's highly optimized for minimum per-request latency, using all of the system's resources provided to it on every request it gets.

2 ⚡ Multi-stream scheduling: the throughput/asynchronous scenario, requests execute in parallel.

multi stream diagram

PRO TIP: The most common use cases for the multi-stream scheduler are where parallelism is low with respect to core count, and where requests need to be made asynchronously without time to batch them.

3 ⚡ Elastic scheduling: requests execute in parallel, but not multiplexed on individual NUMA nodes.

Use Case: A workload that might benefit from the elastic scheduler is one in which multiple requests need to be handled simultaneously, but where performance is hindered when those requests have to share an L3 cache.

🧰 CPU Hardware Support

With support for AVX2, AVX-512, and VNNI instruction sets, the DeepSparse Engine is validated to work on x86 Intel (Haswell generation and later) and AMD CPUs running Linux. Mac and Windows require running Linux in a Docker or virtual machine.

Here is a table detailing specific support for some algorithms over different microarchitectures:

x86 Extension Microarchitectures Activation Sparsity Kernel Sparsity Sparse Quantization
AMD AVX2 Zen 2, Zen 3 not supported optimized emulated
Intel AVX2 Haswell, Broadwell, and newer not supported optimized emulated
Intel AVX-512 Skylake, Cannon Lake, and newer optimized optimized emulated
Intel AVX-512 VNNI (DL Boost) Cascade Lake, Ice Lake, Cooper Lake, Tiger Lake optimized optimized optimized






Be Part of the Future... And the Future is Sparse!

Contribute with code, examples, integrations, and documentation as well as bug reports and feature requests! Learn how here.

For user help or questions about DeepSparse, sign up or log in to our Deep Sparse Community Slack. We are growing the community member by member and happy to see you there. Bugs, feature requests, or additional questions can also be posted to our GitHub Issue Queue. You can get the latest news, webinar and event invites, research papers, and other ML Performance tidbits by subscribing to the Neural Magic community.

For more general questions about Neural Magic, complete this form.


The project's binary containing the DeepSparse Engine is licensed under the Neural Magic Engine License. Example files and scripts included in this repository are licensed under the Apache License Version 2.0 as noted.


Find this project useful in your research or other communications? Please consider citing:

    title = {Inducing and Exploiting Activation Sparsity for Fast Inference on Deep Neural Networks}, 
    author = {Kurtz, Mark and Kopinsky, Justin and Gelashvili, Rati and Matveev, Alexander and Carr, John and Goin, Michael and Leiserson, William and Moore, Sage and Nell, Bill and Shavit, Nir and Alistarh, Dan}, 
    booktitle = {Proceedings of the 37th International Conference on Machine Learning}, 
    pages = {5533--5543}, 
    year = {2020}, 
    editor = {Hal Daumé III and Aarti Singh}, 
    volume = {119}, 
    series = {Proceedings of Machine Learning Research}, 
    address = {Virtual}, 
    month = {13--18 Jul}, 
    publisher = {PMLR}, 
    pdf = {},
    url = {}

  author    = {Eugenia Iofinova and
               Alexandra Peste and
               Mark Kurtz and
               Dan Alistarh},
  title     = {How Well Do Sparse Imagenet Models Transfer?},
  journal   = {CoRR},
  volume    = {abs/2111.13445},
  year      = {2021},
  url       = {},
  eprinttype = {arXiv},
  eprint    = {2111.13445},
  timestamp = {Wed, 01 Dec 2021 15:16:43 +0100},
  biburl    = {},
  bibsource = {dblp computer science bibliography,}

Project details

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Huawei Huawei PSF Sponsor Microsoft Microsoft PSF Sponsor NVIDIA NVIDIA PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page