Products

DeepSparse

DeepSparse Community

Edit on GitHub

Ask on Slack

  DeepSparse Community

An inference runtime offering GPU-class performance on CPUs and APIs to integrate ML into your application

DeepSparse is a CPU inference runtime that takes advantage of sparsity within neural networks to execute inference quickly. Coupled with SparseML, an open-source optimization library, DeepSparse enables you to achieve GPU-class performance on commodity hardware.

Installation

DeepSparse is available in two editions:

1. DeepSparse Community is free for evaluation, research, and non-production use with our DeepSparse Community License.
2. DeepSparse Enterprise requires a trial license or can be fully licensed for production, commercial applications.

Install via Docker (Recommended)

DeepSparse Community is available as a container image hosted on GitHub container registry.

1docker pull ghcr.io/neuralmagic/deepsparse:1.4.2
2docker tag ghcr.io/neuralmagic/deepsparse:1.4.2 deepsparse-docker
3docker run -it deepsparse-docker

• Check out the Docker page for more details.

Install via PyPI

DeepSparse Community is also available via PyPI. We recommend using a virtual enviornment.

pip install deepsparse

• Check out the Installation page for optional dependencies.

Deployment APIs

DeepSparse includes three deployment APIs:

• Engine is the lowest-level API. With Engine, you pass tensors and receive the raw logits.
• Pipeline wraps the Engine with pre- and post-processing. With Pipeline, you pass raw data and receive the prediction.
• Server wraps Pipelines with a REST API using FastAPI. With Server, you send raw data over HTTP and receive the prediction.

Engine

The example below downloads a 90% pruned-quantized BERT model for sentiment analysis in ONNX format from SparseZoo, compiles the model, and runs inference on randomly generated input.

1from deepsparse import Engine
2from deepsparse.utils import generate_random_inputs, model_to_path
3
4# download onnx, compile
5zoo_stub = "zoo:nlp/sentiment_analysis/obert-base/pytorch/huggingface/sst2/pruned90_quant-none"
6batch_size = 1
7compiled_model = Engine(model=zoo_stub, batch_size=batch_size)
8
9# run inference (input is raw numpy tensors, output is raw scores)
10inputs = generate_random_inputs(model_to_path(zoo_stub), batch_size)
11output = compiled_model(inputs)
12print(output)
13
14# > [array([[-0.3380675 ,  0.09602544]], dtype=float32)] << raw scores

DeepSparse Pipelines

Pipeline is the default API for interacting with DeepSparse. Similar to Hugging Face Pipelines, DeepSparse Pipelines wrap Engine with pre- and post-processing (as well as other utilities), enabling you to send raw data to DeepSparse and receive the post-processed prediction.

The example below downloads a 90% pruned-quantized BERT model for sentiment analysis in ONNX format from SparseZoo, sets up a pipeline, and runs inference on sample data.

1from deepsparse import Pipeline
2
3# download onnx, set up pipeline
4zoo_stub = "zoo:nlp/sentiment_analysis/obert-base/pytorch/huggingface/sst2/pruned90_quant-none"  
5sentiment_analysis_pipeline = Pipeline.create(
6  task="sentiment-analysis",    # name of the task
7  model_path=zoo_stub,          # zoo stub or path to local onnx file
8)
9
10# run inference (input is a sentence, output is the prediction)
11prediction = sentiment_analysis_pipeline("I love using DeepSparse Pipelines")
12print(prediction)
13# > labels=['positive'] scores=[0.9954759478569031]

Additional Resources

• Check out the Use Cases Page for more details on supported tasks.
• Check out the Pipelines User Guide for more usage details.

DeepSparse Server

Server wraps Pipelines with REST APIs, enabling you to stand up model serving endpoint running DeepSparse. This enables you to send raw data to DeepSparse over HTTP and receive the post-processed predictions.

DeepSparse Server is launched from the command line, configured via arguments or a server configuration file. The following downloads a 90% pruned-quantized BERT model for sentiment analysis in ONNX format from SparseZoo and launches a sentiment analysis endpoint:

1deepsparse.server \
2  --task sentiment-analysis \
3  --model_path zoo:nlp/sentiment_analysis/obert-base/pytorch/huggingface/sst2/pruned90_quant-none

Sending a request:

1import requests
2
3url = "http://localhost:5543/predict" # Server's port default to 5543
4obj = {"sequences": "Snorlax loves my Tesla!"}
5
6response = requests.post(url, json=obj)
7print(response.text)
8# {"labels":["positive"],"scores":[0.9965094327926636]}

Additional Resources

• Check out the Use Cases Page for more details on supported tasks.
• Check out the Server User Guide for more usage details.

ONNX

DeepSparse accepts models in the ONNX format. ONNX models can be passed in one of two ways:

• SparseZoo Stub: SparseZoo is an open-source repository of sparse models. The examples on this page use SparseZoo stubs to identify models and download them for deployment in DeepSparse.

• Local ONNX File: Users can provide their own ONNX models, whether dense or sparse. For example:

wget https://github.com/onnx/models/raw/main/vision/classification/mobilenet/model/mobilenetv2-7.onnx

1from deepsparse import Engine
2from deepsparse.utils import generate_random_inputs
3onnx_filepath = "mobilenetv2-7.onnx"
4batch_size = 16
5
6# Generate random sample input
7inputs = generate_random_inputs(onnx_filepath, batch_size)
8
9# Compile and run
10compiled_model = Engine(model=onnx_filepath, batch_size=batch_size)
11outputs = compiled_model(inputs)
12print(outputs[0].shape)
13# (16, 1000) << batch, num_classes

Inference Modes

DeepSparse offers different inference scenarios based on your use case.

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

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

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

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.

• Check out the Scheduler User Guide for more details.

Product Usage Analytics

DeepSparse Community Edition gathers basic usage telemetry including, but not limited to, Invocations, Package, Version, and IP Address for Product Usage Analytics purposes. Review Neural Magic's Products Privacy Policy for further details on how we process this data.

To disable Product Usage Analytics, run the command:

export NM_DISABLE_ANALYTICS=True

Confirm that telemetry is shut off through info logs streamed with engine invocation by looking for the phrase "Skipping Neural Magic's latest package version check." For additional assistance, reach out through the DeepSparse GitHub Issue queue.

Additional Resources

• Benchmarking Performance
• User Guide
• Use Cases
• Cloud Deployments and Demos

Versions

• DeepSparse | stable
• DeepSparse-Nightly | nightly (dev)
• GitHub | releases

Info

• Blog
• Resources

Community

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 Neural Magic 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.

License

DeepSparse Community is licensed under the Neural Magic DeepSparse Community License. Some source code, example files, and scripts included in the deepsparse GitHub repository or directory are licensed under the Apache License Version 2.0 as noted.

DeepSparse Enterprise requires a Trial License or can be fully licensed for production, commercial applications.

Cite

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

1@InProceedings{
2    pmlr-v119-kurtz20a, 
3    title = {Inducing and Exploiting Activation Sparsity for Fast Inference on Deep Neural Networks}, 
4    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}, 
5    booktitle = {Proceedings of the 37th International Conference on Machine Learning}, 
6    pages = {5533--5543}, 
7    year = {2020}, 
8    editor = {Hal Daumé III and Aarti Singh}, 
9    volume = {119}, 
10    series = {Proceedings of Machine Learning Research}, 
11    address = {Virtual}, 
12    month = {13--18 Jul}, 
13    publisher = {PMLR}, 
14    pdf = {http://proceedings.mlr.press/v119/kurtz20a/kurtz20a.pdf},
15    url = {http://proceedings.mlr.press/v119/kurtz20a.html}
16}
17
18@article{DBLP:journals/corr/abs-2111-13445,
19  author    = {Eugenia Iofinova and
20               Alexandra Peste and
21               Mark Kurtz and
22               Dan Alistarh},
23  title     = {How Well Do Sparse Imagenet Models Transfer?},
24  journal   = {CoRR},
25  volume    = {abs/2111.13445},
26  year      = {2021},
27  url       = {https://arxiv.org/abs/2111.13445},
28  eprinttype = {arXiv},
29  eprint    = {2111.13445},
30  timestamp = {Wed, 01 Dec 2021 15:16:43 +0100},
31  biburl    = {https://dblp.org/rec/journals/corr/abs-2111-13445.bib},
32  bibsource = {dblp computer science bibliography, https://dblp.org}
33}