Intel® Optimization for PyTorch*

PyTorch Machine Learning Framework

• Create, train, and deploy deep learning models using a Python or C++ API.
• Transition from interactive development in eager mode to fast batch runtimes with graph mode.
• Speed up model development with built-in support for distributed training on a variety of platforms.
• Deploy PyTorch models to production servers with TorchServe.

Intel® Optimizations

• Accelerate PyTorch model performance with Intel® oneAPI Deep Neural Network Library features such as graph and node optimizations.
• Automatically use Intel® Deep Learning Boost instruction set features to parallelize and accelerate PyTorch workloads.
• Reduce inference latency for models deployed with TorchServe.
• Perform distributed training with oneAPI Collective Communications Library Bindings for Pytorch*.

Intel® Extension for PyTorch* Optimizations and Features

• Apply the newest performance optimizations not yet in PyTorch using Python API commands.
• Vectorize operations to take advantage of larger register sizes in Intel® Advanced Vector Extensions 2, Intel® Advanced Vector Extensions 512, and Intel® Advanced Matrix Extensions instruction sets.
• Parallelize operations without having to analyze task dependencies.
• Further improve vectorization by quantizing to smaller word lengths such as bfloat16 (BF16) or INT8.
• Use built-in recipes to balance quantization efficiency with minimal accuracy loss.
• Fuse common FP32 and BF16 graph operations such as Conv2D+ReLU or Linear+ReLU.
• Fold mathematical graph operations with a convolution.
• Control aspects of the thread runtime such as multistream inference and asynchronous task spawning.

Documentation

• PyTorch Documentation
• PyTorch Performance Tuning Guide
• Intel Extension for PyTorch
  • GitHub Documentation & Tutorials
  • Installation Guide (All Operating Systems)
  • API Documentation
  • Release Notes
  • System Requirements
  • TorchServe with Intel Extension for PyTorch
• Intel® oneAPI Collective Communications Library (oneCCL) Bindings for PyTorch

View All Documentation

Code Samples

• Visual Quality Inspection for the Pharmaceutical Industry
• Single-Instance Training
• bfloat16 Inference—Imperative Mode
• bfloat16 Inference—TorchScript Mode
• INT8 Deployment—Graph Mode
• C++ Dynamic Library

More Samples

Achieve Up to 1.77x Boost Ratio for Your AI Workloads

Learn the difference between stock PyTorch and the Intel Extension for PyTorch, followed by in-depth explanations of the key techniques that power this extension.

Watch

Increase PyTorch Inference Throughput by 4x

See how to accelerate PyTorch-based inferencing by applying optimizations from the Intel Extension for PyTorch and quantizing to INT8.

Watch

Accelerate MedMNIST Training and Inference with Intel Extension for PyTorch

See how to use Intel Extension for PyTorch for training and inference on the MedMNIST datasets. It is compared against stock PyTorch and shows the performance gain that Intel Extension for PyTorch offers.

Watch

Speed Training 8x Using PyTorch with a oneCCL Back End

Compare the performance of distributed training of the deep learning recommendation model (DLRM) using oneCCL and other leading back-ends.

Read

AI-Based Customer Service Automation—Conversations in the Cloud

MindTitan* and Intel worked together to optimize their TitanCS solution using Intel Extension for PyTorch, achieving improvements on inference performance running on Intel® CPUs and driving better real-time call analysis.

Listen

KT Optimizes Performance for Personalized Text-to-Speech

Technologists from KT (formerly Korea Telecom) and Intel worked together to optimize performance of the company’s P-TTS service. The optimized CPU-based solution increased real-time function (RTF) performance by 22 percent while maintaining voice quality and number of connections.

Learn More

Accelerate Deep Learning with Intel Extension for PyTorch

See how to use Intel Extension for PyTorch to take advantage of optimizations before they become part of a stock PyTorch release. Apply the newest developments to optimize your PyTorch models running on Intel® hardware.

Learn More

Accelerate bfloat16 PyTorch Models

Get an introduction to Intel Extension for PyTorch and Intel® oneAPI Deep Neural Network Library (oneDNN) with a close look into the technology behind the Intel Extension for PyTorch API and graph fusion optimizations.

Learn More

Intel and Facebook* Accelerate PyTorch Performance

Facebook* and Intel collaborated to improve PyTorch performance on 3rd generation Intel® Xeon® Scalable processors by harnessing the new bfloat16 capability in Intel® Deep Learning Boost, and deliver training and inference performance boosts for a variety of model and data types.

Learn More

Intel and Facebook Collaborate to Boost PyTorch CPU Performance

Learn how Intel software optimizations accelerate PyTorch on Intel CPU hardware.

Learn More