The PyTorch* for Python* package provides one of the fastest implementations of dynamic neural networks to achieve speed and flexibility. Intel and Facebook* extensively collaborated to:
- Include many Intel optimizations in this popular framework
- Provide superior PyTorch performance on Intel® architectures, most notably Intel® Xeon® Scalable processors
The optimizations are built using oneDNN to provide cross-platform support and acceleration.
Intel also provides Intel® Extension for PyTorch* for more capabilities that have not yet been upstreamed, including:
- Support for automatic mixed precision
- Customized operators
- Fusion patterns
This optimization adds bindings with Intel® oneAPI Collective Communications Library (oneCCL) for efficient distributed training and is a consolidated package. It provides the best out-of-box experience to get all of the performance benefits from PyTorch. The package has the latest versions of:
- Stock PyTorch with Intel® optimizations
- Intel Extension for PyTorch
Intel® Optimization for PyTorch* is made available as part of the AI Kit that provides a comprehensive and interoperable set of AI software libraries to accelerate end-to-end data science and machine-learning workflows.
Intel Extension for PyTorch