Introduction to Getting Faster PyTorch* Programs with TorchDynamo

Learn the principles and techniques for making your PyTorch* programs faster, more useable, and performant with the framework’s easy-to-use, just-in-time (JIT) compiler, TorchDynamo.

Introduced in PyTorch 2.0, TorchDynamo transforms a general Python* program into a computational graph. It works by hooking into the Python frame-evaluation process, and then fundamentally changing and supercharging PyTorch operations at the compiler level in the back end.

This session presents this JIT graph compiler and includes:

• An overview of its design and use, including flexible support for graph acquisition with better performance and usability.
• The novel TorchDynamo technique principles in PyTorch 2.0, including debugging.
• How to use Intel® Extension for PyTorch* for TorchDynamo use.
• Intel’s support for and contributions to the tools.

This session includes demos.

Skill level: Intermediate

Get the following stand-alone versions of the tools:

• Intel Extension for PyTorch
• TorchDynamo (GitHub*)

• Intel Extension for PyTorch (GitHub)
• TorchDynamo Benchmarks (GitHub)

See All Code Samples

• Go Deeper with TorchDynamo
• What’s Behind PyTorch 2.0: TorchDynamo and TorchInductor (Especially for Developers)