Skip to content

v1.1.0

Latest
Compare
Choose a tag to compare
@BillHuang2001 BillHuang2001 released this 01 Mar 11:29
· 60 commits to main since this release

Major Changes

Dynamo Integration 🚀

We have transitioned to torch.compile as our compiler backend, replacing TorchScript. This update simplifies the compilation workflow, making it more user-friendly while maintaining high performance and efficiency.

Why Move to torch.compile?

Previously, EvoX used tracing to obtain the computation graph—JAX tracing in pre-1.0.0 and TorchScript in v1.0.0. However, tracing modes were difficult for users, lacked compatibility with many Python functions (especially non-PyTorch functions), and required manual intervention to handle dynamic workflow. While tracing was fast and produced a full computation graph, its limitations made it less flexible and hard to use.

With torch.compile, EvoX can now provide a smoother and robust compilation process:

Ease of Use – No need for explicit scripting or tracing; simply use torch.compile(workflow.step), and your code gets optimized automatically.
Better Performance – The torch.compile backend is actively maintained and optimized, delivering better performance than previous tracing methods, and works on more hardware backends.
Full Python Compatibility – Works directly with Python code and non-PyTorch functions, eliminating compatibility issues.
Future-Proof – Aligns with PyTorch’s long-term strategy, making torch.compile the recommended approach for graph capture.

This transition enhances usability while ensuring compatibility with PyTorch's evolving ecosystem.

What's Changed

Full Changelog: v1.0.1...v1.1.0