Intel today released a new version of the Intel Extension for PyTorch in order to apply optimizations to PyTorch for benefiting Intel's hardware. With the Intel Extension for PyTorch v2.7 release, there is support for new large language models (LLMs) as well as various performance optimizations and other enhancements.

The Intel Extension for PyTorch 2.7 release adds support for the popular DeepSeek-R1 model, including enabling INT8 precision on modern Intel Xeon hardware. The updated Intel extension also supports the recently released Microsoft Phi-4 model -- including Phi-4-mini and Phi-4-multimodal as well.

Intel Extension for PyTorch 2.7 also includes various optimizations to its large language model support in general, including performance optimizations. There is also improved documentation around handling multi-modal models and DeepSeek-R1. The extension is also rebased against the Intel oneDNN 3.7.2 neural network library.

Downloads and more details on this extension for helping speed-up PyTorch on modern Intel processors can be found via [1]GitHub .



[1] https://github.com/intel/intel-extension-for-pytorch/releases/tag/v2.7.0%2Bcpu