PyTorch — a popular machine learning framework — has been updated to version 2.10.
Among the main innovations are:
Support for AMD ROCm:
- Support for grouped GEMM has been introduced.
- Improved ROCm support for Windows.
- New GPU models GFX1150/GFX1151 (AI 300 series) have been added to the hipblaslt support list.
- Extended functionalities such as support for scaled_mm v2 and AOTriton scaled_dot_product_attention.
Support for Intel GPUs:
- New Torch XPU APIs for Intel have been implemented.
- Support for additional ATen operators.
- Performance optimization for Intel GPUs.
Support for NVIDIA CUDA:
- Enhanced capabilities for writing template kernels.
- Improved support for CUDA 13.
Support for Python 3.14 for torch.compile(), as well as experimental support for building Python 3.14 without a global interpreter lock (free-threaded).
Reduced overhead for launching kernels thanks to horizontal merging of combo kernels in Torch Inductor.
The full changelog is available on github.
Pre-built binaries compiled for different GPUs, OSs, and languages (Python/C++/Java) are listed on the PyTorch website.
Source: linux.org.ru
