PyTorch 2.9 Released With Easier Install Support For AMD ROCm & Intel XPUs
PyTorch 2.9 is out today ahead of the PyTorch Conference happening next week in San Francisco. Notable with PyTorch 2.9 is better AMD ROCm and Intel XPU installation support via expanded wheel variant support.
Building off the initial Python wheel variant support within PyTorch 2.8 that was catering to NVIDIA CUDA on Windows, Python 2.9 now brings wheel variant support to AMD ROCm and Intel XPU platforms on Linux.
The wheel variant support is for better hardware/software platform detection for automatically detecting platform attributes for properly handling the correct Python package installation. This makes it easier installing PyTorch (and other Python packages once the WheelNext standard is widely adopted) without resorting to different package names or specifying the wheel index for obtaining the proper packages for your given hardware/software compute platform.
More background information on this wheel variant support around PyTorch to benefit AMD ROCm deployments can be found via this AMD blog post.
Also enhancing the AMD ROCm support with PyTorch 2.9 is supporting the OCP micro-scaling formats of mx-fp8 and mx-fp4. This is focused on AMD GFX950 hardware with ROCm 7.0.
PyTorch 2.9 also introduces symmetric memory for easier programming of multi-GPU kernels, FlexAttention support for Intel GPUs, ARM platform improvements and optimizations, and a variety of other enhancements.
More details on the many PyTorch 2.9 changes via GitHub and the PyTorch.org announcement.
Building off the initial Python wheel variant support within PyTorch 2.8 that was catering to NVIDIA CUDA on Windows, Python 2.9 now brings wheel variant support to AMD ROCm and Intel XPU platforms on Linux.
The wheel variant support is for better hardware/software platform detection for automatically detecting platform attributes for properly handling the correct Python package installation. This makes it easier installing PyTorch (and other Python packages once the WheelNext standard is widely adopted) without resorting to different package names or specifying the wheel index for obtaining the proper packages for your given hardware/software compute platform.
More background information on this wheel variant support around PyTorch to benefit AMD ROCm deployments can be found via this AMD blog post.
Also enhancing the AMD ROCm support with PyTorch 2.9 is supporting the OCP micro-scaling formats of mx-fp8 and mx-fp4. This is focused on AMD GFX950 hardware with ROCm 7.0.
PyTorch 2.9 also introduces symmetric memory for easier programming of multi-GPU kernels, FlexAttention support for Intel GPUs, ARM platform improvements and optimizations, and a variety of other enhancements.
More details on the many PyTorch 2.9 changes via GitHub and the PyTorch.org announcement.