The latest version of the machine learning library PyTorch is now available. PyTorch 1.8.1 introduces the PyTorch Profiler, a tool for performance analysis and troubleshooting for large-scale deep learning models. According to PyTorch’s team, there was a lack of available tools for such a task, and the tools that did exist missed out on important PyTorch-specific information. Machine learning scientists needed to use a combination of tools or manually add correlation information to meet their needs.
PyTorch Profiler collects both GPU hardware and PyTorch information, correlates them, performs detection of model bottlenecks, and generates recommendations on how to ease those bottlenecks. The information is visualized in TensorBoard so the user can easily see it. In addition, the new Profiler API is natively supported in PyTorch, which means users don’t need to install additional packages to make use of the tool. Other recent improvements in PyTorch 1.8.1 include enabling auto-cast for PyTorch xla, making the torch. submodule import more autocomplete-friendly, and more.
The full release notes for PyTorch 1.8.1 are available here.