A Year in Review: Quansight’s Contributions to PyTorch in 2023 (& Early 2024)
2023 will be remembered as the year when AI and LLMs took the world by storm. PyTorch took center stage during this revolution due to the rise of torch.compile. The combination of having a fully flexible eager execution model, paired with a compiler with a rather flexible tracer that is able to understand complex Python programs semantically, has certainly been one of the core components fueling these advances.
PyTorch 2.1: Quansight’s Improvements to BSR Sparse Matrix Multiplication
This post presents Quansight work implementing new block sparse row (BSR) kernels for sparse matrix multiplication in PyTorch 2.1.