nvidia/cutlass has released its latest version, v4.8.1. It is a library that bundles reusable C++ components for CUDA-based matrix multiplication and linear algebra operations together with the CuTe DSL, which runs through a Python interface. The approach is to pick data formats and tile configurations while writing matrix multiplication (GEMM) kernels directly for the GPU, so it is closer to choosing from an existing skeleton than writing a kernel from scratch every time.
The core has two parts. One is the C++ CUTLASS components, built by combining GPU kernels specialized for low-precision arithmetic and Tensor Cores (the compute units dedicated to matrix operations on GPUs). The other is the CuTe DSL, which lets you write Tensor Core kernels from a Python interface and adjust data formats and tile configurations. The README introduces the CuTe DSL's extension capabilities, optional compilation paths, and initial support for Rubin. The repository, created on November 30, 2017, has collected 10,639 stars, and its primary language is C++.
However, no release items showing what changed in v4.8.1 and no performance figures were given. Only the momentum score is filled in, at 4.0, and the other ratings are empty. The license is likewise listed only as NOASSERTION, so the terms of use have to be confirmed directly in the original repository. It is a library for CUDA GEMM and related operations that has continued since 2017, but what this version actually improved is hard to judge before checking the release notes.