paddlepaddle/halops is an LLM inference operator library targeting NVIDIA Blackwell (SM100). It is a repository published by the ERNIE team, covering GQA attention, FP8/FP4 GEMM, and MegaKernel sparse MLA paths. No latest release is registered in the repository. The language is listed as Cuda and the license is Apache-2.0.
This project focuses on optimizing the attention and matrix multiplication operations used in model inference for Blackwell GPUs. It is built on CUTLASS and CuTe, directly leveraging tcgen05, TMEM, and TMA, and also includes a sparse MLA path. The repository ships a runner and baseline configurations for comparing operators, together with installation and build instructions. So rather than a repository that presents a single operator, it bundles several inference paths and comparison tools in one place. The published description presents performance with DeepGEMM and FlashInfer as comparison targets.
However, the repository's public performance comparisons have not been independently reproduced, so it is difficult to confirm performance differences against other implementations based on the presented results alone. What can be confirmed is that a Blackwell-specific operator path and the tools for comparing it have been made public. When judging actual applicability, one should look separately from the published comparison results at whether it can be reproduced in the target environment.