EvanOLeary/pallasbench-robust
收藏资源简介:
PallasBench是一个包含45个JAX Pallas GPU内核的基准数据集,具有稳健的评估结果、编译工件和GPU性能指标。它采用了来自SakanaAI的Towards Robust Agentic CUDA Kernel Benchmarking中的稳健评估方法,并针对Pallas/JAX编译管道进行了适配,提供了首个专注于Pallas内核的GPU基准。每个内核都经过GPU兼容性修复(通过块大小钳位解决Triton的100万元素限制),并通过五个稳健性过滤器进行评估,以区分真正正确的实现与退化实现。
PallasBench is a benchmark dataset containing 45 JAX Pallas GPU kernels, with robust evaluation results, compiled artifacts, and GPU performance metrics. It adopts the robust evaluation methodology from *Towards Robust Agentic CUDA Kernel Benchmarking* by SakanaAI, and adapts it to the Pallas/JAX compilation pipeline, providing the first GPU benchmark exclusively focused on Pallas kernels. Each kernel has undergone GPU compatibility fixes (resolving Triton's 1 million element limit via block size clamping) and is evaluated through five robustness filters to distinguish truly correct implementations from degenerate ones.




