Infatoshi/kernelbench-hard-problems
收藏资源简介:
KernelBench-Hard Problems 是一个用于自主LLM编码代理的基准测试,专注于在单个Blackwell GPU(RTX PRO 6000,sm_120,CUDA 13.2)上编写GPU内核。数据集包含7个问题定义,每个问题都有特定的任务、形状和运行机制(计算受限或内存受限)。每个问题目录包含参考实现、正确性检查、性能基准测试、元数据、输入形状、SOTA参考和提示文本。评分方法基于峰值分数(peak_fraction),即通过几何平均计算实现的吞吐量与硬件峰值的比率。数据集还详细描述了硬件规格,并指出了两个已知的问题(评分泄漏)。该数据集旨在用于基准测试和提升自主编码代理在GPU内核工程中的性能。
KernelBench-Hard Problems is a benchmark for autonomous LLM coding agents focused on writing GPU kernels on a single Blackwell GPU (RTX PRO 6000, sm_120, CUDA 13.2). The dataset includes 7 problem definitions, each with specific tasks, shapes, and regimes (compute-bound or memory-bound). Each problem directory contains reference implementations, correctness checks, performance benchmarks, metadata, input shapes, SOTA references, and prompt texts. The scoring method is based on peak_fraction, which is the geometric mean of achieved throughput relative to hardware peak. The dataset also details hardware specifications and notes two known issues (rubric leaks). It is designed for benchmarking and improving the performance of autonomous coding agents in GPU kernel engineering.




