robust-kbench
收藏资源简介:
robust-kbench是一个用于严格评估核性能和正确性的新基准,它测试了各种设置下的提案正确性,支持前向和后向核优化,并专注于现实世界的应用。该数据集包含了250个神经网络的PyTorch模块任务及其相应的CUDA内核生成结果,用于评估LLM驱动的CUDA内核优化框架。数据集还提供了编译、内存访问和数值正确性的软验证工作流程,以准确分类错误的内核。此外,该数据集还包含了一个端到端的智能工作流程,能够将PyTorch代码转换为工作CUDA内核,优化CUDA运行时,并自动融合多个操作。该数据集旨在解决现有CUDA内核生成基准的局限性,并提供了一个更强大的评估框架,以评估LLM生成的CUDA内核的性能和正确性。
robust-kbench is a novel benchmark for rigorously evaluating the performance and correctness of CUDA kernels. It tests the correctness of kernel proposals across various configurations, supports both forward and backward kernel optimization, and focuses on real-world applications. This benchmark includes 250 PyTorch module tasks for neural networks, along with their corresponding CUDA kernel generation results, and is designed for evaluating LLM-driven CUDA kernel optimization frameworks. It also provides soft validation workflows for compilation, memory access, and numerical correctness to accurately classify erroneous kernels. Furthermore, the benchmark contains an end-to-end intelligent workflow that can convert PyTorch code into functional CUDA kernels, optimize CUDA runtime, and automatically fuse multiple operations. This benchmark aims to address the limitations of existing CUDA kernel generation benchmarks, and provides a more robust evaluation framework to assess the performance and correctness of CUDA kernels generated by LLMs.

- 1通过Sakana AI · 2025年



