togethercomputer/ParallelKernelBench_Problems
收藏资源简介:
ParallelKernelBench(基准测试)是一个用于评估由大型语言模型(LLM)生成的多GPU CUDA内核的基准测试数据集。该数据集包含87个参考实现,位于reference/目录中,以及输入输出张量生成规范在utils/input_output_tensors.py文件中。输入张量是确定性的,可以通过提供的create_input_tensor函数重现,无需存储.pt文件。数据集以parquet格式存储问题信息,包括问题ID、描述、参考代码路径、默认评估设置(如8个H100 GPU、1024x1024基础形状、bfloat16数据类型和5次试验)等。用户可以使用HuggingFace的datasets库加载数据集,并通过相关工具进行本地输入重现和评估。
ParallelKernelBench (benchmark) is a reference problem dataset for ParallelKernelBench: a benchmark for LLM-generated multi-GPU CUDA kernels. This dataset contains 87 reference implementations in reference/ and the input tensor specification in utils/input_output_tensors.py. Inputs are deterministic and can be reproduced using the create_input_tensor function from that file without needing stored .pt files. The dataset is stored in parquet format with columns including problem_id, stem, reference_code, reference_path, input_tensor_spec_path, and default evaluation settings such as world_size, default_m, default_n, default_dtype, default_trials. It can be loaded via HuggingFace datasets and used for local input reproduction and evaluation.




