makora-ai/triton-gpu-latency
收藏资源简介:
Triton GPU延迟数据集是一个大型数据集,包含大量PyTorch问题(主要来自KernelBench),与候选Triton内核实现及其测量的GPU运行时配对,由MakoraGenerate生成。每一行是一个自包含的Python程序,定义了一个参考的PyTorch模型(使用普通PyTorch操作编写)和一个新模型(通过手写或生成的Triton内核重新实现相同的前向传递),标签是新模型的运行时。该数据集专为训练和评估生成快速GPU内核的模型而构建,例如用于内核合成大型语言模型的监督微调/强化学习数据、代码到延迟奖励模型的监督,或内核自动调优器的基准。数据集包括训练和测试分割,总行数约601k,程序大小在5k-15k字符之间,偶尔可达300k,运行时标签为浮点数(可为空表示失败),并包含问题ID和保留组等测试列。
The Triton GPU Latency Dataset is a large dataset of PyTorch problems (mostly from KernelBench) paired with candidate Triton-kernel implementations and their measured GPU runtimes, generated by MakoraGenerate. Each row is a self-contained Python program that defines a reference Model written with plain PyTorch ops and a ModelNew that re-implements the same forward pass with a hand-written or generated Triton kernel, with the label being the runtime of executing ModelNew. It is built for training and evaluating models that generate fast GPU kernels, such as SFT/RL data for kernel-synthesis LLMs, supervision for code→latency reward models, or a benchmark for kernel autotuners. The dataset includes train and test splits with approximately 601k rows, programs typically 5k–15k characters (occasionally up to ~300k), latency labels as float64 (nullable for failures), and test-only columns like problem_id and holdout_group.




