GraphNet
收藏资源简介:
GraphNet 是一个包含 2.7K 个真实世界深度学习计算图的大型数据集,跨越了多个深度学习框架下的六个主要任务类别。该数据集旨在为张量编译器研究提供全面评估。GraphNet 的样本涵盖了从几千到 10B 参数的各种规模的模型,支持多种数据类型,如 BF16、FP16 和 FP32。GraphNet 通过自动化流程从主流深度学习框架中提取计算图,并确保所有样本均符合可运行、可序列化、可分解、可静态分析和可访问自定义操作符等约束条件。该数据集可用于评估和开发张量编译器,帮助研究人员和开发人员识别性能瓶颈,指导下一代张量编译器的进化。
GraphNet is a large-scale dataset containing 2.7K real-world deep learning computation graphs, spanning six major task categories across multiple deep learning frameworks. This dataset aims to provide comprehensive evaluations for tensor compiler research. Samples in GraphNet cover models of various scales ranging from thousands to 10 billion parameters, and support multiple data types such as BF16, FP16 and FP32. GraphNet extracts computation graphs from mainstream deep learning frameworks via an automated pipeline, and ensures that all samples meet constraints including being runnable, serializable, decomposable, statically analyzable, and supporting access to custom operators. This dataset can be used to evaluate and develop tensor compilers, assisting researchers and developers in identifying performance bottlenecks and guiding the evolution of next-generation tensor compilers.



