Tag
This paper introduces PassNet, a large-scale ecosystem for LLM-based compiler pass generation, including a dataset of over 18K computational graphs and a benchmark (PassBench) with a new metric. Experiments reveal that while LLMs can achieve up to 3x speedup over TorchInductor on individual subgraphs, consistency remains a bottleneck; fine-tuning a small model on PassNet trajectories yields significant improvements.