PassNet: Scaling Large Language Models for Graph Compiler Pass Generation
📰 ArXiv cs.AI
Learn how PassNet scales large language models for graph compiler pass generation to improve performance on long-tail workloads
Action Steps
- Apply PassNet to generate compiler passes for graph optimization
- Use TorchInductor to profile and identify performance bottlenecks in real-world subgraphs
- Configure large language models to author structured graph passes
- Test the generated passes on long-tail workloads to measure performance improvements
- Compare the results with default compilation to evaluate the effectiveness of PassNet
Who Needs to Know This
ML engineers and researchers working on compiler optimization and large language models can benefit from this knowledge to improve the performance of their models
Key Insight
💡 PassNet can automatically generate optimized compiler passes for graph optimization using large language models
Share This
🚀 PassNet scales LLMs for graph compiler pass generation to boost performance on long-tail workloads! 💻
Key Takeaways
Learn how PassNet scales large language models for graph compiler pass generation to improve performance on long-tail workloads
Full Article
Title: PassNet: Scaling Large Language Models for Graph Compiler Pass Generation
Abstract:
arXiv:2605.29357v1 Announce Type: new Abstract: Modern tensor compilers such as TorchInductor deliver substantial speedups on mainstream models, yet face a systematic performance ceiling on long-tail workloads -- our profiling shows that 43% of real-world subgraphs experience end-to-end slowdowns under default compilation. While LLMs offer a path toward automated optimization, existing efforts focus on standalone kernel generation. We argue that pass generation -- where LLMs author structured gr
Abstract:
arXiv:2605.29357v1 Announce Type: new Abstract: Modern tensor compilers such as TorchInductor deliver substantial speedups on mainstream models, yet face a systematic performance ceiling on long-tail workloads -- our profiling shows that 43% of real-world subgraphs experience end-to-end slowdowns under default compilation. While LLMs offer a path toward automated optimization, existing efforts focus on standalone kernel generation. We argue that pass generation -- where LLMs author structured gr
DeepCamp AI