QiMeng-Xpiler: Transcompiling Tensor Programs for Deep Learning Systems with a Neural-Symbolic Approach

The overview of QiMeng-Xpiler, a novel transcompiler for automatic transcompilation of tensor programs across different programming models. QiMeng-Xpiler consists of two parts: (a) neural-symbolic program synthesis, which utilizes LLM to transform code and repair incorrect transformation through symbolic synthesis with limited scales, and (b) hierarchical performance auto-tuning, which systemically explores both the parameters and sequences of transformation passes.

Citation

Please cite using the following bibtex entry:

@article{dong2025qimeng,
  title={QiMeng-Xpiler: Transcompiling Tensor Programs for Deep Learning Systems with a Neural-Symbolic Approach},
  author={Dong, Shouyang and Wen, Yuanbo and Bi, Jun and Huang, Di and Guo, Jiaming and Xu, Jianxing and Xu, Ruibai and Song, Xinkai and Hao, Yifan and Zhou, Xuehai and others},
  journal={arXiv preprint arXiv:2505.02146},
  year={2025}
}