Self-Healing Compilers for Dynamic Shape Workloads

BladeDISC and DISC optimize dynamic shapes but assume compile-time knowledge of shape constraints. However, real-world ML workloads (e.g., Korish et al.’s virtual flow meters) encounter unforeseen shape deviations at runtime. This idea proposes a compiler that inserts lightweight "canary" operations to monitor tensor shapes during execution. If anomalies are detected (e.g., shape mismatches causing convolution failures), the compiler triggers a just-in-time recompilation using alternative optimization paths. This goes beyond TVM’s static auto-tuning by combining Zheng et al.’s symbolic shape reasoning with runtime adaptation. For instance, if a fused kernel fails due to unexpected dimensions, the compiler could fall back to unfused operators—similar to how Lebaku et al.’s stacked LSTM detects trajectory anomalies. This bridges the gap between compile-time guarantees and runtime uncertainty.

References:

1. Cybersecurity-Focused Anomaly Detection in Connected Autonomous Vehicles Using Machine Learning. Prathyush Kumar Reddy Lebaku, Lu Gao, Yunpeng Zhang, Zhixia Li, Yongxin Liu, Tanvir Arafin (2025). International Conference on Transportation and Development 2025.
2. Real-Time Production Optimization: A Machine Learning Approach to Virtual Flow Metering. M. Korish, M. Ibrahim, L. Tealdi, A. Al Hanaee (2025). GOTECH.
3. Optimizing Machine Learning Operators and Models for Specific Hardware Using Apache-TVM. Kausthub Thekke Madathil, Abhinav Dugar, Nagamma Patil, Unnikrishnan Cheramangalath (2023). International Conference on Computing Communication and Networking Technologies.
4. TVM: An Automated End-to-End Optimizing Compiler for Deep Learning. Tianqi Chen, T. Moreau, Ziheng Jiang, Haichen Shen, Eddie Q. Yan, Leyuan Wang, Yuwei Hu, L. Ceze, Carlos Guestrin, A. Krishnamurthy (2018).
5. TVM: An Automated End-to-End Optimizing Compiler for Deep Learning. Tianqi Chen, T. Moreau, Ziheng Jiang, Lianmin Zheng, Eddie Q. Yan, Haichen Shen, M. Cowan, Leyuan Wang, Yuwei Hu, L. Ceze, Carlos Guestrin, A. Krishnamurthy (2018). USENIX Symposium on Operating Systems Design and Implementation.
6. BladeDISC: Optimizing Dynamic Shape Machine Learning Workloads via Compiler Approach. Zhen Zheng, Zaifeng Pan, Dalin Wang, Kai Zhu, Wenyi Zhao, Tianyou Guo, Xiafei Qiu, Minmin Sun, Junjie Bai, Feng Zhang, Xiaoyong Du, Jidong Zhai, Wei Lin (2023). Proc. ACM Manag. Data.
7. BladeDISC: Optimizing Dynamic Shape Machine Learning Workloads via Compiler Approach. Zhen Zheng, Zaifeng Pan, Dalin Wang, Kai Zhu, Wenyi Zhao, Tianyou Guo, Xiafei Qiu, Minmin Sun, Junjie Bai, Feng Zhang, Xiaoyong Du, Jidong Zhai, Wei Lin (2023). Proc. ACM Manag. Data.
8. DISC: A Dynamic Shape Compiler for Machine Learning Workloads. Kai Zhu, Wenyi Zhao, Zhen Zheng, Tianyou Guo, Pengzhan Zhao, Junjie Bai, Jun Yang, Xiaoyong Liu, Lansong Diao, Wei Lin (2021). EuroMLSys@EuroSys.