TorchQL: A Programming Framework for Integrity Constraints in Machine Learning
summary
TorchQL is a PyTorch-integrated query framework that expresses model errors as integrity-constraint violations and ships eight core operators to write those checks. It finds sparse failures like object permanence (≈627/15,000 frames), runs hash-join style evaluation over model predictions instead of materializing tables, and supports arbitrary user-defined functions for vision, time-series, labels, and constrained LLM generation. Benchmarks report 13x faster execution vs. Pandas/MongoDB and queries that are 40% shorter than native Python; constrained generation is more accurate and faster than LMQL.
AI-written summary · about
chapters (13)
- 0:00TorchQL introduction
- 2:10Errors as integrity violations
- 3:58Limitations of existing frameworks
- 4:42TorchQL setup and queries
- 6:10Object permanence in TorchQL
- 8:00Complex constraints in TorchQL
- 9:26Constrained LLM generation
- 11:14TorchQL evaluation
- 13:02Performance vs conciseness
- 14:05Future TorchQL plans
- 15:11Conclusion and availability
- 15:34Q&A: TorchQL vs pandas
- 17:42Q&A: Database concepts
related talks
25mTorchBench: Guarding the Performance of the PyTorch Ecosystem with Continuous Benchmarking
18mLightning Talk: TorchRL - RLHF Support
11mUnlock the Future of Generative AI: TorchTitan's Latest...
36mDL Compiler Panel Discussion
16mLightning Talk: Tensor Query Processing
29mOwned By Statistics: How Kubeflow & MLOps Can Help Secure...
33mAccelerating the LLM Lifecycle on the Cloud
32mTaming the Machine: Basics of ML Models Training and Inference Optimization