QiMeng-Xpiler: Transcompiling Tensor Programs for Deep Learning Systems with a...
summary
QiMeng-Xpiler automatically transcompiles tensor programs by combining LLM-assisted transformation passes with SMT-based repair of small faulty snippets, and it applies a hierarchical auto-tuning pipeline (uses MCTS for interpass search) to pick pass orders and tune tiling/loop params. Evaluation across four DLS (Intel DL Boost/VNNI, NVIDIA CUDA, AMD MI/HIP, Cambricon MLU/BANG) on 168 cases reports 95% functional-correctness on average, 0.78× median performance versus hand-tuned kernels with up to 2× wins on some operators, and 30–96× developer productivity speedups.
AI-written summary · about
chapters (11)
- 0:00QiMeng-Xpiler introduction
- 1:22Cross-platform DL challenges
- 2:50Code generation methods
- 3:34Neuro-symbolic approach
- 4:19Transcompiler components
- 5:48Performance auto-tuning
- 6:11Evaluation setup
- 7:18Correctness results
- 8:46FA1 and FA2 case study
- 9:28Productivity improvements
- 10:10Conclusion
related talks
24mVerified Tensor-Program Optimization Via High-Level Scheduling Rewrites
28mFunctional Sparse Tensor Compilation
15mPET: Optimizing Tensor Programs with Partially Equivalent Transformations and Automated
15mEnabling Tensor Language Model to Assist in Generating High-Performance Tensor...
12mCoding Agents for Compiler Construction: Beyond the...
26mThunder: Distribute and Optimize Your PyTorch Models With Zero...
41mHelion's New CuTe DSL and Pallas Backends and How We Built Them Via Agents
11mDemo: Running Transformers on Semidynamic's "All-In-One" Vector and Tensor Unit