Luminal - Search-Based Deep Learning Compilers
summary
Search-first deep learning compiler that represents models as graphs of 12 primitive ops and compiles by enumerating equivalent GPU kernels in an e-graph, profiling them and selecting the fastest. It directly emits CUDA, runs inference-first (training via an external autograd), and reports the repo is under 5,000 lines of code. The system actually rediscovers complex optimizations — notably found flash attention automatically — and applies deterministic passes like buffer reuse and kernel-launch batching to make the kernels practical.
AI-written summary · about
chapters (12)
- 0:00The problem with ML software
- 1:05Radical simplification
- 2:26Graph-based model design
- 7:15Compiling for performance
- 8:36The compiler complexity trap
- 11:58Search-driven optimization
- 14:39Automatic kernel fusion
- 17:48Post-search optimizations
- 18:07Memory and execution efficiency
- 20:21Training capabilities
- 21:31Future directions and cloud
- 23:41Community and invitation
related talks
11mQiMeng-Xpiler: Transcompiling Tensor Programs for Deep Learning Systems with a...
15mDesign Compact Deep Learning Models: Small is the New Big
25mOpen Source Model Performance Optimization With SGLang
54mProgramming Language Design and Implementation for the Machine Learning Era: A Personal(…)
01h 02mScheme as a framework for Deep Learning
15mGALA: A High Performance Graph Neural Network Acceleration LAnguage and Compiler
19mOptimizations and abstractions for sparse machine learning
12mLightning Talk: Accelerating Inference on CPU with Torch.Compile