Bayesian Code Diffusion for Efficient Automatic Deep Learning Program Optimization
summary
Introduces Bayesian code diffusion that clusters similar subgraphs, picks a prior via cosine similarity of tensor shapes, and propagates loop-splitting parameters (initial tile size / split length) with rescaling and nearest-candidate heuristics into posterior subgraphs implemented in Ansor; adds pre-training and fine-tuning of the cost model. Builds an autotuner that reduces fine-tuning iterations and end-to-end compilation time, reporting averages of ~2.52× speedup on CPU, 2.0× on GPU and up to 3.31× in some setups.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:00Deep learning compilers and autotuning
- 3:11Ansor’s autotuning process
- 4:36Challenges in Ansor’s search space
- 5:18Improving autotuning efficiency
- 6:45Bayesian Code Diffusion method
- 8:14Parameter propagation between subgraphs
- 8:58Subgraph clustering for similarity
- 9:20Selecting prior and posterior subgraphs
- 10:03Optimizing the prior subgraph
- 10:44Fine-tuning posterior subgraphs
- 11:51Adapting split parameters
- 13:36Experimental setup and evaluation
- 14:21Compilation time speedup over Ansor
- 15:46Summary
related talks
34mBayesian Optimization and it's application to Neural Networks
01h 22mWriting effective bayesian programs using TensorFlow and TFP
15mEffectively Scheduling Computational Graphs of Deep Neural Networks toward Their...
26mAddressing Some of Challenges When Optimizing Long-to-Train-Models
45mScale Learning and Reasoning Across Heterogeneous Gradients and Semantics
28mBayesian Optimization- Fundamentals, Implementation, and Practice
01h 53mMastering Agentic Coding and GPUs
19mOptimizations and abstractions for sparse machine learning