Keynote
summary
Announcement of PyTorch 2.0 unveils a new compiler pipeline: graph acquisition via Dynamo, lowering to primitives and a Python-written backend TorchInductor, exposed as torch.compile. Goal is ~30% speedup with no model changes; on A100 aggregated runs show ~42% faster overall, 93% of 163 curated models work out of the box, and AMP runs hit up to ~86% on some suites. Dynamic shapes, distributed speedups and reduced memory use are supported early.
AI-written summary · about
chapters (14)
- 0:00Intro
- 5:35Performance vs Flexibility
- 6:04How to think about these problems
- 6:41Graph acquisition
- 8:57Dynamo
- 9:53Gradients
- 10:34Backend
- 11:43TorchInductor
- 12:27Triton
- 13:43AMP Precision
- 14:34AMP Precision Results
- 18:45Full Graph Mode
- 19:43Torture Export
- 20:12Distributed Mode
related talks
24mAdvancements in High-Performance AI/ML through PyTorch's Python Compiler
13mPyTorch 2.0
54mKeynote: PyTorch 2.1 Technical Deep Dive
15mPyTorch 2.0: Unlocking the Power of Deep Learning with the Torch Compile API
32mKeynote: Fireside Chat with Soumith Chintala, Co-Creator of PyTorch & Merve Noyan
36mDL Compiler Panel Discussion
21mThe New Vision of Compiler-Accelerated
14mLightning Talk: Accelerating PyTorch Models With Torch.compile's C++ Wrapper Mode