PyTorch 2.0: Unlocking the Power of Deep Learning with the Torch Compile API
summary
PyTorch 2.0 introduces torch.compile as a one-line API that combines Dynamo, AOTAutograd and torch.inductor to capture graphs, lower ~2–3k ops to ~200–300, fuse operators and emit device‑specific code. Benchmarks on NVIDIA GPUs show 43% speedup over eager mode across ~120 models, with successful compilation about 93% of the time; the stack supports training and inference, custom backends for vendors, and is backed by ~3,000 contributors.
AI-written summary · about
chapters (8)
- 0:00PyTorch overview and evolution
- 3:09PyTorch 2.0 and Torch Compile API
- 4:12How deep learning compilers work
- 7:47JIT Trace to TorchDynamo
- 9:16TorchInductor performance
- 11:00Custom back-ends and contributions
- 12:04Contributing to PyTorch
- 13:27Q&A
related talks
24mAdvancements in High-Performance AI/ML through PyTorch's Python Compiler
14mLightning Talk: Accelerating PyTorch Models With Torch.compile's C++ Wrapper Mode
36mDL Compiler Panel Discussion
29mKeynote
23mBlazing Fast GenAI Inference With Torch.compile
13mPyTorch 2.0
16mtorch.compile Practice and Optimization in Different Scenarios
32mKeynote: Fireside Chat with Soumith Chintala, Co-Creator of PyTorch & Merve Noyan