PyTorch 2.0: TorchDynamo
summary
TorchDynamo is presented as PyTorch 2.0’s partial guarded graph capture that lets models fall back to eager when needed and supports data‑dependent control flow (the toy example emits three graphs). A single torch.compile one-liner hands captured graphs to custom backends; the system does just-in-time recapture on guard mismatches. After five years of work it runs on 7K+ models, integrates with 20+ inference backends and multiple training backends, and already brings more model regions under compiler optimization.
AI-written summary · about
chapters (6)
- 0:00Introduction
- 1:40Why graph capture
- 2:17TorchDynamo
- 3:08Partial Graph Capture
- 4:18Numbers
- 6:36Conclusion
related talks
13mPyTorch 2.0
21mThe New Vision of Compiler-Accelerated
13mLightning Talk: Accelerated Inference in PyTorch 2.X with Torch...
24mAdvancements in High-Performance AI/ML through PyTorch's Python Compiler
15mPyTorch 2.0: Unlocking the Power of Deep Learning with the Torch Compile API
43mExploring Deep Learning Framework PyTorch
10mLightning Talk: Streamlining Model Export with the New ONNX Exporter
22mModel-Changing Transforms With Torch.compile