Research to Production: PyTorch JIT/TorchScript Updates
summary
TorchScript serializes PyTorch models to run without Python and the JIT compiles and optimizes them for production, released in PyTorch 1.3. A single call torch.jit.script single-call conversion turns an nn.Module into a statically typed IR that preserves autograd and control flow. Runtime includes a lightweight thread-safe C++ interpreter usable on mobile and servers. Language support covers Python control flow, lists, dictionaries and classes. Roadmap items: automatic quantization, mobile deployment, and lowering to TVM, Glow and XLA.
AI-written summary · about
chapters (9)
- 0:00Intro
- 0:15What is the JIT
- 0:51Requirements
- 2:14TorchScript
- 3:52Intermediate Representation
- 4:36Recurrent Neural Network
- 6:11Bridge Version
- 7:15Language Features
- 8:34Whats Next
related talks
22mIncreasing the Impact of AI Through Better Software
27mBeyond TorchScript: The Future of C++ Inference in PyTorch
09mPyTorch Mobile: Past, present and 2.x future
06mTorchVision
24mAdvancements in High-Performance AI/ML through PyTorch's Python Compiler
43mExploring Deep Learning Framework PyTorch
13mLightning Talk: Accelerated Inference in PyTorch 2.X with Torch...
14mLightning Talk: The Fastest Path to Production: PyTorch Inference in Python