PyTorch 2.x: What’s coming up next for export-path
summary
A workflow that produces a single, inspectable model graph using torch Dynamo bytecode capture and export, returning explicit assumptions (shapes, device, dtype) and rejecting unsupported patterns; a reduced operator target A10 operator subset (~750 ops) and an even smaller primitive operator set (150–250 ops) with purely functional, contiguous semantics to eliminate mutation/aliasing/strides; bundled graph utilities for DCE, CSE and fusion so backends avoid reinventing passes; APIs appear in the nightly now and aim to cover the majority of benchmark models.
AI-written summary · about
chapters (6)
- 0:00PyTorch 2.x export path intro
- 1:50Export vision components
- 4:59Export API and graph capture
- 6:28Simplified operator set A10
- 8:56Graph transformation tools
- 10:01Future plans and API
related talks
10mLightning Talk: Streamlining Model Export with the New ONNX Exporter
15mLightning Talk: PT2 Export - A Sound Full Graph Capture Mechanism for PyTorch
11mWhat's New in Torch.onnx
13mPyTorch 2.0
05mPyTorch ONNX Export Support
20mPyTorch Edge: Developer Journey for Deploying AI Models Onto Edge Devices
27mBeyond TorchScript: The Future of C++ Inference in PyTorch
24mAdvancements in High-Performance AI/ML through PyTorch's Python Compiler