PyTorch 2.0: TorchInductor
summary
TorchInductor is a PyTorch-native compiler backend written in Python that maps PyTorch models into generated Triton kernels on NVIDIA GPUs and C++/OpenMP on CPUs. It implements a pythonic define-by-run loop IR (the IR is a Python callable), reasons about dynamic shapes with Senpai symbolic math, lowers ~250 primitives to ~50 loop-level operators, and reports up to 1.86x GPU and 1.26x CPU geomean speedups on realistic benchmarks.
AI-written summary · about
related talks
15mLightning Talk: AOTInductor: Ahead-of-Time Compilation for PT2 Exported Models
13mLightning Talk: Adding Backends for TorchInductor: Case Study with Intel GPU
08mA Halide Backend for TorchInductor
24mAdvancements in High-Performance AI/ML through PyTorch's Python Compiler
36mDL Compiler Panel Discussion
16mLightning Talk: Triton Compiler
25mRace to Superintelligence: Building Low-Friction AI Infrastructure for a Heterogen...
13mPyTorch 2.0