Keynote: PyTorch 2.1 Technical Deep Dive
summary
PyTorch 2.1 shows concrete gains: torch.compile reaches broad adoption and reduces graph breaks to boost single‑GPU training 1.8x training on 180+ models and adds ~30% more geometric speed from recent nightly runs, while inference sees about 2x inference vs eager. Numpy interop via torch._npy yields huge wins (examples: 57x numpy on K‑means, 200x when moved to CUDA). Distributed now captures functional collectives for overlapping comms, export IR hits ~98% HuggingFace coverage, and executors already runs on Meta smart glasses and Quest 3.
AI-written summary · about
chapters (7)
- 0:00Introduction to the keynote
- 0:31State of torch.compile
- 10:39PyTorch and NumPy integration
- 19:31Distributed training
- 27:28Inference optimization
- 36:13Torch Export and IR
- 43:28ExecuTorch for edge AI
related talks
13mKeynote
29mKeynote
32mKeynote: Fireside Chat with Soumith Chintala, Co-Creator of PyTorch & Merve Noyan
43mPyTorch Technical Deep Dive
13mPyTorch 2.0
19mIncreasing the Impact of AI Through Better Software (Pytorch Updates)
24mAdvancements in High-Performance AI/ML through PyTorch's Python Compiler
31mPyTorch 2.1 - New Developments