DL Compiler Panel Discussion
summary
PyTorch 2’s torch.compile() drives concrete changes: teams report rebuilding inference stacks around torch.compile() and integrating Triton, TorchInductor, TVM, OpenXLA and even Mojo into deployment paths. Panelists push for tangible projects: a GPU superoptimizer that searches PTX for faster kernels, automation of DMA/tensor core mappings, and an op authoring tool layered on inductor to make custom kernels easier. Reproducible benchmarking, negative-result publishing, and tuning to hit ~80% generality then specialize are pragmatic priorities.
AI-written summary · about
chapters (9)
- 0:00Panel introduction
- 2:05Professional backgrounds
- 5:59Need for ML compilers
- 11:54Future compiler projects
- 17:44Heterogeneous performance
- 20:30Evolution of the landscape
- 22:42ML vs traditional compilers
- 25:30Academia and research
- 30:27Future outlook
related talks
15mPyTorch 2.0: Unlocking the Power of Deep Learning with the Torch Compile API
24mAdvancements in High-Performance AI/ML through PyTorch's Python Compiler
14mLightning Talk: Accelerating PyTorch Models With Torch.compile's C++ Wrapper Mode
32mKeynote: Fireside Chat with Soumith Chintala, Co-Creator of PyTorch & Merve Noyan
29mKeynote
23mBlazing Fast GenAI Inference With Torch.compile
12mLightning Talk: Accelerating Inference on CPU with Torch.Compile
11mLightning Talk: Lessons from Using Pytorch 2.0 Compile in IBM's Watsonx.AI Inference