The New Vision of Compiler-Accelerated
summary
Claim: an out-of-box front end funnels most PyTorch models into optimizing backends with no model changes, using TorchDynamo's partial graph capture, runtime guarded graphs and just-in-time recapture. Evidence: crawled 7K+ GitHub models run through Dynamo, integrated with 20+ inference backends and several training backends, and benchmarked >150 models (TorchBench/TIM/HuggingFace) showing a >30% geomean speedup when paired with capable back ends.
AI-written summary · about
chapters (5)
- 0:00Introduction and challenges
- 9:46Understanding Torch Dynamo
- 12:30Performance and integration
- 15:19Future development plans
- 17:21Vision and conclusion
related talks
08mPyTorch 2.0: TorchDynamo
29mKeynote
24mAdvancements in High-Performance AI/ML through PyTorch's Python Compiler
13mLightning Talk: Accelerated Inference in PyTorch 2.X with Torch...
13mPyTorch 2.0
23mBlazing Fast GenAI Inference With Torch.compile
12mLightning Talk: Accelerating Inference on CPU with Torch.Compile
26mThunder: Distribute and Optimize Your PyTorch Models With Zero...