Jason Ansel

jasonansel4.jpg

I am currently working on PyTorch at Meta where I started the TorchDynamo and TorchInductor projects, both of which are foundational parts of PyTorch 2.0. TorchDynamo is a Python-level JIT compiler designed to make unmodified PyTorch programs faster. TorchDynamo hooks into the frame evaluation API in CPython (PEP 523) to dynamically modify Python bytecode right before it is executed. TorchInductor is a new compier backend for TorchDynamo that maps arbitrary PyTorch programs to Triton or C++/OpenMP.

Before Meta, I was at GoDaddy helping build a deep learning platform for predicting small business behavior and personalizing experiences across the company. I also created GoDaddy Domain Appraisals, which uses neural networks to predict the resale value of a domain name better than a human expert. I joined GoDaddy in 2013 as part of the acquisition of the startup Locu, which I joined in 2011 while I was simultaneously getting my Ph.D. at MIT CSAIL.

I did my Ph.D. dissertation in the Commit group lead by Saman Amarasinghe. I started the OpenTuner project, an extensible framework for program autotuning. I also created the PetaBricks programming language, a language that incorporates algorithmic choices to allow an integrated autotuner to explore search spaces of program implementations. As an undergraduate, I did research with Gene Cooperman and helped create DMTCP, a user-level distributed checkpoint/restart system.