PyTorch's ezyang on formalizing a trivial in-place kernel under PTX semantics

ezyang · x · 2026-11-20

PyTorch core developer Edward Z. Yang explains at a composability sync what it takes to formalize a trivial in-place kernel with PTX semantics and a weak memory model. Even this simplest case turns out to be insanely complicated; the video is framed as a helpful primer rather than novel research.

Original post →

More from Research

Research channel →