Researcher Praises Together's Torch Compile vs Megakernel Tradeoff Study

simran_s_arora · x · 2026-11-17

Weili amplifies and comments on research from the Together Compute team (Simran Arora, Dylan Lim, Reyna Abhyankar) that investigated tradeoffs between Torch Compile and Megakernels, leading to Together Compile.

The comment stresses investigating both sides, citing the authors' note about a Hazy lab tradition of comparing both approaches, tracing back to S4 no-attention vs FlashAttention.

Original post →

More from Infra

Infra channel →