TurboPairFormer open-source kernels speed up OpenFold3 training up to 1.73x
MoAlQuraishi · x · 2026-11-14
Lambda, Stevens Institute of Technology, and the OpenFold team released TurboPairFormer, open-source GPU kernels targeting OpenFold3's cubically scaling triangle attention (doubling the training crop costs 8x):
- Up to 1.73× faster per optimizer step than the original Triton backend;
- Lowest mean Q/K and pair-bias gradient errors vs an FP64 reference;
- Bitwise-identical gradients across all four gradient types in all 600 fixed-input cases;
- Supports H100/H200: on H200 at 768 pixel inputs, 14.67s vs 22.53s per step (1.54×).
Open-sourced, useful for protein structure model fine-tuning.
More from Infra
- llama.cpp hits v1 with MoE offloading and LRU cache for VRAM-bound models — danielhanchen · 2026-11-14
- Ethan Mollick: recent big outages were not caused by AI, despite viral blame — emollick · 2026-11-14
- Joke: "Claude said it's good code" as L3Harris network update crashes US air traffic control — tekbog · 2026-11-14
- KianV open-source RISC-V CPU series spans single-core to superscalar out-of-order — blaizedsouza · 2026-11-14
- Intel Foundry exec: gate-all-around, backside power and a packaging boom arriving at once — ryanshrout · 2026-11-14
- Intel CEO Lip-Bu Tan joins SGLang Summit to discuss silicon and energy efficiency — BanghuaZ · 2026-11-14