All news

Open-source TurboPairFormer speeds up OpenFold3 fine-tuning by up to 1.73 times

Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .

Biological models can be fine-tuned faster and more reproducibly on existing GPUs, allowing scientists to test more hypotheses on the same equipment.

Lambda, Stevens Institute of Technology and the OpenFold team released TurboPairFormer, open-source GPU kernels for triangle attention and multiplication operations in biomolecular structure prediction models. On 16 Nvidia H100 GPUs with a crop size of 768 tokens, OpenFold3 training became up to 1.73 times faster than with the original Triton backend. BF16 rounding compensation reduced mean gradient error by 28–47% in controlled experiments, while a fixed accumulation order produced bit-identical gradients in all 600 repeated-run tests. The package is available through PyPI and supports H100 and H200; the speedup enables more frequent fine-tuning of models for proteins, nucleic acids and drug discovery on the same GPUs.

Sources 1