Search papers, labs, and topics across Lattice.
Sigmoid attention is proposed as a drop-in replacement for softmax attention in biological foundation models, demonstrating improved learned representations (25% higher cell-type separation), faster training (up to 10%), and more stable training. Theoretical analysis reveals that sigmoid attention's globally bounded derivatives and diagonal Jacobian structure contribute to its stability compared to softmax. An optimized GPU kernel, TritonSigmoid, achieves 515 TFLOPS on H100 GPUs, outperforming FlashAttention-2 and FlashSigmoid, further solidifying sigmoid attention's practical advantages.
Ditch softmax attention for sigmoid: it unlocks 25% better cell-type separation, 10% faster training, and rock-solid stability for biological foundation models.
Training stable biological foundation models requires rethinking attention mechanisms: we find that using sigmoid attention as a drop in replacement for softmax attention a) produces better learned representations: on six diverse single-cell datasets, sigmoid achieves 25% higher cell-type separation, better cell-type cohesion metrics, and lower validation loss, b) faster training, models with sigmoid attention train up to 10% faster than their softmax counterparts, and c) more stable training by eliminating inherent sources of instability in softmax attention. We establish that sigmoid attention has globally bounded derivatives ($\leq 0.25$) as opposed to softmax, and a diagonal Jacobian structure in contrast with softmax's dense coupling, which together help alleviate training instabilities. In stress tests on 160M-parameter bidirectional attention models trained without gradient clipping on 8K-token sequences, softmax diverges catastrophically, with gradients exploding by four orders of magnitude, while sigmoid remains stable. Finally, we implement and open-source TritonSigmoid, an efficient GPU kernel that achieves 515 TFLOPS on H100 GPUs, outperforming both FlashAttention-2 and FlashSigmoid, with native padding support, which is essential for biological sequences. Our results establish sigmoid attention as both theoretically grounded and empirically superior for biological foundation models. Code is available at https://github.com/MSDLLCpapers/triton-sigmoid