Beyond the KV cache: Training and serving i14’s Laplace 3B linear attention model on AWS Trainium2 with native PyTorch

i14 developed Laplace 3B, a 3-billion-parameter linear attention model with fixed-size recurrent state, trained and served efficiently on AWS Trainium2 using native PyTorch. This approach reduces memory growth during inference compared to traditional transformer softmax attention, enabling faster training and serving at scale.