Machine learning group i14 has trained and served Laplace 3B, a linear attention model, on AWS Trainium2 using native PyTorch. The implementation focuses on serving efficiency by moving past conventional transformer memory dynamics. This approach enables the system to run on dedicated hardware without relying on traditional attention structures.
A primary operational characteristic of Laplace 3B is that its recurrent state stays fixed at 29.5 MB per sequence regardless of context length. This mechanism avoids the growing key-value cache commonly associated with transformer models. As a consequence, hardware requirements remain stable and operators are able to keep serving memory predictable over extended inputs.