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

i14 trains and serves Laplace 3B, a linear‑attention model, on AWS Trainium 2 with native PyTorch. The model’s recurrent state remains 29.5 MB per sequence, independent of context length, eliminating the expanding KV cache typical of transformers and keeping serving memory predictable. This demonstrates efficient, scalable inference on custom ASICs.

Opening Kapyn…