We provide a decentralized framework for collaborators to run large model computation (training/inference) without any collaborator getting full access to the model.

We propose a method to perform compression in pipeline parallel setting.