0.13.2

• Fixed issue with multi-gpu inference with models that have a bias in their attention linear layers.