[P] A library for linear RNNs
A new PyTorch library promises faster training and inference for state-space models.
Deep Dive
A team of developers has released LRNNX, a fully open-source PyTorch library implementing several popular Linear Recurrent Neural Networks (RNNs). It features accelerated kernels for both inference and training, similar to the performance gains seen with Mamba. The library, released under an MIT license, includes a technical report and is available on GitHub, inviting community feedback and contributions to advance efficient sequence modeling.
Why It Matters
This could significantly lower the barrier to developing and deploying efficient, long-context AI models for real-time applications.