linear-attention

RWKV (pronounced RwaKuv) is an RNN with great LLM performance, which can also be directly trained like a GPT transformer (parallelizable). We are at RWKV-7 "Goose". So it's combining the best of RNN and transformer - great performance, linear time, constant space (no kv-cache), fast training, infinite ctx_len, and free sentence embedding.

[NeurIPS 2024] Official code of ”LION: Linear Group RNN for 3D Object Detection in Point Clouds“

Explorations into the recently proposed Taylor Series Linear Attention

Speed Always Wins: A Survey on Efficient Architectures for Large Language Models

Implementation of Agent Attention in Pytorch

The semantic segmentation of remote sensing images

The semantic segmentation of remote sensing images

Exploring Diffusion Transformer Designs via Grafting

CUDA implementation of autoregressive linear attention, with all the latest research findings

Offical implementation of "MetaLA: Unified Optimal Linear Approximation to Softmax Attention Map" (NeurIPS2024 Oral)

Reference implementation of "Softmax Attention with Constant Cost per Token" (Heinsen, 2024)

Code for the paper "Cottention: Linear Transformers With Cosine Attention"

Implementation of: Hydra Attention: Efficient Attention with Many Heads (https://arxiv.org/abs/2209.07484)

[ICML 2024] Official implementation of "LeaPformer: Enabling Linear Transformers for Autoregressive and Simultaneous Tasks via Learned Proportions."

Official Implementation of SEA: Sparse Linear Attention with Estimated Attention Mask (ICLR 2024)

RWKV Wiki website (archived, please visit official wiki)

LEAP: Linear Explainable Attention in Parallel for causal language modeling with O(1) path length, and O(1) inference

SAUTE is a lightweight transformer-based architecture adapted for dialog modeling

Pure PyTorch implementations of Popular Linear Attention models