RWKV: an RNN that matches Transformer performance and trains in parallel

RWKV is a recurrent (RNN) architecture that achieves performance on par with Transformers. Standard Transformers require compute and memory that grow quadratically as input length increases, but RWKV, being RNN-based, scales only linearly with in theory.

At the same time, RWKV is designed to support like a Transformer, overcoming the slow sequential training that normally limits RNNs. A user posted introducing this architecture.

Key points

  • RWKV is RNN-based but reaches Transformer-level LLM performance
  • Transformers scale quadratically with input length; RWKV does not
  • RWKV supports , unlike traditional RNNs
Read original