blinkdl/rwkv-lm

RWKV (pronounced "RwaKuv") is a research project that designs a new kind of large language model. Most modern chat-style models, like GPT and similar systems, use an architecture called the transformer. RWKV takes a different route: it is built as a recurrent neural network (RNN), which means it reads text one token at a time and keeps a small running "state" instead of looking back at the whole conversation. The claim of this repository is that RWKV can match transformer-level quality while keeping the speed and memory advantages of an RNN. The README is centered on RWKV-7, nicknamed "Goose", which the author calls the strongest linear-time, constant-space, attention-free, fully RNN architecture available at the time of writing. Linear-time means the work grows in proportion to the length of the input, and constant-space means it does not need a growing key-value cache the way a transformer does. The project is hosted under the Linux Foundation AI umbrella so the code and weights are free to use, and the README notes that RWKV is already shipped inside Windows and Office. The repository is mostly training code and reference implementations. There are demo scripts for RWKV-7 in GPT-like mode, in pure RNN mode, and in a faster combined mode, with similar files for RWKV-6 and RWKV-5. A simplified training script in RWKV-v7/train_temp can be run on a single GPU with about 7 GB of VRAM, and a fuller script trains a model on the MiniPile dataset using PyTorch, PyTorch Lightning 1.9.5, DeepSpeed, and CUDA. The README is firm about a few training details, like using PreLN LayerNorm, applying weight decay only to large projection matrices, and following the supplied initialization. The README also lists a wide surrounding ecosystem: pre-trained weights and GGUF conversions on Hugging Face, a pip package called rwkv, Gradio and WebGPU chat demos, a graphical runner, an inference server called Ai00, a PEFT and LoRA tuning project, an RLHF project, fast CUDA kernels, and a mobile inference library. A successor architecture, RWKV-8 "ROSA", is mentioned at the end. The full README is longer than what was shown.