ChatRWKV

ChatRWKV (pronounced as "RwaKuv" (rʌkuv in IPA), from 4 major params: R W K V)

RWKV homepage: https://www.rwkv.com

RWKV-7 code: https://github.com/BlinkDL/RWKV-LM/tree/main/RWKV-v7

ChatRWKV is like ChatGPT but powered by my RWKV (100% RNN) language model, which is the only RNN (as of now) that can match transformers in quality and scaling, while being faster and saves VRAM.

Our latest version is RWKV-7 https://arxiv.org/abs/2503.14456 (Preview models: https://huggingface.co/BlinkDL/temp )

Gradio Demo 1: https://huggingface.co/spaces/BlinkDL/RWKV-Gradio-1

Gradio Demo 2: https://huggingface.co/spaces/BlinkDL/RWKV-Gradio-2

RWKV-LM main repo: https://github.com/BlinkDL/RWKV-LM (explanation, fine-tuning, training, etc.)

Chat Demo for developers: https://github.com/BlinkDL/ChatRWKV/blob/main/API_DEMO_CHAT.py

Efficient inference project: https://github.com/BlinkDL/Albatross

RWKV APP: https://github.com/RWKV-APP/RWKV_APP (local inference for Android / iOS)

RWKV Discord: https://discord.gg/bDSBUMeFpc (7k+ members)

Twitter: https://twitter.com/BlinkDL_AI

Homepage: https://www.rwkv.com/

Raw cutting-edge RWKV weights: https://huggingface.co/BlinkDL

GGUF: https://huggingface.co/collections/shoumenchougou/rwkv7-gxx-gguf

HF-compatible RWKV weights: https://huggingface.co/RWKV

Use v2/convert_model.py to convert a model for a strategy, for faster loading & saves CPU RAM.

Note RWKV_CUDA_ON will build a CUDA kernel (much faster & saves VRAM). Here is how to build it ("pip install ninja" first):

# How to build in Linux: set these and run v2/chat.py
export PATH=/usr/local/cuda/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
# How to build in win:
Install VS2022 build tools (https://aka.ms/vs/17/release/vs_BuildTools.exe select Desktop C++). Reinstall CUDA 11.7 (install VC++ extensions). Run v2/chat.py in "x64 native tools command prompt". 

RWKV pip package: https://pypi.org/project/rwkv/ (please always check for latest version and upgrade)

https://github.com/cgisky1980/ai00_rwkv_server Fastest GPU inference API with vulkan (good for nvidia/amd/intel)

https://github.com/cryscan/web-rwkv backend for ai00_rwkv_server

https://github.com/saharNooby/rwkv.cpp Fast CPU/cuBLAS/CLBlast inference: int4/int8/fp16/fp32

https://github.com/JL-er/RWKV-PEFT lora/pissa/Qlora/Qpissa/state tuning

https://github.com/RWKV/RWKV-infctx-trainer Infctx trainer

World demo script: https://github.com/BlinkDL/ChatRWKV/blob/main/API_DEMO_WORLD.py

Raven Q&A demo script: https://github.com/BlinkDL/ChatRWKV/blob/main/v2/benchmark_more.py

RWKV in 150 lines (model, inference, text generation): https://github.com/BlinkDL/ChatRWKV/blob/main/RWKV_in_150_lines.py

🔥 RWKV v5 in 250 lines 🔥 (with tokenizer too): https://github.com/BlinkDL/ChatRWKV/blob/main/RWKV_v5_demo.py

🔥 Building your own RWKV inference engine 🔥: begin with https://github.com/BlinkDL/ChatRWKV/blob/main/src/model_run.py which is easier to understand (used by https://github.com/BlinkDL/ChatRWKV/blob/main/chat.py).

RWKV preprint https://arxiv.org/abs/2305.13048

RWKV v6 illustrated:

Cool Community RWKV Projects:

https://github.com/saharNooby/rwkv.cpp fast i4 i8 fp16 fp32 CPU inference using ggml

https://github.com/harrisonvanderbyl/rwkv-cpp-cuda fast windows/linux & cuda/rocm/vulkan GPU inference (no need for python & pytorch)

https://github.com/Blealtan/RWKV-LM-LoRA LoRA fine-tuning

https://github.com/josStorer/RWKV-Runner cool GUI

More RWKV projects: https://github.com/search?o=desc&q=rwkv&s=updated&type=Repositories

ChatRWKV v2: with "stream" and "split" strategies, and INT8. 3G VRAM is enough to run RWKV 14B :) https://github.com/BlinkDL/ChatRWKV/tree/main/v2

os.environ["RWKV_JIT_ON"] = '1'
os.environ["RWKV_CUDA_ON"] = '0' # if '1' then use CUDA kernel for seq mode (much faster)
from rwkv.model import RWKV                         # pip install rwkv
model = RWKV(model='/fsx/BlinkDL/HF-MODEL/rwkv-4-pile-1b5/RWKV-4-Pile-1B5-20220903-8040', strategy='cuda fp16')

out, state = model.forward([187, 510, 1563, 310, 247], None)   # use 20B_tokenizer.json
print(out.detach().cpu().numpy())                   # get logits
out, state = model.forward([187, 510], None)
out, state = model.forward([1563], state)           # RNN has state (use deepcopy if you want to clone it)
out, state = model.forward([310, 247], state)
print(out.detach().cpu().numpy())                   # same result as above

Here is https://huggingface.co/BlinkDL/rwkv-4-raven/blob/main/RWKV-4-Raven-14B-v7-Eng-20230404-ctx4096.pth in action:

When you build a RWKV chatbot, always check the text corresponding to the state, in order to prevent bugs.

1. Never call raw forward() directly. Instead, put it in a function that will record the text corresponding to the state.

(For v4-raven models, use Bob/Alice. For v4/v5/v6-world models, use User/Assistant)

2. The best chat format (check whether your text is of this format): Bob: xxxxxxxxxxxxxxxxxx\n\nAlice: xxxxxxxxxxxxx\n\nBob: xxxxxxxxxxxxxxxx\n\nAlice:

• There should not be any space after the final "Alice:". The generation result will have a space in the beginning, and you can simply strip it.
• You can use \n in xxxxx, but avoid \n\n. So simply do xxxxx = xxxxx.strip().replace('\r\n','\n').replace('\n\n','\n')

If you are building your own RWKV inference engine, begin with https://github.com/BlinkDL/ChatRWKV/blob/main/src/model_run.py which is easier to understand (used by https://github.com/BlinkDL/ChatRWKV/blob/main/chat.py)

The lastest "Raven"-series Alpaca-style-tuned RWKV 14B & 7B models are very good (almost ChatGPT-like, good at multiround chat too). Download: https://huggingface.co/BlinkDL/rwkv-4-raven

Previous old model results:

中文模型

QQ群 553456870（加入时请简单自我介绍）。有研发能力的朋友加群 325154699。

中文使用教程：https://zhuanlan.zhihu.com/p/618011122 https://zhuanlan.zhihu.com/p/616351661

推荐UI：https://github.com/l15y/wenda

Star History