Currently being rewritten in the rewrite branch.

tinyrwkv: A tinier port of RWKV-LM

A port of the RWKV-LM family of large language models to the tinygrad framework.

Roadmap

• Implement the WKV kernel as a custom function
• Implement the backwards of the WKV kernel as a custom function
• Add support for the world model and tokenizer
• Add support for the MIDI models
• Add initial support for RWKV-5 models

Dependencies

Currently, requires tinygrad from git or just use the nix flake.

Python

numpy
pydot (only for GRAPH=1)
tinygrad
tokenizers
torch (only for loading pytorch weights)
tqdm
wandb (optional during training)

System

rust (only for compiling)
clang (only for compiling)
graphviz (only for GRAPH=1)

Usage

Run the CLI with python -m cli.

Also, usable as a python package to embed in other projects. It's also possible to compile the model to portable C code and embed it that way.

usage: tinyrwkv-cli [-h] [--seed SEED] {pre,gen,cht,cmp,bch,ptr,gpt,tra,bpt,wkv,mus} ...

CLI for tinyrwkv

positional arguments:
  {pre,gen,cht,cmp,bch,ptr,gpt,tra,bpt,wkv,mus}
    pre                 preprocess either tinyrwkv trained weights or pytorch trained weights into RNN form
    gen                 freeform generation using the RNN mode (requires a preprocessed model using `pre`)
    cht                 chat with a model in RNN mode (requires a preprocessed model using `pre`)
    cmp                 compile a RNN model into c source code and a compiled executable (need to run with CLANG=1)
    bch                 benchmark the rnn mode
    ptr                 preprocess pytorch weights weights into GPT form for training or inference
    gpt                 freeform generation using the GPT mode (requires a preprocessed model using `ptr`)
    tra                 pretrain or finetune a model (if finetuning the model needs to be preprocessed with `ptr`)
    bpt                 benchmark the gpt mode
    wkv                 benchmark/test each wkv module
    mus                 music generation using the RNN mode (requires a preprocessed model using `pre`)

options:
  -h, --help            show this help message and exit
  --seed SEED           seed for random

License

See the LICENSE and NOTICE files.