rwkv.cpp

This is a port of BlinkDL/RWKV-LM to ggerganov/ggml.

This is NOT the original repo

As the AI research is evolving at a rapid pace, I, edbrz9, made a "snapshot" of saharNooby's original repo. This snapshot was made on June 12th 2023.

I also added a flask server with some hardcoded values. This is still a work in progress.

So far I've tried it with the 1b5 model quantized in Q8_0.

https://file.brz9.dev/model/rwkv.cpp-eng98o2-1B5-v12-Q8_0.bin

You can run the flask server with the following command:

$ python rwkv/flask_server.py --model ../model/rwkv.cpp-eng98o2-1B5-v12-Q8_0.bin --port 5349

Then, you can test the API with the following command:

curl -X POST -H "Content-Type: application/json" -d '{"prompt":"Write a hello world program in python", "temperature":0.8, "top_p":0.2, "max_length":250}' http://127.0.0.1:5349/chat

Below is the rest of the original README file.

---

Besides the usual FP32, it supports FP16, quantized INT4, INT5 and INT8 inference. This project is CPU only.

This project provides a C library rwkv.h and a convinient Python wrapper for it.

RWKV is a novel large language model architecture, with the largest model in the family having 14B parameters. In contrast to Transformer with O(n^2) attention, RWKV requires only state from previous step to calculate logits. This makes RWKV very CPU-friendly on large context lenghts.

Loading LoRA checkpoints in Blealtan's format is supported through merge_lora_into_ggml.py script.

Quality and performance

If you use rwkv.cpp for anything serious, please test all available formats for perplexity and latency on a representative dataset, and decide which trade-off is best for you.

Below table is for reference only. Measurements were made on 4C/8T x86 CPU with AVX2, 4 threads.

Format	Perplexity (169M)	Latency, ms (1.5B)	File size, GB (1.5B)
Q4_0	17.507	76	1.53
Q4_1	17.187	72	1.68
Q5_0	16.194	78	1.60
Q5_1	15.851	81	1.68
Q8_0	15.652	89	2.13
FP16	15.623	117	2.82
FP32	15.623	198	5.64

With cuBLAS

Measurements were made on 3060Ti 8G + i7 13700K. Latency per token shown.

Model	Layers on GPU	Format	24 Threads	8 Threads	4 Threads	2 Threads	1 Threads
RWKV-4-Pile-169M	12	Q4_0	20.6 ms	8.6 ms	6.9 ms	6.2 ms	7.9 ms
RWKV-4-Pile-169M	12	Q4_1	21.4 ms	8.6 ms	6.9 ms	6.7 ms	7.8 ms
RWKV-4-Pile-169M	12	Q5_1	22.2 ms	9.0 ms	6.9 ms	6.7 ms	8.1 ms
RWKV-4-Raven-7B-v11	32	Q4_0	94.9 ms	54.3 ms	50.2 ms	51.6 ms	59.2 ms
RWKV-4-Raven-7B-v11	32	Q4_1	94.5 ms	54.3 ms	49.7 ms	51.8 ms	59.2 ms
RWKV-4-Raven-7B-v11	32	Q5_1	101.6 ms	72.3 ms	67.2 ms	69.3 ms	77.0 ms

Note: since there is only ggml_mul_mat() supported with cuBLAS, we still need to assign few CPU resources to execute remaining operations.

How to use

1. Clone the repo

Requirements: git.

git clone --recursive https://github.com/saharNooby/rwkv.cpp.git
cd rwkv.cpp

2. Get the rwkv.cpp library

Option 2.1. Download a pre-compiled library

Windows / Linux / MacOS

Check out Releases, download appropriate ZIP for your OS and CPU, extract rwkv library file into the repository directory.

On Windows: to check whether your CPU supports AVX2 or AVX-512, use CPU-Z.

Option 2.2. Build the library yourself

This option is recommended for maximum performance, because the library would be built specifically for your CPU and OS.

Windows

Requirements: CMake or CMake from anaconda, MSVC compiler.

cmake .
cmake --build . --config Release

If everything went OK, bin\Release\rwkv.dll file should appear.

Windows + cuBLAS

Important: Since there is no cuBLAS static libraries for Windows, after compiling with dynamic libraries following DLLs should be copied from {CUDA}/bin into build/bin/Release: cudart64_12.dll, cublas64_12.dll, cublasLt64_12.dll.

mkdir build
cd build
cmake .. -DRWKV_CUBLAS=ON
cmake --build . --config Release

Linux / MacOS

Requirements: CMake (Linux: sudo apt install cmake, MacOS: brew install cmake, anaconoda: cmake package).

cmake .
cmake --build . --config Release

Anaconda & M1 users: please verify that CMAKE_SYSTEM_PROCESSOR: arm64 after running cmake . — if it detects x86_64, edit the CMakeLists.txt file under the # Compile flags to add set(CMAKE_SYSTEM_PROCESSOR "arm64").

If everything went OK, librwkv.so (Linux) or librwkv.dylib (MacOS) file should appear in the base repo folder.

Linux / MacOS + cuBLAS

mkdir build
cd build
cmake .. -DRWKV_CUBLAS=ON
cmake --build . --config Release

If everything went OK, librwkv.so (Linux) or librwkv.dylib (MacOS) file should appear in the base repo folder.

3. Get an RWKV model

Option 3.1. Download pre-quantized Raven model

There are pre-quantized Raven models available on Hugging Face. Check that you are downloading .bin file, NOT .pth.

Option 3.2. Convert and quantize PyTorch model

Requirements: Python 3.x with PyTorch.

This option would require a little more manual work, but you can use it with any RWKV model and any target format.

First, download a model from Hugging Face like this one.

Second, convert it into rwkv.cpp format using following commands:

# Windows
python rwkv\convert_pytorch_to_ggml.py C:\RWKV-4-Pile-169M-20220807-8023.pth C:\rwkv.cpp-169M.bin float16

# Linux / MacOS
python rwkv/convert_pytorch_to_ggml.py ~/Downloads/RWKV-4-Pile-169M-20220807-8023.pth ~/Downloads/rwkv.cpp-169M.bin float16

Optionally, quantize the model into one of quantized formats from the table above:

# Windows
python rwkv\quantize.py C:\rwkv.cpp-169M.bin C:\rwkv.cpp-169M-Q5_1.bin Q5_1

# Linux / MacOS
python rwkv/quantize.py ~/Downloads/rwkv.cpp-169M.bin ~/Downloads/rwkv.cpp-169M-Q5_1.bin Q5_1

4. Run the model

Requirements: Python 3.x with PyTorch and tokenizers.

Note: change the model path with the non-quantized model for the full weights model.

To generate some text, run:

# Windows
python rwkv\generate_completions.py C:\rwkv.cpp-169M-Q5_1.bin

# Linux / MacOS
python rwkv/generate_completions.py ~/Downloads/rwkv.cpp-169M-Q5_1.bin

To chat with a bot, run:

# Windows
python rwkv\chat_with_bot.py C:\rwkv.cpp-169M-Q5_1.bin

# Linux / MacOS
python rwkv/chat_with_bot.py ~/Downloads/rwkv.cpp-169M-Q5_1.bin

Edit generate_completions.py or chat_with_bot.py to change prompts and sampling settings.

---

Example of using rwkv.cpp in your custom Python script:

import rwkv_cpp_model
import rwkv_cpp_shared_library

# Change to model paths used above (quantized or full weights)
model_path = r'C:\rwkv.cpp-169M.bin'


model = rwkv_cpp_model.RWKVModel(
    rwkv_cpp_shared_library.load_rwkv_shared_library(),
    model_path,
    thread_count=4,    #need to adjust when use cuBLAS
    gpu_layers_count=5 #only enabled when use cuBLAS
)

logits, state = None, None

for token in [1, 2, 3]:
    logits, state = model.eval(token, state)

    print(f'Output logits: {logits}')

# Don't forget to free the memory after you've done working with the model
model.free()

Compatibility

ggml moves fast, and can occasionally break compatibility with older file formats.

rwkv.cpp will attempt it's best to explain why a model file can't be loaded and what next steps are available to the user.

For reference only, here is a list of latest versions of rwkv.cpp that have supported older formats. No support will be provided for these versions.

• Q4_2, old layout of quantized formats
  • commit 3ca9c7f, release with prebuilt binaries
• Q4_3, Q4_1_O
  • commit c736ef5, release with prebuilt binaries

See also FILE_FORMAT.md for version numbers of rwkv.cpp model files and their changelog.

Contributing

There is no complete contributor guide yet; but we have CODE_STYLE.md.