The C#/.NET binding of llama.cpp. It provides higher-level APIs to inference the LLaMA Models and deploy it on local device with C#/.NET. It works on both Windows, Linux and MAC without requirement for compiling llama.cpp yourself. Even without a GPU or not enough GPU memory, you can still apply LLaMA models well with this repo. 🤗

Furthermore, it provides integrations with other projects such as semantic-kernel, kernel-memory and BotSharp to provide higher-level applications.

Discussions about the roadmap to v1.0.0: #287

Table of Contents

• Documentation
• Examples
• Installation
• Quick Start
  • Model Inference and Chat Session
  • Quantization
  • Web API
• Features
• Console Demo
• FAQ
• Contributing
• Contact us
• Apendix
  • Mapping from LLamaSharp to llama.cpp

Documentation

• Quick start
• Tricks for FAQ
• Full documentation
• API reference

Examples

• Official Console Examples
• Unity Demo
• LLamaStack (with WPF and Web support)

Installation

Firstly, search LLamaSharp in nuget package manager and install it.

PM> Install-Package LLamaSharp

Then, search and install one of the following backends. (Please don't install two or more)

LLamaSharp.Backend.Cpu  # cpu for windows, linux and mac (mac metal is also supported)
LLamaSharp.Backend.Cuda11  # cuda11 for windows and linux
LLamaSharp.Backend.Cuda12  # cuda12 for windows and linux
LLamaSharp.Backend.MacMetal  # Removed after v0.8.0, metal support has been moved to cpu version now

We publish these backends because they are the most popular ones. If none of them matches, please compile the llama.cpp yourself. In this case, please DO NOT install the backend packages, instead, add your DLL to your project and ensure it will be copied to the output directory when compiling your project. For more informations please refer to (this guide).

For microsoft semantic-kernel integration, please search and install the following package:

LLamaSharp.semantic-kernel

For microsoft kernel-memory integration, please search and install the following package (currently kernel-memory only supports net6.0):

LLamaSharp.kernel-memory

Tips for choosing a version

In general, there may be some break changes between two minor releases, for example 0.5.1 and 0.6.0. On the contrary, we don't introduce API break changes in patch release. Therefore it's recommended to keep the highest patch version of a minor release. For example, keep 0.5.6 instead of 0.5.3.

Quick Start

Model Inference and Chat Session

LLamaSharp provides two ways to run inference: LLamaExecutor and ChatSession. The chat session is a higher-level wrapping of the executor and the model. Here's a simple example to use chat session.

using LLama.Common;
using LLama;

string modelPath = "<Your model path>"; // change it to your own model path
var prompt = "Transcript of a dialog, where the User interacts with an Assistant named Bob. Bob is helpful, kind, honest, good at writing, and never fails to answer the User's requests immediately and with precision.\r\n\r\nUser: Hello, Bob.\r\nBob: Hello. How may I help you today?\r\nUser: Please tell me the largest city in Europe.\r\nBob: Sure. The largest city in Europe is Moscow, the capital of Russia.\r\nUser:"; // use the "chat-with-bob" prompt here.

// Load a model
var parameters = new ModelParams(modelPath)
{
    ContextSize = 1024,
    Seed = 1337,
    GpuLayerCount = 5
};
using var model = LLamaWeights.LoadFromFile(parameters);

// Initialize a chat session
using var context = model.CreateContext(parameters);
var ex = new InteractiveExecutor(context);
ChatSession session = new ChatSession(ex);

// show the prompt
Console.WriteLine();
Console.Write(prompt);

// run the inference in a loop to chat with LLM
while (prompt != "stop")
{
    await foreach (var text in session.ChatAsync(prompt, new InferenceParams() { Temperature = 0.6f, AntiPrompts = new List<string> { "User:" } }))
    {
        Console.Write(text);
    }
    prompt = Console.ReadLine();
}

// save the session
session.SaveSession("SavedSessionPath");

Quantization

The following example shows how to quantize the model. With LLamaSharp you needn't to compile c++ project and run scripts to quantize the model, instead, just run it in C#.

string srcFilename = "<Your source path>";
string dstFilename = "<Your destination path>";
string ftype = "q4_0";
if(Quantizer.Quantize(srcFileName, dstFilename, ftype))
{
    Console.WriteLine("Quantization succeed!");
}
else
{
    Console.WriteLine("Quantization failed!");
}

For more usages, please refer to Examples.

Web API

We provide the integration of ASP.NET core and a web app demo. Please clone the repo to have a try.

Since we are in short of hands, if you're familiar with ASP.NET core, we'll appreciate it if you would like to help upgrading the Web API integration.

Features

---

✅: completed. ⚠️: outdated for latest release but will be updated. 🔳: not completed

---

✅ LLaMa model inference

✅ Embeddings generation, tokenization and detokenization

✅ Chat session

✅ Quantization

✅ Grammar

✅ State saving and loading

⚠️ BotSharp Integration

✅ ASP.NET core Integration

✅ Semantic-kernel Integration

🔳 Fine-tune

✅ Local document search (enabled by kernel-memory now)

🔳 MAUI Integration

Console Demo

FAQ

1. GPU out of memory: Please try setting n_gpu_layers to a smaller number.
2. Unsupported model: llama.cpp is under quick development and often has break changes. Please check the release date of the model and find a suitable version of LLamaSharp to install, or generate gguf format weights from original weights yourself.
3. Cannot load native lirary: 1) ensure you installed one of the backend packages. 2) Run NativeLibraryConfig.WithLogs() at the very beginning of your code to print more informations. 3) check if your system supports avx2, which is the default settings of official runtimes now. If not, please compile llama.cpp yourself and specify it with NativeLibraryConfig.WithLibrary.
4. How to find a model: Models in format gguf are valid for LLamaSharp (and ggml before v0.5.1). If you're new to LLM/LLaMA, it's a good choice to search LLama and gguf on huggingface to find a model. Another choice is generate gguf format file yourself with a pytorch weight (or any other), pleae refer to convert.py and convert-llama-ggml-to-gguf.py to get gguf file through a ggml transformation.

Contributing

Any contribution is welcomed! There's a TODO list in LLamaSharp Dev Project and you could pick an interested one to start. Please read the contributing guide for more informations.

You can also do one of the followings to help us make LLamaSharp better:

• Submit a feature request.
• Star and share LLamaSharp to let others know it.
• Write a blog or demo about LLamaSharp.
• Help to develop Web API and UI integration.
• Just open an issue about the problem you met!

Contact us

Join our chat on Discord (please contact Rinne to join the dev channel if you want to be a contributor).

Join QQ group

Apendix

Mapping from LLamaSharp to llama.cpp

Here's the mapping of them and corresponding model samples provided by LLamaSharp. If you're not sure which model is available for a version, please try our sample model.

The llama.cpp commit id will help if you want to compile a DLL yourself.

LLamaSharp	Verified Model Resources	llama.cpp commit id
v0.2.0	This version is not recommended to use.	-
v0.2.1	WizardLM, Vicuna (filenames with "old")	-
v0.2.2, v0.2.3	WizardLM, Vicuna (filenames without "old")	63d2046
v0.3.0, v0.4.0	LLamaSharpSamples v0.3.0, WizardLM	7e4ea5b
v0.4.1-preview	Open llama 3b, Open Buddy	aacdbd4
v0.4.2-preview	Llama2 7b GGML	3323112
v0.5.1	Llama2 7b GGUF	6b73ef1
v0.6.0		cb33f43
v0.7.0, v0.8.0	Thespis-13B, LLaMA2-7B	207b519
v0.8.1		e937066

License

This project is licensed under the terms of the MIT license.