LlamaPure 1.0.3

LlamaPure

A dependency-free .NET Standard 2.0 P/Invoke wrapper for llama.cpp (CPU inference, Completion, and Embeddings).

Design Philosophy

LlamaPure is designed for environments where the addition of an extensive package dependency graph is restricted or undesirable — such as large legacy enterprise systems, strict compliance-driven build pipelines, or projects where tight control over every binary in the deployment package is required.

The library achieves this through a single architectural choice: a direct, thin P/Invoke layer over the official llama.dll. It introduces no third-party NuGet dependencies and references nothing beyond System and System.Runtime.InteropServices. The public API is intentionally narrow, covering only the two inference operations most commonly needed in line-of-business applications:

• Text Completion — autoregressive token generation from a prompt
• Text Embedding — dense float-vector extraction from text

Everything else is left to the caller, keeping the integration footprint as small as possible.

Requirements

Installation

Add the NuGet package

dotnet add package LlamaPure

Or via the Package Manager Console:

Install-Package LlamaPure

Quick Start

Text Completion

using LlamaPure;

using (var client = new LlamaPureClient(@"C:\models\llama-3.2-1b.Q4_K_M.gguf"))
{
    string result = client.Complete("The capital of France is", maxNewTokens: 32);
    Console.WriteLine(result);
}

Text Embedding

using LlamaPure;

using (var client = new LlamaPureClient(@"C:\models\nomic-embed-text.Q4_K_M.gguf"))
{
    float[] vector = client.GetEmbedding("Hello, world!");
    Console.WriteLine("Embedding dimensions: " + vector.Length);
}

Configuring Context and Threads

using LlamaPure;

using (var client = new LlamaPureClient(
    modelPath: @"C:\models\llama-3.2-1b.Q4_K_M.gguf",
    contextSize: 4096,
    threads: 8))
{
    string answer = client.Complete("Explain the Pythagorean theorem.", maxNewTokens: 128);
    Console.WriteLine(answer);
}

All examples use string and standard managed arrays and are compatible with C# 6.0 and later.

API Reference

LlamaPureClient

public sealed class LlamaPureClient : IDisposable

Constructor

LlamaPureClient(string modelPath, uint contextSize = 2048, int threads = 4)

Initializes the llama backend, loads the GGUF model, and creates an inference context. Throws InvalidOperationException if the model cannot be loaded.

Complete

string Complete(string prompt, int maxNewTokens = 128)

Tokenizes prompt, performs a prefill decode, then generates up to maxNewTokens tokens using greedy sampling. Generation stops early when an end-of-generation token is produced. Returns only the generated continuation; the prompt text is not included in the return value.

GetEmbedding

float[] GetEmbedding(string text)

Tokenizes text, executes a single forward pass with embeddings enabled, and returns the last-token hidden state as a float[] of length equal to the model's embedding dimension (n_embd). Intended for use with embedding-capable models.

Dispose

void Dispose()

Releases the inference context and model, and frees the llama backend. Idempotent — safe to call multiple times.

Building from Source

git clone --recurse-submodules https://github.com/taqu/LlamaPure
cd LlamaPure
dotnet build LlamaPure/LlamaPure.csproj -c Release

To build the native DLLs from the included llama.cpp submodule:

cd LlamaWrapper
build.bat

License

This software is distributed under two licenses 'The MIT License' or 'Public Domain', choose whichever you like.