llama

llama-go: Run LLMs locally with Go

Go bindings for llama.cpp, enabling you to run large language models locally with hardware acceleration support. Integrate LLM inference directly into Go applications with a clean, idiomatic API.

This is an active fork of go-skynet/go-llama.cpp, which hasn't been maintained since October 2023. The goal is keeping Go developers up-to-date with llama.cpp whilst offering a lighter, more performant alternative to Python-based ML stacks like PyTorch and/or vLLM.

Note: Historical tags use the original module path github.com/go-skynet/go-llama.cpp. For new development, use github.com/tcpipuk/llama-go.

Releases: This fork's tags follow llama.cpp releases using the format llama.cpp-{tag} (e.g. llama.cpp-b6603). This ensures compatibility tracking with the underlying C++ library.

Quick start

# Clone with submodules
git clone --recurse-submodules https://github.com/tcpipuk/llama-go
cd llama-go

# Build the library
make libbinding.a

# Download a test model
wget https://huggingface.co/Qwen/Qwen3-0.6B-GGUF/resolve/main/Qwen3-0.6B-Q8_0.gguf

# Run an example
export LIBRARY_PATH=$PWD C_INCLUDE_PATH=$PWD LD_LIBRARY_PATH=$PWD
go run ./examples/simple -m Qwen3-0.6B-Q8_0.gguf -p "Hello world" -n 50

See getting started guide for detailed instructions and examples for more usage patterns.

Basic usage

package main

import (
    "fmt"
    llama "github.com/tcpipuk/llama-go"
)

func main() {
    model, err := llama.LoadModel(
        "/path/to/model.gguf",
        llama.WithF16Memory(),
        llama.WithContext(512),
    )
    if err != nil {
        panic(err)
    }
    defer model.Close()

    response, err := model.Generate("Hello world", llama.WithMaxTokens(50))
    if err != nil {
        panic(err)
    }

    fmt.Println(response)
}

When building, set these environment variables:

export LIBRARY_PATH=$PWD C_INCLUDE_PATH=$PWD LD_LIBRARY_PATH=$PWD

Thread safety

The library is not thread-safe. For concurrent usage, implement a pool pattern as shown in the API guide.

Documentation

Essential guides

• Getting started - Complete walkthrough from installation to first inference
• Building guide - Build options, hardware acceleration, and troubleshooting
• API guide - Detailed usage examples, configuration options, and thread safety

Quick references

• Examples - CLI example with build instructions and usage options
• Go package documentation - Full API reference
• RELEASE.md - Release process and compatibility tracking

Requirements

• Go 1.25+
• Docker (recommended) or C++ compiler with CMake
• Git with submodule support

The library uses the GGUF model format, which is the current standard for llama.cpp. For legacy GGML format support, use the pre-gguf tag.

Architecture

The library bridges Go and C++ using CGO, keeping the heavy computation in llama.cpp's optimised C++ code whilst providing a clean Go API. This minimises CGO overhead whilst maximising performance.

Key components:

• wrapper.cpp/wrapper.h - CGO interface to llama.cpp
• model.go - Go API using functional options pattern
• llama.cpp/ - Git submodule containing upstream llama.cpp

The design uses functional options for configuration, resource management with finalizers, and streaming callbacks via cgo.Handle for safe Go-C interaction.

Licence

MIT