LLMUnity allows to integrate LLMs (Large Language Models) in the Unity engine.
It allows to build create intelligent characters that your players can interact for an immersive experience.
LLMUnity is built on top of the awesome llama.cpp and llamafile libraries.

• 💻 Cross-platform! Supports Windows, Linux and macOS (supported versions)
• 🏠 Runs locally without internet access but also supports remote servers
• ⚡ Fast inference on CPU and GPU
• 🤗 Support of the major LLM models (supported models)
• 🔧 Easy to setup, call with a single line code
• 💰 Free to use for both personal and commercial purposes

🧪 Tested on Unity: 2021 LTS, 2022 LTS, 2023
🚦 Upcoming Releases

• Join us at Discord and say hi!
• ⭐ Star the repo and spread the word about the project!
• Submit feature requests or bugs as issues or even submit a PR and become a collaborator!

• Open the Package Manager in Unity: Window > Package Manager
• Click the + button and select Add package from git URL
• Use the repository URL https://github.com/undreamai/LLMUnity.git and click Add

For a step-by-step tutorial you can have a look at our guide:

How to Use LLMs in Unity

Create a GameObject for the LLM ♟️:

• Create an empty GameObject. In the GameObject Inspector click Add Component and select the LLM script (Scripts>LLM).
• Download the default model with the Download Model button (this will take a while as it is ~4GB).
  You can also load your own model in .gguf format with the Load model button (see Use your own model).
• Define the role of your AI in the Prompt. You can also define the name of the AI (AI Name) and the player (Player Name).
• (Optional) By default the LLM script is set up to receive the reply from the model as is it is produced in real-time (recommended). If you prefer to receive the full reply in one go, you can deselect the Stream option.
• (Optional) Adjust the server or model settings to your preference (see Options).

In your script you can then use it as follows 🦄:

using LLMUnity;

public class MyScript {
  public LLM llm;
  
  void HandleReply(string reply){
    // do something with the reply from the model
    Debug.Log(reply);
  }
  
  void Game(){
    // your game function
    ...
    string message = "Hello bot!"
    _ = llm.Chat(message, HandleReply);
    ...
  }
}

You can also specify a function to call when the model reply has been completed.
This is useful if the Stream option is selected for continuous output from the model (default behaviour):

  void ReplyCompleted(){
    // do something when the reply from the model is complete
    Debug.Log("The AI replied");
  }
  
  void Game(){
    // your game function
    ...
    string message = "Hello bot!"
    _ = llm.Chat(message, HandleReply, ReplyCompleted);
    ...
  }

• Finally, in the Inspector of the GameObject of your script, select the LLM GameObject created above as the llm property.

That's all ✨!

You can also:

  void Game(){
    // your game function
    ...
    string message = "Hello bot!"
    _ = llm.Chat(message, HandleReply, ReplyCompleted, false);
    ...
  }

  async void Game(){
    // your game function
    ...
    string message = "Hello bot!"
    string reply = await llm.Chat(message, HandleReply, ReplyCompleted);
    Debug.Log(reply);
    ...
  }

  void WarmupCompleted(){
    // do something when the warmup is complete
    Debug.Log("The AI is warm");
  }

  void Game(){
    // your game function
    ...
    _ = llm.Warmup(WarmupCompleted);
    ...
  }

using UnityEngine;
using LLMUnity;

public class MyScript : MonoBehaviour
{
    LLM llm;
    LLMClient llmclient;

    async void Start()
    {
        // Add and setup a LLM object
        gameObject.SetActive(false);
        llm = gameObject.AddComponent<LLM>();
        await llm.SetModel("mistral-7b-instruct-v0.2.Q4_K_M.gguf");
        llm.prompt = "A chat between a curious human and an artificial intelligence assistant.";
        gameObject.SetActive(true);
        // or a LLMClient object
        gameObject.SetActive(false);
        llmclient = gameObject.AddComponent<LLMClient>();
        llmclient.prompt = "A chat between a curious human and an artificial intelligence assistant.";
        gameObject.SetActive(true);
    }
}

The Samples~ folder contains several examples of interaction 🤖:

• SimpleInteraction: Demonstrates simple interaction between a player and a AI
• ServerClient: Demonstrates simple interaction between a player and multiple AIs using a LLM and a LLMClient
• ChatBot: Demonstrates interaction between a player and a AI with a UI similar to a messaging app (see image below)

To install a sample:

• Open the Package Manager: Window > Package Manager
• Select the LLMUnity Package. From the Samples Tab, click Import next to the sample you want to install.

The samples can be run with the Scene.unity scene they contain inside their folder.
In the scene, select the LLM GameObject and click the Download Model button to download the default model.
You can also load your own model in .gguf format with the Load model button (see Use your own model).
Save the scene, run and enjoy!

LLMUnity uses the Mistral 7B Instruct model by default, quantised with the Q4 method (link).
Alternative models can be downloaded from HuggingFace.
The required model format is .gguf as defined by the llama.cpp.
The easiest way is to download gguf models directly by TheBloke who has converted an astonishing number of models 🌈!
Otherwise other model formats can be converted to gguf with the convert.py script of the llama.cpp as described here.

❕ Before using any model make sure you check their license ❕

LLMUnity allows to have multiple AI characters efficiently.
Each character can be implemented with a different client with its own prompt (and other parameters), and all of the clients send their requests to a single server.
This is essential as multiple server instances would require additional compute resources.

In addition to the LLM server functionality, we define the LLMClient class that handles the client functionality.
The LLMClient contains a subset of options of the LLM class described in the Options.
To use multiple instances, you can define one LLM GameObject (as described in How to use) and then multiple LLMClient objects. See the ServerClient sample for a server-client example.

The LLMClient can be configured to connect to a remote instance by providing the IP address of the server in the host property.
The server can be either a LLMUnity server or a standard llama.cpp server.

• Show/Hide Advanced Options Toggle to show/hide advanced options from below
• Show/Hide Expert Options Toggle to show/hide expert options from below

• Num Threads number of threads to use (default: -1 = all)

• Num GPU Layers number of model layers to offload to the GPU. If set to 0 the GPU is not used. Use a large number i.e. >30 to utilise the GPU as much as possible. If the user's GPU is not supported, the LLM will fall back to the CPU

• Stream select to receive the reply from the model as it is produced (recommended!).
  If it is not selected, the full reply from the model is received in one go

• Advanced options

  • Parallel Prompts number of prompts that can happen in parallel (default: -1 = number of LLM/LLMClient objects)
  • Debug select to log the output of the model in the Unity Editor
  • Port port to run the server

• Download model click to download the default model (Mistral 7B Instruct)

• Load model click to load your own model in .gguf format

• Model the model being used (inside the Assets/StreamingAssets folder)

• Advanced options

  • Load lora click to load a LORA model in .bin format
  • Load grammar click to load a grammar in .gbnf format
  • Lora the LORA model being used (inside the Assets/StreamingAssets folder)
  • Grammar the grammar being used (inside the Assets/StreamingAssets folder)
  • Context Size Size of the prompt context (0 = context size of the model)
  • Batch Size Batch size for prompt processing (default: 512)
  • Seed seed for reproducibility. For random results every time select -1
  • Cache Prompt save the ongoing prompt from the chat (default: true)

    Saves the prompt as it is being created by the chat to avoid reprocessing the entire prompt every time
  • Num Predict number of tokens to predict (default: 256, -1 = infinity, -2 = until context filled)

    This is the amount of tokens the model will maximum predict. When N predict is reached the model will stop generating. This means words / sentences might not get finished if this is too low.
  • Temperature LLM temperature, lower values give more deterministic answers

    The temperature setting adjusts how random the generated responses are. Turning it up makes the generated choices more varied and unpredictable. Turning it down makes the generated responses more predictable and focused on the most likely options.
  • Top K top-k sampling (default: 40, 0 = disabled)

    The top k value controls the top k most probable tokens at each step of generation. This value can help fine tune the output and make this adhere to specific patterns or constraints.
  • Top P top-p sampling (default: 0.9, 1.0 = disabled)

    The top p value controls the cumulative probability of generated tokens. The model will generate tokens until this theshold (p) is reached. By lowering this value you can shorten output & encourage / discourage more diverse output.
  • Min P minimum probability for a token to be used (default: 0.05)

    The probability is defined relative to the probability of the most likely token.
  • Repeat Penalty Control the repetition of token sequences in the generated text (default: 1.1)

    The penalty is applied to repeated tokens.
  • Presence Penalty repeated token presence penalty (default: 0.0, 0.0 = disabled)

    Positive values penalize new tokens based on whether they appear in the text so far, increasing the model's likelihood to talk about new topics.
  • Frequency Penalty repeated token frequency penalty (default: 0.0, 0.0 = disabled)

    Positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model's likelihood to repeat the same line verbatim.
• Expert options

  • tfs_z: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled).

  • typical_p: Enable locally typical sampling with parameter p (default: 1.0, 1.0 = disabled).

  • repeat_last_n: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx-size).

  • penalize_nl: Penalize newline tokens when applying the repeat penalty (default: true).

  • penalty_prompt: Prompt for the purpose of the penalty evaluation. Can be either null, a string or an array of numbers representing tokens (default: null = use original prompt).

  • mirostat: Enable Mirostat sampling, controlling perplexity during text generation (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0).

  • mirostat_tau: Set the Mirostat target entropy, parameter tau (default: 5.0).

  • mirostat_eta: Set the Mirostat learning rate, parameter eta (default: 0.1).

  • n_probs: If greater than 0, the response also contains the probabilities of top N tokens for each generated token (default: 0)

  • ignore_eos: Ignore end of stream token and continue generating (default: false).

• Player Name the name of the player
• AI Name the name of the AI
• Prompt a description of the AI role

• Verbal Verdict

The license of LLMUnity is MIT (LICENSE.md) and uses third-party software with MIT and Apache licenses (Third Party Notices.md).