Dllama

A simple and easy to use library for doing LLM inference directly from Delphi. It can load GGUF formatted LLMs into CPU or GPU memory. Uses Vulkan back end for acceleration.

Installation

• Download Dllama and extract to a desired location.
• Download a GGUF model from Hugging Face (supported by llama.cpp). I've been testing using Hermes-2-Pro-Mistral-7B-GGUF.
• If you have a Vulkan supported GPU, it will be accelerated for faster inference, otherwise if will use the CPU. You will not be able to use a model larger than your available resources, so take note of the amount of memory that it requires.
• See the examples in installdir\examples folder on how to use Dllama in Delphi. Be sure to update the CModelPath and CModelFilename constants used by the examples to valid values on your system.
• This project was built using Delphi 12.1 (latest), Windows 11 (latest), Intel Core i5-12400F 2500 Mhz 6 Cores, 12 logical, 36GB RAM, NVIDIA RTX 3060 GPU 12GB RAM.
• Please test it and feel free to submit pull requests. I want to make it into something very cool for us Delphi developers.
• If this project is useful to you, consider starring the repo, sponsoring it, spreading the word, etc. Any help is greatly welcomed and appreciated.

Examples

A query example:

uses
  System.SysUtils,
  Dllama.Utils,
  Dllama;
  
const
  // update to your model path
  CModelPath = 'C:\LLM\gguf';

  // update to your model filename
  CModelFilename = 'Hermes-2-Pro-Mistral-7B.Q4_0.gguf';  
  
var
  LOllama: TOllama;
  LUsage: TOllama.Usage;
  LResponse: string;
  LUsage: TDllama.Usage;  
begin
  // clear console
  Console.Clear();

  // display title
  Console.PrintLn('Dllama - Query Example'+Console.CRLF, Console.MAGENTA);

  LDllama := TDllama.Create(); 

  // set model path
  LDllama.SetModelPath(CModelPath);

  // try to load model
  if LDllama.LoadModel(CModelFilename, 1024) then
    begin
      // show loaded model filename
      Console.ClearLine(Console.WHITE);
      Console.Print('Loaded model: "%s"'+Console.CRLF+Console.CRLF, [GetModelFilename()], Console.CYAN);

      // add messages
      LDllama.AddSystemMessage('You are a helpful AI assistant.');
      LDllama.AddUserMessage('How to make KNO3?');

      // display user message
      Console.Print(LDllama.GetUserMessage(), Console.DARKGREEN);

      // do inference
      if LDllama.Inference(LResponse, @LUsage) then
      begin
        // display usage
        Console.PrintLn();
        Console.PrintLn();
        Console.PrintLn('Tokens :: Input: %d, Output: %d, Total: %d', 
          [LUsage.InputTokens, LUsage.OutputTokens, LUsage.TotalTokens], Console.BRIGHTYELLOW);
        Console.PrintLn('Speed  :: Input: %3.2f t/s, Output: %3.2f t/s', 
          [LUsage.TokenInputSpeed, LUsage.TokenOutputSpeed], Console.BRIGHTYELLOW);
      end;
      // unload model
      LDllama.UnloadModel();
    end
  else
    // failed to load model
    Console.Print(Console.CRLF+'failed to load model!', Console.RED)
    
  LDllama.Free();   
end.

Media

Dllama01.mp4

Support

Our development motto:

• We will not release products that are buggy, incomplete, adding new features over not fixing underlying issues.
• We will strive to fix issues found with our products in a timely manner.
• We will maintain an attitude of quality over quantity for our products.
• We will establish a great rapport with users/customers, with communication, transparency and respect, always encouragingng feedback to help shape the direction of our products.
• We will be decent, fair, remain humble and committed to the craft.

Links

• Issues
• Discussions
• Discord
• YouTube
• X (Twitter)
• learndelphi.org
• tinyBigGAMES

Proud to be an Embarcadero Technical Partner.

_{As an Embarcadero Technical Partner, I'm committed to providing high-quality Delphi components and tools that enhance developer productivity and harness the power of Embarcadero's developer tools.}