using Microsoft.ML.Tokenizers;
using Microsoft.ML.GenAI.Phi;
using Microsoft.ML.GenAI.Core;
using Microsoft.Extensions.AI;


public class Program
{
    public async static void TestPhi3()
    {
        string device = "cuda";
        var weightFolder = @"G:\model\Phi-3-mini-128k-instruct";
        var model = Phi3ForCausalLM.FromPretrained(weightFolder, "config.json", layersOnTargetDevice: -1, quantizeToInt4: true, targetDevice: device);
        var modelPath = Path.Join(weightFolder, "tokenizer.model");
        var tokenizer = Phi3TokenizerHelper.FromPretrained(modelPath);
        var pipeline = new CausalLMPipeline<Tokenizer, Phi3ForCausalLM>(tokenizer, model, device);
        var client = new Phi3CausalLMChatClient(pipeline);
        var task = """
            你能讲一个有趣的笑话吗?
            """;
        List<ChatMessage> _chatHistory = new();
        _chatHistory.Add(new ChatMessage(ChatRole.System, "你是一个助手,用中文回答用户的问题"));
        _chatHistory.Add(new ChatMessage(ChatRole.User, task));
        var options = new ChatOptions
        {
            StopSequences = ["<|end_of_text|>"],//phi3
            AdditionalProperties = new() { { "max_length", 2048 } },
        };
        await foreach (var response in client.GetStreamingResponseAsync(_chatHistory, options))
        {
            Console.Write(response.Text);
        }

        Console.WriteLine();
        Console.WriteLine("End!");
    }

    public static void Main()
    {
        TestPhi3();
    }
}