AI Toolkit for Visual Studio Code 入门

下面是 REST 请求的示例正文：

{
    "model": "Phi-3-mini-4k-directml-int4-awq-block-128-onnx",
    "messages": [
        {
            "role": "user",
            "content": "what is the golden ratio?"
        }
    ],
    "temperature": 0.7,
    "top_p": 1,
    "top_k": 10,
    "max_tokens": 100,
    "stream": true
}'

可以使用 REST 等 API 工具或 CURL 实用工具测试 终结点：

curl -vX POST http://127.0.0.1:5272/v1/chat/completions -H 'Content-Type: application/json' -d @body.json

安装 OpenAI Python 库：

pip install openai

from openai import OpenAI

client = OpenAI(
    base_url="http://127.0.0.1:5272/v1/",
    api_key="x" # required by API but not used
)

chat_completion = client.chat.completions.create(
    messages=[
        {
            "role": "user",
            "content": "what is the golden ratio?",
        }
    ],
    model="Phi-3-mini-4k-directml-int4-awq-block-128-onnx",
)

print(chat_completion.choices[0].message.content)

使用 NuGet 将 适用于 .NET 的 Azure OpenAI 客户端库 添加到项目中：

dotnet add {project_name} package Azure.AI.OpenAI --version 1.0.0-beta.17

将名为 OverridePolicy.cs的 C# 文件添加到项目，并粘贴以下代码：

// OverridePolicy.cs
using Azure.Core.Pipeline;
using Azure.Core;

internal partial class OverrideRequestUriPolicy(Uri overrideUri)
    : HttpPipelineSynchronousPolicy
{
    private readonly Uri _overrideUri = overrideUri;

    public override void OnSendingRequest(HttpMessage message)
    {
        message.Request.Uri.Reset(_overrideUri);
    }
}

接下来，将以下代码粘贴到 Program.cs 文件中：

// Program.cs
using Azure.AI.OpenAI;

Uri localhostUri = new("http://localhost:5272/v1/chat/completions");

OpenAIClientOptions clientOptions = new();
clientOptions.AddPolicy(
    new OverrideRequestUriPolicy(localhostUri),
    Azure.Core.HttpPipelinePosition.BeforeTransport);
OpenAIClient client = new(openAIApiKey: "unused", clientOptions);

ChatCompletionsOptions options = new()
{
    DeploymentName = "Phi-3-mini-4k-directml-int4-awq-block-128-onnx",
    Messages =
    {
        new ChatRequestSystemMessage("You are a helpful assistant. Be brief and succinct."),
        new ChatRequestUserMessage("What is the golden ratio?"),
    }
};

StreamingResponse<StreamingChatCompletionsUpdate> streamingChatResponse
    = await client.GetChatCompletionsStreamingAsync(options);

await foreach (StreamingChatCompletionsUpdate chatChunk in streamingChatResponse)
{
    Console.Write(chatChunk.ContentUpdate);
}

请参阅 本地 REST API Web 服务器中显示的示例。 AI Toolkit REST 服务器是使用 ONNX Runtime 生成的。

安装 Numpy：

pip install numpy

接下来，根据平台和 GPU 可用性将 ONNX Runtime Python 包安装到项目中：

接下来，将以下代码复制并粘贴到名为 app.py 的 Python 文件中：

# app.py
import onnxruntime_genai as og
import argparse

def main(args):
    print("Loading model...")
    model = og.Model(f'{args.model}')
    print("Model loaded")
    tokenizer = og.Tokenizer(model)
    tokenizer_stream = tokenizer.create_stream()
    search_options = {
        'max_length': 2048
    }

    chat_template = '<|user|>\n{input} <|end|>\n<|assistant|>'

    # Keep asking for input prompts in a loop
    while True:
        text = input("Input: ")
    
        # If there is a chat template, use it
        prompt = f'{chat_template.format(input=text)}'

        input_tokens = tokenizer.encode(prompt)

        params = og.GeneratorParams(model)
        params.set_search_options(**search_options)
        params.input_ids = input_tokens
        
        generator = og.Generator(model, params)
        print("\nOutput: ", end='', flush=True)
        while not generator.is_done():
            generator.compute_logits()
            generator.generate_next_token()
            new_token = generator.get_next_tokens()[0]
            print(tokenizer_stream.decode(new_token), end='', flush=True)
              
        print()
        print()

        # Delete the generator to free the captured graph for the next generator, if graph capture is enabled
        del generator


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument('-m', '--model', type=str, required=True, help='Onnx model folder path (must contain config.json and model.onnx)')
    args = parser.parse_args()
    main(args)

若要运行 Python 应用，请使用以下代码：

python app.py --model ~/.aitk/models/{path_to_folder_containing_onnx_file}

根据平台和 GPU 可用性将 ONNX Runtime NuGet 包安装到项目中：

将以下代码复制并粘贴到 C# 文件中：

using Microsoft.ML.OnnxRuntimeGenAI;

// update user_name and path placeholders
string modelPath = "C:\\Users\\{user_name}\\.aitk\\models\\{path}"; 
Console.Write("Loading model from " + modelPath + "...");
using Model model = new(modelPath);
Console.Write("Done\n");
using Tokenizer tokenizer = new(model);
using TokenizerStream tokenizerStream = tokenizer.CreateStream();

while (true)
{
    Console.Write("User:");
   
    string? input = Console.ReadLine();
    string prompt = "<|user|>\n" + input + "<|end|>\n<|assistant|>";

    var sequences = tokenizer.Encode(prompt);

    using GeneratorParams generatorParams = new GeneratorParams(model);
    generatorParams.SetSearchOption("max_length", 200);
    generatorParams.SetInputSequences(sequences);

    Console.Out.Write("\nAI:");
    using Generator generator = new(model, generatorParams);
    while (!generator.IsDone())
    { 
        generator.ComputeLogits();
        generator.GenerateNextToken();
        Console.Out.Write(tokenizerStream.Decode(generator.GetSequence(0)[^1]));
        Console.Out.Flush();
    }
    Console.WriteLine();
}