[GenAI] Add generateEmbedding API to CausalLMPipeline (#7227)

* add embedding

* add frompretrain api to phi3 model

* fix bug

* Update CausalLMPipeline.cs

Author: LittleLittleCloud

docs/samples/Microsoft.ML.GenAI.Samples/Phi3Mini/AutoGenSample.cs

@@ -9,6 +9,7 @@
 using TorchSharp;
 using Microsoft.ML.GenAI.Core;
 using Microsoft.ML.GenAI.Core.Extension;
+using Microsoft.ML.Tokenizers;
 
 namespace Microsoft.ML.GenAI.Samples.Phi3Mini;
 
@@ -26,12 +27,15 @@ public static async Task RunAsync()
         torch.manual_seed(1);
         torch.set_default_dtype(defaultType);
         var weightFolder = @"C:\Users\xiaoyuz\source\repos\Phi-3-mini-4k-instruct";
-        var pipeline = Utils.LoadPhi3Mini4KFromFolder(weightFolder, device: device, quantizeToInt8: false);
+        var tokenizerPath = Path.Combine(weightFolder, "tokenizer.model");
+        var tokenizer = Phi3TokenizerHelper.FromPretrained(tokenizerPath);
+        var model = Phi3ForCasualLM.FromPretrained(weightFolder, "config.json", layersOnTargetDevice: -1, quantizeToInt8: true);
+        var pipeline = new CausalLMPipeline<LlamaTokenizer, Phi3ForCasualLM>(tokenizer, model, device);
+        var question = @"write a C# program to calculate the factorial of a number";
 
         // agent
         var agent = new Phi3Agent(pipeline, "assistant")
             .RegisterPrintMessage();
-        var question = @"write a C# program to calculate the factorial of a number";
 
         // chat with the assistant
         await agent.SendAsync(question);

docs/samples/Microsoft.ML.GenAI.Samples/Phi3Mini/SemanticKernelSample.cs

@@ -1,4 +1,7 @@
-using Microsoft.ML.GenAI.Phi.Extension;
+using Microsoft.ML.GenAI.Core;
+using Microsoft.ML.GenAI.Phi;
+using Microsoft.ML.GenAI.Phi.Extension;
+using Microsoft.ML.Tokenizers;
 using Microsoft.SemanticKernel;
 using Microsoft.SemanticKernel.ChatCompletion;
 using TorchSharp;
@@ -20,8 +23,10 @@ public static async Task RunChatCompletionSample()
         torch.manual_seed(1);
         torch.set_default_dtype(defaultType);
         var weightFolder = @"C:\Users\xiaoyuz\source\repos\Phi-3-mini-4k-instruct";
-        var pipeline = Utils.LoadPhi3Mini4KFromFolder(weightFolder, device: device);
-
+        var tokenizerPath = Path.Combine(weightFolder, "tokenizer.model");
+        var tokenizer = Phi3TokenizerHelper.FromPretrained(tokenizerPath);
+        var model = Phi3ForCasualLM.FromPretrained(weightFolder, "config.json", layersOnTargetDevice: -1, quantizeToInt8: true);
+        var pipeline = new CausalLMPipeline<LlamaTokenizer, Phi3ForCasualLM>(tokenizer, model, device);
 
         var kernel = Kernel.CreateBuilder()
             .AddGenAIChatCompletion(pipeline)
@@ -49,8 +54,10 @@ public static async Task RunTextGenerationSample()
         torch.manual_seed(1);
         torch.set_default_dtype(defaultType);
         var weightFolder = @"C:\Users\xiaoyuz\source\repos\Phi-3-mini-4k-instruct";
-        var pipeline = Utils.LoadPhi3Mini4KFromFolder(weightFolder, device);
-
+        var tokenizerPath = Path.Combine(weightFolder, "tokenizer.model");
+        var tokenizer = Phi3TokenizerHelper.FromPretrained(tokenizerPath);
+        var model = Phi3ForCasualLM.FromPretrained(weightFolder, "config.json", layersOnTargetDevice: -1, quantizeToInt8: true);
+        var pipeline = new CausalLMPipeline<LlamaTokenizer, Phi3ForCasualLM>(tokenizer, model, device);
 
         var kernel = Kernel.CreateBuilder()
             .AddGenAITextGeneration(pipeline)

docs/samples/Microsoft.ML.GenAI.Samples/Phi3Mini/Utils.cs

@@ -1,4 +1,4 @@
 // See https://aka.ms/new-console-template for more information
 using Microsoft.ML.GenAI.Samples.Phi3Mini;
 
-await SemanticKernelSample.RunChatCompletionSample();
+await AutoGenSample.RunAsync();

docs/samples/Microsoft.ML.GenAI.Samples/Program.cs

@@ -32,6 +32,11 @@ string Generate(
         float topP = CausalLMPipeline.Defaults.TopP,
         string[]? stopSequences = CausalLMPipeline.Defaults.StopSequence);
 
+    /// <summary>
+    /// Generate the embedding(last hidden state of the last token) for the prompt. The embedding is normalized by L2 norm.
+    /// </summary>
+    float[] GenerateEmbeddingFromLastTokenPool(string prompt);
+
     IEnumerable<string> GenerateStreaming(
         string prompt,
         int maxLen = CausalLMPipeline.Defaults.MaxLen,
@@ -281,4 +286,23 @@ protected torch.Tensor SampleTopP(torch.Tensor logits, float topP)
         nextToken = torch.gather(probsIndex, dim: -1, index: nextToken);
         return nextToken;
     }
+
+    public float[] GenerateEmbeddingFromLastTokenPool(string prompt)
+    {
+        using var scope = NewDisposeScope();
+        using var noGrad = torch.no_grad();
+        var inputIds = this.Tokenizer.EncodeToIds(prompt);
+        var inputTensor = torch.tensor(inputIds.ToArray(), dtype: ScalarType.Int64, device: this.Device).unsqueeze(0);
+        var attentionMask = torch.ones_like(inputTensor, device: this.Device);
+        var input = new CausalLMModelInput(inputTensor, attentionMask, pastKeyValuesLength: 0);
+        var output = this.Model.forward(input);
+        var lastTokenHiddenState = output.LastHiddenState[0, ^1];
+
+        // shape of lastTokenHiddenState: [hidden_size]
+        // L2 norm
+        var norm = lastTokenHiddenState.norm();
+        var normalized = lastTokenHiddenState / norm;
+
+        return normalized.to_type(ScalarType.Float32).data<float>().ToArray();
+    }
 }

src/Microsoft.ML.GenAI.Core/Pipeline/CausalLMPipeline.cs

@@ -9,6 +9,7 @@
 using System.Text.Json;
 using System.Threading.Tasks;
 using Microsoft.ML.GenAI.Core;
+using Microsoft.ML.GenAI.Core.Extension;
 using Microsoft.ML.GenAI.Phi.Module;
 using TorchSharp;
 using TorchSharp.Modules;
@@ -66,6 +67,55 @@ public static Phi3ForCasualLM FromPretrained(
         return phi;
     }
 
+    public static Phi3ForCasualLM FromPretrained(
+        string modelFolder,
+        string configName = "config.json",
+        string checkPointName = "model.safetensors.index.json",
+        bool quantizeToInt8 = false,
+        bool quantizeToInt4 = false,
+        int layersOnTargetDevice = -1,
+        ScalarType torchDtype = ScalarType.BFloat16,
+        string targetDevice = "cuda")
+    {
+        if (layersOnTargetDevice == -1 && quantizeToInt4 == false && quantizeToInt8 == false)
+        {
+            return FromPretrained(modelFolder, configName, checkPointName, torchDtype, targetDevice);
+        }
+
+        var originalDefaultDevice = torch.get_default_device();
+        torch.set_default_device("meta");
+        var config = Path.Join(modelFolder, configName);
+        var modelConfig = JsonSerializer.Deserialize<Phi3Config>(File.ReadAllText(config)) ?? throw new ArgumentNullException(nameof(config));
+        modelConfig.DType = torchDtype;
+        var model = new Phi3ForCasualLM(modelConfig);
+
+        if (quantizeToInt8)
+        {
+            model.ToInt8QuantizeModule();
+        }
+        else if (quantizeToInt4)
+        {
+            model.ToInt4QuantizeModule();
+        }
+
+        var deviceMap = model.InferDeviceMapForEachLayer(
+            [
+                KeyValuePair.Create(targetDevice, layersOnTargetDevice),
+                KeyValuePair.Create("cpu", -1)
+            ]);
+
+        torch.set_default_device("cpu");
+        model = new Phi3ForCasualLM(modelConfig);
+
+        model.LoadSafeTensors(modelFolder, checkPointName);
+
+        model = model.ToDynamicLoadingModel(deviceMap, targetDevice);
+
+        torch.set_default_device(originalDefaultDevice);
+
+        return model;
+    }
+
     public void LoadSafeTensors(string modelFolder, string checkPointName = "model.safetensors.index.json")
     {
         this.load_checkpoint(path: modelFolder, checkpointName: checkPointName, strict: false, useTqdm: false);