LoRA loading and LoRA application

OnnxStack.Console/Examples/LoRADebug.cs

@@ -0,0 +1,42 @@
+using Microsoft.ML.OnnxRuntime;
+using OnnxStack.Core.Services;
+
+namespace OnnxStack.Console.Runner
+{
+    public sealed class LoRADebug : IExampleRunner
+    {
+        private readonly string _outputDirectory;
+        private readonly IOnnxModelService _modelService;
+        private readonly IOnnxModelAdaptaterService _modelAdaptaterService;
+
+        public LoRADebug(IOnnxModelService modelService)
+        {
+            _modelService = modelService;
+            _outputDirectory = Path.Combine(Directory.GetCurrentDirectory(), "Examples", nameof(StableDebug));
+        }
+
+        public string Name => "LoRA Debug";
+
+        public string Description => "LoRA Debugger";
+
+        public async Task RunAsync()
+        {
+            string modelPath = "D:\\Repositories\\stable-diffusion-v1-5\\unet\\model.onnx";
+            string loraModelPath = "D:\\Repositories\\LoRAFiles\\model.onnx";
+
+            using (var modelession = new InferenceSession(modelPath))
+            using (var loraModelSession = new InferenceSession(loraModelPath))
+            {
+                try
+                {
+                    _modelAdaptaterService.ApplyLowRankAdaptation(modelession, loraModelSession);
+                }
+                catch (Exception ex)
+                {
+
+                }
+            }
+        }
+
+    }
+}

OnnxStack.Core/Model/OnnxModelAdapter.cs

@@ -0,0 +1,14 @@
+using OnnxStack.Core.Config;
+using System;
+using System.Collections.Generic;
+using System.Linq;
+using System.Text;
+using System.Threading.Tasks;
+
+namespace OnnxStack.Core.Model
+{
+    public class OnnxModelAdapter : IOnnxModel
+    {
+        public string Name { get; set; }
+    }
+}

OnnxStack.Core/Registration.cs

@@ -18,6 +18,7 @@ public static void AddOnnxStack(this IServiceCollection serviceCollection)
         {
             serviceCollection.AddSingleton(ConfigManager.LoadConfiguration());
             serviceCollection.AddSingleton<IOnnxModelService, OnnxModelService>();
+            serviceCollection.AddSingleton<IOnnxModelAdaptaterService, OnnxModelAdaptaterService>();
         }
 
 

OnnxStack.Core/Services/IOnnxModelAdaptaterService.cs

@@ -0,0 +1,9 @@
+using Microsoft.ML.OnnxRuntime;
+
+namespace OnnxStack.Core.Services
+{
+    public interface IOnnxModelAdaptaterService
+    {
+        void ApplyLowRankAdaptation(InferenceSession primarySession, InferenceSession loraSession);
+    }
+}

OnnxStack.Core/Services/OnnxModelAdaptaterService.cs

@@ -0,0 +1,50 @@
+using Microsoft.ML.OnnxRuntime;
+using Microsoft.ML.OnnxRuntime.Tensors;
+using System;
+using System.Linq;
+
+namespace OnnxStack.Core.Services
+{
+    public class OnnxModelAdaptaterService : IOnnxModelAdaptaterService
+    {
+        public void ApplyLowRankAdaptation(InferenceSession primarySession, InferenceSession loraSession)
+        {
+            // For simplicity, let's assume we will replace the weights of the first dense layer
+            string layerName = "layer_name";
+
+            // Get the current weights from the primary model
+            var primaryInputName = primarySession.InputMetadata.Keys.First();
+            var primaryInputTensor = primarySession.InputMetadata[primaryInputName];
+            var primaryWeights = new float[primaryInputTensor.Dimensions.Product()];
+
+            // Get the weights from the LoRA model
+            var lraInputName = loraSession.InputMetadata.Keys.First();
+            var lraInputTensor = loraSession.InputMetadata[lraInputName];
+            var lraWeights = new float[lraInputTensor.Dimensions.Product()];
+
+            // Apply LoRA (replace weights) this is where we will do the mutiplication of the weights
+            // but for testing sake just brute for replacing
+            Array.Copy(lraWeights, primaryWeights, Math.Min(primaryWeights.Length, lraWeights.Length));
+
+            // Update the primary model tensor with the modified weights
+            var tensor = new DenseTensor<float>(primaryWeights, primaryInputTensor.Dimensions.ToArray());
+            var inputs = new NamedOnnxValue[] { NamedOnnxValue.CreateFromTensor(primaryInputName, tensor) };
+
+            // Will it run?
+            primarySession.Run(inputs);
+        }
+    }
+
+    public static class Ext
+    {
+        public static int Product(this int[] array)
+        {
+            int result = 1;
+            foreach (int element in array)
+            {
+                result *= element;
+            }
+            return result;
+        }
+    }
+}