feat: upgrade the native libraries.

LLama.Examples/NewVersion/InstructModeExecute.cs

@@ -22,7 +22,7 @@ namespace LLama.Examples.NewVersion
                 "make friend with human, no less than 200 words.\"");
             Console.ForegroundColor = ConsoleColor.White;
 
-            var inferenceParams = new InferenceParams() { Temperature = 0.8f, MaxTokens = 300 };
+            var inferenceParams = new InferenceParams() { Temperature = 0.8f, MaxTokens = 600 };
 
             while (true)
             {

LLama.Examples/NewVersion/InteractiveModeExecute.cs

@@ -18,12 +18,12 @@ namespace LLama.Examples.NewVersion
             InteractiveExecutor ex = new(new LLamaModel(new ModelParams(modelPath, contextSize: 256)));
 
             Console.ForegroundColor = ConsoleColor.Yellow;
-            Console.WriteLine("The executor has been enabled. In this example, the prompt is printed, the maximum tokens is set to 64 and the context size is 256. (an example for small scale usage)");
+            Console.WriteLine("The executor has been enabled. In this example, the prompt is printed, the maximum tokens is set to 128 and the context size is 256. (an example for small scale usage)");
             Console.ForegroundColor = ConsoleColor.White;
 
             Console.Write(prompt);
 
-            var inferenceParams = new InferenceParams() { Temperature = 0.6f, AntiPrompts = new List<string> { "User:" }, MaxTokens = 64 };
+            var inferenceParams = new InferenceParams() { Temperature = 0.6f, AntiPrompts = new List<string> { "User:" }, MaxTokens = 128 };
 
             while (true)
             {

LLama.Examples/NewVersion/LoadAndSaveSession.cs

@@ -47,7 +47,7 @@ namespace LLama.Examples.NewVersion
 
                     ex.Model.Dispose();
                     ex = new(new LLamaModel(new ModelParams(modelPath, contextSize: 1024, seed: 1337, gpuLayerCount: 5)));
-                    session = new ChatSession(ex).WithOutputTransform(new LLamaTransforms.KeywordTextOutputStreamTransform(new string[] { "User:", "Bob:" }, redundancyLength: 8));
+                    session = new ChatSession(ex);
                     session.LoadSession(statePath);
 
                     Console.ForegroundColor = ConsoleColor.Yellow;

LLama/Common/InferenceParams.cs

@@ -5,6 +5,9 @@ using System.Text;
 namespace LLama.Common
 {
     using llama_token = Int32;
+    /// <summary>
+    /// The paramters used for inference.
+    /// </summary>
     public class InferenceParams
     {
         /// <summary>

LLama/Common/ModelParams.cs

@@ -4,13 +4,23 @@ using System.Text;
 
 namespace LLama.Common
 {
+    /// <summary>
+    /// The parameters for initializing a LLama model.
+    /// </summary>
     public class ModelParams
     {
         /// <summary>
         /// Model context size (n_ctx)
         /// </summary>
         public int ContextSize { get; set; } = 512;
-
+        /// <summary>
+        /// the GPU that is used for scratch and small tensors
+        /// </summary>
+        public int MainGpu { get; set; } = 0;
+        /// <summary>
+        /// if true, reduce VRAM usage at the cost of performance
+        /// </summary>
+        public bool LowVram { get; set; } = false;
         /// <summary>
         /// Number of layers to run in VRAM / GPU memory (n_gpu_layers)
         /// </summary>
@@ -40,6 +50,10 @@ namespace LLama.Common
         /// </summary>
         public string ModelPath { get; set; }
         /// <summary>
+        /// model alias
+        /// </summary>
+        public string ModelAlias { get; set; } = "unknown";
+        /// <summary>
         /// lora adapter path (lora_adapter)
         /// </summary>
         public string LoraAdapter { get; set; } = string.Empty;
@@ -67,6 +81,11 @@ namespace LLama.Common
         /// </summary>
         public bool EmbeddingMode { get; set; } = false;
 
+        /// <summary>
+        /// how split tensors should be distributed across GPUs
+        /// </summary>
+        public float[] TensorSplits { get; set; } = new float[] { 0 };
+
         /// <summary>
         /// 
         /// </summary>

LLama/LLamaInstructExecutor.cs

@@ -17,6 +17,7 @@ namespace LLama
     public class InstructExecutor : StatefulExecutorBase
     {
         bool _is_prompt_run = true;
+        string _instructionPrefix;
         llama_token[] _inp_pfx;
         llama_token[] _inp_sfx;
         /// <summary>
@@ -30,6 +31,7 @@ namespace LLama
         {
             _inp_pfx = _model.Tokenize(instructionPrefix, true).ToArray();
             _inp_sfx = _model.Tokenize(instructionSuffix, false).ToArray();
+            _instructionPrefix = instructionPrefix;
         }
 
         /// <inheritdoc />
@@ -104,6 +106,11 @@ namespace LLama
         /// <inheritdoc />
         protected override void PreprocessInputs(string text, InferStateArgs args)
         {
+            if(args.Antiprompts is null)
+            {
+                args.Antiprompts = new List<string>();
+            }
+            args.Antiprompts.Add(_instructionPrefix);
             if (_is_prompt_run)
             {
                 // When running the first input (prompt) in inteactive mode, we should specially process it.

LLama/LLamaSharp.csproj

@@ -8,7 +8,7 @@
     <Platforms>AnyCPU;x64;Arm64</Platforms>
     <AllowUnsafeBlocks>True</AllowUnsafeBlocks>
 
-    <Version>0.4.0</Version>
+    <Version>0.4.1</Version>
     <Authors>Yaohui Liu, Haiping Chen</Authors>
     <Company>SciSharp STACK</Company>
     <GeneratePackageOnBuild>true</GeneratePackageOnBuild>
@@ -21,7 +21,7 @@
       The .NET binding of LLama.cpp, providing APIs to run the model and deploy it on Web. For model weights to run, please go to https://github.com/SciSharp/LLamaSharp for more information.
     </Description>
     <PackageReleaseNotes>
-      LLamaSharp 0.4.0 supports better APIs than v0.3.0. Note that many break changes were made in this version. APIs of v0.3.0 were moved to LLama.Old namespace.
+      LLamaSharp 0.4.1 followed up the master branch of llama.cpp. (commit id: aacdbd4)
     </PackageReleaseNotes>
     <PackageLicenseExpression>MIT</PackageLicenseExpression>
     <PackageOutputPath>packages</PackageOutputPath>

LLama/Native/NativeApi.cs

@@ -44,6 +44,9 @@ namespace LLama.Native
         [DllImport(libraryName, CallingConvention = CallingConvention.Cdecl)]
         public static extern bool llama_mlock_supported();
 
+        [DllImport(libraryName, CallingConvention = CallingConvention.Cdecl)]
+        public static extern int llama_eval_export(SafeLLamaContextHandle ctx, string fname);
+
         /// <summary>
         /// Various functions for loading a ggml llama model.
         /// Allocate (almost) all memory needed for the model.

LLama/Utils.cs

@@ -18,6 +18,8 @@ namespace LLama
             var lparams = NativeApi.llama_context_default_params();
 
             lparams.n_ctx = @params.ContextSize;
+            lparams.n_batch = @params.BatchSize;
+            lparams.main_gpu = @params.MainGpu;
             lparams.n_gpu_layers = @params.GpuLayerCount;
             lparams.seed = @params.Seed;
             lparams.f16_kv = @params.UseFp16Memory;
@@ -25,6 +27,17 @@ namespace LLama
             lparams.use_mlock = @params.UseMemoryLock;
             lparams.logits_all = @params.Perplexity;
             lparams.embedding = @params.EmbeddingMode;
+            lparams.low_vram = @params.LowVram;
+
+            if(@params.TensorSplits.Length != 1)
+            {
+                throw new ArgumentException("Currently multi-gpu support is not supported by " +
+                    "both llama.cpp and LLamaSharp.");
+            }
+            lparams.tensor_split = new TensorSplits()
+            {
+                Item1 = @params.TensorSplits[0]
+            };
 
             if (!File.Exists(@params.ModelPath))
             {