Release/v2.0.4

.github/doxygen/Doxyfile

@@ -48,7 +48,7 @@ PROJECT_NAME           = "LlamaLib"
 # could be handy for archiving the generated documentation or if some version
 # control system is used.
 
-PROJECT_NUMBER = v2.0.3
+PROJECT_NUMBER = v2.0.4
 
 # Using the PROJECT_BRIEF tag one can provide an optional one line description
 # for a project that appears at the top of each page and should give viewer a

.github/tests/csharp-dotnet/Program.csproj

@@ -8,7 +8,7 @@
   </PropertyGroup>
 
   <ItemGroup>
-    <PackageReference Include="LlamaLib" Version="2.0.3" />
+    <PackageReference Include="LlamaLib" Version="2.0.4" />
   </ItemGroup>
 
 </Project>

.github/workflows/build_library.yaml

@@ -748,6 +748,7 @@ jobs:
             ./.github/scripts/release.sh release
             mv release ${{ env.RELEASE_NAME }}
             zip -r ${{ env.RELEASE_NAME }}.zip ${{ env.RELEASE_NAME }}
+            sha256sum ${{ env.RELEASE_NAME }}.zip | awk '{print $1}' > ${{ env.RELEASE_NAME }}.zip.sha256
 
         - id: test_cmake
           name: Test cmake
@@ -785,6 +786,7 @@ jobs:
           with:
             files: |
               ${{ env.RELEASE_NAME }}.zip
+              ${{ env.RELEASE_NAME }}.zip.sha256
               csharp/bin/Release/LlamaLib*nupkg
             name: Release ${{ env.VERSION }}
 

.github/workflows_template/build_library_steps.yaml

@@ -287,6 +287,7 @@ jobs:
           ./.github/scripts/release.sh release
           mv release ${{ env.RELEASE_NAME }}
           zip -r ${{ env.RELEASE_NAME }}.zip ${{ env.RELEASE_NAME }}
+          sha256sum ${{ env.RELEASE_NAME }}.zip | awk '{print $1}' > ${{ env.RELEASE_NAME }}.zip.sha256
 
       - name: Test cmake
         id: test_cmake
@@ -325,4 +326,5 @@ jobs:
           name: "Release ${{ env.VERSION }}"
           files: | 
             ${{ env.RELEASE_NAME }}.zip
+            ${{ env.RELEASE_NAME }}.zip.sha256
             csharp/bin/Release/LlamaLib*nupkg

README.md

@@ -14,6 +14,7 @@
 [![LinkedIn](https://img.shields.io/badge/LinkedIn-blue?style=flat&logo=linkedin&labelColor=blue)](https://www.linkedin.com/company/undreamai)
 [![GitHub Repo stars](https://img.shields.io/github/stars/undreamai/LlamaLib?style=flat&logo=github&color=f5f5f5)](https://github.com/undreamai/LlamaLib)
 [![Documentation](https://img.shields.io/badge/Docs-white.svg?logo=data:image/png;base64,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)](https://undream.ai/LlamaLib)
+[![llama.cpp](https://img.shields.io/badge/llama.cpp-b7777-blue?style=flat)](https://github.com/ggerganov/llama.cpp)
 
 LlamaLib is a **high-level C++ and C#** library for running Large Language Models (LLMs) **anywhere** - from PCs to mobile devices and VR headsets.<br>
 It is built on top of the awesome [llama.cpp](https://github.com/ggerganov/llama.cpp) library.

VERSION

@@ -1 +1 @@
-2.0.3
+2.0.4

cmake/LlamaLibConfigVersion.cmake

@@ -7,7 +7,7 @@
 # PACKAGE_VERSION_COMPATIBLE if the current version is >= requested version.
 # The variable CVF_VERSION must be set before calling configure_file().
 
-set(PACKAGE_VERSION "2.0.3")
+set(PACKAGE_VERSION "2.0.4")
 
 if (PACKAGE_FIND_VERSION_RANGE)
   # Package version must be in the requested version range

cpp_api.md

@@ -767,6 +767,47 @@ agent.remove_last_message();
 agent.remove_last_message();
 ```
 
+#### Context Overflow Management
+
+When conversation history grows large enough to exceed the model's context window, the agent can automatically handle it using a configurable strategy.
+
+```cpp
+void set_overflow_strategy(
+    ContextOverflowStrategy strategy,                  // None, Truncate, or Summarize
+    float target_ratio = 0.5f,                         // target fill ratio of context after truncation (0.0–1.0)
+    const std::string &summarize_prompt = "..."        // custom summarization prompt (default provided)
+);
+```
+
+**Strategies:**
+
+| Value | Behaviour |
+|-------|-----------|
+| `ContextOverflowStrategy::None` | No protection — may crash if context is exceeded |
+| `ContextOverflowStrategy::Truncate` | Drops oldest message pairs until history fits within `target_ratio` of the context |
+| `ContextOverflowStrategy::Summarize` | Asks the LLM to summarise history, embeds summary in the system message, then truncates if still needed |
+
+**Example:**
+```cpp
+// Truncate: drop oldest pairs when context is full, keeping ≤50% filled
+agent.set_overflow_strategy(ContextOverflowStrategy::Truncate, 0.5f);
+
+// Summarize: condense history into a rolling summary in the system prompt
+agent.set_overflow_strategy(ContextOverflowStrategy::Summarize);
+
+// Summarize with a custom prompt
+agent.set_overflow_strategy(
+    ContextOverflowStrategy::Summarize,
+    0.5f,
+    "Summarize the key facts from this conversation briefly:\n\n"
+);
+
+// The rolling summary is automatically saved with save_history / loaded with load_history.
+// You can also access or replace it directly:
+std::string summary = agent.get_summary();
+agent.set_summary("Alice introduced herself as a software engineer.");
+```
+
 #### System Prompt
 
 ```cpp

csharp/LLMAgent.cs

@@ -60,6 +60,17 @@ public override string ToString()
         }
     }
 
+    /// <summary>Strategy to apply when the chat history would exceed the model's context window.</summary>
+    public enum ContextOverflowStrategy
+    {
+        /// <summary>No automatic handling — may crash if context is exceeded.</summary>
+        None = 0,
+        /// <summary>Remove oldest message pairs from the front until history fits within targetRatio of the context.</summary>
+        Truncate = 1,
+        /// <summary>Summarise history, embed summary in the system message, then truncate if still needed.</summary>
+        Summarize = 2
+    }
+
     // LLMAgent class
     public class LLMAgent : LLMLocal
     {
@@ -214,6 +225,39 @@ public int GetHistorySize()
             return llamaLib.LLMAgent_Get_History_Size(llm);
         }
 
+        // Context overflow management
+
+        /// <summary>
+        /// Configure how the agent handles context overflow.
+        /// </summary>
+        /// <param name="strategy">Overflow strategy to use.</param>
+        /// <param name="targetRatio">Target fill ratio of context after truncation (0.0–1.0, default 0.5).</param>
+        /// <param name="summarizePrompt">Custom prompt for summarization; null uses the built-in default.</param>
+        public void SetOverflowStrategy(ContextOverflowStrategy strategy, float targetRatio = 0.5f, string summarizePrompt = null)
+        {
+            CheckLlamaLib();
+            llamaLib.LLMAgent_Set_Overflow_Strategy(llm, (int)strategy, targetRatio, summarizePrompt);
+        }
+
+        /// <summary>
+        /// Get the current rolling summary generated by the Summarize overflow strategy.
+        /// Returns an empty string if no summary has been generated yet.
+        /// </summary>
+        public string GetSummary()
+        {
+            CheckLlamaLib();
+            return llamaLib.LLMAgent_Get_Summary(llm) ?? string.Empty;
+        }
+
+        /// <summary>
+        /// Set the rolling summary directly, e.g. to restore state after loading from file.
+        /// </summary>
+        public void SetSummary(string summary)
+        {
+            CheckLlamaLib();
+            llamaLib.LLMAgent_Set_Summary(llm, summary ?? string.Empty);
+        }
+
         // Chat functionality
         public string Chat(string userPrompt, bool addToHistory = true, LlamaLib.CharArrayCallback callback = null, bool returnResponseJson = false, bool debugPrompt = false)
         {

csharp/LlamaLib.cs

@@ -185,6 +185,15 @@ public delegate IntPtr LLMAgent_Chat_Delegate(IntPtr llm,
         [UnmanagedFunctionPointer(CallingConvention.Cdecl)]
         public delegate int LLMAgent_Get_History_Size_Delegate(IntPtr llm);
 
+        [UnmanagedFunctionPointer(CallingConvention.Cdecl)]
+        public delegate void LLMAgent_Set_Overflow_Strategy_Delegate(IntPtr llm, int strategy, float targetRatio, [MarshalAs(UnmanagedType.LPStr)] string summarizePrompt);
+
+        [UnmanagedFunctionPointer(CallingConvention.Cdecl)]
+        public delegate IntPtr LLMAgent_Get_Summary_Delegate(IntPtr llm);
+
+        [UnmanagedFunctionPointer(CallingConvention.Cdecl)]
+        public delegate void LLMAgent_Set_Summary_Delegate(IntPtr llm, [MarshalAs(UnmanagedType.LPStr)] string summary);
+
         //################################################## FUNCTION POINTERS ##################################################//
 
         // Main lib
@@ -235,13 +244,16 @@ public delegate IntPtr LLMAgent_Chat_Delegate(IntPtr llm,
         public LLMAgent_Save_History_Delegate LLMAgent_Save_History_Internal;
         public LLMAgent_Load_History_Delegate LLMAgent_Load_History_Internal;
         public LLMAgent_Get_History_Size_Delegate LLMAgent_Get_History_Size_Internal;
+        public LLMAgent_Set_Overflow_Strategy_Delegate LLMAgent_Set_Overflow_Strategy_Internal;
+        public LLMAgent_Get_Summary_Delegate LLMAgent_Get_Summary_Internal;
+        public LLMAgent_Set_Summary_Delegate LLMAgent_Set_Summary_Internal;
 
         //################################################## STATUS CHECKING WRAPPER ##################################################//
 
         public void CheckStatus(bool crashesOnly = false)
         {
             int status = LLM_Status_Code_Internal();
-            if (status > 0 || (status < 0 && !crashesOnly))
+            if (status < 0 || (status > 0 && !crashesOnly))
             {
                 string msg = Marshal.PtrToStringAnsi(LLM_Status_Message_Internal()) ?? "";
                 throw new InvalidOperationException($"LlamaLib error {status}: {msg}");
@@ -328,6 +340,9 @@ public IntPtr LLMAgent_Chat(IntPtr llm, string userPrompt, bool addToHistory = t
         public void LLMAgent_Save_History(IntPtr llm, string filepath) => CallWithStatus(() => LLMAgent_Save_History_Internal(llm, filepath));
         public void LLMAgent_Load_History(IntPtr llm, string filepath) => CallWithStatus(() => LLMAgent_Load_History_Internal(llm, filepath));
         public int LLMAgent_Get_History_Size(IntPtr llm) => CallWithStatus(() => LLMAgent_Get_History_Size_Internal(llm));
+        public void LLMAgent_Set_Overflow_Strategy(IntPtr llm, int strategy, float targetRatio, string summarizePrompt = null) => CallWithStatus(() => LLMAgent_Set_Overflow_Strategy_Internal(llm, strategy, targetRatio, summarizePrompt));
+        public string LLMAgent_Get_Summary(IntPtr llm) => Marshal.PtrToStringAnsi(CallWithStatus(() => LLMAgent_Get_Summary_Internal(llm)));
+        public void LLMAgent_Set_Summary(IntPtr llm, string summary) => CallWithStatus(() => LLMAgent_Set_Summary_Internal(llm, summary));
 
         //################################################## MOBILE IMPLEMENTATION ##################################################//
 
@@ -537,6 +552,15 @@ public static extern IntPtr LLMAgent_Chat_Static(IntPtr llm,
         [DllImport(DllName, CallingConvention = CallingConvention.Cdecl, EntryPoint = "LLMAgent_Get_History_Size")]
         public static extern int LLMAgent_Get_History_Size_Static(IntPtr llm);
 
+        [DllImport(DllName, CallingConvention = CallingConvention.Cdecl, EntryPoint = "LLMAgent_Set_Overflow_Strategy")]
+        public static extern void LLMAgent_Set_Overflow_Strategy_Static(IntPtr llm, int strategy, float targetRatio, [MarshalAs(UnmanagedType.LPStr)] string summarizePrompt);
+
+        [DllImport(DllName, CallingConvention = CallingConvention.Cdecl, EntryPoint = "LLMAgent_Get_Summary")]
+        public static extern IntPtr LLMAgent_Get_Summary_Static(IntPtr llm);
+
+        [DllImport(DllName, CallingConvention = CallingConvention.Cdecl, EntryPoint = "LLMAgent_Set_Summary")]
+        public static extern void LLMAgent_Set_Summary_Static(IntPtr llm, [MarshalAs(UnmanagedType.LPStr)] string summary);
+
         public static IntPtr Available_Architectures([MarshalAs(UnmanagedType.I1)] bool gpu) { return IntPtr.Zero; }
         public static bool Has_GPU_Layers([MarshalAs(UnmanagedType.LPStr)] string command) { return false; }
 
@@ -603,6 +627,9 @@ public LlamaLib(bool gpu = false)
             LLMAgent_Save_History_Internal = (llm, filepath) => LLMAgent_Save_History_Static(llm, filepath);
             LLMAgent_Load_History_Internal = (llm, filepath) => LLMAgent_Load_History_Static(llm, filepath);
             LLMAgent_Get_History_Size_Internal = (llm) => LLMAgent_Get_History_Size_Static(llm);
+            LLMAgent_Set_Overflow_Strategy_Internal = (llm, strategy, targetRatio, summarizePrompt) => LLMAgent_Set_Overflow_Strategy_Static(llm, strategy, targetRatio, summarizePrompt);
+            LLMAgent_Get_Summary_Internal = (llm) => LLMAgent_Get_Summary_Static(llm);
+            LLMAgent_Set_Summary_Internal = (llm, summary) => LLMAgent_Set_Summary_Static(llm, summary);
         }
 
         public void Dispose() {}
@@ -618,7 +645,7 @@ public void Dispose() {}
         private List<IntPtr> dependencyHandles = new List<IntPtr>();
         private static int debugLevelGlobal = 0;
         private static CharArrayCallback loggingCallbackGlobal = null;
-        private string[] availableLibraries = null;
+        private List<Tuple<string, bool>> availableLibraries = null;
         private int currentLibraryIndex = 0;
 
         // Runtime lib
@@ -724,10 +751,7 @@ private string GetRuntimeLibraryPath()
         private string[] GetAvailableArchitectures(bool gpu)
         {
             string architecturesString = Marshal.PtrToStringAnsi(Available_Architectures(gpu));
-            if (string.IsNullOrEmpty(architecturesString))
-            {
-                throw new InvalidOperationException("No architectures available for the specified GPU setting.");
-            }
+            if (string.IsNullOrEmpty(architecturesString)) return new string[0];
 
             string[] librariesOptions = architecturesString.Split(',');
             List<string> libraries = new List<string>();
@@ -751,12 +775,20 @@ private string[] GetAvailableArchitectures(bool gpu)
 
         private void LoadLibraries(bool gpu)
         {
-            availableLibraries = GetAvailableArchitectures(gpu);
+            availableLibraries = new List<Tuple<string, bool>>();
+            bool[] arch_options = gpu ? new bool[] { true, false }: new bool[] { false };
+            foreach (bool arch_gpu in arch_options)
+            {
+                string[] archs = GetAvailableArchitectures(arch_gpu);
+                foreach (string arch in archs) availableLibraries.Add(new Tuple<string, bool>(arch, arch_gpu));
+            }
             currentLibraryIndex = -1;
 
             if (!TryNextLibrary())
             {
-                throw new InvalidOperationException($"Failed to load any library. Available libraries: {string.Join(", ", availableLibraries)}");
+                string libs = "";
+                foreach (Tuple<string, bool> arch in availableLibraries) libs += arch.Item1 + ", ";
+                throw new InvalidOperationException($"Failed to load any library. Available libraries: {libs.TrimEnd(',', ' ')}");
             }
         }
 
@@ -797,9 +829,9 @@ public bool TryNextLibrary()
                 libraryHandle = IntPtr.Zero;
             }
 
-            while (++currentLibraryIndex < availableLibraries.Length)
+            while (++currentLibraryIndex < availableLibraries.Count)
             {
-                string library = availableLibraries[currentLibraryIndex];
+                var (library, is_gpu_library) = availableLibraries[currentLibraryIndex];
                 try
                 {
                     string libraryPath = FindLibrary(library.Trim());
@@ -812,6 +844,8 @@ public bool TryNextLibrary()
                     libraryHandle = LibraryLoader.LoadLibrary(libraryPath);
 
                     LoadFunctionPointers();
+                    if (is_gpu_library && !LLMService_Supports_GPU()) continue;
+
                     architecture = library.Trim();
                     if (debugLevelGlobal > 0) Console.WriteLine("Successfully loaded: " + libraryPath);
                     return true;
@@ -858,6 +892,7 @@ private void LoadFunctionPointers()
             LLM_Debug = LibraryLoader.GetSymbolDelegate<LLM_Debug_Delegate>(libraryHandle, "LLM_Debug");
             LLM_Logging_Callback = LibraryLoader.GetSymbolDelegate<LLM_Logging_Callback_Delegate>(libraryHandle, "LLM_Logging_Callback");
             LLM_Logging_Stop = LibraryLoader.GetSymbolDelegate<LLM_Logging_Stop_Delegate>(libraryHandle, "LLM_Logging_Stop");
+            LLMService_Supports_GPU = LibraryLoader.GetSymbolDelegate<LLMService_Supports_GPU_Delegate>(libraryHandle, "LLMService_Supports_GPU");
 
             LLM_Enable_Reasoning_Internal = LibraryLoader.GetSymbolDelegate<LLM_Enable_Reasoning_Delegate>(libraryHandle, "LLM_Enable_Reasoning");
             LLM_Apply_Template_Internal = LibraryLoader.GetSymbolDelegate<LLM_Apply_Template_Delegate>(libraryHandle, "LLM_Apply_Template");
@@ -908,6 +943,9 @@ private void LoadFunctionPointers()
             LLMAgent_Save_History_Internal = LibraryLoader.GetSymbolDelegate<LLMAgent_Save_History_Delegate>(libraryHandle, "LLMAgent_Save_History");
             LLMAgent_Load_History_Internal = LibraryLoader.GetSymbolDelegate<LLMAgent_Load_History_Delegate>(libraryHandle, "LLMAgent_Load_History");
             LLMAgent_Get_History_Size_Internal = LibraryLoader.GetSymbolDelegate<LLMAgent_Get_History_Size_Delegate>(libraryHandle, "LLMAgent_Get_History_Size");
+            LLMAgent_Set_Overflow_Strategy_Internal = LibraryLoader.GetSymbolDelegate<LLMAgent_Set_Overflow_Strategy_Delegate>(libraryHandle, "LLMAgent_Set_Overflow_Strategy");
+            LLMAgent_Get_Summary_Internal = LibraryLoader.GetSymbolDelegate<LLMAgent_Get_Summary_Delegate>(libraryHandle, "LLMAgent_Get_Summary");
+            LLMAgent_Set_Summary_Internal = LibraryLoader.GetSymbolDelegate<LLMAgent_Set_Summary_Delegate>(libraryHandle, "LLMAgent_Set_Summary");
         }
 
         // Static functions
@@ -920,9 +958,13 @@ private void LoadFunctionPointers()
         [UnmanagedFunctionPointer(CallingConvention.Cdecl)]
         public delegate void LLM_Logging_Stop_Delegate();
 
+        [UnmanagedFunctionPointer(CallingConvention.Cdecl)]
+        public delegate bool LLMService_Supports_GPU_Delegate();
+
         public LLM_Debug_Delegate LLM_Debug;
         public LLM_Logging_Callback_Delegate LLM_Logging_Callback;
         public LLM_Logging_Stop_Delegate LLM_Logging_Stop;
+        public LLMService_Supports_GPU_Delegate LLMService_Supports_GPU;
 
         public static void Debug(int debugLevel)
         {