Merge pull request SciSharp#1335 from martindevans/update_llama_cpp_ff4affb4c1aa7eb4f28a0d9de1b205bd719802f2

Update llama.cpp February 2026

Author: martindevans

LLama.Benchmark/Constants.cs

@@ -14,10 +14,6 @@ public static string ModelDir
         public static string Generative7BModelPath =>  Path.Combine(ModelDir, "llama-2-7b-chat.Q3_K_S.gguf");
         public static string EmbeddingModelPath => Path.Combine(ModelDir, "all-MiniLM-L12-v2.Q8_0.gguf");
 
-        public static string LLavaModelPath => Path.Combine("llava-v1.6-mistral-7b.Q3_K_XS.gguf");
-        public static string LLavaMmpPath => Path.Combine("mmproj-model-f16.gguf");
-        public static string LLavaImage => "Assets/extreme-ironing-taxi-610x427.jpg";
-
         public static string TextCompletionPromptsFilePath => "Assets/TextCompletionPrompts.txt";
     }
 }

LLama.Unittest/MtmdWeightsTests.cs

@@ -1,8 +1,5 @@
-using System;
-using System.IO;
 using LLama.Common;
 using LLama.Native;
-using Xunit;
 
 namespace LLama.Unittest
 {

LLama.Unittest/NativeAbiTests.cs

@@ -1,8 +1,5 @@
-using System;
-using System.Collections.Generic;
 using System.Runtime.InteropServices;
 using LLama.Native;
-using Xunit;
 
 namespace LLama.Unittest
 {
@@ -79,27 +76,6 @@ public void ContextParamsSizeMatchesNative()
             Assert.Equal(expectedSize, Marshal.SizeOf<LLamaContextParams>());
         }
 
-        [Fact]
-        public void MtmdContextParamsSizeMatchesNative()
-        {
-            var pointerSize = IntPtr.Size;
-            var fields = new List<(int size, int align)>
-            {
-                (sizeof(sbyte), 1), // use_gpu
-                (sizeof(sbyte), 1), // print_timings
-                (sizeof(int), 4), // n_threads
-                (pointerSize, pointerSize), // image_marker
-                (pointerSize, pointerSize), // media_marker
-                (sizeof(int), 4), // flash_attn_type
-                (sizeof(sbyte), 1), // warmup
-                (sizeof(int), 4), // image_min_tokens
-                (sizeof(int), 4), // image_max_tokens
-            };
-
-            var expectedSize = ComputeSize(fields);
-            Assert.Equal(expectedSize, Marshal.SizeOf<NativeApi.mtmd_context_params>());
-        }
-
         [Fact]
         public void ModelParamsBoolBlockMatchesNative()
         {

LLama.Web/Common/ModelOptions.cs

@@ -41,6 +41,9 @@ public class ModelOptions
         /// <inheritdoc />
         public bool UseMemorymap { get; set; } = true;
 
+        /// <inheritdoc />
+        public bool UseDirectIO { get; }
+
         /// <inheritdoc />
         public bool UseMemoryLock { get; set; } = false;
 

LLama/Abstractions/ILLamaExecutor.cs

@@ -1,4 +1,4 @@
-using System.Collections.Generic;
+using System.Collections.Generic;
 using System.Threading;
 using LLama.Native;
 
@@ -14,8 +14,8 @@ public interface ILLamaExecutor
         /// </summary>
         public LLamaContext Context { get; }
 
-        // LLava Section
-        //
+        // Multimodal Section
+
         /// <summary>
         /// Identify if it's a multi-modal model and there is a image to process.
         /// </summary>

LLama/Abstractions/IModelParams.cs

@@ -54,6 +54,11 @@ public interface IModelParams
         /// </summary>
         bool UseMemorymap { get; }
 
+        /// <summary>
+        /// Use direct io, takes precedence over use_mmap when supported
+        /// </summary>
+        bool UseDirectIO { get; }
+
         /// <summary>
         /// Use mlock to keep model in memory (use_mlock)
         /// </summary>

LLama/Common/ModelParams.cs

@@ -34,6 +34,9 @@ public record ModelParams
         /// <inheritdoc />
         public bool UseMemorymap { get; set; } = true;
 
+        /// <inheritdoc />
+        public bool UseDirectIO { get; set; }
+
         /// <inheritdoc />
         public bool UseMemoryLock { get; set; }
 

LLama/Extensions/IModelParamsExtensions.cs

@@ -38,6 +38,7 @@ public static IDisposable ToLlamaModelParams(this IModelParams @params, out LLam
 
         result.use_mlock = @params.UseMemoryLock;
         result.use_mmap = @params.UseMemorymap;
+        result.use_direct_io = @params.UseDirectIO;
         result.vocab_only = @params.VocabOnly;
         result.check_tensors = @params.CheckTensors;
 

LLama/LLamaExecutorBase.cs

@@ -69,8 +69,8 @@ public abstract class StatefulExecutorBase : ILLamaExecutor
         /// </summary>
         protected AntipromptProcessor AntipromptProcessor { get; }
 
-        // LLava Section 
-        //
+        // Multimodal Section
+
         /// <inheritdoc />
         public bool IsMultiModal
         {

LLama/LLamaSharp.csproj

@@ -57,7 +57,7 @@
   </ItemGroup>
 
   <PropertyGroup>
-    <BinaryReleaseId>506bb6e01009058f355</BinaryReleaseId>
+    <BinaryReleaseId>ff4affb4c1aa7eb4_v3</BinaryReleaseId>
   </PropertyGroup>
 
   <PropertyGroup>