Updated to llama.cpp ff4affb4c1aa7eb4f28a0d9de1b205bd719802f2

Author: martindevans

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/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/LLamaSharp.csproj

@@ -57,7 +57,7 @@
   </ItemGroup>
 
   <PropertyGroup>
-    <BinaryReleaseId>506bb6e01009058f355</BinaryReleaseId>
+    <BinaryReleaseId>ff4affb4c1aa7eb4f28a0_v2</BinaryReleaseId>
   </PropertyGroup>
 
   <PropertyGroup>

LLama/Native/LLamaModelParams.cs

@@ -82,7 +82,7 @@ public bool use_mmap
         private sbyte _use_mmap;
 
         /// <summary>
-        /// use direct io, takes precedence over use_mmap
+        /// use direct io, takes precedence over use_mmap when supported
         /// </summary>
         public bool use_direct_io
         {

LLama/Native/LLamaParamsFitStatus.cs

@@ -0,0 +1,19 @@
+namespace LLama.Native;
+
+public enum LLamaParamsFitStatus
+{
+    /// <summary>
+    /// Found allocations that are projected to fit
+    /// </summary>
+    LLAMA_PARAMS_FIT_STATUS_SUCCESS = 0,
+
+    /// <summary>
+    /// Could not find allocations that are projected to fit
+    /// </summary>
+    LLAMA_PARAMS_FIT_STATUS_FAILURE = 1,
+
+    /// <summary>
+    /// A hard error occurred, e.g. because no model could be found at the specified path
+    /// </summary>
+    LLAMA_PARAMS_FIT_STATUS_ERROR = 2,
+}

LLama/Native/LoraAdapter.cs

@@ -45,6 +45,7 @@ public void Unload()
 
         // Manually free a LoRA adapter. loaded adapters will be free when the associated model is deleted
         [DllImport(NativeApi.libraryName, CallingConvention = CallingConvention.Cdecl)]
+        [Obsolete("adapters are now freed together with the associated model")]
         static extern void llama_adapter_lora_free(IntPtr adapter);
     }
 }