Updated to 11dd5a44eb180e1d69fac24d3852b5222d66fb7f

Author: martindevans

LLama.Web/Common/ModelOptions.cs

@@ -116,6 +116,9 @@ public class ModelOptions
         /// <inheritdoc />
         public bool? SwaFull { get; set; }
 
+        /// <inheritdoc />
+        public bool? KVUnified { get; set; }
+
         /// <inheritdoc />
         public float? DefragThreshold { get; set; }
 

LLama/Abstractions/IContextParams.cs

@@ -128,6 +128,14 @@ public interface IContextParams
     /// </summary>
     bool? OpOffload { get; }
 
+    /// <summary>
+    /// use a unified buffer across the input sequences when computing the attention.
+    /// try to disable when n_seq_max > 1 for improved performance when the sequences do not share a large prefix
+    /// <br />
+    /// ref: <a href="https://github.com/ggml-org/llama.cpp/pull/14363">https://github.com/ggml-org/llama.cpp/pull/14363</a>
+    /// </summary>
+    bool? KVUnified { get; }
+
     /// <summary>
     /// Use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
     /// </summary>

LLama/Common/ModelParams.cs

@@ -118,6 +118,9 @@ public record ModelParams
         /// <inheritdoc />
         public bool? SwaFull { get; set; }
 
+        /// <inheritdoc />
+        public bool? KVUnified { get; set; }
+
         /// <summary>
         /// `Encoding` cannot be directly JSON serialized, instead store the name as a string which can
         /// </summary>

LLama/LLamaSharp.csproj

@@ -57,7 +57,7 @@
   </ItemGroup>
 
   <PropertyGroup>
-    <BinaryReleaseId>a0374a67e2924f2e</BinaryReleaseId>
+    <BinaryReleaseId>11dd5a44eb180e</BinaryReleaseId>
   </PropertyGroup>
 
   <PropertyGroup>