Merge pull request #712 from martindevans/may-2024-binary-update

May 2024 Binary Update (Take 2)

LLama.Examples/Examples/BatchedExecutorGuidance.cs

@@ -79,7 +79,7 @@ await AnsiConsole
                     guidance.Prompt(g);
 
                     // Early exit if we reach the natural end of the guided sentence
-                    if (g == model.Tokens.EOS)
+                    if (model.Tokens.IsEndOfGeneration(g))
                         break;
 
                     // Update progress bar

LLama.KernelMemory/LLamaSharp.KernelMemory.csproj

@@ -4,7 +4,7 @@
     <TargetFrameworks>net6.0;net7.0;net8.0</TargetFrameworks>
     <ImplicitUsings>enable</ImplicitUsings>
     <Nullable>enable</Nullable>
-    <Version>0.11.2</Version>
+    <Version>0.12.0</Version>
     <Authors>Xbotter</Authors>
     <Company>SciSharp STACK</Company>
     <GeneratePackageOnBuild>true</GeneratePackageOnBuild>
@@ -17,7 +17,7 @@
       The integration of LLamaSharp and Microsoft kernel-memory. It could make it easy to support document search for LLamaSharp model inference.
     </Description>
     <PackageReleaseNotes>
-      v0.11.2 followed the updating of LLamaSharp.
+      v0.12.0 released with v0.12.0 of LLamaSharp.
     </PackageReleaseNotes>
     <PackageLicenseExpression>MIT</PackageLicenseExpression>
     <PackageOutputPath>packages</PackageOutputPath>

LLama.SemanticKernel/LLamaSharp.SemanticKernel.csproj

@@ -10,7 +10,7 @@
 		<ImplicitUsings>enable</ImplicitUsings>
 		<Nullable>enable</Nullable>
 
-		<Version>0.11.2</Version>
+		<Version>0.12.0</Version>
 		<Authors>Tim Miller, Xbotter</Authors>
 		<Company>SciSharp STACK</Company>
 		<GeneratePackageOnBuild>true</GeneratePackageOnBuild>
@@ -23,7 +23,7 @@
 			The integration of LLamaSharp and Microsoft semantic-kernel.
 		</Description>
 		<PackageReleaseNotes>
-			v0.11.2 followed the updating of LLamaSharp.
+			v0.12.0 released with v0.12.0 of LLamaSharp.
 		</PackageReleaseNotes>
 		<PackageLicenseExpression>MIT</PackageLicenseExpression>
 		<PackageOutputPath>packages</PackageOutputPath>

LLama.Web/Common/ModelOptions.cs

@@ -29,12 +29,14 @@ public class ModelOptions
         /// <inheritdoc />
         public int GpuLayerCount { get; set; } = 20;
 
-        public uint SeqMax { get; }
+        /// <inheritdoc />
+        public uint SeqMax { get; set; }
 
         /// <inheritdoc />
         public uint? Seed { get; set; } = 1686349486;
 
-        public bool Embeddings { get; }
+        /// <inheritdoc />
+        public bool Embeddings { get; set; }
 
         /// <inheritdoc />
         public bool UseMemorymap { get; set; } = true;
@@ -102,6 +104,9 @@ public class ModelOptions
         /// <inheritdoc />
         public bool NoKqvOffload { get; set; }
 
+        /// <inheritdoc />
+        public bool FlashAttention { get; set; }
+
         /// <inheritdoc />
         public Encoding Encoding { get; set; } = Encoding.UTF8;
 

LLama/Abstractions/IContextParams.cs

@@ -109,7 +109,14 @@ public interface IContextParams
     bool NoKqvOffload { get; }
 
     /// <summary>
+    /// Whether to use flash attention
+    /// </summary>
+    bool FlashAttention { get; }
+
+    /// <summary>
+    /// defragment the KV cache if holes/size &gt; defrag_threshold, Set to &lt; 0 to disable (default)
     /// defragment the KV cache if holes/size &gt; defrag_threshold, Set to <see langword="null"/> or &lt; 0 to disable (default)
+
     /// </summary>
     float? DefragThreshold { get; }
 

LLama/Abstractions/IModelParams.cs

@@ -4,6 +4,7 @@
 using System.Collections.Generic;
 using System.ComponentModel;
 using System.Linq;
+using System.Text;
 using System.Text.Json;
 using System.Text.Json.Serialization;
 using LLama.Native;
@@ -241,6 +242,7 @@ public sealed record MetadataOverride
         private readonly int _valueInt;
         private readonly float _valueFloat;
         private readonly bool _valueBool;
+        private readonly byte[]? _valueString;
 
         /// <summary>
         /// Create a new override for an int key
@@ -278,6 +280,21 @@ public MetadataOverride(string key, bool value)
             Type = LLamaModelKvOverrideType.Bool;
         }
 
+        /// <summary>
+        /// Create a new override for a string key
+        /// </summary>
+        /// <param name="key"></param>
+        /// <param name="value"></param>
+        public MetadataOverride(string key, string value)
+        {
+            Key = key;
+            _valueString = Encoding.UTF8.GetBytes(value); 
+            Type = LLamaModelKvOverrideType.String;
+
+            if (_valueString.Length > 128)
+                throw new ArgumentException("Value string is too long, must be < 128 UTF8 bytes", nameof(value));
+        }
+
         internal void WriteValue(ref LLamaModelMetadataOverride dest)
         {
             switch (Type)
@@ -291,6 +308,13 @@ internal void WriteValue(ref LLamaModelMetadataOverride dest)
                 case LLamaModelKvOverrideType.Bool:
                     dest.BoolValue = _valueBool ? -1L : 0;
                     break;
+                case LLamaModelKvOverrideType.String:
+                    unsafe
+                    {
+                        fixed (byte* strValPtr = dest.StringValue)
+                            new Span<byte>(_valueString!).CopyTo(new Span<byte>(strValPtr, 128));
+                    }
+                    break;
                 default:
                     throw new InvalidEnumArgumentException($"Unknown {nameof(LLamaModelKvOverrideType)} value: {Type}");
             }

LLama/Common/ModelParams.cs

@@ -99,6 +99,10 @@ public record ModelParams
         /// <inheritdoc />
         public bool NoKqvOffload { get; set; }
 
+        /// <inheritdoc />
+
+        public bool FlashAttention { get; set; }
+
         /// <inheritdoc />
         public float? DefragThreshold { get; set; }
 

LLama/Extensions/IContextParamsExtensions.cs

@@ -50,6 +50,7 @@ public static void ToLlamaContextParams(this IContextParams @params, out LLamaCo
             result.type_k = @params.TypeK ?? GGMLType.GGML_TYPE_F16;
             result.type_k = @params.TypeV ?? GGMLType.GGML_TYPE_F16;
             result.offload_kqv = !@params.NoKqvOffload;
+            result.flash_attention = @params.FlashAttention;
             result.llama_pooling_type = @params.PoolingType;
 
             result.n_threads = Threads(@params.Threads);

LLama/LLamaSharp.csproj

@@ -7,7 +7,7 @@
     <Platforms>AnyCPU;x64;Arm64</Platforms>
     <AllowUnsafeBlocks>True</AllowUnsafeBlocks>
 
-    <Version>0.11.2</Version>
+    <Version>0.12.0</Version>
     <Authors>Rinne, Martin Evans, jlsantiago and all the other contributors in https://github.com/SciSharp/LLamaSharp/graphs/contributors.</Authors>
     <Company>SciSharp STACK</Company>
     <GeneratePackageOnBuild>true</GeneratePackageOnBuild>
@@ -22,7 +22,7 @@
       With the higher-level APIs and RAG support, it's convenient to deploy LLM (Large Language Model) in your application with LLamaSharp.
     </Description>
     <PackageReleaseNotes>
-      LLamaSharp 0.11.2 fixed the performance issue of LLaVA on GPU and improved the log suppression.
+      Updated llama.cpp version to include better support for LLama3 tokenization.
     </PackageReleaseNotes>
     <PackageLicenseExpression>MIT</PackageLicenseExpression>
     <PackageOutputPath>packages</PackageOutputPath>

LLama/LLamaStatelessExecutor.cs

@@ -4,6 +4,7 @@
 using System.Collections.Generic;
 using System.Linq;
 using System.Runtime.CompilerServices;
+using System.Text;
 using System.Threading;
 using LLama.Exceptions;
 using LLama.Native;
@@ -123,8 +124,8 @@ public async IAsyncEnumerable<string> InferAsync(string prompt, IInferenceParams
                     );
                 }
 
-                // Check if this is the EOS token
-                if (id == _weights.Tokens.EOS)
+                // Check if this token should end generation
+                if (_weights.Tokens.IsEndOfGeneration(id))
                     break;
 
                 // Decode this token into text

LLama/Native/LLamaContextParams.cs

@@ -151,6 +151,16 @@ public bool offload_kqv
         }
         private sbyte _offload_kqv;
 
+        /// <summary>
+        /// whether to use flash attention
+        /// </summary>
+        public bool flash_attention
+        {
+            readonly get => Convert.ToBoolean(_flash_attention);
+            set => _flash_attention = Convert.ToSByte(value);
+        }
+        private sbyte _flash_attention;
+
         //todo: implement abort callback support
         /// <summary>
         /// ggml_abort_callback

LLama/Native/LLamaFtype.cs

@@ -171,6 +171,11 @@ public enum LLamaFtype
         /// </summary>
         LLAMA_FTYPE_MOSTLY_IQ1_M = 31,
 
+        /// <summary>
+        /// except 1d tensors
+        /// </summary>
+        LLAMA_FTYPE_MOSTLY_BF16 = 32,
+
         /// <summary>
         /// File type was not specified
         /// </summary>

LLama/Native/LLamaModelMetadataOverride.cs

@@ -43,6 +43,12 @@ public unsafe struct LLamaModelMetadataOverride
     /// </summary>
     [FieldOffset(136)]
     public long BoolValue;
+
+    /// <summary>
+    /// Value, **must** only be used if Tag == String
+    /// </summary>
+    [FieldOffset(136)]
+    public fixed byte StringValue[128];
 }
 
 /// <summary>
@@ -65,4 +71,9 @@ public enum LLamaModelKvOverrideType
     /// Overriding a bool value
     /// </summary>
     Bool = 2,
+
+    /// <summary>
+    /// Overriding a string value
+    /// </summary>
+    String = 3,
 }

LLama/Native/LLamaModelParams.cs

@@ -81,6 +81,16 @@ public bool use_mlock
         }
         private sbyte _use_mlock;
 
+        /// <summary>
+        /// validate model tensor data
+        /// </summary>
+        public bool check_tensors
+        {
+            readonly get => Convert.ToBoolean(_check_tensors);
+            set => _check_tensors = Convert.ToSByte(value);
+        }
+        private sbyte _check_tensors;
+
         /// <summary>
         /// Create a LLamaModelParams with default values
         /// </summary>

LLama/Native/LLamaModelQuantizeParams.cs

@@ -70,6 +70,16 @@ public bool pure
         }
         private sbyte _pure;
 
+        /// <summary>
+        /// quantize to the same number of shards
+        /// </summary>
+        public bool keep_split
+        {
+            get => Convert.ToBoolean(_keep_split);
+            set => _keep_split = Convert.ToSByte(value);
+        }
+        private sbyte _keep_split;
+
         /// <summary>
         /// pointer to importance matrix data
         /// </summary>

LLama/Native/LLamaVocabPreType.cs

@@ -0,0 +1,17 @@
+namespace LLama.Native;
+
+/// <summary>
+/// 
+/// </summary>
+/// <remarks>llama_vocab_pre_type</remarks>
+internal enum LLamaVocabPreType
+{
+    LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0,
+    LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1,
+    LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM = 2,
+    LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER = 3,
+    LLAMA_VOCAB_PRE_TYPE_FALCON = 4,
+    LLAMA_VOCAB_PRE_TYPE_MPT = 5,
+    LLAMA_VOCAB_PRE_TYPE_STARCODER = 6,
+    LLAMA_VOCAB_PRE_TYPE_GPT2 = 7,
+}

LLama/Native/NativeApi.LLava.cs

@@ -13,6 +13,7 @@ public static unsafe partial class NativeApi
     /// <param name="ctxClip">Llava Model</param>
     /// <returns>True if validate successfully</returns>
     [DllImport(llavaLibraryName, EntryPoint = "llava_validate_embed_size", CallingConvention = CallingConvention.Cdecl)]
+    [return: MarshalAs(UnmanagedType.U1)]
     public static extern bool llava_validate_embed_size( SafeLLamaContextHandle ctxLlama, SafeLlavaModelHandle ctxClip);
 
     /// <summary>
@@ -56,7 +57,7 @@ SafeLlavaImageEmbedHandle llava_image_embed_make_with_filename(SafeLlavaModelHan
     /// <param name="embed">Embedding handle</param>
     /// <returns>True on success</returns>
     [DllImport(llavaLibraryName, EntryPoint = "llava_eval_image_embed", CallingConvention = CallingConvention.Cdecl)]
-    public static extern bool llava_eval_image_embed(SafeLLamaContextHandle ctx_llama, SafeLlavaImageEmbedHandle embed,
-        int n_batch, ref int n_past);
+    [return: MarshalAs(UnmanagedType.U1)]
+    public static extern bool llava_eval_image_embed(SafeLLamaContextHandle ctx_llama, SafeLlavaImageEmbedHandle embed, int n_batch, ref int n_past);
 
 }

LLama/Native/NativeApi.Sampling.cs

@@ -176,7 +176,7 @@ public static void llama_sample_apply_guidance(SafeLLamaContextHandle ctx, Span<
         public static extern LLamaToken llama_sample_token_greedy(SafeLLamaContextHandle ctx, ref LLamaTokenDataArrayNative candidates);
 
         /// <summary>
-        /// Randomly selects a token from the candidates based on their probabilities.
+        /// Randomly selects a token from the candidates based on their probabilities using the RNG of ctx.
         /// </summary>
         /// <param name="ctx"></param>
         /// <param name="candidates">Pointer to LLamaTokenDataArray</param>