bug fix: remove flash attention parameter from the model params

Author: krisbiradar

LLama.KernelMemory/BuilderExtensions.cs

@@ -77,7 +77,6 @@ public static IKernelMemoryBuilder WithLLamaSharpDefaults(this IKernelMemoryBuil
                 SplitMode = config.SplitMode,
                 BatchSize = 512,
                 UBatchSize = 512,
-                FlashAttention = true,
                 UseMemorymap = true
             };
 

LLama.KernelMemory/LLamaSharpTextEmbeddingGenerator.cs

@@ -40,7 +40,6 @@ public LLamaSharpTextEmbeddingGenerator(LLamaSharpConfig config)
                 SplitMode = config?.SplitMode ?? LLama.Native.GPUSplitMode.Layer,
                 BatchSize = 512,
                 UBatchSize = 512,
-                FlashAttention = true,
                 UseMemorymap = true,
                 PoolingType = LLamaPoolingType.Mean,
             };
@@ -68,7 +67,6 @@ public LLamaSharpTextEmbeddingGenerator(LLamaSharpConfig config, LLamaWeights we
                 SplitMode = config?.SplitMode ?? LLama.Native.GPUSplitMode.Layer,
                 BatchSize = 512,
                 UBatchSize = 512,
-                FlashAttention = true,
                 UseMemorymap = true,
                 PoolingType = LLamaPoolingType.Mean,
             };

LLama.KernelMemory/LlamaSharpTextGenerator.cs

@@ -38,7 +38,6 @@ public LlamaSharpTextGenerator(LLamaSharpConfig config)
                 SplitMode = config?.SplitMode ?? LLama.Native.GPUSplitMode.Layer,
                 BatchSize = 512,
                 UBatchSize = 512,
-                FlashAttention = true,
                 UseMemorymap = true
             };
             _weights = LLamaWeights.LoadFromFile(@params);
@@ -66,7 +65,6 @@ public LlamaSharpTextGenerator(LLamaWeights weights, LLamaSharpConfig config, St
                 SplitMode = config?.SplitMode ?? LLama.Native.GPUSplitMode.Layer,
                 BatchSize = 512,
                 UBatchSize = 512,
-                FlashAttention = true,
                 UseMemorymap = true
             };
             _executor = executor ?? new StatelessExecutor(_weights, @params);

LLama.Unittest/SamplingTests.cs

@@ -104,7 +104,7 @@ public void BatchedSampling()
                 }
             }
 
-            // Add " repeat" and test whether next tokens will be "this phrase forever.".
+            // Add " repeat" and test whether next tokens will be "this phrase forever."
             for (int i = 0; i < 4; i++)
             {
                 for (int b = 0; b < batch_count; b++)

LLama/Abstractions/IContextParams.cs

@@ -103,11 +103,6 @@ public interface IContextParams
     /// </summary>
     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)
     /// </summary>

LLama/Common/ModelParams.cs

@@ -1,3 +1,4 @@
+using System;
 using LLama.Abstractions;
 using System.Text;
 using System.Text.Json.Serialization;
@@ -97,10 +98,7 @@ public record ModelParams
         public bool NoKqvOffload { get; set; }
 
         /// <inheritdoc />
-
-        public bool FlashAttention { get; set; }
-
-        /// <inheritdoc />
+        [Obsolete]
         public float? DefragThreshold { get; set; }
 
         /// <inheritdoc />

LLama/Extensions/IContextParamsExtensions.cs

@@ -49,7 +49,6 @@ public static void ToLlamaContextParams(this IContextParams @params, out LLamaCo
             result.type_k = @params.TypeK ?? GGMLType.GGML_TYPE_F16;
             result.type_v = @params.TypeV ?? GGMLType.GGML_TYPE_F16;
             result.offload_kqv = !@params.NoKqvOffload;
-            result.flash_attention = @params.FlashAttention;
             result.llama_pooling_type = @params.PoolingType;
             result.attention_type = @params.AttentionType;
 

LLama/Native/SafeLLamaContextHandle.cs

@@ -348,15 +348,15 @@ static SafeLLamaContextHandle()
         /// <param name="key"></param>
         /// <param name="buf"></param>
         /// <param name="buf_size"></param>
-        /// <returns></returns>
+        /// <returns>The length of the value string (on success) -1 otherwise </returns>
         [DllImport(NativeApi.libraryName, CallingConvention = CallingConvention.Cdecl)]
         private static extern int llama_adapter_meta_val_str(IntPtr adapter, string key, StringBuilder buf, UIntPtr buf_size);
 
         /// <summary>
         /// Get the number of metadata key value pairs
         /// </summary>
         /// <param name="adapter"></param>
-        /// <returns></returns>
+        /// <returns>The count of meta key value pairs</returns>
         [DllImport(NativeApi.libraryName, CallingConvention =  CallingConvention.Cdecl)]
         private static extern int llama_adapter_meta_count(IntPtr adapter);
 
@@ -367,7 +367,7 @@ static SafeLLamaContextHandle()
         /// <param name="i"></param>
         /// <param name="buf"></param>
         /// <param name="buf_size"></param>
-        /// <returns></returns>
+        /// <returns>The length of string i.e meta key (on success) -1 otherwise</returns>
         [DllImport(NativeApi.libraryName, CallingConvention = CallingConvention.Cdecl)]
         private static extern int llama_adapter_meta_key_by_index(IntPtr adapter, int i, StringBuilder buf, UIntPtr buf_size);
 
@@ -378,7 +378,7 @@ static SafeLLamaContextHandle()
         /// <param name="i"></param>
         /// <param name="buf"></param>
         /// <param name="buf_size"></param>
-        /// <returns></returns>
+        /// <returns>The length of value string (on success) -1 otherwise</returns>
         [DllImport(NativeApi.libraryName, CallingConvention = CallingConvention.Cdecl)]
         private static extern int llama_adapter_meta_val_by_index(IntPtr adapter, int i, StringBuilder buf,  UIntPtr buf_size);