Fix model saving bug for GQA

zhongkaifu · Oct 7, 2024 · 7e8fee2 · 7e8fee2
1 parent ec3eeb6
commit 7e8fee2
Show file tree

Hide file tree

Showing 3 changed files with 5 additions and 3 deletions.
diff --git a/Seq2SeqSharp/Layers/GroupQueryAttention.cs b/Seq2SeqSharp/Layers/GroupQueryAttention.cs
@@ -108,7 +108,7 @@ public GroupQueryAttention(string name, int num_heads, int num_kv_groups, int d_
         /// <returns>Transformered output tensor</returns>
         public (IWeightTensor, IWeightTensor) Perform(IWeightTensor inputQ, IWeightTensor inputK, IWeightTensor inputV, IWeightTensor keyMask, int batchSize, IComputeGraph graph, bool outputAttenWeights = false, Dictionary<string, IWeightTensor> cachedTensors = null)
         {
-            string keyName = $"{m_name}_MultiHeadAttention_3";
+            string keyName = $"{m_name}_GroupQueryAttention_3";
             using IComputeGraph g = graph.CreateSubGraph(keyName);
             int seqLenQ = inputQ.Rows / batchSize;
 
@@ -340,7 +340,7 @@ public void Load(IModel stream)
 
         public IWeightTensor Perform(IWeightTensor inputQ, IWeightTensor keyMask, int batchSize, IComputeGraph graph, Dictionary<string, IWeightTensor> cachedTensors = null)
         {
-            string keyName = $"{m_name}_MultiHeadAttention_3";
+            string keyName = $"{m_name}_GroupQueryAttention_1";
             using IComputeGraph g = graph.CreateSubGraph(keyName);
             int seqLenQ = inputQ.Rows / batchSize;
 

diff --git a/Seq2SeqSharp/Models/Model_4_ProtoBufSerializer.cs b/Seq2SeqSharp/Models/Model_4_ProtoBufSerializer.cs
@@ -284,6 +284,8 @@ public Model_4_ProtoBufSerializer(Model m)
             ExpertsPerTokenFactor = m.ExpertsPerTokenFactor;              
             PEType= m.PEType;
             NormType = m.NormType;
+            MultiHeadAttentionType = m.MultiHeadAttentionType;
+            KVGroupNum = m.KVGroupNum;
         }
         public static Model_4_ProtoBufSerializer Create(Model m) => new Model_4_ProtoBufSerializer(m);
 

diff --git a/Seq2SeqSharp/Seq2SeqSharp.csproj b/Seq2SeqSharp/Seq2SeqSharp.csproj
@@ -15,7 +15,7 @@
     <PlatformTarget>AnyCPU</PlatformTarget>
     <AppendTargetFrameworkToOutputPath>false</AppendTargetFrameworkToOutputPath>
     <OutputPath>bin\</OutputPath>
-    <Version>2.8.13</Version>
+    <Version>2.8.14</Version>
     <Description>Seq2SeqSharp is a tensor based fast &amp; flexible encoder-decoder deep neural network framework written by .NET (C#). It can be used for sequence-to-sequence task, sequence-labeling task and sequence-classification task and other NLP tasks. Seq2SeqSharp supports both CPUs (x86, x64 and ARM64) and GPUs. It's powered by .NET core, so Seq2SeqSharp can run on both Windows and Linux without any modification and recompilation.</Description>
     <PackageReadmeFile>README.md</PackageReadmeFile>
     <Title>Seq2SeqSharp</Title>