- Added new binaries

- Fixed stateless executor out-of-context handling
 - Fixed token tests

LLama.Unittest/LLamaContextTests.cs

@@ -37,7 +37,7 @@ namespace LLama.Unittest
         {
             var tokens = _context.Tokenize("The quick brown fox", true);
 
-            Assert.Equal(new[] { 1, 450, 4996, 17354, 1701, 29916 }, tokens);
+            Assert.Equal(new[] { 1, 1576, 4996, 17354, 1701, 29916 }, tokens);
         }
 
         [Fact]
@@ -45,7 +45,7 @@ namespace LLama.Unittest
         {
             var tokens = _context.Tokenize("The quick brown fox", false);
 
-            Assert.Equal(new[] { 450, 4996, 17354, 1701, 29916 }, tokens);
+            Assert.Equal(new[] { 1576, 4996, 17354, 1701, 29916 }, tokens);
         }
 
         [Fact]

LLama.Unittest/StatelessExecutorTest.cs

@@ -48,13 +48,13 @@ namespace LLama.Unittest
         {
             var executor = new StatelessExecutor(_weights, _params);
 
-            const string question = " Question. why is a cat the best pet?\nAnswer: ";
+            const string question = " Question. cats or dogs?\nAnswer: ";
 
             // The context size is set to 60. Generate more than that, forcing it to generate a coherent response
             // with a modified context
             var @params = new InferenceParams()
             {
-                MaxTokens = 80,
+                MaxTokens = 70,
                 TokensKeep = question.Length,
             };
 

LLama/LLamaStatelessExecutor.cs

@@ -112,9 +112,6 @@ namespace LLama
                     NativeApi.llama_kv_cache_seq_shift(Context.NativeHandle, (LLamaSeqId)0, inferenceParams.TokensKeep + 1 + n_discard, n_past, -n_discard);
 
                     n_past -= n_discard;
-
-                    tokens.Clear();
-                    tokens.AddRange(lastTokens.Skip(lastTokens.Count - n_left / 2).Take(n_left / 2));
                 }
 
                 n_past = Context.Eval(tokens, n_past);