From 596d51184a8354b2c7f6060c45418fde10109daa Mon Sep 17 00:00:00 2001
From: Jacob Mink <Jacob.Mink@Dell.com>
Date: Fri, 6 Jun 2025 09:25:09 -0500
Subject: [PATCH] Add DecodeSpecialTokens property to IInferenceParams -
 implement it where applicable. Use the new property in StatelessExecutor

---
 LLama.Web/Common/InferenceOptions.cs   | 3 +++
 LLama/Abstractions/IInferenceParams.cs | 8 ++++++++
 LLama/Common/InferenceParams.cs        | 3 +++
 LLama/LLamaStatelessExecutor.cs        | 5 ++++-
 4 files changed, 18 insertions(+), 1 deletion(-)

diff --git a/LLama.Web/Common/InferenceOptions.cs b/LLama.Web/Common/InferenceOptions.cs
index c49d3aa31..1b32a996c 100644
--- a/LLama.Web/Common/InferenceOptions.cs
+++ b/LLama.Web/Common/InferenceOptions.cs
@@ -21,5 +21,8 @@ public class InferenceOptions
 
         /// <inheritdoc />
         public ISamplingPipeline SamplingPipeline { get; set; } = new DefaultSamplingPipeline();
+
+        /// <inheritdoc />
+        public bool DecodeSpecialTokens { get; set; }
     }
 }
diff --git a/LLama/Abstractions/IInferenceParams.cs b/LLama/Abstractions/IInferenceParams.cs
index fe415efed..12a335d08 100644
--- a/LLama/Abstractions/IInferenceParams.cs
+++ b/LLama/Abstractions/IInferenceParams.cs
@@ -28,5 +28,13 @@ public interface IInferenceParams
 		/// Set a custom sampling pipeline to use.
 		/// </summary>
 		ISamplingPipeline SamplingPipeline { get; set; }
+
+		/// <summary>
+		/// If true, special characters will be converted to text. If false they will be invisible.
+		/// </summary>
+		/// <remark>
+		/// Controls the behavior of decoders like <see cref="StreamingTokenDecoder" />
+		/// </remark>
+		public bool DecodeSpecialTokens { get; set; }
 	}
 }
\ No newline at end of file
diff --git a/LLama/Common/InferenceParams.cs b/LLama/Common/InferenceParams.cs
index 8f2a5a14f..aa923bf8f 100644
--- a/LLama/Common/InferenceParams.cs
+++ b/LLama/Common/InferenceParams.cs
@@ -30,6 +30,9 @@ public record InferenceParams
 
         /// <inheritdoc />
         public ISamplingPipeline SamplingPipeline { get; set; } = new DefaultSamplingPipeline();
+
+        /// <inheritdoc />
+        public bool DecodeSpecialTokens { get; set; }
     }
 
     /// <summary>
diff --git a/LLama/LLamaStatelessExecutor.cs b/LLama/LLamaStatelessExecutor.cs
index 8aa705062..817738895 100644
--- a/LLama/LLamaStatelessExecutor.cs
+++ b/LLama/LLamaStatelessExecutor.cs
@@ -88,7 +88,10 @@ public async IAsyncEnumerable<string> InferAsync(string prompt, IInferenceParams
                 throw new ArgumentOutOfRangeException(nameof(inferenceParams), $"TokensKeep ({inferenceParams.TokensKeep}) cannot be larger than ContextSize ({Context.ContextSize})");
 
             // Create decoders for the token stream
-            var decoder = new StreamingTokenDecoder(Context);
+            var decoder = new StreamingTokenDecoder(Context)
+            {
+                DecodeSpecialTokens = inferenceParams.DecodeSpecialTokens,
+            };
             var antiprocessor = new AntipromptProcessor(inferenceParams.AntiPrompts);
 
             if (ApplyTemplate)