diff --git a/LLama/Sampling/DefaultSamplingPipeline.cs b/LLama/Sampling/DefaultSamplingPipeline.cs index 27a4243c8..77d20e36a 100644 --- a/LLama/Sampling/DefaultSamplingPipeline.cs +++ b/LLama/Sampling/DefaultSamplingPipeline.cs @@ -246,7 +246,8 @@ public override LLamaToken Sample(SafeLLamaContextHandle ctx, int index) } // Extended optimization : Apply the grammar to the TopK tokens and check if the selected token is valid - if (GrammarOptimization == GrammarOptimizationMode.Extended) + // Only run if TopK > 0 + if (GrammarOptimization == GrammarOptimizationMode.Extended && TopK > 0) { // Calculate a safe TopK value var safeTopK = Math.Min(TopK, nativeAll.Data.Length);