Merge pull request #1282 from SerialKicked/fix-grammar-topk

martindevans · web-flow · commit c1783d4e24f8 · 2025-10-19T15:49:38.000+01:00
Fixes Extended grammar optimization mode would crash if TopK sampling is set to zero.
diff --git 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);

Original file line number	Diff line number	Diff line change
`@@ -246,7 +246,8 @@ public override LLamaToken Sample(SafeLLamaContextHandle ctx, int index)`
`246`	`246`	`}`
`247`	`247`
`248`	`248`	`// Extended optimization : Apply the grammar to the TopK tokens and check if the selected token is valid`
`249`		`- if (GrammarOptimization == GrammarOptimizationMode.Extended)`
	`249`	`+ // Only run if TopK > 0`
	`250`	`+ if (GrammarOptimization == GrammarOptimizationMode.Extended && TopK > 0)`
`250`	`251`	`{`
`251`	`252`	`// Calculate a safe TopK value`
`252`	`253`	`var safeTopK = Math.Min(TopK, nativeAll.Data.Length);`