Added back the assumption comment.

microsoft · Aug 17, 2024 · d0b0627 · d0b0627
1 parent c8b187f
commit d0b0627
Showing 1 changed file with 1 addition and 0 deletions.
diff --git a/js/web/lib/wasm/jsep/webgpu/ops/attention.ts b/js/web/lib/wasm/jsep/webgpu/ops/attention.ts
@@ -675,6 +675,7 @@ export const applyAttention = (
   parameters: AttentionParameters,
   attributes: AttentionAttrs,
 ) => {
+  // Assumption  is that presentKey/presentValue exists only if pastKey/pastValue exists.
   const outputCount = Math.min(context.outputCount, 1 + (pastKey ? 1 : 0) + (pastValue ? 1 : 0));
   const pastSequenceLength = parameters.kvNumHeads !== undefined || outputCount > 1 ? parameters.pastSequenceLength : 0;
   const totalSequenceLength = pastSequenceLength + parameters.kvSequenceLength;