From 87b29c19ac50bed22239bf558ca212265cdfa4c6 Mon Sep 17 00:00:00 2001
From: "assistant-librarian[bot]"
 <assistant-librarian[bot]@users.noreply.github.com>
Date: Tue, 8 Jul 2025 01:34:41 +0000
Subject: [PATCH] Merge commit 'b2dea90116d1060c67db5edddb6d4498188ebac4' into
 develop

---
 .../block_fmha_batch_prefill_pipeline_qr_ks_vs_async.hpp       | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/include/ck_tile/ops/fmha/pipeline/block_fmha_batch_prefill_pipeline_qr_ks_vs_async.hpp b/include/ck_tile/ops/fmha/pipeline/block_fmha_batch_prefill_pipeline_qr_ks_vs_async.hpp
index 6398bf316e..10daea99d1 100644
--- a/include/ck_tile/ops/fmha/pipeline/block_fmha_batch_prefill_pipeline_qr_ks_vs_async.hpp
+++ b/include/ck_tile/ops/fmha/pipeline/block_fmha_batch_prefill_pipeline_qr_ks_vs_async.hpp
@@ -122,6 +122,9 @@ struct BlockFmhaBatchPrefillPipelineQRKSVSAsync
             {
                 if constexpr(kPadSeqLenK && BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
                     return 1;
+                // use larger K/V LDS buffer size will lower the occupancy
+                else if constexpr(64 <= kK0 || 64 <= kK1)
+                    return 1;
                 else
                     return 2;
             }