From 0a70ca0bc05abeccf78867545555ab6079ec2c4d Mon Sep 17 00:00:00 2001
From: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Date: Tue, 23 Sep 2025 17:25:47 +0300
Subject: [PATCH] Fix #772

---
 ggml/src/ggml-cuda/fattn-mma-f16.cuh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
index 0c1aaf7d..27328319 100644
--- a/ggml/src/ggml-cuda/fattn-mma-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
@@ -1408,7 +1408,7 @@ void launch_fattn_mma(
 
         //const bool use_stream_k = cc >= CC_ADA_LOVELACE || tiles_efficiency_percent < 75;
         //  On my RTX-4080 the above is slightly slower for PP. It would be useful to try and see what happens on Blackwell
-        const bool use_stream_k = tiles_efficiency_percent < 75;
+        const bool use_stream_k = tiles_efficiency_percent < 75 || Q->ne[1] > 2048;
 
         blocks_num.x = use_stream_k ? nblocks_stream_k : ntiles_total;
         blocks_num.y = 1;