From 2bb57b49005a5d56b6915d538dc1690fb5ef9d84 Mon Sep 17 00:00:00 2001
From: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Date: Sun, 9 Nov 2025 10:57:52 +0200
Subject: [PATCH] This seems better

---
 ggml/src/ggml-cuda/fattn-new-mma.cu | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/ggml/src/ggml-cuda/fattn-new-mma.cu b/ggml/src/ggml-cuda/fattn-new-mma.cu
index 68c12dd3..b2285fdd 100644
--- a/ggml/src/ggml-cuda/fattn-new-mma.cu
+++ b/ggml/src/ggml-cuda/fattn-new-mma.cu
@@ -1746,6 +1746,9 @@ static void launch_fattn_new_mma(
     const int nsm_actual = ggml_cuda_info().devices[id].nsm;
     int nsm = 1; while (nsm*2 <= nsm_actual) nsm *= 2;
 
+    if (Q->ne[1] ==  1 && K->ne[1] <= 4096 && nsm > 32) nsm /= 2;
+    if (Q->ne[1] >= 32 && K->ne[1] >= 4096) nsm *= 2;
+
     ggml_cuda_pool_alloc<half>   K_f16(pool);
     ggml_cuda_pool_alloc<half>   V_f16(pool);
     ggml_cuda_pool_alloc<int>    KV_max(pool);