diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index 1fed9d23e..c98b551b3 100644
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -10933,7 +10933,7 @@ inline void ggml_cuda_flash_attn_ext(const ggml_tensor * Q, const ggml_tensor *
 
     const int nwarps_max = 8; // TODO: we don't want to launch too much warps. how much is too much?
                               // TODO: produces wrong results for nwarps > 8 (RTX 2060) - not sure why
-    const int nwarps = Q->ne[1] <= nqpb ? std::max(2, std::min((int) K->ne[1]/ncpw, nwarps_max)) : 2;
+    const int nwarps = Q->ne[1] <= nqpb ? std::max(2, std::min((int) K->ne[1]/ncpw, nwarps_max)) : 1;
 
     dim3 blocks_num((Q->ne[1] + nqpb - 1) / nqpb, Q->ne[2], Q->ne[3]);
     dim3 block_dim(32, nwarps, 1);