From fb24cf84975107c6cdb0ed1f071b0fd68bf6b6ec Mon Sep 17 00:00:00 2001
From: Umberto <umberto.valleriani@gmail.com>
Date: Wed, 14 Oct 2026 12:05:49 +0200
Subject: [PATCH] cuda: exclude sm_75 from the short-query in-place
 quantized-K/V MMA FA path (#348)

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

diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
index 32f6440e667d..d74a591f4ae2 100644
--- a/ggml/src/ggml-cuda/fattn.cu
+++ b/ggml/src/ggml-cuda/fattn.cu
@@ -13,7 +13,7 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_con
 
     // keep type_KV: the F16 instantiation converts K/V to f16 into scratch that
     // ggml_cuda_flash_attn_ext_get_alloc_size does not reserve for the in-place quantized path
-    if (turing_mma_available(cc) && ggml_is_quantized(V->type) && Q->ne[1] > 2 && Q->ne[1] <= 4) {
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) > GGML_CUDA_CC_TURING && ggml_is_quantized(V->type) && Q->ne[1] > 2 && Q->ne[1] <= 4) {
         ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 64/ncols2, ncols2, type_KV>(ctx, dst);
         return;
     }