fix kernel selection logic

JohannesGaessler · JohannesGaessler · commit ec176eef73e8 · 2025-11-30T08:45:35.000+01:00
diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
@@ -306,7 +306,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
         while (gqa_ratio % (2*gqa_ratio_eff) == 0 && gqa_ratio_eff < ncols2_max) {
             gqa_ratio_eff *= 2;
         }
-        if (Q->ne[1] * gqa_ratio_eff <= 2) {
+        if (can_use_vector_kernel && Q->ne[1] * gqa_ratio_eff <= 2) {
             return BEST_FATTN_KERNEL_VEC;
         }
         if (Q->ne[1] * gqa_ratio_eff <= 16) {

Original file line number	Diff line number	Diff line change
`@@ -306,7 +306,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const`
`306`	`306`	`while (gqa_ratio % (2*gqa_ratio_eff) == 0 && gqa_ratio_eff < ncols2_max) {`
`307`	`307`	`gqa_ratio_eff *= 2;`
`308`	`308`	`}`
`309`		`- if (Q->ne[1] * gqa_ratio_eff <= 2) {`
	`309`	`+ if (can_use_vector_kernel && Q->ne[1] * gqa_ratio_eff <= 2) {`
`310`	`310`	`return BEST_FATTN_KERNEL_VEC;`
`311`	`311`	`}`
`312`	`312`	`if (Q->ne[1] * gqa_ratio_eff <= 16) {`