don't use mmvq in pascal and lower

Author: am17an

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -2131,6 +2131,11 @@ static bool ggml_cuda_should_fuse_mul_mat_vec_q(const ggml_tensor * tensor) {
     bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) && !bad_padding_clear && src1->type == GGML_TYPE_F32 &&
                              dst->type == GGML_TYPE_F32 && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
 
+    // fusion is not universally faster on Pascal
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    if (cc <= GGML_CUDA_CC_PASCAL) {
+        return false;
+    }
     //we only support fusion for ncols_dst = 1
     if (tensor->op == GGML_OP_MUL_MAT && dst->ne[1] != 1) {
         return false;