diff --git a/otherarch/sdcpp/lora.hpp b/otherarch/sdcpp/lora.hpp
index 5f458faee..69673a6f4 100644
--- a/otherarch/sdcpp/lora.hpp
+++ b/otherarch/sdcpp/lora.hpp
@@ -140,10 +140,12 @@ struct LoraModel : public GGMLRunner {
             scale_value *= multiplier;
 
             // flat lora tensors to multiply it
-            int64_t lora_up_rows   = lora_up->ne[ggml_n_dims(lora_up) - 1];
-            lora_up                = ggml_reshape_2d(compute_ctx, lora_up, ggml_nelements(lora_up) / lora_up_rows, lora_up_rows);
-            int64_t lora_down_rows = lora_down->ne[ggml_n_dims(lora_down) - 1];
-            lora_down              = ggml_reshape_2d(compute_ctx, lora_down, ggml_nelements(lora_down) / lora_down_rows, lora_down_rows);
+            int64_t lora_up_rows  = lora_up->ne[ggml_n_dims(lora_up) - 1];
+            lora_up               = ggml_reshape_2d(compute_ctx, lora_up, ggml_nelements(lora_up) / lora_up_rows, lora_up_rows);
+            auto lora_down_n_dims = ggml_n_dims(lora_down);
+            // assume n_dims should always be a multiple of 2 (otherwise rank 1 doesn't work)
+            lora_down_n_dims       = (lora_down_n_dims + lora_down_n_dims % 2);
+            int64_t lora_down_rows = lora_down->ne[lora_down_n_dims - 1];
 
             // ggml_mul_mat requires tensor b transposed
             lora_down                  = ggml_cont(compute_ctx, ggml_transpose(compute_ctx, lora_down));