CUDA: fixed row rounding for 0 tensor splits (llama/4594)

diff --git a/src/ggml-cuda.cu b/src/ggml-cuda.cu
index 7c2a834e34382f44ca788ff006d3877f8bab1fa3..490081cac8c1b656aa1401230144cf53cb25b27b 100644 (file)
--- a/src/ggml-cuda.cu
+++ b/src/ggml-cuda.cu
@@ -7937,12 +7937,16 @@ static void ggml_cuda_op_mul_mat(
 
             if (id != 0) {
                 row_low[id]  = ne01*g_tensor_split[id];
-                row_low[id] -= row_low[id] % rounding;
+                if (row_low[id] < ne01) {
+                    row_low[id] -= row_low[id] % rounding;
+                }
             }
 
             if (id != g_device_count - 1) {
                 row_high[id]  = ne01*g_tensor_split[id + 1];
-                row_high[id] -= row_high[id] % rounding;
+                if (row_high[id] < ne01) {
+                    row_high[id] -= row_high[id] % rounding;
+                }
             }
         }
     }