diff --git a/ggml-cuda.cu b/ggml-cuda.cu index ac91ee12e3428..51db2c0457408 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -7929,12 +7929,16 @@ 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; + } } } }