ggml

79ea7a41 - cuda : support flattened GLM-style rope to reduce kernel launch (#477)