This is a feature request. Support for alternative dtypes would be quite helpful, such as float16 or bfloat16.