**Is your feature request related to a problem? Please describe.** based on cuda 11.3 pytorch c3d40fd