Previously models produced garbage output when running on GPU with layers offloaded. Similar to related fix on another repo: 331326a0e3
331326a0e3