added torch_gc() which calls both cuda.empty_cache() and cuda.ipc_collect()
called before and after generation