I understand this. But what else does it contain? I experienced multiple times that there is a significant difference between reserved and allocated memory after calling torch.cuda.empty_cache(). My undertanding is that they shoud be almost the same after the cache is freed. @ptrblck
1 Like