in OpticalFlowLoss.init_GMA, RAFTGMA is initialized wrapped in torch.nn.DataParallel. I'm pretty sure the only thing this accomplishes is providing a target for "module." prefix in the parameter key names. I suspect that nuances with DataParallel that are not respected in the pytti code base are the cause memory leaks like crashing the discord bot and other OOM issues that manifest as if memory is being filled and not released -- even in the presence of errors -- unless the kernel is reset.
in
OpticalFlowLoss.init_GMA, RAFTGMA is initialized wrapped in torch.nn.DataParallel. I'm pretty sure the only thing this accomplishes is providing a target for"module."prefix in the parameter key names. I suspect that nuances with DataParallel that are not respected in the pytti code base are the cause memory leaks like crashing the discord bot and other OOM issues that manifest as if memory is being filled and not released -- even in the presence of errors -- unless the kernel is reset.