sg>pddlZddlZddlmcmcmcmZGddejjZ y)NceZdZdZdej j dejfdZ e dZ dZ dZ dZd Zdd efd Zd Zy )PostLocalSGDOptimizera Wraps an arbitrary :class:`torch.optim.Optimizer` and runs `post-local SGD `_, This optimizer runs local optimizer at every step. After the warm-up stage, it averages parameters periodically afer the local optimizer is applied. Args: optim: The local optimizer. averager: A model averager instance to run post-localSGD algorithm. Example:: >>> # xdoctest: +SKIP("undefined variables") >>> import torch >>> import torch.distributed as dist >>> import torch.distributed.algorithms.model_averaging.averagers as averagers >>> import torch.nn as nn >>> from torch.distributed.optim import PostLocalSGDOptimizer >>> from torch.distributed.algorithms.ddp_comm_hooks.post_localSGD_hook import ( >>> PostLocalSGDState, >>> post_localSGD_hook, >>> ) >>> >>> model = nn.parallel.DistributedDataParallel( >>> module, device_ids=[rank], output_device=rank >>> ) >>> >>> # Register a post-localSGD communication hook. >>> state = PostLocalSGDState(process_group=None, subgroup=None, start_localSGD_iter=100) >>> model.register_comm_hook(state, post_localSGD_hook) >>> >>> # Create a post-localSGD optimizer that wraps a local optimizer. >>> # Note that ``warmup_steps`` used in ``PostLocalSGDOptimizer`` must be the same as >>> # ``start_localSGD_iter`` used in ``PostLocalSGDState``. >>> local_optim = torch.optim.SGD(params=model.parameters(), lr=0.01) >>> opt = PostLocalSGDOptimizer( >>> optim=local_optim, >>> averager=averagers.PeriodicModelAverager(period=4, warmup_steps=100) >>> ) >>> >>> # In the first 100 steps, DDP runs global gradient averaging at every step. >>> # After 100 steps, DDP runs gradient averaging within each subgroup (intra-node by default), >>> # and post-localSGD optimizer runs global model averaging every 4 steps after applying the local optimizer. >>> for step in range(0, 200): >>> opt.zero_grad() >>> loss = loss_fn(output, labels) >>> loss.backward() >>> opt.step() optimaveragercV||_|jj|_||_yN)r param_groupsr)selfrrs b/var/www/html/venv/lib/python3.12/site-packages/torch/distributed/optim/post_localSGD_optimizer.py__init__zPostLocalSGDOptimizer.__init__:s"  JJ33  c.|jjSr)rstater s r rzPostLocalSGDOptimizer.state?szzr c6|jjSr)r__repr__rs r rzPostLocalSGDOptimizer.__repr__Cszz""$$r cl|jj}|jj|d<|S)z This is the same as :class:`torch.optim.Optimizer` :meth:`state_dict`, but adds an extra entry to record model averager's step to the checkpoint to ensure reload does not cause unnecessary warm up again. step)r state_dictrr)r optim_state_dicts r rz PostLocalSGDOptimizer.state_dictFs2  ::002#'==#5#5 r c|jj|d|vr|d|j_yt j dd|j_y)aW This is the same as :class:`torch.optim.Optimizer` :meth:`load_state_dict`, but also restores model averager's step value to the one saved in the provided ``state_dict``. If there is no ``"step"`` entry in ``state_dict``, it will raise a warning and initialize the model averager's step to 0. rz]Loaded state dict does not contain a step counter for an averager. Setting step counter to 0.rN)rload_state_dictrrwarningswarn)r rs r rz%PostLocalSGDOptimizer.load_state_dictPsN "":. Z !+F!3DMM  MM- "#DMM r c|jj|jj|jy)zI Performs a single optimization step (parameter update). )paramsN)rrraverage_parametersr rs r rzPostLocalSGDOptimizer.stepcs-  ((0A0A(Br set_to_nonec<|jj|y)N)r)r zero_grad)r rs r r zPostLocalSGDOptimizer.zero_gradjs 5r c:|jj|yr)radd_param_group)r param_groups r r"z%PostLocalSGDOptimizer.add_param_groupms "";/r N)T)__name__ __module__ __qualname____doc__torchr Optimizer averagers ModelAveragerr propertyrrrrrboolr r"r r rrse/b!ekk33!y?V?V!   % #&C6T60r r) rr(6torch.distributed.algorithms.model_averaging.averagers distributed algorithmsmodel_averagingr*rr)rr.r r r3s) JJf0EKK11f0r