U dE@sdZddlmZddlmZmZmZmZmZm Z m Z m Z m Z ddl Z ddl mZmZddlmZddlZ ddlZ ddlmZdd lmZdd lmZdd lmZdd lmZdd lmZm Z dgZ!e e j"e#e$fZ%e ee%ee%fZ&e eZ'e ee'fZ(erej)e(Z)ee$e)fZ*n ej)Z)eZ*e$e$dddZ+ej,ddddZ-ej,eej,ee j"ddddZ.Gddde/Z0ej)e j"dddZ1Gdddej,Z2Gdddej)Z3eej)d d!d"Z4ej,e eej,ee j"fd#d$d%Z5e6d&Z7Gd'dde)Z8dS)(zThe Pipe interface.) OrderedDict) TYPE_CHECKINGAnyIterableListOptionalUnionSequenceTuplecastN)Tensornn)RRef) microbatch)DeferredBatchNorm)Pipeline)inspect_skip_layout)verify_skippables)AbstractStream new_streamPipe)messagereturncCs |dS)zBExpands a message with recommendation to :mod:`torchpipe.balance`.a If your model is still under development, its optimal balance would change frequently. In this case, we highly recommend 'torch.distributed.pipeline.sync.balance' for naive automatic balancing: from torch.distributed.pipeline.sync import Pipe from torch.distributed.pipeline.sync.balance import balance_by_time partitions = torch.cuda.device_count() sample = torch.empty(...) balance = balance_by_time(partitions, model, sample) model = Pipe(model, balance, ...) )rrrH/tmp/pip-unpacked-wheel-ua33x9lu/torch/distributed/pipeline/sync/pipe.py_recommend_auto_balance*sr)modulercCs<t|tjstdt|}t|t|kr8tddS)Nz.module must be nn.Sequential to be partitionedz/module with duplicate children is not supported) isinstancer Sequential TypeErrorlistnamed_childrenlen ValueError)rr"rrr_verify_module=s   r%)r partitionsdevicesrc Cstt|}tdd|D}||kr2dStt|D]j}t|dt|D]R}||}||}||||krzqT|D]"} |D]} | | krtdqqqTq>dS)Ncss|]}tt|VqdSN)r#r! parameters).0childrrr Jsz$_verify_splitting..rzEmodule with duplicate parameters on distinct devices is not supported)r#r!r)sumchildrenranger$) rr&r'Znum_parametersZnum_child_parametersijZpartiZpartjpqrrr_verify_splittingFs  r4c@s eZdZdS) BalanceErrorN)__name__ __module__ __qualname__rrrrr5Zsr5cCsPd}|D],}|dkr |j}q ||jkr td|q |dk rF|StdS)a=Validates all parameters in the Module have the same device and returns the appropriate device. Args: An ``nn.Module`` to process. Returns: ``torch.Device`` for the entire module. Raises: ValueError: If devices for ``nn.Module`` parameters are not all same. Nzunn.Module: {}, should have all parameters on a single device, please use .to() to place the module on a single devicecpu)r)devicer$formattorch)rr:Z parameterrrr_retrieve_device^s  r=c@seZdZdZddZdS)PipeSequentialzK Pipe variant of ``nn.Sequential`` which supports multiple inputs. cGs*|D] }t|tr||}q||}q|Sr()rr )selfinputsrrrrforward~s    zPipeSequential.forwardN)r6r7r8__doc__rArrrrr>ysr>csLeZdZdZejejdfdd ZddZ e ddZ e d d ZZ S) WithDevicea Wraps an ``nn.Module`` which is part of ``nn.Sequential`` passed into :class:`Pipe` that overrides the device for that module. In cases where :class:`Pipe` can't implicitly determine the device for the module and places it on CPU, this wrapper can be used to override the implicit behavior and explicitly specify which device a module should run on. The provided module is also moved to the given device via ``.to(device)`` by :class:`Pipe` Args: module(:class:`torch.nn.Module`): The module to be wrapped. device(:class:`torch.device`): The device to run the module on. Example:: >>> fc1 = nn.Linear(16, 8).cuda(0) >>> fc2 = nn.Linear(8, 4).cuda(1) >>> dropout = nn.Dropout() >>> >>> # Dropout does not have any parameters/buffers, but we want to >>> # run it on cuda:1 to avoid any GPU to CPU transfers. >>> model = nn.Sequential(fc1, fc2, WithDevice(dropout, 'cuda:1')) >>> model = Pipe(model, chunks=8) )rr:cs$tt|||_t||_dSr()superrC__init___moduler<r:_device)r?rr: __class__rrrEszWithDevice.__init__cOs |j||Sr(rFr?argskwargsrrrrAszWithDevice.forwardcCs|jSr(rJr?rrrrszWithDevice.modulecCs|jSr()rGrNrrrr:szWithDevice.device) r6r7r8rBr Moduler<r:rErApropertyr __classcell__rrrHrrCs rC)modulescCs<g}|D]*}t|tjr(||q||qt|Sr()rr rextendr.appendr>)rRZ modules_listrrrr_assemble_partitions   rU)rRrcCsg}g}g}d}|D]v\}}t|trB|j}|j}||nt|}|dk r||ksd|jdkr|t |||g}|}||q|dk r|t |||t t t j t |}||fS)Nr9)r"rrCr:rtor=typerTrUr rr rZ ModuleList)rRr&r'Zcurrent_partitionZcurrent_devicenamerr:rrr _split_modules*     rYzRdenied to move parameters and buffers, because Pipe should manage device placementcseZdZdZdejeeeddfdd Z ed d d Z eej d d dZ e ej d ddZdeeddddZdd ddZeeddfdd Zeeed ddZed ddZZS)raWraps an arbitrary :class:`nn.Sequential ` module to train on using synchronous pipeline parallelism. If the module requires lots of memory and doesn't fit on a single GPU, pipeline parallelism is a useful technique to employ for training. The implementation is based on the torchgpipe_ paper. .. _torchgpipe: https://arxiv.org/abs/2004.09910 Pipe combines pipeline parallelism with checkpointing to reduce peak memory required to train while minimizing device under-utilization. You should place all the modules on the appropriate devices and wrap them into an :class:`nn.Sequential ` module defining the desired order of execution. If a module does not contain any parameters/buffers, it is assumed this module should be executed on CPU and appropriate input tensors to the module are moved to CPU before execution. This behavior can be overridden by the :class:`WithDevice` wrapper which can be used to explicitly specify which device a module should run on. Args: module (:class:`nn.Sequential `): sequential module to be parallelized using pipelining. Each module in the sequence has to have all of its parameters on a single device. Each module in the sequence has to either be an nn.Module or :class:`nn.Sequential ` (to combine multiple sequential modules on a single device) chunks (int): number of micro-batches (default: ``1``) checkpoint (str): when to enable checkpointing, one of ``'always'``, ``'except_last'``, or ``'never'`` (default: ``'except_last'``). ``'never'`` disables checkpointing completely, ``'except_last'`` enables checkpointing for all micro-batches except the last one and ``'always'`` enables checkpointing for all micro-batches. deferred_batch_norm (bool): whether to use deferred ``BatchNorm`` moving statistics (default: :data:`False`). If set to :data:`True`, we track statistics across multiple micro-batches to update the running statistics per mini-batch. Raises: TypeError: the module is not a :class:`nn.Sequential `. ValueError: invalid arguments Example:: Pipeline of two FC layers across GPUs 0 and 1. >>> # Need to initialize RPC framework first. >>> os.environ['MASTER_ADDR'] = 'localhost' >>> os.environ['MASTER_PORT'] = '29500' >>> torch.distributed.rpc.init_rpc('worker', rank=0, world_size=1) >>> >>> # Build pipe. >>> fc1 = nn.Linear(16, 8).cuda(0) >>> fc2 = nn.Linear(8, 4).cuda(1) >>> model = nn.Sequential(fc1, fc2) >>> model = Pipe(model, chunks=8) >>> input = torch.rand(16, 16).cuda(0) >>> output_rref = model(input) .. note:: You can wrap a :class:`Pipe` model with :class:`torch.nn.parallel.DistributedDataParallel` only when the checkpoint parameter of :class:`Pipe` is ``'never'``. .. note:: :class:`Pipe` only supports intra-node pipelining currently, but will be expanded to support inter-node pipelining in the future. The forward function returns an :class:`~torch.distributed.rpc.RRef` to allow for inter-node pipelining in the future, where the output might be on a remote host. For intra-node pipelinining you can use :meth:`~torch.distributed.rpc.RRef.local_value` to retrieve the output locally. .. warning:: :class:`Pipe` is experimental and subject to change. r except_lastFN)rchunks checkpointdeferred_batch_normrcsttjjstdt|}t|}|dkr>t d|dkrNt dt |t |||_ ||_ |rzt||}t|\|_|_t||j|jg|_t|j|_|}|j |j ddd|j }t|j|j||j||_dS)NzMPlease initialize RPC framework for Pipe using torch.distributed.rpc.init_rpcrz)number of chunks must be positive integer)alwaysrZneverz.)r-r&rNrrr__len__`sz Pipe.__len__)indexrc Cst|j}|dkr|ddd}|D]J}z||WStk rFYnXt|}|dkrb||7}q ||8}q tdS)z1Gets a layer in the underlying sequential module.rN)r& IndexErrorr#)r?rgr& partitionshiftrrr __getitem__ds  zPipe.__getitem__ccs|jD]}|EdHqdS)z;Iterates over children of the underlying sequential module.N)r&)r?rjrrr__iter__ys z Pipe.__iter__)r:rcCstdSr( MOVING_DENIED)r?r:rrrcudasz Pipe.cudacCstdSr(rnrNrrrr9szPipe.cpu)rLrMrcsRd|ksd|krt|rDt|dtjttfr2tt|drDttj||S)Nr:Ztensorr) rorr<r:rarbZ is_tensorrDrVrKrHrrrVs zPipe.tocs8|js2|jD]$|jfddt|jDq |jS)aEnsures that :class:`Pipe` caches CUDA streams for copy. It's worth to cache CUDA streams although PyTorch already manages a pool of pre-allocated CUDA streams, because it may reduce GPU memory fragementation when the number of micro-batches is small. csg|] }tqSr)r)r*_r:rr sz-Pipe._ensure_copy_streams..)rcr'rTr/r[rNrrrrrds "zPipe._ensure_copy_streamscGspt|jdkr|jdntd}tj|f||js@t|Stj|d|ji}|j |t |}t|S)a Processes a single input mini-batch through the pipe and returns an :class:`~torch.distributed.rpc.RRef` pointing to the output. :class:`Pipe` is a fairly transparent module wrapper. It doesn't modify the input and output signature of the underlying module. But there's type restriction. Input and output have to contain at least one tensor. This restriction is applied at partition boundaries too. The sequence of inputs are fed into the first stage of the pipeline as ``*inputs``. As a result the positional args for this function should match the positional args for the first stage of the pipeline. The same condition applies for output of one stage of the pipeline which is the input for the next stage. The input tensor is split into multiple micro-batches based on the ``chunks`` parameter used to initialize :class:`Pipe`. The batch size is assumed to be the first dimension of the tensor and if the batch size is less than ``chunks``, the number of micro-batches is equal to the batch size. Only tensors are split into multiple micro-batches, non-Tensor inputs are just replicated as-is in each micro-batch. For non-Tensor outputs in the last stage of the pipeline, they are aggregated as a ``List`` and returned the user. For example, if you have 2 micro-batches returning the integer 5, the user would receive the consolidated output of `[5, 5]` All the input tensors need to be on the same device as the first partition of the pipeline. If a tensor is wrapped with the :class:`NoChunk` wrapper, the tensor is not split across micro-batches and is replicated as-is similar to non-tensors. Args: inputs: input mini-batch Returns: :class:`~torch.distributed.rpc.RRef` to the output of the mini-batch Raises: TypeError: input doesn't contain at least one tensor rr9r[) r#r'r<r:rcheckrZscatterr[rerunZgather)r?r@Zfirst_partition_deviceZbatchesoutputrrrrAs-"  z Pipe.forward)rrZF)N)r6r7r8rBr rrarbboolrErfrOrlrrmrDevicerpr9rrVrrrdrrArQrrrHrrs&U1)9rB collectionsrtypingrrrrrrr r r r<r r Ztorch.distributed.rpcrZtorch.autogradZ torch.cudarZ batchnormrrerZ skip.layoutrZskip.skippablerstreamrr__all__r:rarbrxZDevicesZTensorsZTensorOrTensorsrOZ NamedModulesrrr%r4r$r5r=r>rCrUrYr rorrrrrsL ,          * &