o Zh} @sddlZddlmZddlmZmZddlZddlmZddlm Z m Z ddl m Z ddl mZdgZ  dddd ejd ee d eeeeeeffd eed ejf ddZdS)Nfnmatch)OptionalUnion)_mesh_resources DeviceMesh)_validate_tp_mesh_dim) ParallelStyleparallelize_module src_data_rankmodule device_meshparallelize_planr returnc stjd|p t}t||durtd|St|t r)||_ | ||St|t r}| D]H\}}|d}t|dkrEtd|rz|dtfdd|}|D]\}} |rod|} t| || |i|d qYt| |||d qY|sGq2|Std t|d ) a Apply Tensor Parallelism in PyTorch by parallelizing modules or sub-modules based on a user-specified plan. We parallelize module or sub_modules based on a parallelize_plan. The parallelize_plan contains :class:`ParallelStyle`, which indicates how user wants the module or sub_module to be parallelized. User can also specify different parallel style per module fully qualified name (FQN). Note that ``parallelize_module`` only accepts a 1-D :class:`DeviceMesh`, if you have a 2-D or N-D :class:`DeviceMesh`, slice the DeviceMesh to a 1-D sub DeviceMesh first then pass to this API(i.e. ``device_mesh["tp"]``) Args: module (:class:`nn.Module`): Module to be parallelized. device_mesh (:class:`DeviceMesh`, optional): Object which describes the mesh topology of devices for the DTensor. If not specified, the call must be under a DeviceMesh context. parallelize_plan (Union[:class:`ParallelStyle`, Dict[str, :class:`ParallelStyle`]], optional): The plan used to parallelize the module. It can be either a :class:`ParallelStyle` object which contains how we prepare input/output for Tensor Parallelism or it can be a dict of module FQN and its corresponding :class:`ParallelStyle` object. If not specified, the call will do nothing at the moment. Keyword args: src_data_rank (int, optional): the rank of the source data for the logical/global tensor, it is used by :meth:`distribute_tensor` to scatter/broadcast the shards/replicas to other ranks. By default, we use ``group_rank=0`` on each DeviceMesh dimension as the source data to preserve the single-device semantic. If passing ``None`` explicitly, :meth:`parallelize_module` simply uses its local data instead of trying to preserve the single-device semantic via scatter/broadcast. Default: 0 Return: A :class:`nn.Module` object parallelized. Example:: >>> # xdoctest: +SKIP("distributed") >>> from torch.distributed.tensor.parallel import parallelize_module, ColwiseParallel >>> from torch.distributed.device_mesh import init_device_mesh >>> >>> # Define the module. >>> m = Model(...) >>> tp_mesh = init_device_mesh("cuda", (8,)) >>> m = parallelize_module(m, tp_mesh, {"w1": ColwiseParallel(), "w2": RowwiseParallel()}) >>> .. note:: For complex module architecture like Attention, MLP layers, we recommend composing different ParallelStyles together (i.e. ``ColwiseParallel`` and ``RowwiseParallel``) and pass as a parallelize_plan, to achieves the desired sharding computation. z4torch.distributed.tensor.parallel.parallelize_moduleNzNo parallelize_plan is provided and auto-parallel is not supported at the moment, so this parallelize_module call will do nothing..rz9Expect module path to be non-empty, but got empty string!cst|dS)Nrr)tZatomT/var/www/auris/lib/python3.10/site-packages/torch/distributed/tensor/parallel/api.pydsz$parallelize_module..r zLExpect Union[ParallelStyle, Dict[str, ParallelStyle]] for parallelize_plan, z found!)torchZ_CZ_log_api_usage_oncerZget_current_meshrwarningswarn isinstancer r _applydictitemssplitlen ValueErrorpopfilterZnamed_childrenjoinr TypeErrortype) r rrr module_pathZparallelize_styleZ path_splitsZmatched_children_ submoduleZ leaf_pathrrrr sb 7         )NN)rrtypingrrrZtorch.nnnnZtorch.distributed.device_meshrrZ(torch.distributed.tensor.parallel._utilsrZ'torch.distributed.tensor.parallel.styler __all__Modulerstrintr rrrrs0