o * i@sddlmZddlmZmZer(ddlmZddlmZddl m Z ddl m Z ddl ZddlZddlmZddlmZdd lmZd d lmZmZ   d"d#ddZ  d$d%d d!ZdS)&) annotations) TYPE_CHECKINGAny)Sequence)Tensor)task)GroupN) framework)stream)convert_object_to_tensorconvert_tensor_to_objectTtensorr tensor_listSequence[Tensor] | Nonesrcintgroup Group | Nonesync_opboolreturn task | NonecCst|||||S)ak Scatter a tensor to all participators. As shown below, one process is started with a GPU and the source of the scatter is GPU0. Through scatter operator, the data in GPU0 will be sent to all GPUs averagely. .. image:: https://githubraw.cdn.bcebos.com/PaddlePaddle/docs/develop/docs/api/paddle/distributed/img/scatter.png :width: 800 :alt: scatter :align: center Args: tensor (Tensor): The output Tensor. Its data type should be float16, float32, float64, int32, int64, int8, uint8, bool or bfloat16. tensor_list (list|tuple): A list/tuple of Tensors to scatter. Every element in the list must be a Tensor whose data type should be float16, float32, float64, int32, int64, int8, uint8, bool or bfloat16. Default value is None. src (int): The source rank id. Default value is 0. group (Group, optional): The group instance return by new_group or None for global default group. sync_op (bool, optional): Whether this op is a sync op. The default value is True. Returns: None. Examples: .. code-block:: python >>> # doctest: +REQUIRES(env: DISTRIBUTED) >>> import paddle >>> import paddle.distributed as dist >>> dist.init_parallel_env() >>> if dist.get_rank() == 0: ... data1 = paddle.to_tensor([7, 8, 9]) ... data2 = paddle.to_tensor([10, 11, 12]) ... dist.scatter(data1, src=1) >>> else: ... data1 = paddle.to_tensor([1, 2, 3]) ... data2 = paddle.to_tensor([4, 5, 6]) ... dist.scatter(data1, tensor_list=[data1, data2], src=1) >>> print(data1, data2) >>> # [1, 2, 3] [10, 11, 12] (2 GPUs, out for rank 0) >>> # [4, 5, 6] [4, 5, 6] (2 GPUs, out for rank 1) )r scatter)rrrrrrt/home/app/PaddleOCR-VL-test/.venv_paddleocr/lib/python3.10/site-packages/paddle/distributed/communication/scatter.pyr's1rout_object_list list[Any]in_object_listlist[Any] | NoneNonecCs,tsJdt}g}g}||kr.|D]}t|\}} |||| qt|} ntjgdd} t | |t | } g} |D]} | }t|| g}t|}| |qEtj| gdd}t|||krm| nd||tjgdd}t|||kr|nd||||t|| dS)a Scatter picklable objects from the source to all others. Similar to scatter(), but python object can be passed in. Args: out_object_list (list): The list of objects to store the scattered objects. in_object_list (list): The list of objects to scatter. Only objects on the src rank will be scattered. src (int): The source rank in global view. group (Group): The group instance return by new_group or None for global default group. Returns: None. Warning: This API only supports the dygraph mode. Examples: .. code-block:: python >>> # doctest: +REQUIRES(env: DISTRIBUTED) >>> import paddle.distributed as dist >>> dist.init_parallel_env() >>> out_object_list = [] # type: ignore >>> if dist.get_rank() == 0: ... in_object_list = [{'foo': [1, 2, 3]}, {'foo': [4, 5, 6]}] >>> else: ... in_object_list = [{'bar': [1, 2, 3]}, {'bar': [4, 5, 6]}] >>> dist.scatter_object_list(out_object_list, in_object_list, src=1) >>> print(out_object_list) >>> # [{'bar': [1, 2, 3]}] (2 GPUs, out for rank 0) >>> # [{'bar': [4, 5, 6]}] (2 GPUs, out for rank 1) z6scatter_object_list doesn't support static graph mode.Zint64)ZdtypeZuint8N)r Zin_dynamic_modedistZget_rankr appendmaxpaddleemptyr broadcastritemnumpynpresizeZ to_tensorrclearr )rrrrZrankZin_obj_tensorsZ in_obj_sizesobjZ obj_tensorZobj_sizeZmax_obj_size_tensorZ max_obj_sizeZin_tensor_listrZ numpy_dataZ in_tensorZ out_tensorZout_tensor_sizerrrscatter_object_list[s: '         r-)NrNT) rrrrrrrrrrrr)NrN) rrrrrrrrrr ) __future__rtypingrrcollections.abcrr$rZpaddle.base.corerZ&paddle.distributed.communication.grouprr(r)Zpaddle.distributed distributedr!r Z paddle.distributed.communicationr Zserialization_utilsr r rr-rrrrs,         6