mindspore 2.3.0__cp310-cp310-win_amd64.whl → 2.4.1__cp310-cp310-win_amd64.whl

mindspore/ops/operations/array_ops.py

@@ -39,11 +39,12 @@ from ..auto_generate import (ExpandDims, Reshape, TensorShape, Transpose, Gather
                              OnesLike, ZerosLike, Argmax, ArgMaxExt,
                              ReverseV2, Diag, Eye, ScatterNd, ResizeNearestNeighborV2,
                              GatherNd, GatherD, Range, MaskedFill, RightShift, NonZero,
-                             ResizeNearestNeighbor, Identity, Split, CumSum, CumProd,
+                             ResizeNearestNeighbor, Identity, Split, CumSum, CumProd, MaskedSelect,
                              Cummax, Cummin, Argmin, Concat, UnsortedSegmentSum, ScalarToTensor,
                              Triu, BroadcastTo, StridedSlice, Select, TopkExt, SearchSorted)
 from .manually_defined import Rank, Shape, Tile, Cast, Ones, Zeros
 from ..auto_generate import ArgMaxWithValue, ArgMinWithValue
+from ..auto_generate import TensorScatterElements as TensorScatterElementsExt
 
 class _ScatterOp(PrimitiveWithInfer):
     """
@@ -769,41 +770,15 @@ class Padding(Primitive):
 
 class UniqueWithPad(Primitive):
     """
-    Returns unique elements and relative indexes in 1-D tensor, filled with padding num.
-
-    The basic function is the same as the Unique operator, but the UniqueWithPad operator adds a Pad function.
-    The returned tuple(`y`, `idx`) after the input Tensor `x` is processed by the unique operator,
-    in which the shapes of `y` and `idx` are mostly not equal. Therefore, in order to solve the above situation,
-    the UniqueWithPad operator will fill the `y` Tensor with the `pad_num` specified by the user
-    to make it have the same shape as the Tensor `idx`.
-
-    Refer to :func:`mindspore.ops.unique_with_pad` for more details.
-
-    Inputs:
-        - **x** (Tensor) - The tensor need to be unique. Must be 1-D vector with types: int32, int64.
-        - **pad_num** (int) - Pad num. The data type is an int.
-
-    Outputs:
-        tuple(Tensor), tuple of 2 tensors, `y` and `idx`.
-
-        - y (Tensor) - The unique elements filled with pad_num, the shape and data type same as `x`.
-        - idx (Tensor) - The index of each value of `x` in the unique output `y`, the shape and data type same as `x`.
+    'ops.UniqueWithPad' is deprecated from version 2.4 and will be removed in a future version.
+    Please use the :func:`mindspore.ops.unique` combined with :func:`mindspore.ops.pad` to realize
+    the same function.
 
     Supported Platforms:
-        ``Ascend`` ``GPU`` ``CPU``
-
-    Examples:
-        >>> import mindspore
-        >>> import numpy as np
-        >>> from mindspore import Tensor, ops
-        >>> x = Tensor(np.array([1, 1, 2, 2, 3, 3, 4, 5]), mindspore.int32)
-        >>> pad_num = 8
-        >>> output = ops.UniqueWithPad()(x, pad_num)
-        >>> print(output)
-        (Tensor(shape=[8], dtype=Int32, value= [1, 2, 3, 4, 5, 8, 8, 8]),
-         Tensor(shape=[8], dtype=Int32, value= [0, 0, 1, 1, 2, 2, 3, 4]))
+        Deprecated
     """
 
+    @deprecated("2.4", "ops.unique and ops.pad", False)
     @prim_attr_register
     def __init__(self):
         """init UniqueWithPad"""
@@ -819,7 +794,7 @@ class Size(Primitive):
 
     Inputs:
         - **input_x** (Tensor) - Input parameters, the shape of tensor is :math:`(x_1, x_2, ..., x_R)`. The data type is
-          `number <https://www.mindspore.cn/docs/en/master/api_python/mindspore.html#mindspore.dtype>`_.
+          `number <https://www.mindspore.cn/docs/en/master/api_python/mindspore/mindspore.dtype.html>`_.
 
     Outputs:
         int. A scalar representing the elements' size of `input_x`, tensor is the number of elements
@@ -2112,60 +2087,6 @@ class Rint(Primitive):
         self.init_prim_io_names(inputs=['x'], outputs=['output'])
 
 
-class StridedSliceV2(Primitive):
-    r"""
-    StridedSliceV2 will be deprecated by StridedSlice in the future.
-    Extracts a strided slice of a tensor.
-    Refer to class StridedSlice for more details.
-
-    Args:
-        begin_mask (int): Starting index of the slice. Default: ``0`` .
-        end_mask (int): Ending index of the slice. Default: ``0`` .
-        ellipsis_mask (int): An int mask. Default: ``0`` .
-        new_axis_mask (int): An int mask. Default: ``0`` .
-        shrink_axis_mask (int): An int mask. Default: ``0`` .
-
-    Inputs:
-        - **input_x** (Tensor) - The input Tensor.
-        - **begin** (tuple[int]) - A tuple which represents the location where to start. Only
-          constant value is allowed.
-        - **end** (tuple[int]) - A tuple or which represents the maximum location where to end.
-          Only constant value is allowed.
-        - **strides** (tuple[int]) - A tuple which represents the stride is continuously added
-          before reaching the maximum location. Only constant value is allowed.
-
-    Outputs:
-        Tensor, The output is explained by following example.
-
-    Raises:
-        TypeError: If `begin_mask`, `end_mask`, `ellipsis_mask`, `new_axis_mask` or `shrink_axis_mask` is not an int.
-        TypeError: If `begin`, `end` or `strides` is not a tuple.
-        ValueError: If `begin_mask`, `end_mask`, `ellipsis_mask`, `new_axis_mask` or `shrink_axis_mask` is less than 0.
-
-    Supported Platforms:
-        ``Ascend`` ``CPU``
-
-    Examples:
-        >>> input_x = Tensor([[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]],
-        ...                   [[5, 5, 5], [6, 6, 6]]], mindspore.float32)
-        >>> strided_slice_v2 = ops.StridedSliceV2()
-        >>> output = strided_slice_v2(input_x, (1, 0, 2), (3, 1, 3), (1, 1, 1))
-        >>> print(output)
-        [[[3.]]
-         [[5.]]]
-    """
-
-    @prim_attr_register
-    def __init__(self,
-                 begin_mask=0,
-                 end_mask=0,
-                 ellipsis_mask=0,
-                 new_axis_mask=0,
-                 shrink_axis_mask=0):
-        """Initialize StridedSliceV2"""
-        self.init_prim_io_names(inputs=['x', 'begin', 'end', 'strides'], outputs=['output'])
-
-
 class DiagPart(PrimitiveWithCheck):
     r"""
 
@@ -4356,53 +4277,6 @@ class MaskedScatter(Primitive):
         self.init_prim_io_names(inputs=['x', 'mask', 'updates'], outputs=['y'])
 
 
-class MaskedSelect(PrimitiveWithCheck):
-    """
-    Returns a new 1-D Tensor which indexes the `x` tensor according to the boolean `mask`.
-    The shapes of the `mask` tensor and the `x` tensor don't need to match, but they must be broadcastable.
-
-    Inputs:
-        - **x** (Tensor) - Input Tensor of any dimension.
-        - **mask** (Tensor[bool]) - Boolean mask Tensor, has the same shape as `x`.
-
-    Outputs:
-        A 1-D Tensor, with the same type as x.
-
-    Raises:
-        TypeError: If `x` or `mask` is not a Tensor.
-        TypeError: If dtype of `mask` is not bool.
-
-    Supported Platforms:
-        ``Ascend`` ``GPU`` ``CPU``
-
-    Examples:
-        >>> import mindspore
-        >>> import numpy as np
-        >>> from mindspore import Tensor, ops
-        >>> x = Tensor(np.array([1, 2, 3, 4]), mindspore.int32)
-        >>> mask = Tensor(np.array([1, 0, 1, 0]), mindspore.bool_)
-        >>> output = ops.MaskedSelect()(x, mask)
-        >>> print(output)
-        [1 3]
-        >>> x = Tensor(2.1, mindspore.float32)
-        >>> mask = Tensor(True, mindspore.bool_)
-        >>> output = ops.MaskedSelect()(x, mask)
-        >>> print(output)
-        [2.1]
-    """
-
-    @prim_attr_register
-    def __init__(self):
-        self.init_prim_io_names(inputs=['x', 'mask'], outputs=['output'])
-
-    def check_shape(self, x_shape, mask_shape):
-        get_broadcast_shape(x_shape, mask_shape, self.name, arg_name1="x", arg_name2="mask")
-
-    def check_dtype(self, x_dtype, mask_dtype):
-        validator.check_tensor_dtype_valid('mask', mask_dtype, [mstype.bool_], self.name)
-        validator.check_tensor_dtype_valid('x', x_dtype, (mstype.bool_,) + mstype.number_type, self.name)
-
-
 class _TensorScatterOp(PrimitiveWithInfer):
     """
     Defines TensorScatter Base Operators
@@ -4962,7 +4836,7 @@ class SplitV(Primitive):
         self.init_prim_io_names(inputs=['input_x'], outputs=['output'])
 
 
-class TensorScatterElements(Primitive):
+class TensorScatterElements(TensorScatterElementsExt):
     """
     Write all elements in `updates` to the index specified by `indices` in `input_x` according to the reduction
     operation specified by `reduction`.
@@ -4977,6 +4851,9 @@ class TensorScatterElements(Primitive):
     .. warning::
         This is an experimental API that is subject to change or deletion.
 
+    Note:
+        The backward is supported only for the case `updates.shape == indices.shape`.
+
     Args:
         axis (int, optional): Specify which axis to do scatter operation. Default: ``0`` .
         reduction (str, optional): Which reduction operation to scatter, default is ``"none"`` . Other option: "add".
@@ -4986,7 +4863,7 @@ class TensorScatterElements(Primitive):
         - **indices** (Tensor) - The index of `input_x` to do scatter operation whose data type must be int32 or
           int64. It has the same rank as `data`. And accepted range is [-s, s) where s is the size along axis.
         - **updates** (Tensor) - The tensor doing the scatter operation with `data`,
-          it has the same type as `data` and the same shape as `indices`.
+          it has the same type as `data`.
 
     Outputs:
         Tensor, has the same shape and type as `data`.
@@ -5021,16 +4898,7 @@ class TensorScatterElements(Primitive):
 
     @prim_attr_register
     def __init__(self, axis=0, reduction="none"):
-        """Initialize TensorScatterElements"""
-        validator.check_value_type("axis", axis, [int], self.name)
-        validator.check_value_type("reduction", reduction, [str], self.name)
-        validator.check_string(reduction, ["none", "add"], "reduction", self.name)
-        self.init_prim_io_names(inputs=['data', 'indices', 'updates'], outputs=['y'])
-        target = context.get_context("device_target")
-        if reduction != 'none' and target.lower() == "ascend":
-            raise ValueError(f"For '{self.name}', "
-                             f"Currently Ascend device_target only support `reduction`='none', "
-                             f"but got {reduction}")
+        super().__init__(axis, reduce=reduction)
 
 
 class ExtractVolumePatches(Primitive):

mindspore/ops/operations/comm_ops.py

@@ -54,7 +54,7 @@ class ReduceOp:
             For Ascend/GPU/CPU devices, it is recommended to use the msrun startup method
             without any third-party or configuration file dependencies.
             Please see the `msrun start up
-            <https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/msrun_launcher.html>`_
+            <https://www.mindspore.cn/docs/zh-CN/master/model_train/parallel/msrun_launcher.html>`_
             for more details.
 
             This example should be run with multiple devices.
@@ -141,7 +141,7 @@ class AllReduce(Primitive):
             For Ascend/GPU/CPU devices, it is recommended to use the msrun startup method
             without any third-party or configuration file dependencies.
             Please see the `msrun start up
-            <https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/msrun_launcher.html>`_
+            <https://www.mindspore.cn/docs/zh-CN/master/model_train/parallel/msrun_launcher.html>`_
             for more details.
 
             This example should be run with 2 devices.
@@ -178,14 +178,15 @@ class AllReduce(Primitive):
     @prim_attr_register
     def __init__(self, op=ReduceOp.SUM, group=GlobalComm.WORLD_COMM_GROUP):
         """Initialize AllReduce."""
+        self.group = _get_group(group)
         if not isinstance(op, type(ReduceOp.SUM)):
             raise TypeError(f"For '{self.name}', the 'op' must be str, but got {type(op).__name__}.")
-        if not isinstance(_get_group(group), str):
+        if not isinstance(self.group, str):
             raise TypeError(f"For '{self.name}', the 'group' must be str, "
-                            f"but got {type(_get_group(group)).__name__}.")
-        check_hcom_group_valid(group, prim_name=self.name)
+                            f"but got {type(self.group).__name__}.")
+        check_hcom_group_valid(self.group, prim_name=self.name)
         self.op = op
-        self.add_prim_attr('group', _get_group(group))
+        self.add_prim_attr('group', self.group)
         self.add_prim_attr('fusion', 0)
         self.add_prim_attr('index', 0)
         self.add_prim_attr('no_eliminate', True)
@@ -230,7 +231,7 @@ class Reduce(PrimitiveWithInfer):
             For Ascend/GPU/CPU devices, it is recommended to use the msrun startup method without any third-party
             or configuration file dependencies.
             Please see the `msrun start up
-            <https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/msrun_launcher.html>`_
+            <https://www.mindspore.cn/docs/zh-CN/master/model_train/parallel/msrun_launcher.html>`_
             for more details.
 
             This example should be run with 4 devices.
@@ -314,7 +315,7 @@ class AllGather(PrimitiveWithInfer):
             For Ascend/GPU/CPU devices, it is recommended to use the msrun startup method
             without any third-party or configuration file dependencies.
             Please see the `msrun start up
-            <https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/msrun_launcher.html>`_
+            <https://www.mindspore.cn/docs/zh-CN/master/model_train/parallel/msrun_launcher.html>`_
             for more details.
 
             This example should be run with 2 devices.
@@ -354,12 +355,13 @@ class AllGather(PrimitiveWithInfer):
     @prim_attr_register
     def __init__(self, group=GlobalComm.WORLD_COMM_GROUP):
         """Initialize AllGather."""
-        validator.check_value_type('group', _get_group(group), (str,), self.name)
-        self.rank = get_rank(_get_group(group))
-        self.rank_size = get_group_size(_get_group(group))
+        self.group = _get_group(group)
+        validator.check_value_type('group', self.group, (str,), self.name)
+        self.rank = get_rank(self.group)
+        self.rank_size = get_group_size(self.group)
         validator.check('rank', self.rank, 'rank_size', self.rank_size, validator.LT, self.name)
         self.add_prim_attr('rank_size', self.rank_size)
-        self.add_prim_attr('group', _get_group(group))
+        self.add_prim_attr('group', self.group)
         self.add_prim_attr('fusion', 0)
         self.add_prim_attr('mean_flag', False)
         self.add_prim_attr('no_eliminate', True)
@@ -375,25 +377,6 @@ class AllGather(PrimitiveWithInfer):
         return x_dtype
 
 
-class AShardIdentity(PrimitiveWithInfer):
-    """
-    Auto parallel virtual operator. Identity operator only for shard function.
-    Do nothing in terms of infer_shape, infer_dtype, and the tensor.
-
-    It is only for internal use of parallel modules and cannot be called by users.
-    """
-
-    @prim_attr_register
-    def __init__(self):
-        pass
-
-    def infer_shape(self, x_shape):
-        return x_shape
-
-    def infer_dtype(self, x_dtype):
-        return x_dtype
-
-
 class _MiniStepAllGather(PrimitiveWithInfer):
     """
     Auto parallel virtual operator. Do nothing in forward, do reducescatter in backward in mini-step. It is only for
@@ -555,7 +538,7 @@ class ReduceScatter(Primitive):
             For Ascend/GPU/CPU devices, it is recommended to use the msrun startup method
             without any third-party or configuration file dependencies.
             Please see the `msrun start up
-            <https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/msrun_launcher.html>`_
+            <https://www.mindspore.cn/docs/zh-CN/master/model_train/parallel/msrun_launcher.html>`_
             for more details.
 
             This example should be run with 2 devices.
@@ -597,11 +580,12 @@ class ReduceScatter(Primitive):
     def __init__(self, op=ReduceOp.SUM, group=GlobalComm.WORLD_COMM_GROUP):
         """Initialize ReduceScatter."""
         validator.check_value_type('op', op, (type(ReduceOp.SUM),), self.name)
-        validator.check_value_type('group', _get_group(group), (str,), self.name)
+        self.group = _get_group(group)
+        validator.check_value_type('group', self.group, (str,), self.name)
         self.op = op
-        self.rank_size = get_group_size(_get_group(group))
+        self.rank_size = get_group_size(self.group)
         self.add_prim_attr('rank_size', self.rank_size)
-        self.add_prim_attr('group', _get_group(group))
+        self.add_prim_attr('group', self.group)
         self.add_prim_attr('fusion', 0)
         self.add_prim_attr('no_eliminate', True)
 
@@ -692,7 +676,7 @@ class Broadcast(PrimitiveWithInfer):
             For Ascend/GPU/CPU devices, it is recommended to use the msrun startup method
             without any third-party or configuration file dependencies.
             Please see the `msrun start up
-            <https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/msrun_launcher.html>`_
+            <https://www.mindspore.cn/docs/zh-CN/master/model_train/parallel/msrun_launcher.html>`_
             for more details.
 
             This example should be run with 2 devices.
@@ -922,7 +906,7 @@ class AlltoAll(PrimitiveWithInfer):
             For Ascend/GPU/CPU devices, it is recommended to use the msrun startup method
             without any third-party or configuration file dependencies.
             Please see the `msrun start up
-            <https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/msrun_launcher.html>`_
+            <https://www.mindspore.cn/docs/zh-CN/master/model_train/parallel/msrun_launcher.html>`_
             for more details.
 
             This example should be run with 8 devices.
@@ -1041,7 +1025,7 @@ class NeighborExchangeV2(Primitive):
             For Ascend/GPU/CPU devices, it is recommended to use the msrun startup method
             without any third-party or configuration file dependencies.
             Please see the `msrun start up
-            <https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/msrun_launcher.html>`_
+            <https://www.mindspore.cn/docs/zh-CN/master/model_train/parallel/msrun_launcher.html>`_
             for more details.
 
             This example should be run with 2 devices.
@@ -1158,7 +1142,7 @@ class CollectiveScatter(Primitive):
             For Ascend/GPU/CPU devices, it is recommended to use the msrun startup method
             without any third-party or configuration file dependencies.
             Please see the `msrun start up
-            <https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/msrun_launcher.html>`_
+            <https://www.mindspore.cn/docs/zh-CN/master/model_train/parallel/msrun_launcher.html>`_
             for more details.
 
             This example should be run with 2 devices.
@@ -1243,7 +1227,7 @@ class CollectiveGather(Primitive):
             For Ascend/GPU/CPU devices, it is recommended to use the msrun startup method
             without any third-party or configuration file dependencies.
             Please see the `msrun start up
-            <https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/msrun_launcher.html>`_
+            <https://www.mindspore.cn/docs/zh-CN/master/model_train/parallel/msrun_launcher.html>`_
             for more details.
 
             This example should be run with 4 devices.
@@ -1308,8 +1292,6 @@ class Barrier(PrimitiveWithInfer):
     Raises:
         TypeError: If `group` is not a str.
         RuntimeError: If backend is invalid, or distributed initialization fails.
-        ValueError: If the local rank id of the calling process in the group
-                    is larger than the group's rank size.
 
     Supported Platforms:
         ``Ascend``
@@ -1321,7 +1303,7 @@ class Barrier(PrimitiveWithInfer):
             For Ascend/GPU/CPU devices, it is recommended to use the msrun startup method
             without any third-party or configuration file dependencies.
             Please see the `msrun start up
-            <https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/msrun_launcher.html>`_
+            <https://www.mindspore.cn/docs/zh-CN/master/model_train/parallel/msrun_launcher.html>`_
             for more details.
 
             This example should be run with 2 devices.
@@ -1395,7 +1377,7 @@ class Send(PrimitiveWithInfer):
             For Ascend/GPU/CPU devices, it is recommended to use the msrun startup method
             without any third-party or configuration file dependencies.
             Please see the `msrun start up
-            <https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/msrun_launcher.html>`_
+            <https://www.mindspore.cn/docs/zh-CN/master/model_train/parallel/msrun_launcher.html>`_
             for more details.
 
             This example should be run with 2 devices.
@@ -1431,7 +1413,7 @@ class Send(PrimitiveWithInfer):
     def __init__(self, sr_tag, dest_rank, group=GlobalComm.WORLD_COMM_GROUP, group_back=GlobalComm.WORLD_COMM_GROUP):
         self.rank = dest_rank
         self.sr_tag = sr_tag
-        self.group = group
+        self.group = _get_group(group)
         self.add_prim_attr("no_eliminate", True)
 
     def infer_shape(self, x_shape):
@@ -1479,7 +1461,7 @@ class Receive(PrimitiveWithInfer):
             For Ascend/GPU/CPU devices, it is recommended to use the msrun startup method
             without any third-party or configuration file dependencies.
             Please see the `msrun start up
-            <https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/msrun_launcher.html>`_
+            <https://www.mindspore.cn/docs/zh-CN/master/model_train/parallel/msrun_launcher.html>`_
             for more details.
 
             This example should be run with 2 devices.
@@ -1517,7 +1499,7 @@ class Receive(PrimitiveWithInfer):
         self.tag = sr_tag
         self.shape = shape
         self.dtype = dtype
-        self.group = group
+        self.group = _get_group(group)
         self.add_prim_attr("no_eliminate", True)
         valid_type = [mstype.float16, mstype.float32, mstype.float64, mstype.bfloat16,
                       mstype.int8, mstype.int16, mstype.int32, mstype.int64,
@@ -1695,9 +1677,32 @@ class _VirtualAssignAdd(PrimitiveWithInfer):
 
     def infer_dtype(self, x_dtype, y_dtype):
         return x_dtype
+
+
 virtual_assign_add = _VirtualAssignAdd()
 
 
+class _VirtualAssignKvCache(PrimitiveWithInfer):
+    """
+    Auto parallel virtual operator. Do nothing in forward, do Assign kv cache in backward. It is only for
+    internal use of parallel modules and cannot be called by users.
+
+    """
+
+    @prim_attr_register
+    def __init__(self):
+        """Initialize _VirtualAssignAdd."""
+        self.add_prim_attr('order_enforce_skip', True)
+        self.add_prim_attr('side_effect_backprop_mem', True)
+
+    def infer_shape(self, x_shape, y_shape, kv_equal_shape):
+        return x_shape
+
+    def infer_dtype(self, x_dtype, y_dtype, kv_equal_dtype):
+        return x_dtype
+virtual_assign_kv_cache = _VirtualAssignKvCache()
+
+
 class _VirtualAccuGrad(PrimitiveWithInfer):
     """
     Auto parallel virtual operator. Do nothing in forward, return y in backward. It is only for
@@ -1834,7 +1839,7 @@ class BatchISendIRecv(PrimitiveWithInfer):
             without any third-party or configuration file dependencies.
 
             Please see the `msrun start up
-            <https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/msrun_launcher.html>`_
+            <https://www.mindspore.cn/docs/zh-CN/master/model_train/parallel/msrun_launcher.html>`_
             for more details.
 
             This example should be run with 2 devices.
@@ -1924,6 +1929,7 @@ class AlltoAllV(PrimitiveWithInfer):
         recv_numel_list(Union[tuple[int], list[int]]): split numel to gather from different remote rank.
         group (str): The communication group to work on. Default: ``GlobalComm.WORLD_COMM_GROUP``, which
           means ``"hccl_world_group"`` in Ascend, and ``"nccl_world_group"`` in GPU.
+        TODO:
 
     Inputs:
         - **input_x** (Tensor) - flatten tensor to scatter. The shape of tensor is :math:`(x_1)`.
@@ -1946,7 +1952,7 @@ class AlltoAllV(PrimitiveWithInfer):
             without any third-party or configuration file dependencies.
 
             Please see the `msrun start up
-            <https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/msrun_launcher.html>`_
+            <https://www.mindspore.cn/docs/zh-CN/master/model_train/parallel/msrun_launcher.html>`_
             for more details.
 
             This example should be run with 2 devices.
@@ -1986,11 +1992,15 @@ class AlltoAllV(PrimitiveWithInfer):
     """
 
     @prim_attr_register
-    def __init__(self, send_numel_list, recv_numel_list, group=None):
+    def __init__(self, send_numel_list, recv_numel_list, group=None, split_sizes_empty=False):
         validator.check_value_type("send_numel_list", send_numel_list, [tuple, list], self.name)
         validator.check_value_type("recv_numel_list", recv_numel_list, [tuple, list], self.name)
-        if group is None:
-            group = GlobalComm.WORLD_COMM_GROUP
-        self.add_prim_attr('group', group)
+        self.group = GlobalComm.WORLD_COMM_GROUP if group is None else _get_group(group)
+        self.send_numel_list = send_numel_list
+        self.recv_numel_list = recv_numel_list
+        self.split_sizes_empty = split_sizes_empty
+        self.rank_size = get_group_size(self.group)
+
+        self.add_prim_attr('group', self.group)
         self.add_prim_attr('send_numel_list', send_numel_list)
         self.add_prim_attr('recv_numel_list', recv_numel_list)