mindspore 2.4.10__cp310-cp310-win_amd64.whl → 2.6.0rc1__cp310-cp310-win_amd64.whl

mindspore/ops/function/other_func.py

@@ -15,9 +15,11 @@
 """Defines other operators with functional form."""
 from mindspore.ops import operations as P
 from mindspore.ops.auto_generate import rotary_position_embedding
+from mindspore.ops.auto_generate.gen_ops_prim import moe_init_routing_v2_op
 
 partial_ = P.Partial()
 depend_ = P.Depend()
+move_to_ = P.MoveTo()
 
 
 def partial(func, *args):
@@ -107,9 +109,165 @@ def depend(value, expr):
     return depend_(value, expr)
 
 
+def move_to(input, to="CPU", blocking=True):  # pylint: disable=redefined-outer-name
+    """
+    Copy tensor to target device synchronously or asynchronously, default synchronously.
+
+    .. note::
+        This interface currently only supports Graph mode with jit_level of O0 or O1.
+
+    Args:
+        input (Union[Tensor, list[int], tuple[int]]): The input tensor. When the input is list and tuple, it will be
+                                                      converted to tensor before copying.
+        to (str, optional): Specify the target device, with optional values of ``"Ascend"`` and ``"CPU"``.
+                            Default ``"CPU"`` .
+        blocking (bool, optional): Whether use synchronous copying. Default ``True``.
+
+    Returns:
+        A new tensor on target device.
+
+    Supported Platforms:
+        ``Ascend`` ``CPU``
+
+    Examples:
+        >>> import mindspore
+        >>> from mindspore import nn, ops, Tensor
+        >>> mindspore.set_context(mode=mindspore.GRAPH_MODE)
+        >>> class MoveToNet(nn.Cell):
+        ...     def __init__(self):
+        ...         super().__init__()
+        ...
+        ...     def construct(self, x):
+        ...         cpu_x = ops.move_to(x, "CPU")
+        ...         npu_x = ops.move_to(cpu_x, "Ascend")
+        ...         return npu_x
+        ...
+        >>> net = MoveToNet()
+        >>> x = Tensor([1, 2, 3], mindspore.int64)
+        >>> y = net(x)
+        >>> print(y)
+        [1 2 3]
+    """
+    return move_to_(input, to, blocking)
+
+
+def moe_init_routing_v2(x, expert_idx, active_num, expert_capacity, expert_num, drop_pad_mode,
+                        expert_tokens_count_or_cumsum_flag, expert_tokens_before_capacity_flag):
+    """
+    Performs routing on the computation result of MoeGatingTopKSoftmaxV2.
+
+    Notes:
+        - NUM_ROWS: The number of rows in 'x', which represents the number of original input tokens.
+        - H: The number of cols in 'x', which denotes for the hiddens of input tokens.
+        - K: The number of experts corresponding to each row of features in the output of MoeGatingTopKSoftmaxV2.
+        - Currently, MoeInitRoutingV2 does not support mutable inputs.
+
+    Args:
+        x (Tensor): A 2D tensor, which contains the input feature tokens. The shape of the tensor
+        is (NUM_ROWS, H). Supported dtypes: Float16, BFloat16, Float32.
+        expert_idx (Tensor): A 2D tensor, representing K experts corresponding to each row of features
+        in the output of MoeGatingTopKSoftmaxV2. The shape of the tensor is (NUM_ROWS, K).
+        Supported dtype: int32. In the Drop/Pad scenario or when the output 'expert_tokens_count_or_cumsum'
+        is required in the non-Drop/Pad scenario, the value range of this tensor is [0, 'expert_num' - 1].
+        In other scenarios, the value must be greater than or equal to 0.
+        active_num (int64): Indicates whether the scenario is Active, this value works only
+        when 'drop_pad_mode' = 0. The value must be greater than or equal to 0 where 0 is the Dropless scenario
+        and others represent the Active scenario.
+        expert_capacity (int64): The number of tokens that each expert can process.
+        The value must be greater than or equal to 0. In the Drop/Pad scenario, the value range is (0, NUM_ROWS].
+        expert_num (int64): The number of experts. The value must be greater than or equal to 0.
+        In the Drop/Pad scenario or when 'expert_tokens_count_or_cumsum_flag' > 0, the value must be greater than 0.
+        drop_pad_mode (int64): Indicates whether the scenario is Drop/Pad. The value must be 0 or 1:
+
+          - 0: non-Drop/Pad scenario.
+          - 1: Drop/Pad scenario.
+        expert_tokens_count_or_cumsum_flag (int64): A flag which controls whether the
+        output 'expert_tokens_count_or_cumsum' is required. The value must be 0, 1 or 2:
+
+          - 0: The output 'expert_tokens_count_or_cumsum' is not required.
+          - 1: The output 'expert_tokens_count_or_cumsum' is required,
+            which represents the accumulated number of tokens processed by each expert.
+          - 2: The output 'expert_tokens_count_or_cumsum' is required,
+            which represents the number of tokens processed by each expert.
+        expert_tokens_before_capacity_flag (bool): A flag which controls whether the
+        output 'expert_tokens_before_capacity' is required.
+
+          - False: The output 'expert_tokens_before_capacity' is not required.
+          - True: The output 'expert_tokens_before_capacity' is required, which represents the
+            number of tokens processed by each expert before the drop.
+
+    Returns:
+        A tuple of tensors.
+        expanded_x (Tensor): A 2D/3D tensor which indicates features extended based on 'expert_idx'.
+        The shape of the tensor depends on scenarios:
+
+          - Dropless scenario: The shape is (NUM_ROWS * K, H).
+          - Active scenario: The shape is (min('active_num', NUM_ROWS * K), H).
+          - Drop/Pad scenario: The shape is ('expert_num', 'expert_capacity', H).
+        Supported dtypes: Float16, BFloat16, Float32.
+        expanded_row_idx (Tensor): A 1D tensor which represents the mapping between 'expanded_x' and 'x'.
+        The shape of the tensor is (NUM_ROWS * K). Supported dtype: int32.
+        expert_tokens_count_or_cumsum (Tensor): A 1D tensor which indicates the statistics on the
+        number of tokens processed by each expert and the accumulated value.
+        The value of the tensor is valid only in the non-Drop/Pad scenario which is controlled by the
+        'expert_tokens_count_or_cumsum_flag'.
+        The value of this tensor is dirty data from the memory when it is not required.
+        The shape of the tensor is ('expert_num'). Supported dtype: int32.
+        expert_tokens_before_capacity (Tensor): A 1D tensor which indicates the statistics on the
+        number of tokens processed by each expert before the drop.
+        The value of the tensor is valid only in the Drop/Pad scenario which is controlled by the
+        'expert_tokens_before_capacity_flag'.
+        The value of this tensor is dirty data from the memory when it is not required.
+        The shape of the tensor is ('expert_num'). Supported dtype: int32.
+
+    Raises:
+        TypeError: If the data type of input Tensor does not match the description in args.
+        ShapeError: If the shape of input Tensor does not match the description in args.
+        ValueError: If the value of the inputs do not match the description in args.
+
+    Supported Platforms:
+        ``Ascend``
+
+    Examples:
+        >>> import mindspore as ms
+        >>> from mindspore import Tensor
+        >>> from mindspore import ops
+        >>> import numpy as np
+        >>> x = Tensor(np.array([[0.1, 0.2, 0.3], [0.2, 0.7, 0.8], [0.3, 0.3, 0.5]]), ms.float16)
+        >>> expert_idx = Tensor(np.array([[0, 1, 1], [2, 1, 1], [0, 0, 0]]), ms.int32)
+        >>> active_num = 3
+        >>> expert_capacity = 2
+        >>> expert_num = 3
+        >>> drop_pad_mode = 1
+        >>> out1, out2 = ops.moe_init_routing_v2(x, expert_idx, active_num, expert_capacity,
+        expert_num, drop_pad_mode, 0, False)
+        >>> print(out1)
+        [[[0.1  0.2  0.3]
+          [0.3  0.3  0.5]]
+         [[0.1  0.2  0.3]
+          [0.1  0.2  0.3]]
+         [[0.2  0.7  0.8]
+          [0.   0.   0. ]]]
+        >>> print(out2)
+        [ 0  2  3  4 -1 -1  1 -1 -1 ]
+    """
+    expanded_x, expanded_row_idx, \
+    expert_tokens_count_or_cumsum, \
+    expert_tokens_before_capacity = moe_init_routing_v2_op(x, expert_idx, active_num, expert_capacity,
+                                                           expert_num, drop_pad_mode,
+                                                           expert_tokens_count_or_cumsum_flag,
+                                                           expert_tokens_before_capacity_flag)
+    if drop_pad_mode == 1 and expert_tokens_before_capacity_flag:
+        return expanded_x, expanded_row_idx, expert_tokens_before_capacity
+    if drop_pad_mode == 0 and expert_tokens_count_or_cumsum_flag != 0:
+        return expanded_x, expanded_row_idx, expert_tokens_count_or_cumsum
+    return expanded_x, expanded_row_idx
+
 __all__ = [
     'depend',
     'partial',
-    'rotary_position_embedding'
+    'rotary_position_embedding',
+    'move_to',
+    'moe_init_routing_v2'
 ]
 __all__.sort()

mindspore/ops/function/parameter_func.py

@@ -17,105 +17,39 @@
 
 from mindspore.ops import operations as P
 from mindspore.ops._primitive_cache import _get_cache_prim
-from mindspore.ops.auto_generate import assign, assign_add
-
-
-assign_sub_ = P.AssignSub()
-assign_add_ = P.AssignAdd()
-
-
-def assign_sub(variable, value):
-    """
-    Updates a `Parameter` by subtracting a value from it.
-
-    Args of `variable` and `value` comply with the implicit type conversion rules to make the data types consistent.
-    If they have different data types, the lower priority data type will be converted to
-    the relatively highest priority data type.
-    If `value` is a number, the number is automatically converted to Tensor,
-    and the data type is consistent with the Tensor data type involved in the operation.
-
-    Note:
-        Since `variable` is a data type Parameter, the data type cannot be changed,
-        so only the type of `value` is allowed to be promoted to the type of `variable`.
-        And the conversion type supported by different devices will be different,
-        it is recommended to use the same data type when using this operator.
-
-    Args:
-        variable (Parameter): The `Parameter`.
-            :math:`(N,*)` where :math:`*` means, any number of additional dimensions.
-        value (Tensor): The value to be subtracted from the `variable`.
-            It must have the same shape as `variable`.
-            it is recommended to use the same data type when using this operator.
-
-    Returns:
-        Tensor, has the same data type and shape as original `variable`.
-
-    Raises:
-        TypeError: If `value` is neither Number nor Tensor.
-        RuntimeError: If the data type of `x`, `y` conversion of Parameter is required
-                      when data type conversion of Parameter is not supported.
-
-    Supported Platforms:
-        ``Ascend`` ``GPU`` ``CPU``
-
-    Examples:
-        >>> import mindspore
-        >>> import numpy as np
-        >>> from mindspore import Tensor, ops
-        >>> from mindspore.common.initializer import initializer
-        >>> variable = mindspore.Parameter(initializer(1, [1], mindspore.int32), name="global_step")
-        >>> value = Tensor(np.ones([1]).astype(np.int32) * 100)
-        >>> ops.assign_sub(variable, value)
-        >>> print(variable.asnumpy())
-        [-99]
-    """
-    return assign_sub_(variable, value)
+from mindspore.ops.auto_generate import assign, assign_add, assign_sub
 
 
 def index_add(x, indices, y, axis, use_lock=True, check_index_bound=True):
     """
-    Adds tensor `y` to specified axis and indices of Parameter `x`. The axis should be in [0,  len(x.dim) - 1],
-    and indices should be in [0, x.shape[axis] - 1] at the axis dimension.
+    Add the elements of input `y` into input `x` along the given axis and indices.
+
+    .. note::
+        - `indices` is a one-dimensional tensor, and :math:`indices.shape[0] = y.shape[axis]` .
+        - The value range of the elements in `indices` is :math:`[0, x.shape[axis] - 1]` .
 
     Args:
-        x (Parameter): The input Parameter to add to.
-        indices (Tensor): Add the value of `x` and `y` along the dimension of the `axis` according to the
-            specified index value, with data type int32.
-            The `indices` must be 1D with the same size as the size of `y` in the `axis` dimension. The values
-            of `indices` should be in [0, b), where the b is the size of `x` in the `axis` dimension.
-        y (Tensor): The input tensor with the value to add. Must have same data type as `x`.
-            The shape must be the same as `x` except the `axis` th dimension.
-        axis (int): The dimension along which to index.
+        x (Union[Parameter, Tensor]): The input parameter or tensor.
+        indices (Tensor): The specified indices.
+        y (Tensor): The input tensor to add to `x`.
+        axis (int): The specified axis.
         use_lock (bool, optional): Whether to enable a lock to protect the updating process of variable tensors.
-            If ``True`` , when updating the value of `x`, this process will be protected by a lock by using atomic
-            operation.
-            If ``False`` , the result may be unpredictable. Default: ``True`` .
-        check_index_bound (bool, optional): If ``True``, check index boundary. If ``False`` ,
-            don't check index boundary. Default: ``True`` .
+           Default ``True`` .
+        check_index_bound (bool, optional): Whether to check index boundary. Default ``True`` .
 
     Returns:
-        Tensor, has the same shape and dtype as `x`.
-
-    Raises:
-        TypeError: If `x` is not a Parameter.
-        TypeError: If neither `indices` nor `y` is a Tensor.
-        ValueError: If axis is out of `x` rank's range.
-        ValueError: If `x` rank is not the same as `y` rank.
-        ValueError: If shape of `indices` is not 1D or size of `indices` is not equal to dimension of y[axis].
-        ValueError: If `y`'s shape is not the same as `x` except the `axis` th dimension.
+        Tensor
 
     Supported Platforms:
         ``Ascend`` ``GPU`` ``CPU``
 
     Examples:
-        >>> import numpy as np
         >>> import mindspore
-        >>> from mindspore import Tensor, Parameter
-        >>> from mindspore import ops
-        >>> x = Parameter(Tensor(np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]]), mindspore.float32), name="name_x")
-        >>> indices = Tensor(np.array([0, 2]), mindspore.int32)
-        >>> y = Tensor(np.array([[0.5, 1.0], [1.0, 1.5], [2.0, 2.5]]), mindspore.float32)
-        >>> output = ops.index_add(x, indices, y, 1)
+        >>> x = mindspore.Parameter(mindspore.tensor([[1, 2, 3], [4, 5, 6], [7, 8, 9]], mindspore.float32),
+        ...                         name="name_x")
+        >>> indices = mindspore.tensor([0, 2], mindspore.int32)
+        >>> y = mindspore.tensor([[0.5, 1.0], [1.0, 1.5], [2.0, 2.5]], mindspore.float32)
+        >>> output = mindspore.ops.index_add(x, indices, y, 1)
         >>> print(output)
         [[ 1.5  2.   4. ]
          [ 5.   5.   7.5]