mindspore 2.2.10__cp38-none-any.whl → 2.2.14__cp38-none-any.whl

mindspore/ops/function/array_func.py

@@ -268,7 +268,7 @@ def cat(tensors, axis=0):
 
     Returns:
         Tensor, the shape is :math:`(x_1, x_2, ..., \sum_{i=1}^Nx_{mi}, ..., x_R)`.
-            The data type is the same with `tensors`.
+        The data type is the same with `tensors`.
 
     Raises:
         TypeError: If `axis` is not an int.
@@ -660,7 +660,7 @@ def one_hot(indices, depth, on_value=1, off_value=0, axis=-1):
 
     Note:
         If the input indices is rank `N`, the output will have rank `N+1`. The new axis is created at dimension `axis`.
-        On Ascend, if `on_value` is Int64 dtype, `indices` must be Int64 dtype.
+        On Ascend, if `on_value` is int64 dtype, `indices` must be int64 dtype.
 
     Args:
         indices(Tensor): A tensor of indices. Tensor of shape :math:`(X_0, \ldots, X_n)`.
@@ -4228,6 +4228,7 @@ def space_to_batch_nd(input_x, block_size, paddings):
 
     Examples:
         >>> import numpy as np
+        >>> import mindspore
         >>> from mindspore import Tensor, ops
         >>> block_size = [2, 2]
         >>> paddings = [[0, 0], [0, 0]]
@@ -5395,8 +5396,8 @@ def masked_select(input, mask):
 
     Examples:
         >>> import numpy as np
-        >>> import mindspore.ops as ops
-        >>> from mindspore import Tensor
+        >>> import mindspore
+        >>> from mindspore import Tensor, ops
         >>> x = Tensor(np.array([1, 2, 3, 4]), mindspore.int64)
         >>> mask = Tensor(np.array([1, 0, 1, 0]), mindspore.bool_)
         >>> output = ops.masked_select(x, mask)
@@ -6520,7 +6521,7 @@ def topk(input, k, dim=None, largest=True, sorted=True):
 
     Args:
         input (Tensor): Input to be computed, data type must be float16, float32 or int32.
-        k (int): The number of top or bottom elements to be computed along the last dimension, constant input is needed.
+        k (int): The number of top or bottom elements to be computed along the last dimension.
         dim (int, optional): The dimension to sort along. Default: ``None`` .
         largest (bool, optional): If largest is ``False``  then the k smallest elements are returned.
             Default: ``True`` .

mindspore/ops/function/debug_func.py

@@ -51,7 +51,7 @@ def print_(*input_x):
 
     Examples:
         >>> import numpy as np
-        >>> from mindspore import Tensor
+        >>> from mindspore import Tensor, ops
         >>> x = Tensor(np.ones([2, 1]).astype(np.int32))
         >>> y = Tensor(np.ones([2, 2]).astype(np.int32))
         >>> result = ops.print_('Print Tensor x and Tensor y:', x, y)

mindspore/ops/function/linalg_func.py

@@ -59,6 +59,8 @@ def cond(A, p=None):
 
     Args:
         A (Tensor): Tensor of shape :math:`(*, n)` or :math:`(*, m, n)` where * is zero or more batch dimensions.
+            If `p` is one of Union[1, -1, inf, -inf, 'fro', 'nuc'], the function uses
+            :class:`mindspore.ops.MatrixInverse` , therefore, :math:`(*, m, n)` has to be square and ivertible.
         p (Union[int, float, inf, -inf, 'fro', 'nuc'], optional): norm's mode. Refer to the table above for
             behavior. Default: ``None``.
 
@@ -84,8 +86,8 @@ def cond(A, p=None):
     matrix_inverse = _get_cache_prim(P.MatrixInverse)(adjoint=False)
     if p is None:
         p = 2
-    norm_a = F.norm(A, p)
-    norm_inv_a = F.norm(matrix_inverse(A), p)
+    norm_a = F.matrix_norm(A, p)
+    norm_inv_a = F.matrix_norm(matrix_inverse(A), p)
     return norm_a * norm_inv_a
 
 
@@ -194,6 +196,8 @@ def geqrf(input):
         ``Ascend`` ``GPU`` ``CPU``
 
     Examples:
+        >>> from mindspore import Tensor, ops
+        >>> import numpy as np
         >>> input_x = Tensor(np.array([[-2.0, -1.0], [1.0, 2.0]]).astype(np.float32))
         >>> y, tau = ops.geqrf(input_x)
         >>> print(y)
@@ -266,6 +270,16 @@ def svd(input, full_matrices=False, compute_uv=True):
     return s
 
 
+def _check_pinv_shape(x):
+    if not isinstance(x, (Tensor, Tensor_)):
+        raise TypeError("The input x must be tensor")
+    if x.shape == ():
+        raise TypeError("For pinv, the 0-D input is not supported")
+    x_shape = F.shape(x)
+    if len(x_shape) < 2:
+        raise ValueError("input x should have 2 or more dimensions, " f"but got {len(x_shape)}.")
+
+
 def pinv(x, *, atol=None, rtol=None, hermitian=False):
     r"""
     Computes the (Moore-Penrose) pseudo-inverse of a matrix.
@@ -318,19 +332,15 @@ def pinv(x, *, atol=None, rtol=None, hermitian=False):
         ``CPU``
 
     Examples:
+        >>> import mindspore
+        >>> from mindspore import Tensor, ops
         >>> x = Tensor([[4., 0.], [0., 5.]], mindspore.float32)
         >>> output = ops.pinv(x)
         >>> print(output)
-        [[0.25  0. ]
-        [0.  0.2 ]]
+        [[0.25 0.  ]
+         [0.   0.2 ]]
     """
-    if not isinstance(x, (Tensor, Tensor_)):
-        raise TypeError("The input x must be tensor")
-    if x.shape == ():
-        raise TypeError("For pinv, the 0-D input is not supported")
-    x_shape = F.shape(x)
-    if len(x_shape) < 2:
-        raise ValueError("input x should have 2 or more dimensions, " f"but got {len(x_shape)}.")
+    _check_pinv_shape(x)
     x_dtype = _get_cache_prim(P.DType)()(x)
     _check_input_dtype("x", x_dtype, [mstype.float32, mstype.float64], "pinv")
     _check_attr_dtype("hermitian", hermitian, [bool], "pinv")

mindspore/ops/function/math_func.py

@@ -343,6 +343,7 @@ def add(input, other):
 
     Examples:
         >>> import numpy as np
+        >>> import mindspore
         >>> from mindspore import Tensor, ops
         >>> # case 1: x and y are both Tensor.
         >>> x = Tensor(np.array([1, 2, 3]).astype(np.float32))
@@ -12784,6 +12785,7 @@ def count_nonzero(x, axis=(), keep_dims=False, dtype=mstype.int32):
     Examples:
         >>> from mindspore import Tensor, ops
         >>> import numpy as np
+        >>> import mindspore
         >>> # case 1: each value specified.
         >>> x = Tensor(np.array([[0, 1, 0], [1, 1, 0]]).astype(np.float32))
         >>> nonzero_num = ops.count_nonzero(x=x, axis=[0, 1], keep_dims=True, dtype=mindspore.int32)
@@ -13392,6 +13394,7 @@ def batch_dot(x1, x2, axes=None):
         ``Ascend`` ``GPU`` ``CPU``
 
     Examples:
+        >>> import mindspore
         >>> from mindspore import Tensor, ops
         >>> import numpy as np
         >>> x1 = Tensor(np.ones(shape=[2, 2, 3]), mindspore.float32)

mindspore/ops/function/nn_func.py

@@ -2748,6 +2748,7 @@ def softshrink(x, lambd=0.5):
         ``Ascend`` ``GPU`` ``CPU``
 
     Examples:
+        >>> import mindspore
         >>> from mindspore import Tensor
         >>> from mindspore import ops
         >>> import numpy as np
@@ -3000,11 +3001,12 @@ def dense(input, weight, bias=None):
 
     Examples:
         >>> import numpy as np
+        >>> import mindspore
         >>> from mindspore import Tensor, ops
-        >>> input = mindspore.Tensor([[-1., 1., 2.], [-3., -3., 1.]], mindspore.float32)
-        >>> weight = mindspore.Tensor([[-2., -2., -2.], [0., -1., 0.]], mindspore.float32)
-        >>> bias = mindspore.Tensor([0., 1.], mindspore.float32)
-        >>> output = mindspore.ops.dense(input, weight, bias)
+        >>> input = Tensor([[-1., 1., 2.], [-3., -3., 1.]], mindspore.float32)
+        >>> weight = Tensor([[-2., -2., -2.], [0., -1., 0.]], mindspore.float32)
+        >>> bias = Tensor([0., 1.], mindspore.float32)
+        >>> output = ops.dense(input, weight, bias)
         >>> print(output)
         [[-4.  0.]
          [10.  4.]]
@@ -3992,8 +3994,8 @@ def l1_loss(input, target, reduction='mean'):
     Examples:
         >>> from mindspore import Tensor, ops
         >>> from mindspore import dtype as mstype
-        >>> x = ms.Tensor([[1, 2, 3], [4, 5, 6]], mstype.float32)
-        >>> target = ms.Tensor([[6, 5, 4], [3, 2, 1]], mstype.float32)
+        >>> x = Tensor([[1, 2, 3], [4, 5, 6]], mstype.float32)
+        >>> target = Tensor([[6, 5, 4], [3, 2, 1]], mstype.float32)
         >>> output = ops.l1_loss(x, target, reduction="mean")
         >>> print(output)
         3.0
@@ -5342,7 +5344,7 @@ def conv1d(input, weight, bias=None, stride=1, pad_mode="valid", padding=0, dila
         >>> from mindspore import Tensor, ops
         >>> x = Tensor(np.arange(64).reshape((4, 4, 4)), mindspore.float32)
         >>> weight = Tensor(np.arange(8).reshape((2, 2, 2)), mindspore.float32)
-        >>> bias = Tensor([-0.12345, 2.7683], ms.float32)
+        >>> bias = Tensor([-0.12345, 2.7683], mindspore.float32)
         >>> output = ops.conv1d(x, weight, pad_mode='pad', padding=(1,), bias=bias, groups=2)
         >>> print(output.shape)
         (4, 2, 5)
@@ -7444,8 +7446,8 @@ def max_pool2d(x, kernel_size, stride=None, padding=0, dilation=1, return_indice
     return out
 
 
-def prompt_flash_attention(query, key, value, padding_mask, attn_mask, actual_seq_lengths,
-                           actual_seq_lengths_kv, deq_scale1, quant_scale1,
+def prompt_flash_attention(query, key, value, attn_mask, actual_seq_lengths,
+                           actual_seq_lengths_kv, pse_shift, deq_scale1, quant_scale1,
                            deq_scale2, quant_scale2, quant_offset2, num_heads, scale_value=1.0, pre_tokens=2147483547,
                            next_tokens=0, input_layout='BSH',
                            num_key_value_heads=0, sparse_mode=0):
@@ -7468,11 +7470,11 @@ def prompt_flash_attention(query, key, value, padding_mask, attn_mask, actual_se
           Input tensor of shape :math:`(B, S, H)` / `(B, N, S, D)`.
         value (Tensor) - The value tensor with data type of float16 or float32.
           Input tensor of shape :math:`(B, S, H)` / `(B, N, S, D)`.
-        padding_mask (Tensor) - The padding mask tensor with data type of float16 or float32
         attn_mask (Tensor) - The attention mask tensor with data type of float16 or float32.
           For each element, 0 indicates retention and 1 indicates discard. Input tensor of shape :math:`(B, 1, S, S)`.
         actual_seq_lengths (list[int]): Describe actual sequence length of each input with data type of int.
         actual_seq_lengths_kv (list[int]): Describe actual sequence length of each input with data type of int.
+        pse_shift (Tensor) - The position encoding tensor with data type of float16 or float32.
         dep_scale1 (Tensor)
         quant_scale1 (Tensor)
         deq_scale2 (Tensor)
@@ -7516,7 +7518,7 @@ def prompt_flash_attention(query, key, value, padding_mask, attn_mask, actual_se
 
     pfa = _get_cache_prim(NN_OPS.PromptFlashAttention)(num_heads, scale_value, pre_tokens, next_tokens, input_layout,
                                                        num_key_value_heads, sparse_mode)
-    return pfa(query, key, value, padding_mask, attn_mask, actual_seq_lengths, actual_seq_lengths_kv, deq_scale1,
+    return pfa(query, key, value, attn_mask, actual_seq_lengths, actual_seq_lengths_kv, pse_shift, deq_scale1,
                quant_scale1, deq_scale2, quant_scale2, quant_offset2)
 
 

mindspore/ops/function/parameter_func.py

@@ -99,6 +99,7 @@ def assign_sub(variable, value):
         >>> 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)
@@ -149,6 +150,7 @@ def assign_add(variable, value):
         >>> 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_add(variable, value)

mindspore/ops/function/sparse_unary_func.py

@@ -28,7 +28,7 @@ def csr_cos(x: CSRTensor) -> CSRTensor:
         out_i = \cos(x_i)
 
     .. warning::
-        Currently support data types float16 and float32. If use Float64, there may be a problem of missing precision.
+        Currently support data types float16 and float32. If use float64, there may be a problem of missing precision.
 
     Args:
         x (CSRTensor): Input CSRTensor.
@@ -69,7 +69,7 @@ def coo_cos(x: COOTensor) -> COOTensor:
         out_i = \cos(x_i)
 
     .. warning::
-        If use Float64, there may be a problem of missing precision.
+        If use float64, there may be a problem of missing precision.
 
     Args:
         x (COOTensor): Input COOTensor.

mindspore/ops/function/vmap_func.py

@@ -81,6 +81,7 @@ def vmap(fn, in_axes=0, out_axes=0):
         ``Ascend`` ``GPU`` ``CPU``
 
     Examples:
+        >>> import numpy as np
         >>> from mindspore import Tensor
         >>> from mindspore import vmap
         >>> def test_vmap(x, y, z):                                              # ([a],[a],[a]) -> [a]

mindspore/ops/operations/__init__.py

@@ -118,7 +118,7 @@ from .nn_ops import (LSTM, SGD, Adam, AdamWeightDecay, FusedSparseAdam, FusedSpa
                      Dilation2D, DataFormatVecPermute, DeformableOffsets, Dense, FractionalAvgPool,
                      FractionalMaxPool, FractionalMaxPool3DWithFixedKsize, FractionalMaxPoolWithFixedKsize,
                      GridSampler2D, TripletMarginLoss, UpsampleNearest3D, UpsampleTrilinear3D, PadV3, ChannelShuffle,
-                     GLU, MaxUnpool3D, Pdist, RmsNorm)
+                     GLU, MaxUnpool3D, Pdist, RmsNorm, PagedAttention, PagedAttentionMask, ReshapeAndCache)
 from .other_ops import (Assign, IOU, BoundingBoxDecode, BoundingBoxEncode,
                         ConfusionMatrix, UpdateState, Load, StopGradient,
                         CheckValid, Partial, Depend, Push, Pull, PyExecute, PyFunc, _DynamicLossScale,
@@ -692,7 +692,10 @@ __all__ = [
     "MaskedScatter",
     "Ormqr",
     "RandpermV2",
-    "RmsNorm"
+    "RmsNorm",
+    "PagedAttention",
+    "PagedAttentionMask",
+    "ReshapeAndCache"
 ]
 
 __custom__ = [

mindspore/ops/operations/_embedding_cache_ops.py

@@ -65,7 +65,7 @@ class SubAndFilter(PrimitiveWithCheck):
 
     Inputs:
         - **input_x** (Tensor) - Input tensor.
-        - **max_num** (Int) - The max value of element that after sub `offset`.
+        - **max_num** (int) - The max value of element that after sub `offset`.
         - **offset** (int) - Specifies the offset value of this `input_x`.
 
     Outputs:

mindspore/ops/operations/_grad_ops.py

@@ -3861,10 +3861,10 @@ class FlashAttentionScoreGrad(Primitive):
         validator.check_value_type('input_layout', input_layout, [str], self.name)
         if input_layout not in ["BSH", "BNSD"]:
             raise ValueError(f"Attribute 'input_layout' must be either 'BSH' or 'BNSD', but got {input_layout}")
-        self.init_prim_io_names(inputs=['query', 'key', 'value', 'attn_mask', 'attention_in', 'softmax_max',
-                                        'softmax_sum', 'dy', 'drop_mask', 'real_shift', "padding_mask", 'softmax_out',
+        self.init_prim_io_names(inputs=['query', 'key', 'value', 'dy', 'pse_shift', 'drop_mask', "padding_mask",
+                                        'attn_mask', 'softmax_max', 'softmax_sum', 'softmax_out', 'attention_in',
                                         'prefix'],
-                                outputs=['dq', 'dk', 'dv'])
+                                outputs=['dq', 'dk', 'dv', 'dpse'])
 
 
 class RmsNormGrad(Primitive):
@@ -3882,4 +3882,3 @@ class RmsNormGrad(Primitive):
         """Initialize RmsNormGrad."""
         self.init_prim_io_names(inputs=["dy", "x", "rstd", "gamma"],
                                 outputs=["dx", "dgamma"])
-        

mindspore/ops/operations/_inner_ops.py

@@ -16,6 +16,7 @@
 """Inner operators."""
 from types import FunctionType, MethodType
 from collections.abc import Iterable
+import os
 import numpy as np
 
 from mindspore.common import Tensor
@@ -1008,7 +1009,7 @@ class Centralization(PrimitiveWithInfer):
 
     Inputs:
         - **input_x** (Tensor) - The input tensor. The data type mast be float16 or float32.
-        - **axis** (Union[Int, Tuple(Int), List(Int)]) - The dimensions to reduce. Default: (), reduce all dimensions.
+        - **axis** (Union[int, Tuple(int), List(int)]) - The dimensions to reduce. Default: (), reduce all dimensions.
           Only constant value is allowed. Must be in the range [-rank(input_x), rank(input_x)).
 
     Outputs:
@@ -2858,6 +2859,60 @@ class DecoderKVCache(Primitive):
         self.add_prim_attr('side_effect_mem', True)
 
 
+class _MirrorSilentCheck(PrimitiveWithInfer):
+    """
+    The operator _MirrorSilentCheck implements accuracy-sensitive detection on the tensor input in backpropagator.
+    Call _MirrorSilentCheck in method __call__ of derived class to implement accuracy-sensitive detection.
+
+    Inputs:
+        - **input** (Tensor) : The tensor used for detection.
+          Its data type must be mindspore.float16, mindspore.float32 or mindspore.bfloat16.
+        - **pre_val** (Parameter(Tensor)) : Support parameter in accuracy-sensitive detection.
+          Please only generated by method generate_params() of ASDBase.
+        - **min_val** (Parameter(Tensor)) : Support parameter in accuracy-sensitive detection.
+          Please only generated by method generate_params() of ASDBase.
+        - **max_val** (Parameter(Tensor)) : Support parameter in accuracy-sensitive detection.
+          Please only generated by method generate_params() of ASDBase.
+        - **cnt** (Parameter(Tensor)) : Support parameter in accuracy-sensitive detection.
+          Please only generated by method generate_params() of ASDBase.
+          After each invocation of _MirrorSilentCheck, increment the value of cnt by one.
+
+    Outputs:
+        - **output** (Tensor) - Same shape, type and value as `input`.
+    """
+    @prim_attr_register
+    def __init__(self, min_steps=8):
+        upper_thresh, sigma_thresh = self.get_thresh()
+        self.min_steps = min_steps
+        self.thresh_l1 = upper_thresh[0]
+        self.coeff_l1 = sigma_thresh[0]
+        self.thresh_l2 = upper_thresh[1]
+        self.coeff_l2 = sigma_thresh[1]
+        self.add_prim_attr('side_effect_mem', True)
+
+    def parse_thresh(self, env_var_name, default_value, min_value):
+        env_var = os.environ.get(env_var_name, default=default_value)
+        thresh = [value.strip() for value in env_var.split(",")]
+        if len(thresh) != 2 or not all(value.isdigit() for value in thresh):
+            thresh = default_value.split(",")
+        thresh = [float(max(int(value), min_value)) for value in thresh]
+        if thresh[0] <= thresh[1]:
+            thresh = [float(value) for value in default_value.split(",")]
+
+        return thresh
+
+    def get_thresh(self):
+        upper_thresh = self.parse_thresh("NPU_ASD_UPPER_THRESH", "1000000,10000", 3)
+        sigma_thresh = self.parse_thresh("NPU_ASD_SIGMA_THRESH", "100000,5000", 3)
+        return upper_thresh, sigma_thresh
+
+    def infer_shape(self, x_shape, pre_shape, min_shape, max_shape, n_step, loss_scale_shape):
+        return x_shape
+
+    def infer_dtype(self, x_dtype, pre_dtype, min_dtype, max_dtype, n_dtype, loss_scale_dtype):
+        return x_dtype
+
+
 class PromptKVCache(Primitive):
     r"""
     The PromptKVCache is used for prefill the KVCache of transformer network.

mindspore/ops/operations/_quant_ops.py

@@ -579,7 +579,7 @@ class FakeQuantWithMinMaxVars(PrimitiveWithInfer):
             range is [1, 2^num_bits-1]. Default: ``False``.
 
     Inputs:
-        - **x** (Tensor) - Float32 tensor representing the shape of the output tensor.
+        - **x** (Tensor) - float32 tensor representing the shape of the output tensor.
         - **min** (Tensor) - Value of the min range of the input data x.
         - **max** (Tensor) - Value of the max range of the input data x.
 
@@ -638,7 +638,7 @@ class FakeQuantWithMinMaxVarsGradient(PrimitiveWithInfer):
 
     Inputs:
         - **gradients** (Tensor) - The gradient above the FakeQuantWithMinMaxVars.
-        - **x** (Tensor) - Float32 tensor representing the shape of the output tensor.
+        - **x** (Tensor) - float32 tensor representing the shape of the output tensor.
         - **min** (Tensor) - Value of the min range of the input data x.
         - **max** (Tensor) - Value of the max range of the input data x.
 
@@ -702,7 +702,7 @@ class FakeQuantWithMinMaxVarsPerChannel(PrimitiveWithInfer):
             range is [1, 2^num_bits-1]. Default: ``False``.
 
     Inputs:
-        - **x** (Tensor) - Float32 tensor representing the shape of the output tensor.
+        - **x** (Tensor) - float32 tensor representing the shape of the output tensor.
         - **min** (Tensor) - Value of the min range of the input data x.
         - **max** (Tensor) - Value of the max range of the input data x.
 
@@ -754,7 +754,7 @@ class FakeQuantWithMinMaxVarsPerChannelGradient(PrimitiveWithInfer):
 
     Inputs:
         - **gradients** (Tensor) - The gradient above the FakeQuantWithMinMaxVars.
-        - **x** (Tensor) - Float32 tensor representing the shape of the output tensor.
+        - **x** (Tensor) - float32 tensor representing the shape of the output tensor.
         - **min** (Tensor) - Value of the min range of the input data x.
         - **max** (Tensor) - Value of the max range of the input data x.
 

mindspore/ops/operations/_rl_inner_ops.py

@@ -341,7 +341,7 @@ class LSTMV2(Primitive):
         - **h** (Tensor) - Tensor of shape (num_directions * `num_layers`, batch_size, `hidden_size`).
         - **c** (Tensor) - Tensor of shape (num_directions * `num_layers`, batch_size, `hidden_size`).
         - **w** (Tensor) - The input tensor which states for weights.
-        - **seq_lengths** (Tensor) - The Tensor[Int32] of shape (batch_size, ),
+        - **seq_lengths** (Tensor) - The Tensor[int32] of shape (batch_size, ),
           indicates the seq_length of each batch dim.
 
     Outputs:

mindspore/ops/operations/array_ops.py

@@ -227,7 +227,11 @@ class ExpandDims(PrimitiveWithCheck):
     def infer_value(self, input_x, axis):
         value = None
         if input_x is not None and axis is not None:
-            value = Tensor(np.expand_dims(input_x.asnumpy(), axis))
+            dtype = input_x.dtype
+            if input_x.dtype == mstype.bfloat16:
+                cpu_cast = Cast().set_device("CPU")
+                input_x = cpu_cast(input_x, mstype.float32)
+            value = Tensor(np.expand_dims(input_x.asnumpy(), axis), dtype)
         return value
 
 
@@ -375,6 +379,9 @@ class Cast(PrimitiveWithCheck):
         if isinstance(x, (int, float)):
             value = Tensor(np.array(x).astype(np_dst_type), dtype=dst_type)
         else:
+            if x.dtype == mstype.bfloat16:
+                cpu_cast = Cast().set_device("CPU")
+                x = cpu_cast(x, mstype.float32)
             value = Tensor(x.asnumpy().astype(np_dst_type), dtype=dst_type)
         return value
 
@@ -5837,6 +5844,9 @@ class SpaceToBatchND(Primitive):
         ``Ascend`` ``GPU`` ``CPU``
 
     Examples:
+        >>> import mindspore
+        >>> from mindspore import Tensor, ops
+        >>> import numpy as np
         >>> block_shape = [2, 2]
         >>> paddings = [[0, 0], [0, 0]]
         >>> space_to_batch_nd = ops.SpaceToBatchND(block_shape, paddings)
@@ -6279,8 +6289,8 @@ class Sort(Primitive):
     Sorts the elements of the input tensor along the given dimension in the specified order.
 
     .. warning::
-        Currently, the data types of Float16 is well supported.
-        Using Float32 might cause loss of accuracy.
+        Currently, the data types of float16, uint8, int8, int16, int32, int64 are well supported.
+        If use float32, it may cause loss of accuracy.
 
     Args:
         axis (int, optional): The dimension to sort along. Default: ``-1``, means the last dimension.
@@ -8719,7 +8729,8 @@ class TopK(Primitive):
           - GPU: float16, float32.
           - CPU: all numeric types.
 
-        - **k** (int) - The number of top elements to be computed along the last dimension, constant input is needed.
+        - **k** (Union(Tensor, int)) - The number of top elements to be computed along the last dimension.
+          If `k` is a Tensor, the supported dtype is int32 and it should be 0-D or 1-D with shape :math:`(1, )` .
 
     Outputs:
         A tuple consisting of `values` and `indexes`.

mindspore/ops/operations/custom_ops.py

@@ -470,7 +470,7 @@ class Custom(ops.PrimitiveWithInfer):
     op_path_in_cache = []  # Save paths for op functions created in the cached.
     custom_aot_warning = True  # Flag to enable warnings about custom aot path white list
 
-    def __init__(self, func, out_shape=None, out_dtype=None, func_type=HYBRID_TYPE, bprop=None, reg_info=None):
+    def __init__(self, func, out_shape=None, out_dtype=None, func_type="hybrid", bprop=None, reg_info=None):
         super().__init__("Custom")
 
         self.supported_targets = [ASCEND, GPU, CPU]

mindspore/ops/operations/debug_ops.py

@@ -472,7 +472,7 @@ class Print(Primitive):
 
     Examples:
         >>> import numpy as np
-        >>> from mindspore import Tensor, nn
+        >>> from mindspore import Tensor, nn, ops
         >>> class PrintDemo(nn.Cell):
         ...     def __init__(self):
         ...         super(PrintDemo, self).__init__()

mindspore/ops/operations/image_ops.py

@@ -388,7 +388,7 @@ class NonMaxSuppressionV3(Primitive):
           single score associated with each box (i.e., each row of the `boxes` Tensor).
           It is required that the number of scores in `scores` must be equal to the number of boxes in `boxes`.
           The supported data type is float32.
-        - **max_output_size** (Union[Tensor, Number.Int]) - A scalar integer Tensor representing the maximum
+        - **max_output_size** (Union[Tensor, Number.int]) - A scalar integer Tensor representing the maximum
           number of boxes to be selected by non max suppression. The supported data type is int32.
         - **iou_threshold** (Union[Tensor, Number.Float]) - A scalar float Tensor represents the threshold
           used for determining if the intersection over union (IOU) between boxes is too high.
@@ -459,7 +459,7 @@ class NonMaxSuppressionWithOverlaps(Primitive):
           single score associated with each box (i.e., each row of the `boxes` Tensor).
           It is required that the number of scores in `scores` must be equal to the number of boxes in `boxes`.
           The supported data type is float32.
-        - **max_output_size** (Union[Tensor, Number.Int]) - A scalar integer Tensor representing the maximum
+        - **max_output_size** (Union[Tensor, Number.int]) - A scalar integer Tensor representing the maximum
           number of boxes to be selected by non max suppression, and max_output_size must be equal to or greater
           than 0.
           Types allowed:int32.
@@ -816,7 +816,7 @@ class ResizeBicubic(Primitive):
     Examples:
         >>> import mindspore
         >>> import numpy as np
-        >>> from mindspore import Tensor, ops
+        >>> from mindspore import Tensor, ops, nn
         >>> class NetResizeBicubic(nn.Cell):
         ...     def __init__(self):
         ...         super(NetResizeBicubic, self).__init__()

mindspore/ops/operations/inner_ops.py

@@ -642,6 +642,55 @@ class FusedAdaFactorWithGlobalNorm(FusedAdaFactor):
         return param_type
 
 
+class GenerateEodMask(Primitive):
+    r"""
+    Given the input `inputs_ids`, if found eod_token_id, the output position and attention mask matrix will be reset.
+    This means the `position_id` will start counting from 0, and the corresponding mask matrix will be filled with 0.
+
+    Args:
+        eod_token_id (int) - In the NLP scenario, this value corresponds to the id of
+            the symbol of 'EodOfDocument' in the vocabulary.
+
+    Inputs:
+      - **inputs_ids** (Tensor) - token id, a 2-D Tensor with shape :math:`(batch\_size, seq\_length)`.
+
+    Outputs:
+      - **position_id** (Tensor) - position id matrix with same shape and type as original `inputs_ids`.
+      - **attention_mask** (Tensor) - attention mask matrix with type
+            float16 and shape :math:`(batch\_size, seq\_length)`.
+
+    Supported Platforms:
+        ``Ascend``
+
+    Examples:
+        >>> op = ops.GenerateEodMask(eod_token_id=0)
+        >>> position, mask = op(Tensor([[1, 0, 3], [1, 0, 0]], dtype=mindspore.int32))
+        >>> print(position)
+        [[0 1 0] [0 0 1]]
+        >>> print(mask)
+        [[[ 1. 0. 0.]
+          [1. 1. 0.]
+          [0. 0. 1.]]
+         [[1. 0. 0.]
+          [0. 1. 0.]
+          [0. 1. 1.]]]
+
+    Raises:
+        - **TypeError** - If `eod_token_id` is not int.
+        - **TypeError** - If `inputs_ids` is not int.
+        - **ValueError** - If `inputs_ids` is not a 2-D Tensor.
+    """
+    @prim_attr_register
+    def __init__(self, n_pos, eod_token_id, n_step, n_error_mode='specific'):
+        """Initialize GenerateEodMask"""
+        validator.check_value_type("eod_token_id", eod_token_id, [int], self.name)
+        validator.check_value_type("n_pos", n_pos, [int], self.name)
+        validator.check_value_type("n_step", n_step, [list], self.name)
+        validator.check_value_type("n_error_mode", n_error_mode, [str], self.name)
+        self.init_prim_io_names(inputs=['inputs_ids'],
+                                outputs=['position_ids'])
+
+
 class ScaleGrad(PrimitiveWithInfer):
     """
     Scale the input grad according to the loss scale.