mindspore 2.2.0__cp37-cp37m-manylinux1_x86_64.whl → 2.2.11__cp37-cp37m-manylinux1_x86_64.whl

mindspore/nn/layer/flash_attention.py

@@ -21,9 +21,7 @@ import mindspore.common.dtype as mstype
 from mindspore.common.tensor import Tensor
 from mindspore import ops
 from mindspore.nn.cell import Cell
-from mindspore.ops._op_impl._custom_op.flash_attention.flash_attention_impl import get_flash_attention
 from mindspore.ops.operations.nn_ops import FlashAttentionScore
-from mindspore._c_expression import MSContext
 
 __all__ = ['FlashAttention']
 
@@ -46,25 +44,25 @@ class FlashAttention(Cell):
             Default 65536.
         next_block_num(int): A integer to define the number of blocks to look behind for local block sparse attention.
             Default 65536.
-        tiling_stgy_name(str): A str to define tiling strategy of flash attention.
         dp(int): data parallel.
             Default 1.
         mp(int): model parallel.
             Default 1.
-        high_precision(bool): This mode has higher precision but some performance loss.
+        high_precision(bool): This mode has higher precision but some performance loss. Only take effect on Ascend910A.
             Default False.
         have_attention_mask_batch(bool): indicates whether attention_mask contains the batch dimension.
             Default True
         alibi(bool): This parameter indicates whether the flashattention supports the Alibi.
             Default: False
+        use_mqa(bool): Using MQA if True, only take effect under 910B. Default: False.
 
 
     Inputs:
       - **query** (Tensor) - Tensor query (:class:`mstype.fp16` [batch_size, head_num, seq_length, head_dim])
       - **key** (Tensor) - Tensor key (:class:`mstype.fp16` [batch_size, head_num, seq_length, head_dim])
       - **value** (Tensor) - Tensor value (:class:`mstype.fp16` [batch_size, head_num, seq_length, head_dim])
-      - **attention_mask** (Tensor) - Float Tensor the mask of (:class:`mstype.fp16` [batch_size, seq_length,
-          seq_length]): A matrix to pass masked information.
+      - **attention_mask** (Tensor) - Float Tensor the mask of (:class:`mstype.fp16` `mstype.uint8`
+        [batch_size, seq_length, seq_length]): A matrix to pass masked information.
 
     Outputs:
         A Tensor. The output of the attention with shape [batch_size, head_num, seq_length, head_dim]
@@ -97,56 +95,51 @@ class FlashAttention(Cell):
                  dropout_rate=0.0,
                  prev_block_num=65536,
                  next_block_num=65536,
-                 tiling_stgy_name="sparse",
                  dp=1,
                  mp=1,
                  high_precision=False,
                  have_attention_mask_batch=True,
-                 alibi=False
+                 alibi=False,
+                 use_mqa=False
                  ):
         super(FlashAttention, self).__init__()
 
         scaling_constant = math.sqrt(head_dim)
         if scaling_constant == 0:
             raise ValueError("the scaling constant must not be 0.")
-        self.scale_factor = Tensor([1. / scaling_constant], dtype=mstype.float16)
+        self.dropout_rate = dropout_rate
+        self.alibi = alibi
+        self.have_attention_mask_batch = have_attention_mask_batch
 
-        self.is_910A = MSContext.get_instance().get_ascend_soc_version() == "Ascend910"
-        if self.is_910A:
-            self.flash_attention = get_flash_attention(
-                prev_block_num=prev_block_num,
-                next_block_num=next_block_num,
-                tiling_stgy_name=tiling_stgy_name,
-                high_precision=high_precision
-            )
-            self.flash_attention.add_prim_attr("primitive_target", "Ascend")
-        else:
-            if alibi:
-                raise ValueError(f"When soc_version is not Ascend910A, alibi must be False")
-            self.transpose_4d_pre = ops.Transpose().shard(((dp, mp, 1, 1),))
-            self.transpose_4d_post = ops.Transpose().shard(((dp, 1, mp, 1),))
-            self.reshape = ops.Reshape()
-            self.zeros_like = ops.ZerosLike().shard(((dp, mp, 1, 1),))
-            self.zeros = ops.Zeros()
-            self.attn_expand_dims = ops.ExpandDims().shard(((dp, 1, 1),))
-            fa_strategies = ((dp, 1, mp),
-                             (dp, 1, mp),
-                             (dp, 1, mp),
+        self.transpose_4d_pre = ops.Transpose().shard(((dp, mp, 1, 1),))
+        self.transpose_4d_post = ops.Transpose().shard(((dp, 1, mp, 1),))
+        self.reshape = ops.Reshape()
+        self.zeros_like = ops.ZerosLike().shard(((dp, mp, 1, 1),))
+        self.zeros = ops.Zeros()
+        self.attn_cast = ops.Cast()
+        if use_mqa:
+            fa_strategies = ((dp, mp, 1, 1),
+                             (dp, 1, 1, 1),
                              (dp, 1, 1, 1))
-            if dropout_rate > 1e-5:
-                fa_strategies += ((dp, mp, 1, 1),)
-            self.flash_attention = FlashAttentionScore(head_num=head_num, pre_tokens=prev_block_num,
-                                                       next_tokens=next_block_num,
-                                                       keep_prob=1 - dropout_rate,
-                                                       scale_value=1.0,
-                                                       inner_precise=0 if high_precision else 1).shard(fa_strategies)
+        else:
+            fa_strategies = ((dp, mp, 1, 1),
+                             (dp, mp, 1, 1),
+                             (dp, mp, 1, 1))
+        if self.alibi:
+            self.alibi_rescale_mul = ops.Mul().shard(((dp, mp, 1, 1), (1,)))
+            self.alibi_rescale_factor = Tensor([scaling_constant], dtype=mstype.float16)
+            fa_strategies += ((dp, mp, 1, 1),)
+        if dropout_rate > 1e-5:
+            fa_strategies += ((dp, mp, 1, 1),)
+        fa_strategies += ((dp, 1, 1, 1),)
+        self.flash_attention = FlashAttentionScore(head_num=head_num, pre_tokens=prev_block_num,
+                                                   next_tokens=next_block_num,
+                                                   keep_prob=1 - dropout_rate,
+                                                   scale_value=1. / scaling_constant,
+                                                   inner_precise=0,
+                                                   input_layout="BNSD").shard(fa_strategies)
 
-        self.ones = ops.Ones()
-        self.dim_mask = Tensor([1 for _ in range(head_dim)], dtype=mstype.int8)
-        self.scale_mul = ops.Mul().shard(((dp, mp, 1, 1), (1,)))
         self.dropout_rate = dropout_rate
-        self.have_attention_mask_batch = have_attention_mask_batch
-        self.alibi = alibi
         if self.dropout_rate > 1e-5:
             self.keep_prob = Tensor(1 - self.dropout_rate, dtype=mstype.float16)
             self.fill_v2 = ops.FillV2().shard(((dp, mp, 1, 1), ()))
@@ -162,46 +155,7 @@ class FlashAttention(Cell):
                                   such as MatMul. Default: None.
         :return:
         """
-        if in_strategy is None:
-            # default: dp=1, mp=1, construct inputs only contain query, key, value
-            in_strategy = (
-                (1, 1, 1, 1),
-                (1, 1, 1, 1),
-                (1, 1, 1, 1),
-            )
         self.flash_attention.shard(in_strategy)
-        dp = in_strategy[0][0]
-        mp = in_strategy[0][1]
-        self.flash_attention.add_prim_attr("dev_matrix_shape", [dp, mp, 1, 1])
-        inputs_tensor_map = [
-            [3, 2, 1, 0],
-            [3, 2, 1, 0],
-            [3, 2, 1, 0],
-        ]
-        if self.have_attention_mask_batch:
-            inputs_tensor_map.append([3, 1, 0])
-        else:
-            inputs_tensor_map.append([-1, 1, 0])
-
-        input_empty_args_num = 2
-        # dropout_mask
-        if self.dropout_rate > 1e-5:
-            input_empty_args_num -= 1
-            inputs_tensor_map.append([3, 2, 1, 0])
-
-        if self.alibi:
-            input_empty_args_num -= 1
-            inputs_tensor_map.append([3, 2, 1, 0])
-
-        self.flash_attention.add_prim_attr("inputs_tensor_map", inputs_tensor_map)
-
-        self.flash_attention.add_prim_attr("outputs_tensor_map", [
-            [3, 2, 1, 0],  # O
-            [3, 2, 1],  # L
-            [3, 2, 1]  # M
-        ])
-        self.flash_attention.add_prim_attr("as_loss_divisor", 0)
-        self.flash_attention.add_prim_attr("empty_mirror_ops", input_empty_args_num)
 
     def construct(self, query, key, value, attn_mask=None, alibi_mask=None):
         """FlashAttention forward
@@ -212,53 +166,24 @@ class FlashAttention(Cell):
         :param alibi_mask: [bsz, head_num, 1, seq_len], if not None
         :return: output          [bsz, head_num, seq_len, head_dim]
         """
-        query = self.scale_mul(query, self.scale_factor)
-        bsz, head_num, seq_len, head_dim = query.shape
-        _, k_head_num, k_seq_len, _ = key.shape
-        _, v_head_num, v_seq_len, _ = value.shape
-        if head_num != k_head_num or head_num != v_head_num:
-            raise ValueError(
-                "the head_num of query, key and value must be the same, "
-                "If different head_num are used, users need to change themselves to be same by tile.")
-        if seq_len % 16 != 0 or k_seq_len % 16 != 0 or k_seq_len != v_seq_len:
-            raise ValueError(
-                "query, key, value seq_len must be a multiple of 16, and key seq_len, value seq_len must be the same.")
-
-        if head_dim > 304:
-            raise ValueError(
-                "the head_dim must be less than 304, otherwise the ub would be OOM.")
-
-        if self.is_910A:
-            # 910A -- FlashAttentionPrimtive
-            if self.dropout_rate > 1e-5:
-                drop_mask_bits = self.drop_gen_mask((bsz, head_num, seq_len, seq_len), self.keep_prob)
-                tensor_shape = Tensor((bsz, head_num, seq_len, seq_len), mstype.int32)
-                ones = self.fill_v2(tensor_shape, self.tensor_one)
-                ones = self.depend(ones, query)
-                drop_mask = self.do_dropout(ones, drop_mask_bits, self.keep_prob)
-            else:
-                drop_mask = None
-            output, _, _ = self.flash_attention(query, key, value, attn_mask, drop_mask, alibi_mask)
+        bsz, head_num, seq_len, _ = query.shape
+        # 910B -- FlashAttentionScore
+        if self.dropout_rate > 1e-5:
+            drop_mask_bits = self.reshape(self.drop_gen_mask((bsz, head_num, seq_len, seq_len), self.keep_prob),
+                                          (bsz, head_num, seq_len, seq_len // 8))
         else:
-            # FlashAttentionScore
-            # Useless input, just for binary calls.
-            if self.dropout_rate > 1e-5:
-                drop_mask_bits = self.reshape(self.drop_gen_mask((bsz, head_num, seq_len, seq_len), self.keep_prob),
-                                              (bsz, head_num, seq_len, seq_len // 8))
-            else:
-                drop_mask_bits = None
-            # (B, N, S, D) -> (B, S, H)
-            query = self.reshape(self.transpose_4d_pre(query, (0, 2, 1, 3)), (bsz, seq_len, -1))
-            key = self.reshape(self.transpose_4d_pre(key, (0, 2, 1, 3)), (bsz, seq_len, -1))
-            value = self.reshape(self.transpose_4d_pre(value, (0, 2, 1, 3)), (bsz, seq_len, -1))
-            attn_mask = self.attn_expand_dims(attn_mask, 1)
-            output, _, _ = self.flash_attention(query,
-                                                key,
-                                                value,
-                                                attn_mask,
-                                                drop_mask_bits,
-                                                None,
-                                                None)
-            output = self.transpose_4d_post(self.reshape(output, (bsz, seq_len, head_num, head_dim)), (0, 2, 1, 3))
-
+            drop_mask_bits = None
+        if self.alibi:
+            alibi_mask = self.alibi_rescale_mul(alibi_mask, self.cast(self.alibi_rescale_factor, alibi_mask.dtype))
+        # (B, S, S) -> (B, 1, S, S)
+        if self.have_attention_mask_batch:
+            attn_mask = self.cast(self.reshape(attn_mask, (bsz, 1, seq_len, seq_len)), mstype.uint8)
+        _, _, _, output = self.flash_attention(query,
+                                               key,
+                                               value,
+                                               alibi_mask,
+                                               drop_mask_bits,
+                                               None,
+                                               attn_mask,
+                                               None)
         return output

mindspore/nn/layer/math.py

@@ -375,9 +375,6 @@ class DiGamma(Cell):
                            nan, real_result)
 
 
-eps_fp32 = Tensor(np.finfo(np.float32).eps, mstype.float32)
-
-
 def _while_helper_func(cond, body, vals):
     while cond(vals).any():
         vals = body(vals)
@@ -394,7 +391,7 @@ def _igamma_series(ax, x, a, enabled):
     select = P.Select()
 
     # If more data types are supported, this epsilon need to be selected.
-    epsilon = eps_fp32
+    epsilon = Tensor(np.finfo(np.float32).eps, mstype.float32)
 
     def cond(vals):
         enabled = vals[0]
@@ -443,7 +440,7 @@ def _igammac_continued_fraction(ax, x, a, enabled):
     select = P.Select()
 
     # If more data types are supported, this epsilon need to be selected.
-    epsilon = eps_fp32
+    epsilon = Tensor(np.finfo(np.float32).eps, mstype.float32)
 
     def cond(vals):
         enabled = vals[0]
@@ -620,8 +617,7 @@ class IGamma(Cell):
             x = F.broadcast_to(x, para_shape)
             a = F.broadcast_to(a, para_shape)
         x_is_zero = self.equal(x, 0)
-        log_maxfloat = self.log_maxfloat32
-        underflow = self.less(ax, self.neg(log_maxfloat))
+        underflow = self.less(ax, self.neg(self.log_maxfloat32))
         ax = self.exp(ax)
         enabled = self.logicalnot(self.logicalor(self.logicalor(x_is_zero, domain_error), underflow))
         output = self.select(use_igammac,

mindspore/nn/layer/rnn_cells.py

@@ -83,7 +83,7 @@ def _check_lstmcell_init(func):
 
 
 def _rnn_tanh_cell(inputs, hidden, w_ih, w_hh, b_ih, b_hh):
-    '''RNN cell function with tanh activation'''
+    """RNN cell function with tanh activation"""
     if b_ih is None:
         igates = P.MatMul(False, True)(inputs, w_ih)
         hgates = P.MatMul(False, True)(hidden, w_hh)
@@ -94,7 +94,7 @@ def _rnn_tanh_cell(inputs, hidden, w_ih, w_hh, b_ih, b_hh):
 
 
 def _rnn_relu_cell(inputs, hidden, w_ih, w_hh, b_ih, b_hh):
-    '''RNN cell function with relu activation'''
+    """RNN cell function with relu activation"""
     if b_ih is None:
         igates = P.MatMul(False, True)(inputs, w_ih)
         hgates = P.MatMul(False, True)(hidden, w_hh)
@@ -105,7 +105,7 @@ def _rnn_relu_cell(inputs, hidden, w_ih, w_hh, b_ih, b_hh):
 
 
 def _lstm_cell(inputs, hidden, w_ih, w_hh, b_ih, b_hh):
-    '''LSTM cell function'''
+    """LSTM cell function"""
     hx, cx = hidden
     if b_ih is None:
         gates = P.MatMul(False, True)(inputs, w_ih) + P.MatMul(False, True)(hx, w_hh)
@@ -125,7 +125,7 @@ def _lstm_cell(inputs, hidden, w_ih, w_hh, b_ih, b_hh):
 
 
 def _gru_cell(inputs, hidden, w_ih, w_hh, b_ih, b_hh):
-    '''GRU cell function'''
+    """GRU cell function"""
     if b_ih is None:
         gi = P.MatMul(False, True)(inputs, w_ih)
         gh = P.MatMul(False, True)(hidden, w_hh)
@@ -144,7 +144,7 @@ def _gru_cell(inputs, hidden, w_ih, w_hh, b_ih, b_hh):
 
 
 class RNNCellBase(Cell):
-    '''Basic class for RNN Cells'''
+    """Basic class for RNN Cells"""
     def __init__(self, input_size: int, hidden_size: int, has_bias: bool, num_chunks: int,
                  dtype=mstype.float32):
         super().__init__()

mindspore/nn/wrap/__init__.py

@@ -20,7 +20,8 @@ Use the Wrapper to combine the loss or build the training steps.
 from __future__ import absolute_import
 
 from mindspore.nn.wrap.cell_wrapper import ForwardValueAndGrad, TrainOneStepCell, WithLossCell, WithGradCell, \
-    WithEvalCell, ParameterUpdate, GetNextSingleOp, VirtualDatasetCellTriple, MicroBatchInterleaved, PipelineCell
+    WithEvalCell, ParameterUpdate, GetNextSingleOp, VirtualDatasetCellTriple, MicroBatchInterleaved, PipelineCell, \
+    GradAccumulationCell
 from mindspore.nn.wrap.loss_scale import TrainOneStepWithLossScaleCell,\
     DynamicLossScaleUpdateCell, FixedLossScaleUpdateCell
 from mindspore.nn.wrap.grad_reducer import DistributedGradReducer
@@ -42,5 +43,6 @@ __all__ = [
     "ParameterUpdate",
     "DynamicLossScaleUpdateCell",
     "FixedLossScaleUpdateCell",
-    "VirtualDatasetCellTriple"
+    "VirtualDatasetCellTriple",
+    "GradAccumulationCell"
     ]

mindspore/nn/wrap/cell_wrapper.py

@@ -644,6 +644,9 @@ class PipelineCell(Cell):
         self.micro_inputs = nn.CellList()
         self.micro_size = micro_size
         self.add_list = []
+        if not isinstance(network, Cell):
+            raise TypeError("For 'PipelineCell', the argument 'network' must cell type, "
+                            "but got the type : {}.".format(type(network)))
         if not isinstance(micro_size, int):
             raise TypeError("For 'PipelineCell', the argument 'micro_size' must be integer, "
                             "but got the type : {}.".format(type(micro_size)))
@@ -670,7 +673,7 @@ class PipelineCell(Cell):
 
 class GradAccumulationCell(Cell):
     """
-    Wrap the network with Micro Batch.
+    Wrap the network with Micro Batch to enable the grad accumulation in semi_auto_parallel/auto_parallel mode.
 
     Args:
         network (Cell): The target network to wrap.
@@ -680,8 +683,11 @@ class GradAccumulationCell(Cell):
         ``Ascend`` ``GPU``
 
     Examples:
-        >>> net = Net()
-        >>> net = GradAccumulationCell(net, 4)
+        >>> import mindspore.nn as nn
+        >>> # Define the network structure of LeNet5. Refer to
+        >>> # https://gitee.com/mindspore/docs/blob/r2.2/docs/mindspore/code/lenet.py
+        >>> net = LeNet5()
+        >>> net = nn.GradAccumulationCell(net, 4)
     """
     def __init__(self, network, micro_size):
         super(GradAccumulationCell, self).__init__(auto_prefix=False)
@@ -689,6 +695,9 @@ class GradAccumulationCell(Cell):
         self.micro_inputs = nn.CellList()
         self.micro_size = micro_size
         self.add_list = []
+        if not isinstance(network, Cell):
+            raise TypeError("For 'GradAccumulationCell', the argument 'network' must cell type, "
+                            "but got the type : {}.".format(type(network)))
         if not isinstance(micro_size, int):
             raise TypeError("For 'GradAccumulationCell', the argument 'micro_size' must be integer, "
                             "but got the type : {}.".format(type(micro_size)))

mindspore/numpy/utils_const.py

@@ -143,8 +143,8 @@ def _infer_out_shape(*shapes):
     shape_out = list()
     max_len = max([len(it) for it in shapes])
     for i in range(max_len):
-        items = [it[i-max_len+len(it)] if i-max_len +
-                 len(it) >= 0 else 1 for it in shapes]
+        items = [
+            it[i - max_len + len(it)] if i - max_len + len(it) >= 0 else 1 for it in shapes]
         max_size = 0 if 0 in items else max(items)
         _check()
         shape_out.append(max_size)
@@ -158,8 +158,8 @@ def _can_broadcast(*shapes):
     """
     max_len = max([len(it) for it in shapes])
     for i in range(max_len):
-        items = [it[i-max_len+len(it)] if i-max_len +
-                 len(it) >= 0 else 1 for it in shapes]
+        items = [
+            it[i - max_len + len(it)] if i - max_len + len(it) >= 0 else 1 for it in shapes]
         max_size = 0 if 0 in items else max(items)
         if any(item not in (1, max_size) for item in items):
             return False
@@ -399,7 +399,7 @@ def _broadcast_tuples(tup1, tup2):
         if not isinstance(tup1, (tuple, list)) or not isinstance(tup2, (tuple, list)):
             raise TypeError("input shift and axis must be tuple or list or int.")
         if len(tup1) == len(tup2) or len(tup1) == 1 or len(tup2) == 1:
-            return None
+            return
         raise ValueError("shape mismatch: objects cannot be broadcast to a single shape")
 
     tup1 = (tup1,) if isinstance(tup1, int) else tup1

mindspore/ops/_grad_experimental/grad_array_ops.py

@@ -203,7 +203,7 @@ def get_bprop_index_put(self):
         if is_ascend:
             indices_ms = [convert_idx_positive(indices_ms[i], x1.shape[i]) for i in range(len(indices_ms))]
         indices_me = stack(indices_ms)
-        indices_grad = F.transpose(indices_me, F.make_range(F.rank(indices_me)-1, -1, -1))
+        indices_grad = F.transpose(indices_me, F.make_range(F.rank(indices_me) - 1, -1, -1))
         values_grad = gather_nd(dout, indices_grad)
         if equal(cast(x2.shape[0], mstype.int32), Tensor(1)):
             values_grad = values_grad.sum().reshape(1)

mindspore/ops/_grad_experimental/grad_implementations.py

@@ -19,7 +19,7 @@ from mindspore.ops import functional as F
 from mindspore.ops import operations as P
 from mindspore.ops.composite import multitype_ops as C
 from mindspore.ops.composite.multitype_ops.zeros_like_impl import zeros_like
-from mindspore.ops._grad_experimental.grad_base import bprops
+from mindspore.ops._grad_experimental.grad_base import bprops, bprop_getters
 from mindspore.common import dtype as mstype
 
 get_dtype = P.DType()
@@ -193,7 +193,7 @@ def bprop_tensor_move(x, out, dout):
     return (dout,)
 
 
-@bprops.register("DictInplaceSetItem")
+@bprop_getters.register("DictInplaceSetItem")
 def get_bprop_dict_inplace_setitem(self):
     """Generate bprop for dict inplace pop"""
 

mindspore/ops/_grad_experimental/grad_math_ops.py

@@ -135,7 +135,7 @@ def get_bprop_matrix_triangular_solve(self):
 
     def bprop(matrix, rhs, out, dout):
         grad_rhs = matrix_triangular_solve_op(matrix, dout)
-        if matrix.dtype == mstype.complex64 or matrix.dtype == mstype.complex128:
+        if matrix.dtype in (mstype.complex64, mstype.complex128):
             grad_rhs_temp = _adjoint(grad_rhs)
             out_temp = _adjoint(out)
         else:
@@ -156,14 +156,14 @@ def get_bprop_matrix_triangular_solve(self):
                 grad_matrix = mat_mul_op(grad_rhs, out_temp)
                 grad_matrix = neg_op(grad_matrix)
         if lower_a:
-            if grad_matrix.dtype == mstype.complex64 or grad_matrix.dtype == mstype.complex128:
+            if grad_matrix.dtype in (mstype.complex64, mstype.complex128):
                 grad_matrix_real = matrix_band_part_op(real_op(grad_matrix), -1, 0)
                 grad_matrix_imag = matrix_band_part_op(imag_op(grad_matrix), -1, 0)
                 grad_matrix = complex_op(grad_matrix_real, grad_matrix_imag)
             else:
                 grad_matrix = matrix_band_part_op(grad_matrix, -1, 0)
         else:
-            if grad_matrix.dtype == mstype.complex64 or grad_matrix.dtype == mstype.complex128:
+            if grad_matrix.dtype in (mstype.complex64, mstype.complex128):
                 grad_matrix_real = matrix_band_part_op(real_op(grad_matrix), 0, -1)
                 grad_matrix_imag = matrix_band_part_op(imag_op(grad_matrix), 0, -1)
                 grad_matrix = complex_op(grad_matrix_real, grad_matrix_imag)
@@ -219,7 +219,7 @@ def get_bprop_matrix_solve(self):
 @_primexpr
 def _generate_perm_matrix_solve_ls(x_dim):
     perm = tuple(range(x_dim - 2))
-    perm = perm + (x_dim-1, x_dim-2)
+    perm = perm + (x_dim - 1, x_dim - 2)
     return perm
 
 
@@ -647,20 +647,21 @@ def _fft_rank_offset(norm_shape, rank):
 @_primexpr
 def _fft_with_size_back_norm(norm_shape, norm, inverse, rank):
     """generate reverse term for fft_with_size"""
+    norm_ = None
     if inverse is False:
         if norm == "forward":
-            norm_ = 1 / _fft_rank_offset(norm_shape, rank)
-        if norm == "backward":
-            norm_ = 1 * _fft_rank_offset(norm_shape, rank)
-        if norm == "ortho":
-            norm_ = 1
-    if inverse is True:
+            norm_ = 1.0 / _fft_rank_offset(norm_shape, rank)
+        elif norm == "backward":
+            norm_ = 1.0 * _fft_rank_offset(norm_shape, rank)
+        elif norm == "ortho":
+            norm_ = 1.0
+    else:
         if norm == "forward":
-            norm_ = 1 * _fft_rank_offset(norm_shape, rank)
-        if norm == "backward":
-            norm_ = 1 / _fft_rank_offset(norm_shape, rank)
-        if norm == "ortho":
-            norm_ = 1
+            norm_ = 1.0 * _fft_rank_offset(norm_shape, rank)
+        elif norm == "backward":
+            norm_ = 1.0 / _fft_rank_offset(norm_shape, rank)
+        elif norm == "ortho":
+            norm_ = 1.0
     return norm_
 
 
@@ -670,9 +671,9 @@ def _rfft_norm(norm_shape, norm, rank):
     norm_ = 1.0
     if norm == "forward":
         norm_ = 1 / _fft_rank_offset(norm_shape, rank)
-    if norm == "backward":
-        norm_ = 1
-    if norm == "ortho":
+    elif norm == "backward":
+        norm_ = 1.0
+    elif norm == "ortho":
         norm_ = 1 / np.sqrt(_fft_rank_offset(norm_shape, rank))
     return norm_
 

mindspore/ops/_grad_experimental/grad_sparse_ops.py

@@ -358,10 +358,10 @@ def get_bprop_ragged_tensor_to_sparse(self):
                 split.append(zeros_like(i))
             all_d = (split, ragged_values_grad)
             return all_d
-        split = ()
+        split_ = ()
         for i in enumerate(rt_nested_splits):
-            split = split + (zeros_like(i),)
-        all_d = (split, ragged_values_grad)
+            split_ = split_ + (zeros_like(i),)
+        all_d = (split_, ragged_values_grad)
         return all_d
 
     return bprop

mindspore/ops/_op_impl/aicpu/add.py

@@ -29,9 +29,9 @@ add_op_info = AiCPURegOp("Add") \
     .dtype_format(DataType.I32_Default, DataType.I32_Default, DataType.I32_Default) \
     .dtype_format(DataType.I64_Default, DataType.I64_Default, DataType.I64_Default) \
     .dtype_format(DataType.U8_Default, DataType.U8_Default, DataType.U8_Default) \
-    .dtype_format(DataType.U16_Default, DataType.I16_Default, DataType.I16_Default) \
-    .dtype_format(DataType.U32_Default, DataType.I32_Default, DataType.I32_Default) \
-    .dtype_format(DataType.U64_Default, DataType.I64_Default, DataType.I64_Default) \
+    .dtype_format(DataType.U16_Default, DataType.U16_Default, DataType.U16_Default) \
+    .dtype_format(DataType.U32_Default, DataType.U32_Default, DataType.U32_Default) \
+    .dtype_format(DataType.U64_Default, DataType.U64_Default, DataType.U64_Default) \
     .dtype_format(DataType.C64_Default, DataType.C64_Default, DataType.C64_Default) \
     .dtype_format(DataType.C128_Default, DataType.C128_Default, DataType.C128_Default) \
     .get_op_info()

mindspore/ops/_op_impl/aicpu/linear_sum_assignment.py

@@ -1,4 +1,4 @@
-# Copyright 2022 Huawei Technologies Co., Ltd
+# Copyright 2023 Huawei Technologies Co., Ltd
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -24,11 +24,30 @@ lsap_op_info = AiCPURegOp("LinearSumAssignment") \
     .input(2, 'maximize', "required") \
     .output(0, "row_ind", "required") \
     .output(1, "col_ind", "required") \
-    .attr("cust_aicpu", "str") \
     .dtype_format(DataType.F64_Default, DataType.I64_Default,
                   DataType.BOOL_Default, DataType.I64_Default, DataType.I64_Default,) \
     .dtype_format(DataType.F32_Default, DataType.I64_Default,
                   DataType.BOOL_Default, DataType.I64_Default, DataType.I64_Default,) \
+    .dtype_format(DataType.F16_Default, DataType.I64_Default,
+                  DataType.BOOL_Default, DataType.I64_Default, DataType.I64_Default,) \
+    .dtype_format(DataType.BOOL_Default, DataType.I64_Default,
+                  DataType.BOOL_Default, DataType.I64_Default, DataType.I64_Default,) \
+    .dtype_format(DataType.I16_Default, DataType.I64_Default,
+                  DataType.BOOL_Default, DataType.I64_Default, DataType.I64_Default,) \
+    .dtype_format(DataType.I32_Default, DataType.I64_Default,
+                  DataType.BOOL_Default, DataType.I64_Default, DataType.I64_Default,) \
+    .dtype_format(DataType.I64_Default, DataType.I64_Default,
+                  DataType.BOOL_Default, DataType.I64_Default, DataType.I64_Default,) \
+    .dtype_format(DataType.I8_Default, DataType.I64_Default,
+                  DataType.BOOL_Default, DataType.I64_Default, DataType.I64_Default,) \
+    .dtype_format(DataType.U16_Default, DataType.I64_Default,
+                  DataType.BOOL_Default, DataType.I64_Default, DataType.I64_Default,) \
+    .dtype_format(DataType.U32_Default, DataType.I64_Default,
+                  DataType.BOOL_Default, DataType.I64_Default, DataType.I64_Default,) \
+    .dtype_format(DataType.U64_Default, DataType.I64_Default,
+                  DataType.BOOL_Default, DataType.I64_Default, DataType.I64_Default,) \
+    .dtype_format(DataType.U8_Default, DataType.I64_Default,
+                  DataType.BOOL_Default, DataType.I64_Default, DataType.I64_Default,) \
     .get_op_info()
 
 

mindspore/ops/_utils/utils.py

@@ -77,10 +77,12 @@ def get_broadcast_shape(x_shape, y_shape, prim_name, arg_name1="x", arg_name2="y
     broadcast_shape = list(broadcast_shape_front) + broadcast_shape_back
     return broadcast_shape
 
+
 def dim_not_equal(dim1, dim2):
     """Compare dim in shape"""
     return dim1 != dim2 and dim1 >= 0 and dim2 >= 0
 
+
 def get_concat_offset(x_shp, x_type, axis, prim_name):
     """for concat and concatoffset check args and compute offset"""
     validator.check_value_type("shape", x_shp, [tuple, list], prim_name)

mindspore/ops/composite/multitype_ops/_compile_utils.py

@@ -1255,7 +1255,8 @@ def _tensor_setitem_by_bool_tensor_with_tensor(data, index, value):
     index = index.reshape(const_utils.generate_padding_shape(index.shape, len(data.shape)))
     index = F.broadcast_to(index, data.shape)
     value = F.cast(value, F.dtype(data))
-    value = value.reshape(const_utils.generate_padding_shape(value.shape, len(data.shape)))
+    while value.ndim < data.ndim:
+        value = value.unsqueeze(-1)
     value = F.broadcast_to(value, data.shape)
     result = F.select(index, value, data)
     return result

mindspore/ops/composite/multitype_ops/getitem_impl.py

@@ -161,7 +161,7 @@ def _tuple_getitem_by_slice(data, slice_index):
         if start is None:
             start = 0 if step >= 1 else -1
         if stop is None:
-            stop = (2**31-1) if step >= 1 else -(2**31-1)
+            stop = (2**31 - 1) if step >= 1 else -(2**31 - 1)
         return sequence_slice(data, start, stop, step)
     return _tuple_slice(data, slice_index)
 
@@ -236,7 +236,7 @@ def _list_getitem_by_slice(data, slice_index):
         if start is None:
             start = 0 if step >= 1 else -1
         if stop is None:
-            stop = (2**31-1) if step >= 1 else -(2**31-1)
+            stop = (2**31 - 1) if step >= 1 else -(2**31 - 1)
         return sequence_slice(data, start, stop, step)
     return _list_slice(data, slice_index)