mindspore 2.2.0__cp39-cp39-win_amd64.whl → 2.2.10__cp39-cp39-win_amd64.whl

mindspore/include/api/model.h

@@ -136,6 +136,13 @@ class MS_API Model {
   /// \return Status.
   Status UpdateWeights(const std::vector<MSTensor> &new_weights);
 
+  /// \brief Change the size and or content of weight tensors
+  ///
+  /// \param[in]  A vector where model constant are arranged in sequence
+  ///
+  /// \return Status.
+  Status UpdateWeights(const std::vector<std::vector<MSTensor>> &new_weights);
+
   /// \brief Inference model API. If use this API in train mode, it's equal to RunStep API.
   ///
   /// \param[in] inputs A vector where model inputs are arranged in sequence.
@@ -358,6 +365,13 @@ class MS_API Model {
 
   const std::shared_ptr<ModelImpl> impl() const { return impl_; }
 
+  /// \brief Get model info by key
+  ///
+  /// \param[in] key The key of model info key-value pair
+  ///
+  /// \return The value of the model info associated with the given key.
+  inline std::string GetModelInfo(const std::string &key);
+
  private:
   friend class Serialization;
   // api without std::string
@@ -374,6 +388,7 @@ class MS_API Model {
                const std::vector<char> &cropto_lib_path);
   Status Build(const std::vector<char> &model_path, ModelType model_type, const std::shared_ptr<Context> &model_context,
                const Key &dec_key, const std::vector<char> &dec_mode, const std::vector<char> &cropto_lib_path);
+  std::vector<char> GetModelInfo(const std::vector<char> &key);
   std::shared_ptr<ModelImpl> impl_;
 };
 
@@ -416,5 +431,7 @@ Status Model::Build(const std::string &model_path, ModelType model_type,
                     const std::shared_ptr<Context> &model_context) {
   return Build(StringToChar(model_path), model_type, model_context);
 }
+
+inline std::string Model::GetModelInfo(const std::string &key) { return CharToString(GetModelInfo(StringToChar(key))); }
 }  // namespace mindspore
 #endif  // MINDSPORE_INCLUDE_API_MODEL_H

mindspore/include/api/status.h

@@ -83,9 +83,14 @@ enum StatusCode : uint32_t {
   kLiteModelRebuild = kLite | (0x0FFFFFFF & -12),    /**< Model has been built. */
 
   // Executor error code, range: [-100,-200)
-  kLiteOutOfTensorRange = kLite | (0x0FFFFFFF & -100), /**< Failed to check range. */
-  kLiteInputTensorError = kLite | (0x0FFFFFFF & -101), /**< Failed to check input tensor. */
-  kLiteReentrantError = kLite | (0x0FFFFFFF & -102),   /**< Exist executor running. */
+  kLiteOutOfTensorRange = kLite | (0x0FFFFFFF & -100),           /**< Failed to check range. */
+  kLiteInputTensorError = kLite | (0x0FFFFFFF & -101),           /**< Failed to check input tensor. */
+  kLiteReentrantError = kLite | (0x0FFFFFFF & -102),             /**< Exist executor running. */
+  kLiteLLMWaitProcessTimeOut = kLite | (0x0FFFFFFF & -103),      /**< Wait to be processed time out. */
+  kLiteLLMKVCacheNotExist = kLite | (0x0FFFFFFF & -104),         /**< KV Cache not exist. */
+  kLiteLLMRepeatRequest = kLite | (0x0FFFFFFF & -105),           /**< repeat request. */
+  kLiteLLMRequestAlreadyCompleted = kLite | (0x0FFFFFFF & -106), /**< request already complete!. */
+  kLiteLLMEngineFinalized = kLite | (0x0FFFFFFF & -107),         /**< llm engine finalized. */
 
   // Graph error code, range: [-200,-300)
   kLiteGraphFileError = kLite | (0x0FFFFFFF & -200), /**< Failed to verify graph file. */

mindspore/nn/cell.py

@@ -1081,9 +1081,6 @@ class Cell(Cell_):
         if not isinstance(param, Parameter) and param is not None:
             raise TypeError(f"For 'insert_param_to_cell', the argument 'param' must be 'Parameter' if not None, "
                             f"but got {type(param)}.")
-        if param is None:
-            raise TypeError(f"For 'insert_param_to_cell', the argument 'param' must not be None, "
-                            f"but got None.")
         if isinstance(param, Parameter) and param.name == PARAMETER_NAME_DEFAULT:
             param.name = param_name
         self._params[param_name] = param

mindspore/nn/layer/activation.py

@@ -932,10 +932,8 @@ class GELU(Cell):
         """Initialize GELU."""
         super(GELU, self).__init__()
         validator.check_bool(approximate, 'approximate', self.cls_name)
-        self.approximate = approximate
-        if approximate:
-            self.approximate = 'tanh'
-        else:
+        self.approximate = 'tanh'
+        if not approximate:
             self.approximate = 'none'
 
     def construct(self, x):
@@ -1335,7 +1333,8 @@ class LRN(Cell):
 
     .. warning::
         LRN is deprecated on Ascend due to potential accuracy problem. It's recommended to use other
-        normalization methods, e.g. :class:`mindspore.nn.BatchNorm`.
+        normalization methods, e.g. :class:`mindspore.nn.BatchNorm1d` ,
+        :class:`mindspore.nn.BatchNorm2d` , :class:`mindspore.nn.BatchNorm3d`.
 
     Refer to :func:`mindspore.ops.lrn` for more details.
 

mindspore/nn/layer/conv.py

@@ -718,9 +718,9 @@ class Conv3d(_Conv):
 
         .. math::
             \begin{array}{ll} \\
-                D_{out} ＝ \left \lceil{\frac{D_{in}}{\text{stride[0]}}} \right \rceil \\
-                H_{out} ＝ \left \lceil{\frac{H_{in}}{\text{stride[1]}}} \right \rceil \\
-                W_{out} ＝ \left \lceil{\frac{W_{in}}{\text{stride[2]}}} \right \rceil \\
+                D_{out} = \left \lceil{\frac{D_{in}}{\text{stride[0]}}} \right \rceil \\
+                H_{out} = \left \lceil{\frac{H_{in}}{\text{stride[1]}}} \right \rceil \\
+                W_{out} = \left \lceil{\frac{W_{in}}{\text{stride[2]}}} \right \rceil \\
             \end{array}
 
 
@@ -728,11 +728,11 @@ class Conv3d(_Conv):
 
         .. math::
             \begin{array}{ll} \\
-                D_{out} ＝ \left \lfloor{\frac{D_{in} - \text{dilation[0]} \times (\text{kernel_size[0]} - 1) }
+                D_{out} = \left \lfloor{\frac{D_{in} - \text{dilation[0]} \times (\text{kernel_size[0]} - 1) }
                 {\text{stride[0]}} + 1} \right \rfloor \\
-                H_{out} ＝ \left \lfloor{\frac{H_{in} - \text{dilation[1]} \times (\text{kernel_size[1]} - 1) }
+                H_{out} = \left \lfloor{\frac{H_{in} - \text{dilation[1]} \times (\text{kernel_size[1]} - 1) }
                 {\text{stride[1]}} + 1} \right \rfloor \\
-                W_{out} ＝ \left \lfloor{\frac{W_{in} - \text{dilation[2]} \times (\text{kernel_size[2]} - 1) }
+                W_{out} = \left \lfloor{\frac{W_{in} - \text{dilation[2]} \times (\text{kernel_size[2]} - 1) }
                 {\text{stride[2]}} + 1} \right \rfloor \\
             \end{array}
 
@@ -740,11 +740,11 @@ class Conv3d(_Conv):
 
         .. math::
             \begin{array}{ll} \\
-                D_{out} ＝ \left \lfloor{\frac{D_{in} + padding[0] + padding[1] - (\text{dilation[0]} - 1) \times
+                D_{out} = \left \lfloor{\frac{D_{in} + padding[0] + padding[1] - (\text{dilation[0]} - 1) \times
                 \text{kernel_size[0]} - 1 }{\text{stride[0]}} + 1} \right \rfloor \\
-                H_{out} ＝ \left \lfloor{\frac{H_{in} + padding[2] + padding[3] - (\text{dilation[1]} - 1) \times
+                H_{out} = \left \lfloor{\frac{H_{in} + padding[2] + padding[3] - (\text{dilation[1]} - 1) \times
                 \text{kernel_size[1]} - 1 }{\text{stride[1]}} + 1} \right \rfloor \\
-                W_{out} ＝ \left \lfloor{\frac{W_{in} + padding[4] + padding[5] - (\text{dilation[2]} - 1) \times
+                W_{out} = \left \lfloor{\frac{W_{in} + padding[4] + padding[5] - (\text{dilation[2]} - 1) \times
                 \text{kernel_size[2]} - 1 }{\text{stride[2]}} + 1} \right \rfloor \\
             \end{array}
 
@@ -812,7 +812,7 @@ class Conv3d(_Conv):
             bias_init,
             data_format,
             dtype=dtype)
-        out_channels = self.out_channels
+        out_channels = self.out_channels // group
         self.conv3d = P.Conv3D(out_channel=out_channels,
                                kernel_size=self.kernel_size,
                                mode=1,
@@ -820,17 +820,33 @@ class Conv3d(_Conv):
                                pad=self.padding,
                                stride=self.stride,
                                dilation=self.dilation,
-                               group=group,
+                               group=1,
                                data_format=self.data_format)
         self.bias_add = P.BiasAdd(data_format=self.data_format)
         self.shape = P.Shape()
+        self.concat = P.Concat(1)
+        self.split_0 = P.Split(0, self.group)
+        self.split_1 = P.Split(1, self.group)
 
     def construct(self, x):
         x_shape = self.shape(x)
         _check_input_5dims(x_shape, self.cls_name)
-        out = self.conv3d(x, self.weight)
-        if self.has_bias:
-            out = self.bias_add(out, self.bias)
+        if self.group == 1:
+            out = self.conv3d(x, self.weight)
+            if self.has_bias:
+                out = self.bias_add(out, self.bias)
+        else:
+            features = self.split_1(x)
+            weights = self.split_0(self.weight)
+            outputs = ()
+            for i in range(self.group):
+                output = self.conv3d(features[i], weights[i])
+                outputs = outputs + (output,)
+            out = self.concat(outputs)
+            if self.bias is not None:
+                new_shape = [1 for _ in range(out.ndim)]
+                new_shape[1] = self.out_channels
+                out = out + self.bias.reshape(new_shape)
         return out
 
 
@@ -921,9 +937,9 @@ class Conv3dTranspose(_Conv):
 
         .. math::
             \begin{array}{ll} \\
-                D_{out} ＝ \left \lfloor{\frac{D_{in}}{\text{stride[0]}} + 1} \right \rfloor \\
-                H_{out} ＝ \left \lfloor{\frac{H_{in}}{\text{stride[1]}} + 1} \right \rfloor \\
-                W_{out} ＝ \left \lfloor{\frac{W_{in}}{\text{stride[2]}} + 1} \right \rfloor \\
+                D_{out} = \left \lfloor{\frac{D_{in}}{\text{stride[0]}} + 1} \right \rfloor \\
+                H_{out} = \left \lfloor{\frac{H_{in}}{\text{stride[1]}} + 1} \right \rfloor \\
+                W_{out} = \left \lfloor{\frac{W_{in}}{\text{stride[2]}} + 1} \right \rfloor \\
             \end{array}
 
 
@@ -931,11 +947,11 @@ class Conv3dTranspose(_Conv):
 
         .. math::
             \begin{array}{ll} \\
-                D_{out} ＝ \left \lfloor{\frac{D_{in} - \text{dilation[0]} \times (\text{kernel_size[0]} - 1) }
+                D_{out} = \left \lfloor{\frac{D_{in} - \text{dilation[0]} \times (\text{kernel_size[0]} - 1) }
                 {\text{stride[0]}} + 1} \right \rfloor \\
-                H_{out} ＝ \left \lfloor{\frac{H_{in} - \text{dilation[1]} \times (\text{kernel_size[1]} - 1) }
+                H_{out} = \left \lfloor{\frac{H_{in} - \text{dilation[1]} \times (\text{kernel_size[1]} - 1) }
                 {\text{stride[1]}} + 1} \right \rfloor \\
-                W_{out} ＝ \left \lfloor{\frac{W_{in} - \text{dilation[2]} \times (\text{kernel_size[2]} - 1) }
+                W_{out} = \left \lfloor{\frac{W_{in} - \text{dilation[2]} \times (\text{kernel_size[2]} - 1) }
                 {\text{stride[2]}} + 1} \right \rfloor \\
             \end{array}
 
@@ -943,11 +959,11 @@ class Conv3dTranspose(_Conv):
 
         .. math::
             \begin{array}{ll} \\
-                D_{out} ＝ \left \lfloor{\frac{D_{in} + padding[0] + padding[1] - (\text{dilation[0]} - 1) \times
+                D_{out} = \left \lfloor{\frac{D_{in} + padding[0] + padding[1] - (\text{dilation[0]} - 1) \times
                 \text{kernel_size[0]} - 1 }{\text{stride[0]}} + 1} \right \rfloor \\
-                H_{out} ＝ \left \lfloor{\frac{H_{in} + padding[2] + padding[3] - (\text{dilation[1]} - 1) \times
+                H_{out} = \left \lfloor{\frac{H_{in} + padding[2] + padding[3] - (\text{dilation[1]} - 1) \times
                 \text{kernel_size[1]} - 1 }{\text{stride[1]}} + 1} \right \rfloor \\
-                W_{out} ＝ \left \lfloor{\frac{W_{in} + padding[4] + padding[5] - (\text{dilation[2]} - 1) \times
+                W_{out} = \left \lfloor{\frac{W_{in} + padding[4] + padding[5] - (\text{dilation[2]} - 1) \times
                 \text{kernel_size[2]} - 1 }{\text{stride[2]}} + 1} \right \rfloor \\
             \end{array}
 

mindspore/nn/layer/flash_attention.py

@@ -57,14 +57,15 @@ class FlashAttention(Cell):
             Default True
         alibi(bool): This parameter indicates whether the flashattention supports the Alibi.
             Default: False
+        use_mqa(bool): Using MHA 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]
@@ -102,17 +103,23 @@ class FlashAttention(Cell):
                  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.is_910A = MSContext.get_instance().get_ascend_soc_version() == "Ascend910"
+        self.dropout_rate = dropout_rate
+        self.is_910A = MSContext.get_instance().get_ascend_soc_version() == "ascend910"
         if self.is_910A:
+            self.scale_factor = Tensor([1. / math.sqrt(scaling_constant)], dtype=mstype.float16)
+            self.scale_mul = ops.Mul().shard(((dp, mp, 1, 1), (1,)))
+            self.ones = ops.Ones()
+            self.dim_mask = Tensor([1 for _ in range(head_dim)], dtype=mstype.int8)
+            self.have_attention_mask_batch = have_attention_mask_batch
+            self.alibi = alibi
             self.flash_attention = get_flash_attention(
                 prev_block_num=prev_block_num,
                 next_block_num=next_block_num,
@@ -120,6 +127,10 @@ class FlashAttention(Cell):
                 high_precision=high_precision
             )
             self.flash_attention.add_prim_attr("primitive_target", "Ascend")
+            fa_strategies = ((dp, mp, 1, 1),
+                             (dp, mp, 1, 1),
+                             (dp, mp, 1, 1))
+            self.shard(fa_strategies)
         else:
             if alibi:
                 raise ValueError(f"When soc_version is not Ascend910A, alibi must be False")
@@ -128,25 +139,27 @@ class FlashAttention(Cell):
             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),
-                             (dp, 1, 1, 1))
+            self.attn_cast = ops.Cast()
+            if use_mqa:
+                fa_strategies = ((dp, mp, 1, 1),
+                                 (dp, 1, 1, 1),
+                                 (dp, 1, 1, 1),
+                                 (dp, 1, 1, 1))
+            else:
+                fa_strategies = ((dp, mp, 1, 1),
+                                 (dp, mp, 1, 1),
+                                 (dp, mp, 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)
+                                                       scale_value=1. / scaling_constant,
+                                                       inner_precise=0 if high_precision else 1,
+                                                       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 +175,49 @@ 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])
+        if self.is_910A:
+            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])
+            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])
+            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("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)
+            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)
+        else:
+            self.flash_attention.shard(in_strategy)
 
     def construct(self, query, key, value, attn_mask=None, alibi_mask=None):
         """FlashAttention forward
@@ -212,24 +228,22 @@ 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:
+            _, 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 the seq_len between key and value must be equal.")
             # 910A -- FlashAttentionPrimtive
+            if head_dim > 304:
+                raise ValueError(
+                    "the head_dim must be less than 304, otherwise the ub would be OOM.")
             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)
@@ -238,27 +252,25 @@ class FlashAttention(Cell):
                 drop_mask = self.do_dropout(ones, drop_mask_bits, self.keep_prob)
             else:
                 drop_mask = None
+            query = self.scale_mul(query, self.scale_factor)
+            key = self.scale_mul(key, self.scale_factor)
+            attn_mask = self.cast(attn_mask, mstype.float16)
             output, _, _ = self.flash_attention(query, key, value, attn_mask, drop_mask, alibi_mask)
         else:
-            # FlashAttentionScore
-            # Useless input, just for binary calls.
+            # 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:
                 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)
+            # (B, S, S) -> (B, 1, S, S)
+            attn_mask = self.cast(self.reshape(attn_mask, (bsz, 1, seq_len, seq_len)), mstype.uint8)
             output, _, _ = self.flash_attention(query,
                                                 key,
                                                 value,
                                                 attn_mask,
                                                 drop_mask_bits,
                                                 None,
+                                                None,
                                                 None)
-            output = self.transpose_4d_post(self.reshape(output, (bsz, seq_len, head_num, head_dim)), (0, 2, 1, 3))
-
         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/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)))
@@ -689,6 +692,9 @@ 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)))