mindspore 2.2.14__cp37-cp37m-manylinux1_x86_64.whl → 2.3.0rc2__cp37-cp37m-manylinux1_x86_64.whl

mindspore/ops/operations/comm_ops.py

@@ -53,14 +53,14 @@ class ReduceOp:
 
             For the Ascend devices, users need to prepare the rank table, set rank_id and device_id.
             Please see the `rank table Startup
-            <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/rank_table.html>`_
+            <https://www.mindspore.cn/tutorials/experts/en/master/parallel/rank_table.html>`_
             for more details.
 
             For the GPU devices, users need to prepare the host file and mpi, please see the `mpirun Startup
-            <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/mpirun.html>`_ .
+            <https://www.mindspore.cn/tutorials/experts/en/master/parallel/mpirun.html>`_ .
 
             For the CPU device, users need to write a dynamic cluster startup script, please see the `Dynamic Cluster
-            Startup <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/dynamic_cluster.html>`_ .
+            Startup <https://www.mindspore.cn/tutorials/experts/en/master/parallel/dynamic_cluster.html>`_ .
 
             This example should be run with multiple devices.
 
@@ -94,7 +94,7 @@ class ReduceOp:
 
 def check_collective_target_dtype(data_name, data_dtype, prim_name):
     """Check if data type is valid."""
-    default_target_dtypes = (mstype.int8, mstype.uint8, mstype.int32, mstype.float16, mstype.bfloat16, mstype.float32)
+    default_target_dtypes = (mstype.int8, mstype.int32, mstype.float16, mstype.float32, mstype.bfloat16)
     gpu_target_dtypes = (mstype.bool_, mstype.int8, mstype.int32, mstype.int64, mstype.uint32, mstype.uint64,
                          mstype.float16, mstype.float32, mstype.float64)
 
@@ -144,14 +144,14 @@ class AllReduce(Primitive):
 
             For the Ascend devices, users need to prepare the rank table, set rank_id and device_id.
             Please see the `rank table Startup
-            <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/rank_table.html>`_
+            <https://www.mindspore.cn/tutorials/experts/en/master/parallel/rank_table.html>`_
             for more details.
 
             For the GPU devices, users need to prepare the host file and mpi, please see the `mpirun Startup
-            <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/mpirun.html>`_ .
+            <https://www.mindspore.cn/tutorials/experts/en/master/parallel/mpirun.html>`_ .
 
             For the CPU device, users need to write a dynamic cluster startup script, please see the `Dynamic Cluster
-            Startup <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/dynamic_cluster.html>`_ .
+            Startup <https://www.mindspore.cn/tutorials/experts/en/master/parallel/dynamic_cluster.html>`_ .
 
             This example should be run with 2 devices.
 
@@ -180,7 +180,7 @@ class AllReduce(Primitive):
 
     Tutorial Examples:
         - `Distributed Set Communication Primitives - AllReduce
-          <https://www.mindspore.cn/docs/en/r2.2/api_python/samples/ops/communicate_ops.html#allreduce>`_
+          <https://www.mindspore.cn/docs/en/master/api_python/samples/ops/communicate_ops.html#allreduce>`_
 
     """
 
@@ -233,14 +233,14 @@ class AllGather(PrimitiveWithInfer):
 
             For the Ascend devices, users need to prepare the rank table, set rank_id and device_id.
             Please see the `rank table Startup
-            <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/rank_table.html>`_
+            <https://www.mindspore.cn/tutorials/experts/en/master/parallel/rank_table.html>`_
             for more details.
 
             For the GPU devices, users need to prepare the host file and mpi, please see the `mpirun Startup
-            <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/mpirun.html>`_ .
+            <https://www.mindspore.cn/tutorials/experts/en/master/parallel/mpirun.html>`_ .
 
             For the CPU device, users need to write a dynamic cluster startup script, please see the `Dynamic Cluster
-            Startup <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/dynamic_cluster.html>`_ .
+            Startup <https://www.mindspore.cn/tutorials/experts/en/master/parallel/dynamic_cluster.html>`_ .
 
             This example should be run with 2 devices.
 
@@ -272,7 +272,7 @@ class AllGather(PrimitiveWithInfer):
 
     Tutorial Examples:
         - `Distributed Set Communication Primitives - AllGather
-          <https://www.mindspore.cn/docs/en/r2.2/api_python/samples/ops/communicate_ops.html#allgather>`_
+          <https://www.mindspore.cn/docs/en/master/api_python/samples/ops/communicate_ops.html#allgather>`_
 
     """
 
@@ -458,14 +458,14 @@ class ReduceScatter(Primitive):
 
             For the Ascend devices, users need to prepare the rank table, set rank_id and device_id.
             Please see the `rank table Startup
-            <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/rank_table.html>`_
+            <https://www.mindspore.cn/tutorials/experts/en/master/parallel/rank_table.html>`_
             for more details.
 
             For the GPU devices, users need to prepare the host file and mpi, please see the `mpirun Startup
-            <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/mpirun.html>`_ .
+            <https://www.mindspore.cn/tutorials/experts/en/master/parallel/mpirun.html>`_ .
 
             For the CPU device, users need to write a dynamic cluster startup script, please see the `Dynamic Cluster
-            Startup <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/dynamic_cluster.html>`_ .
+            Startup <https://www.mindspore.cn/tutorials/experts/en/master/parallel/dynamic_cluster.html>`_ .
 
             This example should be run with 2 devices.
 
@@ -498,7 +498,7 @@ class ReduceScatter(Primitive):
 
     Tutorial Examples:
         - `Distributed Set Communication Primitives - ReduceScatter
-          <https://www.mindspore.cn/docs/en/r2.2/api_python/samples/ops/communicate_ops.html#reducescatter>`_
+          <https://www.mindspore.cn/docs/en/master/api_python/samples/ops/communicate_ops.html#reducescatter>`_
 
     """
 
@@ -600,14 +600,14 @@ class Broadcast(PrimitiveWithInfer):
 
             For the Ascend devices, users need to prepare the rank table, set rank_id and device_id.
             Please see the `rank table Startup
-            <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/rank_table.html>`_
+            <https://www.mindspore.cn/tutorials/experts/en/master/parallel/rank_table.html>`_
             for more details.
 
             For the GPU devices, users need to prepare the host file and mpi, please see the `mpirun Startup
-            <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/mpirun.html>`_ .
+            <https://www.mindspore.cn/tutorials/experts/en/master/parallel/mpirun.html>`_ .
 
             For the CPU device, users need to write a dynamic cluster startup script, please see the `Dynamic Cluster
-            Startup <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/dynamic_cluster.html>`_ .
+            Startup <https://www.mindspore.cn/tutorials/experts/en/master/parallel/dynamic_cluster.html>`_ .
 
             This example should be run with multiple devices.
 
@@ -638,7 +638,7 @@ class Broadcast(PrimitiveWithInfer):
 
     Tutorial Examples:
         - `Distributed Set Communication Primitives - Broadcast
-          <https://www.mindspore.cn/docs/en/r2.2/api_python/samples/ops/communicate_ops.html#broadcast>`_
+          <https://www.mindspore.cn/docs/en/master/api_python/samples/ops/communicate_ops.html#broadcast>`_
 
     """
 
@@ -651,16 +651,6 @@ class Broadcast(PrimitiveWithInfer):
         self.add_prim_attr('group', _get_group(group))
         self.add_prim_attr('no_eliminate', True)
 
-    def infer_shape(self, x_shape):
-        return x_shape
-
-    def infer_dtype(self, x_dtype):
-        if not isinstance(x_dtype, tuple):
-            raise TypeError(f"For '{self.name}', the 'input_x' must be a tuple, but got {type(x_dtype).__name__}!")
-        for _ele in x_dtype:
-            check_collective_target_dtype('tuple input_x', _ele, self.name)
-        return x_dtype
-
 
 class _AllSwap(PrimitiveWithCheck):
     """
@@ -728,11 +718,11 @@ class NeighborExchange(Primitive):
         The user needs to preset
         communication environment variables before running the following example, please check the details on the
         official website of `MindSpore \
-        <https://www.mindspore.cn/docs/en/r2.2/api_python/mindspore.ops.primitive.html#communication-operator>`_.
+        <https://www.mindspore.cn/docs/en/master/api_python/mindspore.ops.primitive.html#communication-operator>`_.
 
         This operator requires a full-mesh network topology, each device has the same vlan id, and the ip & mask are
         in the same subnet, please check the `details \
-        <https://www.mindspore.cn/docs/en/r2.2/api_python/samples/ops/communicate_ops.html#notes>`_.
+        <https://www.mindspore.cn/docs/en/master/api_python/samples/ops/communicate_ops.html#notes>`_.
 
     Args:
         send_rank_ids (list(int)): Ranks which the data is sent to.
@@ -781,7 +771,7 @@ class NeighborExchange(Primitive):
 
     Tutorial Examples:
         - `Distributed Set Communication Primitives - NeighborExchange
-          <https://www.mindspore.cn/docs/en/r2.2/api_python/samples/ops/communicate_ops.html#neighborexchange>`_
+          <https://www.mindspore.cn/docs/en/master/api_python/samples/ops/communicate_ops.html#neighborexchange>`_
 
     """
 
@@ -814,7 +804,7 @@ class AlltoAll(PrimitiveWithInfer):
     Note:
         This operator requires a full-mesh network topology, each device has the same vlan id, and the ip & mask are
         in the same subnet, please check the `details \
-        <https://www.mindspore.cn/docs/en/r2.2/api_python/samples/ops/communicate_ops.html#notes>`_.
+        <https://www.mindspore.cn/docs/en/master/api_python/samples/ops/communicate_ops.html#notes>`_.
 
     Args:
         split_count (int): On each process, divide blocks into split_count number.
@@ -845,14 +835,14 @@ class AlltoAll(PrimitiveWithInfer):
 
             For the Ascend devices, users need to prepare the rank table, set rank_id and device_id.
             Please see the `rank table Startup
-            <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/rank_table.html>`_
+            <https://www.mindspore.cn/tutorials/experts/en/master/parallel/rank_table.html>`_
             for more details.
 
             For the GPU devices, users need to prepare the host file and mpi, please see the `mpirun Startup
-            <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/mpirun.html>`_ .
+            <https://www.mindspore.cn/tutorials/experts/en/master/parallel/mpirun.html>`_ .
 
             For the CPU device, users need to write a dynamic cluster startup script, please see the `Dynamic Cluster
-            Startup <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/dynamic_cluster.html>`_ .
+            Startup <https://www.mindspore.cn/tutorials/experts/en/master/parallel/dynamic_cluster.html>`_ .
 
             This example should be run with 8 devices.
 
@@ -883,7 +873,7 @@ class AlltoAll(PrimitiveWithInfer):
 
     Tutorial Examples:
         - `Distributed Set Communication Primitives - AlltoAll
-          <https://www.mindspore.cn/docs/en/r2.2/api_python/samples/ops/communicate_ops.html#alltoall>`_
+          <https://www.mindspore.cn/docs/en/master/api_python/samples/ops/communicate_ops.html#alltoall>`_
 
     """
 
@@ -931,7 +921,7 @@ class NeighborExchangeV2(Primitive):
     Note:
         This operator requires a full-mesh network topology, each device has the same vlan id, and the ip & mask are
         in the same subnet, please check the `details \
-        <https://www.mindspore.cn/docs/en/r2.2/api_python/samples/ops/communicate_ops.html#notes>`_.
+        <https://www.mindspore.cn/docs/en/master/api_python/samples/ops/communicate_ops.html#notes>`_.
 
     Args:
         send_rank_ids (list(int)): Ranks which the data is sent to. 8 rank_ids represents 8 directions, if one
@@ -969,14 +959,14 @@ class NeighborExchangeV2(Primitive):
 
             For the Ascend devices, users need to prepare the rank table, set rank_id and device_id.
             Please see the `rank table Startup
-            <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/rank_table.html>`_
+            <https://www.mindspore.cn/tutorials/experts/en/master/parallel/rank_table.html>`_
             for more details.
 
             For the GPU devices, users need to prepare the host file and mpi, please see the `mpirun Startup
-            <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/mpirun.html>`_ .
+            <https://www.mindspore.cn/tutorials/experts/en/master/parallel/mpirun.html>`_ .
 
             For the CPU device, users need to write a dynamic cluster startup script, please see the `Dynamic Cluster
-            Startup <https://www.mindspore.cn/tutorials/experts/en/r2.2/parallel/dynamic_cluster.html>`_ .
+            Startup <https://www.mindspore.cn/tutorials/experts/en/master/parallel/dynamic_cluster.html>`_ .
 
             This example should be run with 2 devices.
 
@@ -998,7 +988,6 @@ class NeighborExchangeV2(Primitive):
         ...     def construct(self, x):
         ...         out = self.neighbor_exchangev2(x)
         ...         return out
-        >>>
         ... class Net1(nn.Cell):
         ...     def __init__(self):
         ...         super(Net1, self).__init__()
@@ -1028,7 +1017,7 @@ class NeighborExchangeV2(Primitive):
 
     Tutorial Examples:
         - `Distributed Set Communication Primitives - NeighborExchangeV2
-          <https://www.mindspore.cn/docs/en/r2.2/api_python/samples/ops/communicate_ops.html#neighborexchangev2>`_
+          <https://www.mindspore.cn/docs/en/master/api_python/samples/ops/communicate_ops.html#neighborexchangev2>`_
 
     """
 
@@ -1306,8 +1295,11 @@ class _GetTensorSlice(PrimitiveWithInfer):
         """Initialize _GetTensorSlice."""
         self.add_prim_attr('order_enforce_skip', True)
 
-    def infer_value(self, x, dev_mat, tensor_map):
+    def infer_value(self, x, dev_mat, tensor_map, slice_shape, full_shape):
         from mindspore.parallel._tensor import _load_tensor
         validator.check_value_type("dev_mat", dev_mat, [tuple], self.name)
         validator.check_value_type("tensor_map", tensor_map, [tuple], self.name)
-        return Tensor(_load_tensor(x, dev_mat, tensor_map), x.dtype)
+        tensor_slice = _load_tensor(x, dev_mat, tensor_map, full_shape)
+        if tensor_slice.shape != slice_shape:
+            tensor_slice = tensor_slice.reshape(slice_shape)
+        return Tensor(tensor_slice, x.dtype)

mindspore/ops/operations/custom_ops.py

@@ -1,4 +1,4 @@
-# Copyright 2021-2022 Huawei Technologies Co., Ltd
+# Copyright 2021-2024 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.
@@ -164,7 +164,7 @@ class Custom(ops.PrimitiveWithInfer):
     function if needed. Then these `Custom` objects can be directly used in neural networks.
     Detailed description and introduction of user-defined operators, including correct writing of parameters,
     please refer to `Custom Operators Tutorial
-    <https://www.mindspore.cn/tutorials/experts/en/r2.2/operation/op_custom.html>`_ .
+    <https://www.mindspore.cn/tutorials/experts/en/master/operation/op_custom.html>`_ .
 
     .. warning::
         - This is an experimental API that is subject to change.
@@ -172,19 +172,11 @@ class Custom(ops.PrimitiveWithInfer):
     .. note::
         The supported platforms are determined by the input `func_type`. The supported platforms are as follows:
 
-        - "hybrid": supports ["Ascend", "GPU", "CPU"].
-        - "akg": supports ["Ascend", "GPU", "CPU"].
-        - "tbe": supports ["Ascend"].
+        - "hybrid": supports ["GPU", "CPU"].
+        - "akg": supports ["GPU", "CPU"].
         - "aot": supports ["GPU", "CPU"].
         - "pyfunc": supports ["CPU"].
         - "julia": supports ["CPU"].
-        - "aicpu": supports ["Ascend"].
-
-        If run on ge backend, use `CustomRegOp` to generate the registration information of "aicpu" and "tbe" operator,
-        use `custom_info_register` to bind the registration information to the `func` of the "tbe" operator,
-        then save the registration information of "aicpu" operator and the `func` implementation of "tbe" operator to
-        a file or separate files, keep these files in a separate directory, and set the absolute path of this directory
-        to environment variable "MS_DEV_CUSTOM_OPP_PATH" before running the network.
 
     Args:
         func (Union[function, str]):
@@ -193,9 +185,8 @@ class Custom(ops.PrimitiveWithInfer):
               computation logic of a user defined operator. The function can be one of the following:
 
               1. A AKG operator implementation function, which can use ir builder/tvm compute/hybrid grammar.
-              2. A TBE operator implementation function.
-              3. A pure python function
-              4. An kernel decorated function written by the Hybrid DSL.
+              2. A pure python function
+              3. An kernel decorated function written by the Hybrid DSL.
 
             - str: If func is of str type, then str should be a path of file along with a function name.
               This could be used when func_type is "aot" or "julia".
@@ -299,18 +290,7 @@ class Custom(ops.PrimitiveWithInfer):
 
         func_type (str): The implementation type of `func`, should be one of
 
-            [ ``"hybrid"`` , ``"akg"`` , ``"tbe"`` , ``"aot"`` , ``"pyfunc"`` , ``"julia"`` , ``"aicpu"`` ].
-
-            Each `func_type` only supports specific platforms(targets). Default: ``"hybrid"`` .
-            The supported platforms of `func_type`:
-
-            - ``"hybrid"``: supports ["Ascend", "GPU", "CPU"].
-            - ``"akg"``: supports ["Ascend", "GPU", "CPU"].
-            - ``"tbe"``: supports ["Ascend"].
-            - ``"aot"``: supports ["GPU", "CPU"].
-            - ``"pyfunc"``: supports ["CPU"].
-            - ``"julia"``: supports ["CPU"].
-            - ``"aicpu"``: supports ["Ascend"].
+            [ ``"hybrid"`` , ``"akg"`` , ``"aot"`` , ``"pyfunc"`` , ``"julia"`` ].
 
         bprop (function): The back propagation function of `func`. Default: ``None`` .
         reg_info (Union[str, dict, list, tuple]): Represents the registration information(reg info) of `func` with
@@ -343,7 +323,7 @@ class Custom(ops.PrimitiveWithInfer):
             or the attributes of `func` differs in different targets.
 
     Supported Platforms:
-        ``Ascend`` ``GPU`` ``CPU``
+        ``GPU`` ``CPU``
 
     Examples:
         >>> import numpy as np
@@ -372,68 +352,6 @@ class Custom(ops.PrimitiveWithInfer):
         >>> # the result will be a 16 * 16 tensor with all elements 2
         >>> print(output.shape)
         (16, 16)
-        >>> # Example, func_type = "tbe"
-        >>> square_with_bias_op_info = CustomRegOp() \
-        ...     .fusion_type("OPAQUE") \
-        ...     .attr("bias", "required", "float") \
-        ...     .input(0, "x") \
-        ...     .output(0, "y") \
-        ...     .dtype_format(DataType.F32_Default, DataType.F32_Default) \
-        ...     .dtype_format(DataType.F16_Default, DataType.F16_Default) \
-        ...     .target("Ascend") \
-        ...     .get_op_info()
-        >>>
-        >>> @custom_info_register(square_with_bias_op_info)
-        ... def square_with_bias(input_x, output_y, bias=0.0, kernel_name="square_with_bias"):
-        ...     import te.lang.cce
-        ...     from te import tvm
-        ...     from topi.cce import util
-        ...
-        ...     shape = input_x.get("shape")
-        ...     dtype = input_x.get("dtype").lower()
-        ...
-        ...     shape = util.shape_refine(shape)
-        ...     data = tvm.placeholder(shape, name="data", dtype=dtype)
-        ...
-        ...     with tvm.target.cce():
-        ...         res0 = te.lang.cce.vmul(data, data)
-        ...         res = te.lang.cce.vadds(res0, bias)
-        ...         sch = te.lang.cce.auto_schedule(res)
-        ...
-        ...     config = {"print_ir": False,
-        ...               "name": kernel_name,
-        ...               "tensor_list": [data, res]}
-        ...
-        ...     te.lang.cce.cce_build_code(sch, config)
-        >>>
-        >>> def test_tbe():
-        ...     square_with_bias = ops.Custom(square_with_bias, out_shape=lambda x, _: x, \
-        ...                                   out_dtype=lambda x, _: x, func_type="tbe")
-        ...     res = self.square_with_bias(input_x, 1.0)
-        ...     return res
-        >>>
-        >>> # Example, func_type = "aicpu"
-        >>> resize_bilinear_op_info = CustomRegOp("ResizeBilinear") \
-        ...     .fusion_type("OPAQUE") \
-        ...     .input(0, "input", "required") \
-        ...     .output(1, "output", "required") \
-        ...     .attr("align_corners", "required", "bool") \
-        ...     .attr("cust_aicpu", "optional", "str", "aicpu_kernels") \
-        ...     .dtype_format(DataType.F32_Default, DataType.F32_Default) \
-        ...     .dtype_format(DataType.F16_Default, DataType.F32_Default) \
-        ...     .target("Ascend") \
-        ...     .get_op_info()
-        >>>
-        >>> @custom_info_register(resize_bilinear_op_info)
-        ... def resize_bilinear_aicpu():
-        ...     return
-        >>>
-        >>> def test_aicpu(x):
-        ...     resize_bilinear_op = ops.Custom(resize_bilinear_aicpu, out_shape=[1, 1, 9, 9], \
-        ...                                     out_dtype=mstype.float32, func_type="aicpu")
-        ...     res = resize_bilinear_op(x, True, "aicpu_kernels")
-        ...     return res
-        >>>
         >>> # Example, func_type = "aot"
         >>> def test_aot(x, y, out_shapes, out_types):
         ...     program = ops.Custom("./reorganize.so:CustomReorganize", out_shapes, out_types, "aot")
@@ -541,9 +459,9 @@ class Custom(ops.PrimitiveWithInfer):
         # deal with the case of ms script
         # enable auto infer function if any infer information is missing
         if self._is_ms_kernel and (infer_dtype is None or infer_shape is None):
-            logger.warning("{}, 'out_shape' or 'out_dtype' is None, infer the output shape and output dtype "
-                           "automatically. There might be some Python RuntimeWarning but it wouldn't influence the "
-                           "result.".format(self.log_prefix))
+            logger.info("{}, 'out_shape' or 'out_dtype' is None, infer the output shape and output dtype "
+                        "automatically. There might be some Python RuntimeWarning but it wouldn't influence the "
+                        "result.".format(self.log_prefix))
 
             auto_infer_result = self._auto_infer(*args)
 
@@ -634,9 +552,9 @@ class Custom(ops.PrimitiveWithInfer):
             file_path = os.path.abspath(file_name_list[0])
             if os.environ.get('MS_CUSTOM_AOT_WHITE_LIST') is None:
                 if Custom.custom_aot_warning:
-                    logger.warning("{}, no white list is set and it might cause problems. "
-                                   "Set the legal path of the file in MS_CUSTOM_AOT_WHITE_LIST"
-                                   .format(self.log_prefix))
+                    logger.info("{}, no white list is set and it might cause problems. "
+                                "Set the legal path of the file in MS_CUSTOM_AOT_WHITE_LIST"
+                                .format(self.log_prefix))
                     Custom.custom_aot_warning = False
             else:
                 legal_path = os.path.abspath(os.environ.get('MS_CUSTOM_AOT_WHITE_LIST'))

mindspore/ops/operations/debug_ops.py

@@ -13,8 +13,11 @@
 # limitations under the License.
 # ============================================================================
 """debug_ops"""
+import os
+import stat
 from types import FunctionType, MethodType
 
+import numpy as np
 from mindspore import log as logger
 from mindspore._c_expression import security
 from mindspore._c_expression import Tensor as Tensor_
@@ -26,6 +29,7 @@ from mindspore.ops.primitive import prim_attr_register, Primitive, PrimitiveWith
 
 
 SUMMARY_TENSOR_CACHE = []
+TENSORDUMP_ID = 0
 
 
 def _cache_summary_data(op_name, define_name, tensor):
@@ -55,10 +59,11 @@ SUMMARY_RETURN_VALUE = {'dtype': mstype.int32, 'shape': [1], 'value': None}
 
 class ScalarSummary(Primitive):
     """
-    This operator will put a scalar to a summary file with protocol buffer format. It must be used with SummaryRecord
-    or SummaryCollector, which specify the directory of the summary file. The summary file can
+    This operator will put a scalar to a summary file with protocol buffer format.
+    It must be used with :class:`mindspore.SummaryRecord` or :class:`mindspore.SummaryCollector`,
+    which specify the directory of the summary file. The summary file can
     be loaded and shown by MindInsight, see `MindInsight documents <https://www.mindspore.cn/
-    mindinsight/docs/en/r2.2/index.html>`_ for details.
+    mindinsight/docs/en/master/index.html>`_ for details.
 
     Inputs:
         - **name** (str) - The name of the input variable, it must not be an empty string.
@@ -105,17 +110,18 @@ class ScalarSummary(Primitive):
 
         self.add_prim_attr("side_effect_io", True)
         self.add_prim_attr("channel_name", "ms_scalar_summary")
+        self.add_prim_attr("dyn_input_sizes", [-1, 1])
 
     def __call__(self, *args):
         _cache_summary_data(self.name, args[0], args[1])
 
 
-class ImageSummary(PrimitiveWithInfer):
+class ImageSummary(Primitive):
     """
     This operator will put an image tensor to a summary file with protocol buffer format. It must be used with
     SummaryRecord or SummaryCollector, which specify the directory of the summary file. The summary file can
     be loaded and shown by MindInsight, see `MindInsight documents <https://www.mindspore.cn/
-    mindinsight/docs/en/r2.2/index.html>`_ for details.
+    mindinsight/docs/en/master/index.html>`_ for details.
 
     Inputs:
         - **name** (str) - The name of the input variable, it must not be an empty string.
@@ -155,18 +161,7 @@ class ImageSummary(PrimitiveWithInfer):
 
         self.add_prim_attr("side_effect_io", True)
         self.add_prim_attr("channel_name", "ms_image_summary")
-
-    def __infer__(self, name, value):
-        _check_summary_param(name, value, self.__class__.__name__)
-
-        # The shape dim of image should be 4.
-        v_shape = value['shape']
-        image_dim = 4
-        if len(v_shape) != image_dim:
-            raise ValueError(f"For '{self.name}', the dimension of 'value' must be {image_dim},"
-                             f" but got {len(v_shape)}.")
-
-        return SUMMARY_RETURN_VALUE
+        self.add_prim_attr("dyn_input_sizes", [-1, 1])
 
     def __call__(self, *args):
         _cache_summary_data(self.name, args[0], args[1])
@@ -177,7 +172,7 @@ class TensorSummary(Primitive):
     This operator will put a tensor to a summary file with protocol buffer format. It must be used with SummaryRecord
     or SummaryCollector, which specify the directory of the summary file. The summary file can
     be loaded and shown by MindInsight, see `MindInsight documents <https://www.mindspore.cn/
-    mindinsight/docs/en/r2.2/index.html>`_ for details.
+    mindinsight/docs/en/master/index.html>`_ for details.
 
     Inputs:
         - **name** (str) - The name of the input variable.
@@ -229,12 +224,96 @@ class TensorSummary(Primitive):
         _cache_summary_data(self.name, args[0], args[1])
 
 
-class HistogramSummary(PrimitiveWithInfer):
+class TensorDump(Primitive):
+    """
+    Save the Tensor as an npy file in numpy format.
+
+    The file name will automatically have a prefix added based on the execution order. For example, if `file` is `a`,
+    the first saved file will be named `0_a.npy`, and the second one will be named `1_a.npy`, and so on.
+
+    .. warning::
+        - If a large amount of data is stored within a short period, it may lead to memory overflow on the device side.
+          Consider slicing the data to reduce the data scale.
+        - Since data saving is processed asynchronously, when the amount of data is too large or the main process exits
+          too quickly, data loss may occur. You need to actively control the destruction time of the main process,
+          such as using sleep.
+
+    Inputs:
+        - **file** (str) - The path of the file to be saved.
+        - **input_x** (Tensor) - Input Tensor of any dimension.
+
+    Raises:
+        TypeError: If `file` is not a str.
+        TypeError: If `input_x` is not a Tensor.
+
+    Supported Platforms:
+        ``Ascend``
+
+    Examples:
+        >>> import numpy as np
+        >>> import mindspore as ms
+        >>> import time
+        >>> from mindspore import nn, Tensor, ops
+        >>> ms.set_context(mode=ms.GRAPH_MODE, device_target="Ascend")
+        >>> class Net(nn.Cell):
+        ...     def __init__(self):
+        ...         super(Net, self).__init__()
+        ...         self.dump = ops.TensorDump()
+        ...
+        ...     def construct(self, x):
+        ...         x += 1.
+        ...         self.dump('add', x)
+        ...         x /= 2.
+        ...         self.dump('div', x)
+        ...         x *= 5.
+        ...         self.dump('mul', x)
+        ...         return x
+        ...
+        >>> x = np.array([[1, 2, 3, 4], [5, 6, 7, 8]]).astype(np.float32)
+        >>> input_x = Tensor(x)
+        >>> net = Net()
+        >>> out = net(input_x)
+        >>> time.sleep(0.5)
+        >>> add = np.load('0_add.npy')
+        >>> print(add)
+        [[2. 3. 4. 5.]
+         [6. 7. 8. 9.]]
+    """
+    @prim_attr_register
+    def __init__(self):
+        """Initialize TensorDump."""
+        if security.enable_security():
+            raise ValueError('The TensorDump is not supported, please without `-s on` and recompile source.')
+        self.add_prim_attr("side_effect_io", True)
+        self.add_prim_attr("channel_name", "ms_tensor_dump")
+
+    def __call__(self, file, input_x):
+        validator.check_value_type('file', file, [str], self.__class__.__name__)
+        if not file:
+            raise ValueError("For 'TensorDump', the input argument[file] cannot be an empty string.")
+        validator.check_value_type('input_x', input_x, [Tensor], self.__class__.__name__)
+        global TENSORDUMP_ID
+        npy_suffix = ".npy"
+        directory, filename = os.path.split(file)
+        if directory and not os.path.exists(directory):
+            os.makedirs(directory, mode=0o700, exist_ok=True)
+        new_filename = f"{TENSORDUMP_ID}_{filename}"
+        new_file = os.path.join(directory, new_filename)
+        if not new_file.endswith(npy_suffix):
+            new_file += npy_suffix
+        if os.path.exists(new_file):
+            os.chmod(new_file, stat.S_IWUSR)
+        np.save(new_file, input_x.asnumpy())
+        os.chmod(new_file, stat.S_IRUSR)
+        TENSORDUMP_ID += 1
+
+
+class HistogramSummary(Primitive):
     """
     This operator will calculate the histogram of a tensor and put it to a summary file with protocol buffer format.
     It must be used with SummaryRecord or SummaryCollector, which specify the directory of the summary file.
     The summary file can be loaded and shown by MindInsight, see `MindInsight documents <https://www.mindspore.cn/
-    mindinsight/docs/en/r2.2/index.html>`_ for details.
+    mindinsight/docs/en/master/index.html>`_ for details.
 
     Inputs:
         - **name** (str) - The name of the input variable.
@@ -280,17 +359,7 @@ class HistogramSummary(PrimitiveWithInfer):
 
         self.add_prim_attr("side_effect_io", True)
         self.add_prim_attr("channel_name", "ms_histogram_summary")
-
-    def __infer__(self, name, value):
-        _check_summary_param(name, value, self.__class__.__name__)
-
-        v_shape = value['shape']
-        # In the summary, the histogram value should be a tensor whose shape is not [].
-        if not v_shape:
-            raise ValueError(f"For '{self.name}', the type of 'value' must be tensor, "
-                             f"its shape should not be [], but got {v_shape}.")
-
-        return SUMMARY_RETURN_VALUE
+        self.add_prim_attr("dyn_input_sizes", [-1, 1])
 
     def __call__(self, *args):
         _cache_summary_data(self.name, args[0], args[1])