mindspore 2.3.0rc1__cp38-cp38-manylinux1_x86_64.whl → 2.3.0rc2__cp38-cp38-manylinux1_x86_64.whl

mindspore/parallel/checkpoint_transform.py

@@ -26,18 +26,18 @@ from mindspore.parallel._utils import _is_in_auto_parallel_mode
 from mindspore.parallel._parallel_serialization import _rank_list_for_transform_parallel_checkpoint, \
     _transform_parallel_checkpoint, _get_device_num_from_strategy, _make_dir, \
     _extract_layout_map, _extract_src_dst_layout_map, _parameter_not_in_local_stage, _extract_pipeline_stage_num, \
-    _merge_protobuf_strategy, _merge_json_strategy
+    _merge_protobuf_strategy, _merge_json_strategy, _extract_src_dst_layout_map_by_src
 
 
 __all__ = ["merge_pipeline_strategys", "rank_list_for_transform", "transform_checkpoint_by_rank",
-           "transform_checkpoints", "sync_pipeline_shared_parameters"]
+           "transform_checkpoints", "sync_pipeline_shared_parameters", "load_segmented_checkpoints"]
 
 
 def merge_pipeline_strategys(src_strategy_dirs, dst_strategy_file):
     """
     Merge parallel strategy between all pipeline stages in pipeline parallel mode.
     For more details about converting distributed Checkpoint, please refer to
-    `Model Transformation <https://www.mindspore.cn/tutorials/experts/en/r2.3.q1/parallel/model_transformation.html>`_.
+    `Model Transformation <https://www.mindspore.cn/tutorials/experts/en/master/parallel/model_transformation.html>`_.
 
     Note:
         Strategy file of each pipeline stage should be included in src_strategy_dirs.
@@ -77,7 +77,7 @@ def rank_list_for_transform(rank_id, src_strategy_file=None, dst_strategy_file=N
     """
     List of original distributed checkpoint rank index for obtaining the target checkpoint of a rank_id during the
     distributed checkpoint conversion. For more details about converting distributed Checkpoint, please refer to
-    `Model Transformation <https://www.mindspore.cn/tutorials/experts/en/r2.3.q1/parallel/model_transformation.html>`_.
+    `Model Transformation <https://www.mindspore.cn/tutorials/experts/en/master/parallel/model_transformation.html>`_.
 
     Args:
         rank_id (int): The rank of which distributed checkpoint needs to be obtained after conversion.
@@ -140,7 +140,7 @@ def transform_checkpoint_by_rank(rank_id, checkpoint_files_map, save_checkpoint_
     """
     Transform distributed checkpoint from source sharding strategy to destination sharding strategy by rank
     for a network. For more details about converting distributed Checkpoint, please refer to
-    `Model Transformation <https://www.mindspore.cn/tutorials/experts/en/r2.3.q1/parallel/model_transformation.html>`_.
+    `Model Transformation <https://www.mindspore.cn/tutorials/experts/en/master/parallel/model_transformation.html>`_.
 
     Args:
         rank_id (int): The rank of which distributed checkpoint needs to be obtained after conversion.
@@ -225,49 +225,63 @@ def transform_checkpoint_by_rank(rank_id, checkpoint_files_map, save_checkpoint_
     ms.save_checkpoint(transform_param_list, save_checkpoint_file_name)
 
 
-def transform_checkpoints(src_checkpoints_dir, dst_checkpoints_dir, ckpt_prefix, src_strategy_file=None,
-                          dst_strategy_file=None):
-    """
-    Transform distributed checkpoint from source sharding strategy to destination sharding strategy for a rank.
-    For more details about converting distributed Checkpoint, please refer to
-    `Model Transformation <https://www.mindspore.cn/tutorials/experts/en/r2.3.q1/parallel/model_transformation.html>`_.
-
-    Note:
-        The `src_checkpoints_dir` directory structure should be organized like "src_checkpoints_dir/rank_0/a.ckpt", the
-        rank number should be set to a subdirectory and the checkpoint file is stored in this subdirectory. If multiple
-        files exist in a rank directory, the last file in the lexicgraphic order would be selected.
-
-    Args:
-        src_checkpoints_dir (str): The source checkpoints directory.
-        dst_checkpoints_dir (str): The destination checkpoints directory to save the converted checkpoints.
-        ckpt_prefix (str): The destination checkpoint name prefix.
-        src_strategy_file (str): Name of source sharding strategy file which saved by
-                                 'mindspore.set_auto_parallel_context(strategy_ckpt_save_file)'.
-                                 when the `src_strategy_file` is ``None``, it means that the source sharding strategy is
-                                 without any sharing for each parameter. Default: ``None``.
-        dst_strategy_file (str): Name of destination sharding strategy file which saved by
-                                 'mindspore.set_auto_parallel_context(strategy_ckpt_save_file)'.
-                                 when the `dst_strategy_file` is ``None``,
-                                 it means that the destination sharding strategy
-                                 is without any sharing for each parameter. Default: ``None``.
-
-    Raises:
-        ValueError: `src_strategy_file` or `dst_strategy_file` is incorrect.
-        NotADirectoryError: `src_checkpoints_dir` or `dst_checkpoints_dir` is not a directory.
-        ValueError: The checkpoint file is missing in `src_checkpoints_dir`.
-        TypeError: `src_strategy_file` or `dst_strategy_file` is not a string.
-
-    Examples:
-        >>> import mindspore as ms
-        >>> ms.transform_checkpoints(src_checkpoints_dir, dst_checkpoints_dir, "dst_checkpoint",
-        ...                       "./src_strategy.ckpt", "./dst_strategy.ckpt")
-
-    """
-    if not os.path.isdir(src_checkpoints_dir):
-        raise NotADirectoryError("src_checkpoints_dir {} is not a directory.".format(src_checkpoints_dir))
-    _make_dir(dst_checkpoints_dir, "path")
-    if not isinstance(ckpt_prefix, str):
-        raise TypeError("The ckpt_prefix should be a str.")
+def _transform_checkpoint_by_stage(src_checkpoints_dir, dst_checkpoints_dir, ckpt_prefix, src_strategy_file,
+                                   dst_strategy_file=None):
+    """Transform checkpoint for stage in src_strategy_file"""
+    param_total_dict = defaultdict(dict)
+    param_attr_dict = defaultdict(dict)
+    param_type_dict = defaultdict(dict)
+    src_strategy_list, dst_strategy_list, stage_id = _extract_src_dst_layout_map_by_src(src_strategy_file, \
+                                                                                             dst_strategy_file)
+    src_stage_device_num = np.prod(src_strategy_list.get(list(src_strategy_list.keys())[0])[0]) if src_strategy_list \
+                                                                                                   is not None else 1
+    dst_stage_device_num = np.prod(dst_strategy_list.get(list(dst_strategy_list.keys())[0])[0]) if dst_strategy_list \
+                                                                                                   is not None else 1
+    origin_dst_strategy_list = _extract_layout_map(dst_strategy_file)
+    origin_src_strategy_list = _extract_layout_map(src_strategy_file)
+    checkpoint_files_map = {}
+    src_rank_id_start = stage_id * src_stage_device_num
+    for local_rank in range(src_stage_device_num):
+        rank_id = src_rank_id_start + local_rank
+        checkpoint_file_name = os.path.join(src_checkpoints_dir, "rank_{}".format(rank_id), "*.ckpt")
+        rank_ckpts = glob.glob(checkpoint_file_name)
+        rank_ckpts.sort()
+        for checkpoint_file in rank_ckpts:
+            if not os.path.isfile(checkpoint_file):
+                ms.log.warning("{} is not a checkpoint file.".format(checkpoint_file))
+                continue
+            checkpoint_files_map[rank_id] = checkpoint_file
+    for rank, local_file in checkpoint_files_map.items():
+        if not os.path.exists(local_file):
+            raise ValueError("Checkpoint file {} in rank {} not exits: ".format(local_file, rank))
+    for rank, file_name in checkpoint_files_map.items():
+        ckpt_dict = ms.load_checkpoint(file_name)
+        for param_name, param in ckpt_dict.items():
+            # cut the parameter not in the pipeline stage.
+            if _parameter_not_in_local_stage(param_name, origin_src_strategy_list, src_strategy_list) \
+                    and _parameter_not_in_local_stage(param_name, origin_dst_strategy_list, dst_strategy_list):
+                continue
+            src_rank = rank % src_stage_device_num
+            param_type_dict[param_name][src_rank] = str(param.data.dtype)
+            if param.data.dtype == mstype.bfloat16:
+                param.set_dtype(mstype.float32)
+            param_total_dict[param_name][src_rank] = param.data.asnumpy()
+            param_attr_dict[param_name][src_rank] = (param.requires_grad, param.layerwise_parallel)
+    for local_rank_id in range(dst_stage_device_num):
+        transform_param_list = _transform_parallel_checkpoint(local_rank_id, param_total_dict,
+                                                              param_attr_dict, src_strategy_list, dst_strategy_list,
+                                                              param_type_dict)
+        save_checkpoint_file = "{}{}_part{}.ckpt".format(ckpt_prefix, local_rank_id, stage_id)
+        save_checkpoint_file_dir = os.path.join(dst_checkpoints_dir, "rank_{}".format(local_rank_id))
+        if not os.path.exists(save_checkpoint_file_dir):
+            _make_dir(save_checkpoint_file_dir, "path")
+        save_checkpoint_file_name = os.path.join(save_checkpoint_file_dir, save_checkpoint_file)
+        ms.save_checkpoint(transform_param_list, save_checkpoint_file_name)
+
+
+def _transform_checkpoints(src_checkpoints_dir, dst_checkpoints_dir, ckpt_prefix, src_strategy_file=None,
+                           dst_strategy_file=None):
+    """Transform checkpoints for all stages in src_strategy_file"""
     checkpoints_rank_dir_list = os.path.join(src_checkpoints_dir, "rank_[0-9]*")
     all_checkpoint_files_map = {}
     for checkpoint_dir in glob.glob(checkpoints_rank_dir_list):
@@ -342,6 +356,76 @@ def transform_checkpoints(src_checkpoints_dir, dst_checkpoints_dir, ckpt_prefix,
         del param_total_dict
 
 
+def transform_checkpoints(src_checkpoints_dir, dst_checkpoints_dir, ckpt_prefix, src_strategy_file=None,
+                          dst_strategy_file=None):
+    """
+    Transform distributed checkpoint from source sharding strategy to destination sharding strategy for a rank.
+    For more details about converting distributed Checkpoint, please refer to
+    `Model Transformation <https://www.mindspore.cn/tutorials/experts/en/master/parallel/model_transformation.html>`_.
+
+    Note:
+        The `src_checkpoints_dir` directory structure should be organized like "src_checkpoints_dir/rank_0/a.ckpt", the
+        rank number should be set to a subdirectory and the checkpoint file is stored in this subdirectory. If multiple
+        files exist in a rank directory, the last file in the lexicgraphic order would be selected.
+
+    Args:
+        src_checkpoints_dir (str): The source checkpoints directory.
+        dst_checkpoints_dir (str): The destination checkpoints directory to save the converted checkpoints.
+        ckpt_prefix (str): The destination checkpoint name prefix.
+        src_strategy_file (str): Name of source sharding strategy file which saved by
+                                 'mindspore.set_auto_parallel_context(strategy_ckpt_save_file)'.
+                                 when the 'src_strategy_file' is None, it means that the source sharding strategy is
+                                 without any sharing for each parameter. Default:None.
+        dst_strategy_file (str): Name of destination sharding strategy file which saved by
+                                 'mindspore.set_auto_parallel_context(strategy_ckpt_save_file)'.
+                                 when the 'dst_strategy_file' is None, it means that the destination sharding strategy
+                                 is without any sharing for each parameter. Default:None.
+
+    Raises:
+        ValueError: `src_strategy_file` or `dst_strategy_file` is incorrect.
+        NotADirectoryError: `src_checkpoints_dir` or `dst_checkpoints_dir` is not a directory.
+        ValueError: The checkpoint file is missing in `src_checkpoints_dir`.
+        TypeError: `src_strategy_file` or `dst_strategy_file` is not a string.
+
+    Examples:
+        >>> import mindspore as ms
+        >>> ms.transform_checkpoints(src_checkpoints_dir, dst_checkpoints_dir, "dst_checkpoint",
+        ...                       "./src_strategy.ckpt", "./dst_strategy.ckpt")
+
+    """
+    if not os.path.isdir(src_checkpoints_dir):
+        raise NotADirectoryError("src_checkpoints_dir {} is not a directory.".format(src_checkpoints_dir))
+    _make_dir(dst_checkpoints_dir, "path")
+    if not isinstance(ckpt_prefix, str):
+        raise TypeError("The ckpt_prefix should be a str.")
+    if src_strategy_file and os.path.dirname(src_strategy_file) and not os.path.exists(
+            os.path.dirname(src_strategy_file)):
+        raise ValueError("The director of src_strategy_file: {} is not exists.".
+                         format(os.path.dirname(src_strategy_file)))
+    if dst_strategy_file and os.path.dirname(dst_strategy_file) and not os.path.exists(
+            os.path.dirname(dst_strategy_file)):
+        raise ValueError("The director of dst_strategy_file: {} is not exists.".
+                         format(os.path.dirname(dst_strategy_file)))
+    src_layout_map = _extract_layout_map(src_strategy_file)
+    dst_layout_map = _extract_layout_map(dst_strategy_file)
+    pipeline_stage_num = _extract_pipeline_stage_num(src_strategy_file)
+    if src_layout_map:
+        src_param_keys = {param_name for param_name in src_layout_map if not param_name.startswith("accu_grads")}
+    if dst_layout_map:
+        dst_param_keys = {param_name for param_name in dst_layout_map if not param_name.startswith("accu_grads")}
+    if src_layout_map and dst_layout_map and pipeline_stage_num == 1 \
+        and src_param_keys.issubset(dst_param_keys) and len(src_param_keys) < len(dst_param_keys):
+        dst_stage_num = _extract_pipeline_stage_num(dst_strategy_file)
+        if dst_stage_num > 1:
+            raise NotImplementedError("When using unmerged src strategy, dst strategy doesn't \
+                                       support strategy with pipeline parallel.")
+        _transform_checkpoint_by_stage(src_checkpoints_dir, dst_checkpoints_dir, ckpt_prefix,
+                                       src_strategy_file, dst_strategy_file)
+    else:
+        _transform_checkpoints(src_checkpoints_dir, dst_checkpoints_dir, ckpt_prefix,
+                               src_strategy_file, dst_strategy_file)
+
+
 def _sync_params(name, param, layout):
     """synchronize single parameter"""
     if len(layout) < 10:
@@ -394,53 +478,79 @@ def sync_pipeline_shared_parameters(net):
     Args:
         net (nn.Cell): the inference network.
 
+    Supported Platforms:
+        ``Ascend``
+
     Examples:
+        .. note::
+            Before running the following examples, you need to configure the communication environment variables.
+
+            For the Ascend device, users need to write a dynamic cluster startup script, please see the `Dynamic Cluster
+            Startup <https://www.mindspore.cn/tutorials/experts/en/master/parallel/dynamic_cluster.html>`_ .
+
         >>> import numpy as np
         >>> import mindspore as ms
-        >>> from mindspore import nn, ops, Parameter, Tensor
-        >>> class VocabEmbedding(nn.Cell):
-        ...     def __init__(self, vocab_size, embedding_size):
+        >>> import mindspore.communication.management as D
+        >>> from mindspore import lazy_inline, context, nn, ops, Parameter, Tensor
+        >>> context.set_context(mode=context.GRAPH_MODE)
+        >>> class Embedding(nn.Cell):
+        ...     def __init__(self, shape):
         ...         super().__init__()
-        ...         self.embedding_table = Parameter(Tensor(np.ones([vocab_size, embedding_size]), ms.float32),
-        ...                                          name='embedding')
-        ...         self.gather = ops.Gather()
-        ...
+        ...         self.w = Parameter(Tensor(np.ones(shape), ms.float32), name='w')
+        ...         self.matmul = ops.MatMul().shard(((1, 1), (1, 1)))
         ...     def construct(self, x):
-        ...         output = self.gather(self.embedding_table, x, 0)
-        ...         output = output.squeeze(1)
-        ...         return output, self.embedding_table.value()
+        ...         return self.matmul(x, self.w), self.w
         ...
         >>> class LMHead(nn.Cell):
         ...     def __init__(self):
         ...         super().__init__()
-        ...         self.matmul = ops.MatMul(transpose_b=True)
-        ...
-        ...     def construct(self, state, embed):
-        ...         return self.matmul(state, embed)
+        ...         self.matmul = ops.MatMul(transpose_b=True).shard(((1, 1), (1, 1)))
+        ...     def construct(self, x, w):
+        ...         return self.matmul(x, w)
         ...
         >>> class Network(nn.Cell):
         ...     @lazy_inline
         ...     def __init__(self):
         ...         super().__init__()
-        ...         self.word_embedding = VocabEmbedding(vocab_size=4, embedding_size=4)
-        ...         self.head = LMHead()
-        ...
+        ...         shape = (4, 4)
+        ...         self.word_embedding = Embedding(shape)
+        ...         self.lm_head = LMHead()
+        ...         self.word_embedding.pipeline_stage = 0
+        ...         self.lm_head.pipeline_stage = 1
         ...     def construct(self, x):
         ...         x, embed = self.word_embedding(x)
-        ...         x = self.head(x, embed)
-        ...         return x
-        >>>
+        ...         return self.lm_head(x, embed)
+        ...
+        >>> class PipelineCellInference(nn.Cell):
+        ...     def __init__(self, network, micro_batch_num):
+        ...         super().__init__()
+        ...         self.network = network
+        ...         self.micro_batch_num = micro_batch_num
+        ...         self.concat = ops.Concat()
+        ...     def construct(self, x):
+        ...         ret = ()
+        ...         for i in range(self.micro_batch_num):
+        ...             micro_batch_size = x.shape[0] // self.micro_batch_num
+        ...             start = micro_batch_size * i
+        ...             end = micro_batch_size * (i + 1)
+        ...             micro_input = x[start:end]
+        ...             y = self.network(micro_input)
+        ...             ret = ret + (y,)
+        ...         ret = self.concat(ret)
+        ...         return ret
+        >>> D.init()
+        >>> context.set_auto_parallel_context(parallel_mode='semi_auto_parallel', full_batch=True, pipeline_stages=2)
         >>> net = Network()
-        >>> net.word_embedding.pipeline_stage = 0
-        >>> net.head.pipeline_stage = 1
-        >>> x = Tensor(np.ones((8, 4))
-        >>> net.compile()
+        >>> net = PipelineCellInference(net, 2)
+        >>> net.set_train(False)
+        >>> x = Tensor(np.ones((2, 4)), ms.float32)
+        >>> net.compile(x)
         >>> ms.sync_pipeline_shared_parameters(net)
-        >>> print(net.word_embedding.embedding_table.asnumpy())
-        >>> [[1. 1. 1. 1.]
-             [1. 1. 1. 1.]
-             [1. 1. 1. 1.]
-             [1. 1. 1. 1.]]
+        >>> print(net.network.word_embedding.w.asnumpy())
+        [[1. 1. 1. 1.]
+         [1. 1. 1. 1.]
+         [1. 1. 1. 1.]
+         [1. 1. 1. 1.]]
     """
 
     if not isinstance(net, ms.nn.Cell):
@@ -466,3 +576,65 @@ def sync_pipeline_shared_parameters(net):
 
     # restore parallel context
     ms.context.set_auto_parallel_context(parallel_mode=parallel_mode, full_batch=full_batch)
+
+
+def load_segmented_checkpoints(ckpt_file_dir, net=None, strict_load=False, filter_prefix=None,
+                               dec_key=None, dec_mode="AES-GCM", specify_prefix=None, choice_func=None):
+    """
+    Load checkpoint info from a specified file. If the specified ckpt_file_dir path contains multiple
+    checkpoint files, all checkpoint files will be loaded one by one and the combined dictionary will be return.
+
+    Note:
+        - `specify_prefix` and `filter_prefix` do not affect each other.
+        - If none of the parameters are loaded from checkpoint file, it will throw ValueError.
+        - `specify_prefix` and `filter_prefix` are in the process of being deprecated,
+          `choice_func` is recommended instead.
+          And using either of those two args will override `choice_func` at the same time.
+
+    Args:
+        ckpt_file_dir (str): Checkpoint file directory.
+        net (Cell): The network where the parameters will be loaded. Default: ``None`` .
+        strict_load (bool): Whether to strict load the parameter into net. If ``False`` , it will load parameter
+                            into net when parameter name's suffix in checkpoint file is the same as the
+                            parameter in the network. When the types are inconsistent perform type conversion
+                            on the parameters of the same type, such as float32 to float16. Default: ``False`` .
+        filter_prefix (Union[str, list[str], tuple[str]]): Deprecated(see `choice_func`). Parameters starting with the
+            filter_prefix will not be loaded. Default: ``None`` .
+        dec_key (Union[None, bytes]): Byte type key used for decryption. If the value is ``None`` , the decryption
+                                      is not required. Default: ``None`` .
+        dec_mode (str): This parameter is valid only when dec_key is not set to ``None`` . Specifies the decryption
+                        mode, currently supports ``"AES-GCM"`` and ``"AES-CBC"`` and ``"SM4-CBC"`` .
+                        Default: ``"AES-GCM"`` .
+        specify_prefix (Union[str, list[str], tuple[str]]): Deprecated(see `choice_func`). Parameters starting with the
+            specify_prefix will be loaded. Default: ``None`` .
+        choice_func (Union[None, function]) : Input value of the function is a Parameter name of type string,
+            and the return value is a bool. If returns ``True`` , the Parameter
+            that matches the custom condition will be loaded. If returns ``False`` , the Parameter that
+            matches the custom condition will be removed. Default: ``None`` .
+
+    Returns:
+        Dict, key is parameter name, value is a Parameter or string. When the `append_dict` parameter of
+        :func:`mindspore.save_checkpoint` and the `append_info` parameter of :class:`mindspore.train.CheckpointConfig`
+        are used to save the checkpoint, `append_dict` and `append_info` are dict types, and their value are string,
+        then the return value obtained by loading checkpoint is string, and in other cases the return value is
+        Parameter.
+
+    Raises:
+        TypeError: Input ckpt_file_dir is not a string.
+        ValueError: Checkpoint file directory doesn't exist. Or it's not a directory
+        ValueError: Checkpoint file's format is incorrect.
+        ValueError: Parameter's dict is None after load checkpoint file.
+        TypeError: The type of `specify_prefix` or `filter_prefix` is incorrect.
+    """
+    if not isinstance(ckpt_file_dir, str):
+        raise TypeError("The ckpt_file_dir should be a str.")
+    if not os.path.isdir(ckpt_file_dir):
+        raise ValueError("The dst_strategy_file: {} doesn't exist. Or it's not a directory".
+                         format(ckpt_file_dir))
+    checkpoint_file_name = os.path.join(ckpt_file_dir, "*.ckpt")
+    rank_ckpts = glob.glob(checkpoint_file_name)
+    parameter_dict = {}
+    for checkpoint_file in rank_ckpts:
+        parameter_dict.update(ms.load_checkpoint(checkpoint_file, net, strict_load, filter_prefix, dec_key,
+                                                 dec_mode, specify_prefix, choice_func))
+    return parameter_dict

mindspore/parallel/cluster/process_entity/_api.py

@@ -42,7 +42,7 @@ class _Node:
         os.environ["MS_WORKER_NUM"] = str(self.worker_num)
         os.environ["MS_SCHED_HOST"] = self.sched_host
         os.environ["MS_SCHED_PORT"] = str(self.sched_port)
-        os.environ["MS_CLUSTER_TIMEOUT"] = str(self.timeout)
+        os.environ["MS_TOPO_TIMEOUT"] = str(self.timeout)
 
 class _MetaServerNode(_Node):
     """

mindspore/parallel/parameter_broadcast.py

@@ -84,7 +84,7 @@ def parameter_broadcast(net, layout, cur_rank=0, initial_rank=0):
         >>> net.matmul2.shard(((1, 8), (8, 1)))
         >>> net.relu2.shard(((8, 1),))
         >>> # Create the dataset taking MNIST as an example. Refer to
-        >>> # https://gitee.com/mindspore/docs/blob/r2.3.q1/docs/mindspore/code/mnist.py
+        >>> # https://gitee.com/mindspore/docs/blob/master/docs/mindspore/code/mnist.py
         >>> dataset = create_dataset()
         >>> optim = nn.SGD(net.trainable_params(), 1e-2)
         >>> loss = nn.CrossEntropyLoss()

mindspore/parallel/shard.py

@@ -328,7 +328,7 @@ def shard(fn, in_strategy, out_strategy=None, parameter_plan=None, device="Ascen
 
     Tutorial Examples:
         - `Functional Operator Sharding
-          <https://www.mindspore.cn/tutorials/experts/en/r2.3.q1/parallel/pynative_shard_function_parallel.html>`_
+          <https://www.mindspore.cn/tutorials/experts/en/master/parallel/pynative_shard_function_parallel.html>`_
     """
     if not isinstance(fn, (ms.nn.Cell)):
         logger.warning("'fn' is not a mindspore.nn.Cell, and its definition cannot involve Parameter; "

mindspore/profiler/parser/ascend_analysis/fwk_cann_parser.py

@@ -34,6 +34,7 @@ class FwkCANNParser:
     def __init__(self, source_path: str, msprof_data: List, rank_id: int):
         source_path = validate_and_normalize_path(source_path)
         ProfilerInfoParser.init_source_path(source_path)
+        ProfilerInfoParser.init_rank_id(rank_id)
         fwk_parser = FwkFileParser(source_path, rank_id)
         msprof_timeline_parser = MsprofTimelineParser(msprof_data)
         self._fwk_op_data = fwk_parser.get_op_range_data()

mindspore/profiler/parser/ascend_analysis/profiler_info_parser.py

@@ -20,9 +20,9 @@ from subprocess import CalledProcessError, TimeoutExpired
 from subprocess import Popen, PIPE
 
 from mindspore import log as logger
-import mindspore._c_expression as c_expression
 from mindspore.profiler.common.validator.validate_path import validate_and_normalize_path
 from mindspore.profiler.parser.ascend_analysis.constant import Constant
+from mindspore.profiler.parser.profiler_info import ProfilerInfo
 
 
 class ProfilerInfoParser:
@@ -37,6 +37,7 @@ class ProfilerInfoParser:
     # profiler information related files
     _source_prof_path = None
     _loaded_frequency = False
+    _rank_id = 0
 
     @classmethod
     def init_source_path(cls, source_path: str):
@@ -48,12 +49,15 @@ class ProfilerInfoParser:
             raise RuntimeError("Input source path is invalid!")
         cls._source_prof_path = prof_path
 
+    @classmethod
+    def init_rank_id(cls, rank_id):
+        """initialize the rank id."""
+        cls._rank_id = rank_id
+
     @classmethod
     def get_local_time(cls, syscnt: int) -> Decimal:
         """Convert syscnt to local time."""
         if not cls._loaded_frequency:
-            localtime_stamp = c_expression.get_clock_time()
-            syscnt_stamp = c_expression.get_clock_syscnt()
             outs, _ = cls.__run_cmd(['which', cls._msprof_cmd])
             if not outs:
                 raise FileNotFoundError("Failed to find msprof command!")
@@ -70,8 +74,16 @@ class ProfilerInfoParser:
                 cls._freq = float(cpu_info.get("Frequency", cls._freq))
             except ValueError:
                 pass
-            cls._start_cnt = syscnt_stamp
-            cls._time_offset = localtime_stamp
+            profiler_info_path = os.path.join(cls._source_prof_path, os.path.pardir,
+                                              f"profiler_info_{cls._rank_id}.json")
+            if not os.path.isfile(profiler_info_path):
+                raise RuntimeError(f"Can`t find the file {profiler_info_path}, please check!")
+            with os.fdopen(os.open(profiler_info_path, os.O_RDONLY, 0o600), 'r') as fr:
+                profiler_info_data = json.load(fr)
+            cls._start_cnt = profiler_info_data.get('system_cnt')
+            cls._time_offset = profiler_info_data.get('system_time')
+            ProfilerInfo.set_system_time(cls._time_offset)
+            ProfilerInfo.set_system_cnt(cls._start_cnt)
             cls._loaded_frequency = True
         start_ns = cls.__get_timestamp(syscnt)
         start_us = Decimal(start_ns) / Constant.NS_TO_US

mindspore/profiler/parser/ascend_msprof_exporter.py

@@ -251,10 +251,10 @@ class AscendMsprofExporter:
                 msprof_json.add(f)
 
         if not op_summary:
-            raise RuntimeError("The op_summary csv file was not found, perhaps the original data was not collected.")
+            logger.warning("The op_summary csv file was not found, perhaps the original data was not collected.")
         if not op_statistic:
-            raise RuntimeError("The op_statistics csv file was not found, perhaps the original data was not collected.")
+            logger.warning("The op_statistics csv file was not found, perhaps the original data was not collected.")
         if not msprof_json:
-            raise RuntimeError("The msprof json file was not found, perhaps the original data was not collected.")
+            logger.warning("The msprof json file was not found, perhaps the original data was not collected.")
 
         logger.info("Finish checking files.")

mindspore/profiler/parser/ascend_msprof_generator.py

@@ -88,7 +88,10 @@ class AscendMsprofDataGenerator:
         """read op summary to memory"""
         op_summary = []
         op_summary_name = fr'{self.mindstudio_profiler_output}/op_summary_*.csv'
-        op_summary_file = get_newest_file(glob.glob(op_summary_name))[0]
+        op_summary_files = glob.glob(op_summary_name)
+        if not op_summary_files:
+            return
+        op_summary_file = get_newest_file(op_summary_files)[0]
         with open(op_summary_file, newline='') as csvfile:
             reader = csv.DictReader(csvfile, delimiter=',', quotechar='"')
             for row in reader:
@@ -129,7 +132,10 @@ class AscendMsprofDataGenerator:
         """read op statistic to memory"""
         op_statistic = []
         op_statistic_name = fr'{self.mindstudio_profiler_output}/op_statistic_*.csv'
-        op_statistic_file = get_newest_file(glob.glob(op_statistic_name))[0]
+        op_statistic_files = glob.glob(op_statistic_name)
+        if not op_statistic_files:
+            return
+        op_statistic_file = get_newest_file(op_statistic_files)[0]
         with open(op_statistic_file, newline='') as csvfile:
             reader = csv.DictReader(csvfile, delimiter=',', quotechar='"')
             for row in reader:
@@ -140,7 +146,8 @@ class AscendMsprofDataGenerator:
                 )
                 new_row = tuple(['0' if d == 'N/A' else d for d in new_row])
                 op_statistic.append(new_row)
-
+        if not op_statistic:
+            return
         op_statistic_dt = np.dtype(self.op_statistic_type)
         self.op_statistic = np.array(op_statistic, dtype=op_statistic_dt)
         self.op_statistic['Total Time'] *= 1e-3