mindspore 2.2.14__cp39-cp39-win_amd64.whl → 2.4.0__cp39-cp39-win_amd64.whl

mindspore/parallel/checkpoint_transform.py

@@ -1,4 +1,4 @@
-# Copyright 2022 Huawei Technologies Co., Ltd
+# Copyright 2023 Huawei Technologies Co., Ltd
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -22,35 +22,36 @@ from collections import defaultdict
 import numpy as np
 import mindspore as ms
 from mindspore.common import dtype as mstype
+from mindspore.parallel._utils import _is_in_auto_parallel_mode, _get_pipeline_stages
 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, _extract_src_dst_layout_map_by_src
-
+from mindspore.parallel.transform_safetensors import _transform_safetensors, _collect_safetensor_files
 
 __all__ = ["merge_pipeline_strategys", "rank_list_for_transform", "transform_checkpoint_by_rank",
-           "transform_checkpoints", "load_segmented_checkpoints"]
+           "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/r2.2/parallel/model_transformation.html>`_.
+    `Model Transformation <https://www.mindspore.cn/docs/en/master/model_train/parallel/model_transformation.html>`_.
 
     Note:
         Strategy file of each pipeline stage should be included in src_strategy_dirs.
 
     Args:
         src_strategy_dirs (str): The directory of strategy files including all pipeline stage which is saved by
-                                 'mindspore.set_auto_parallel_context(strategy_ckpt_save_file)'
+                                 'mindspore.set_auto_parallel_context(strategy_ckpt_save_file)'.
         dst_strategy_file (str): The file merged strategy to save.
 
     Raises:
         NotADirectoryError: `src_strategy_dirs` is not a directory.
 
     Examples:
-        >>> import mindspore ms
+        >>> import mindspore as ms
         >>> # src_strategy_dir/stra0.ckpt, src_strategy_dir/stra1.ckpt ... src_strategy_dir/stra127.ckpt
         >>> ms.merge_pipeline_strategys("./src_strategy_dir", "./dst_strategy.ckpt")
 
@@ -73,28 +74,29 @@ def merge_pipeline_strategys(src_strategy_dirs, dst_strategy_file):
 
 def rank_list_for_transform(rank_id, src_strategy_file=None, dst_strategy_file=None):
     """
-    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/r2.2/parallel/model_transformation.html>`_.
+    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/docs/en/master/model_train/parallel/model_transformation.html>`_.
 
     Args:
         rank_id (int): The rank of which distributed checkpoint needs to be obtained after conversion.
         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.
+                                 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.
+                                 when the `dst_strategy_file` is ``None``,
+                                 it means that the destination sharding strategy
+                                 is without any sharing for each parameter. Default: ``None``.
 
     Returns:
         List, the rank list required for converting the distributed checkpoint of rank_id.
 
     Raises:
-        ValueError: `src_strategy_file` or dst_strategy_file is incorrect.
-        TypeError: `src_strategy_file` or dst_strategy_file is not a string.
-        TypeError: `rank_id` is not a int.
+        ValueError: `src_strategy_file` or `dst_strategy_file` is incorrect.
+        TypeError: `src_strategy_file` or `dst_strategy_file` is not a string.
+        TypeError: `rank_id` is not an int.
 
     Examples:
         >>> import mindspore as ms
@@ -129,7 +131,9 @@ def rank_list_for_transform(rank_id, src_strategy_file=None, dst_strategy_file=N
             src_rank_id_start = src_pipeline_stage_id * src_stage_device_num
             result_set.update([src_rank_id_start + rank for rank in needed_rank_list_in_local_stage])
             handled_pipeline_stage.append(src_pipeline_stage_id)
-    return list(result_set)
+    result_list = list(result_set)
+    result_list.sort(reverse=True)
+    return list(result_list)
 
 
 def transform_checkpoint_by_rank(rank_id, checkpoint_files_map, save_checkpoint_file_name,
@@ -137,7 +141,7 @@ 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/r2.2/parallel/model_transformation.html>`_.
+    `Model Transformation <https://www.mindspore.cn/docs/en/master/model_train/parallel/model_transformation.html>`_.
 
     Args:
         rank_id (int): The rank of which distributed checkpoint needs to be obtained after conversion.
@@ -146,12 +150,13 @@ def transform_checkpoint_by_rank(rank_id, checkpoint_files_map, save_checkpoint_
         save_checkpoint_file_name (str): The file name to save the converted checkpoint.
         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.
+                                 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.
+                                 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.
@@ -159,7 +164,7 @@ def transform_checkpoint_by_rank(rank_id, checkpoint_files_map, save_checkpoint_
         ValueError: `save_checkpoint_file_name` is not end with ".ckpt".
         TypeError: `checkpoint_files_map` is not a dict.
         TypeError: `src_strategy_file` or `dst_strategy_file` is not a string.
-        TypeError: `rank_id` is not a int.
+        TypeError: `rank_id` is not an int.
         TypeError: `save_checkpoint_file_name` is not a string.
 
     Examples:
@@ -228,7 +233,7 @@ def _transform_checkpoint_by_stage(src_checkpoints_dir, dst_checkpoints_dir, ckp
     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)
+                                                                                        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 \
@@ -353,29 +358,35 @@ def _transform_checkpoints(src_checkpoints_dir, dst_checkpoints_dir, ckpt_prefix
 
 
 def transform_checkpoints(src_checkpoints_dir, dst_checkpoints_dir, ckpt_prefix, src_strategy_file=None,
-                          dst_strategy_file=None):
+                          dst_strategy_file=None, process_num=1, output_format="ckpt"):
     """
     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/r2.2/parallel/model_transformation.html>`_.
+    `Model Transformation <https://www.mindspore.cn/docs/en/master/model_train/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.
 
+        The number of multiprocess settings is related to the size of the host, and it is not recommended to set it
+        too large, otherwise it may cause freezing.
+
     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
+        src_strategy_file (str, optional): 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
+        dst_strategy_file (str, optional): 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.
+        process_num (int, optional): Number of processes to use for parallel processing. Defaults: 1.
+        output_format (str, optional): Control the format of the output checkpoint after conversion.
+            It can be set to either "ckpt" or "safetensors". Default: "ckpt".
 
     Raises:
         ValueError: `src_strategy_file` or `dst_strategy_file` is incorrect.
@@ -389,6 +400,21 @@ def transform_checkpoints(src_checkpoints_dir, dst_checkpoints_dir, ckpt_prefix,
         ...                       "./src_strategy.ckpt", "./dst_strategy.ckpt")
 
     """
+    all_safetensor_files_map = _collect_safetensor_files(src_checkpoints_dir)
+    all_ckpt_files_map = _collect_safetensor_files(src_checkpoints_dir, format='ckpt')
+    if all_safetensor_files_map and all_ckpt_files_map:
+        raise ValueError("For 'transform_checkpoints', the 'src_checkpoints_dir' cannot contain "
+                         "both ckpt file and safetensors file simultaneously")
+    if all_safetensor_files_map and not all_ckpt_files_map:
+        _transform_safetensors(src_checkpoints_dir, dst_checkpoints_dir, ckpt_prefix, src_strategy_file,
+                               dst_strategy_file, process_num, output_format)
+        return
+    if not all_safetensor_files_map and not all_ckpt_files_map:
+        raise ValueError("For 'transform_checkpoints', the 'src_checkpoints_dir' can not be empty.")
+    if all_ckpt_files_map and not all_safetensor_files_map and output_format == 'safetensors':
+        raise ValueError("For 'transform_checkpoints', 'output_format' can not be 'safetensors' "
+                         "when 'src_checkpoints_dir' only contains ckpt file.")
+
     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")
@@ -405,23 +431,163 @@ def transform_checkpoints(src_checkpoints_dir, dst_checkpoints_dir, ckpt_prefix,
     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)
+    dst_stage_num = _extract_pipeline_stage_num(dst_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")}
+        src_param_keys = {param_name for param_name in src_layout_map if
+                          not param_name.startswith(("accu_grads", "adam_v", "adam_m"))}
     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.")
+        dst_param_keys = {param_name for param_name in dst_layout_map if
+                          not param_name.startswith(("accu_grads", "adam_v", "adam_m"))}
+    layout_is_passed = src_layout_map and dst_layout_map
+
+    if layout_is_passed and pipeline_stage_num == 1 and dst_stage_num == 1 and \
+            src_param_keys.issubset(dst_param_keys) and len(src_param_keys) < len(dst_param_keys):
+        ms.log.info("Transform checkpoint by every pipeline stage.")
         _transform_checkpoint_by_stage(src_checkpoints_dir, dst_checkpoints_dir, ckpt_prefix,
                                        src_strategy_file, dst_strategy_file)
     else:
+        ms.log.info("Transform checkpoints by all pipeline stage.")
         _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:
+        ms.log.warning("The layout dict does not contain the pipeline_shared_param info %s", name)
+        return
+
+    pipeline_shared = layout[8]
+    if not pipeline_shared:
+        return
+
+    is_send = layout[9]
+    peer_rank = layout[10]
+    sr_tag = layout[11]
+    if is_send:
+        ms.ops.Send(sr_tag=sr_tag, dest_rank=peer_rank)(param)
+    else:
+        param.assign_value(ms.ops.Receive(sr_tag=sr_tag,
+                                          src_rank=peer_rank,
+                                          shape=param.shape,
+                                          dtype=param.dtype)(param))
+
+
+def sync_pipeline_shared_parameters(net):
+    """synchronize pipeline parallel stage shared parameters.
+    Parameters may be shared between different stages. For example, `embedding table` is
+    shared by `WordEmbedding` layer and `LMHead` layer, which are usually split into different stages. It is necessary
+    to perform synchronization after `embedding table` changes.
+
+    Note:
+        The network should be compiled before synchronize pipeline parallel stage shared parameters.
+
+    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/docs/en/master/model_train/parallel/dynamic_cluster.html>`_ .
+
+        >>> import numpy as np
+        >>> import mindspore as ms
+        >>> 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.w = Parameter(Tensor(np.ones(shape), ms.float32), name='w')
+        ...         self.matmul = ops.MatMul().shard(((1, 1), (1, 1)))
+        ...     def construct(self, x):
+        ...         return self.matmul(x, self.w), self.w
+        ...
+        >>> class LMHead(nn.Cell):
+        ...     def __init__(self):
+        ...         super().__init__()
+        ...         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__()
+        ...         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)
+        ...         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 = 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.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):
+        ms.log.critical("Failed to synchronize pipeline shared parameters.")
+        msg = ("For 'sync_pipeline_shared_parameters', the argument 'net' should be a Cell, "
+               "but got {}.".format(type(net)))
+        raise TypeError(msg)
+
+    if _get_pipeline_stages() < 2:
+        return
+
+    layout_dict = net.parameter_layout_dict
+    if _is_in_auto_parallel_mode() and not layout_dict:
+        from mindspore.common.api import _get_parameter_layout
+        layout_dict = _get_parameter_layout()
+
+    # switch to standalone mode
+    parallel_mode = ms.context.get_auto_parallel_context("parallel_mode")
+    full_batch = ms.context.get_auto_parallel_context("full_batch")
+    ms.context.set_auto_parallel_context(parallel_mode="stand_alone", full_batch=False)
+
+    # synchronize shared parameter
+    for name, param in net.parameters_and_names():
+        if name in layout_dict:
+            _sync_params(name, param, layout_dict[name])
+
+    # 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):
     """
@@ -473,8 +639,8 @@ def load_segmented_checkpoints(ckpt_file_dir, net=None, strict_load=False, filte
     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 exists. Or it's not a directory.".
-                         format(os.path.dirname(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 = {}

mindspore/parallel/cluster/__init__.py

@@ -0,0 +1,15 @@
+# Copyright 2023 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+"""Entry for ms_run"""

mindspore/parallel/cluster/process_entity/__init__.py

@@ -0,0 +1,18 @@
+# Copyright 2023 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+"""Interfaces for ms_run"""
+from ._api import _Node, _MetaServerNode, _ComputeGraphNode, _ProcessManager
+
+from ._utils import _generate_cmd, _generate_url, _is_local_ip, _send_scale_num