# 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.
# ============================================================================
"""shard"""
import copy
import mindspore as ms
from mindspore import log as logger
from mindspore._c_expression import Shard_
[docs]class Layout:
"""
Parallel layout describes the detailed sharding information.
Note:
- It is valid only in semi auto parallel or auto parallel mode.
- The multiplication result of the `device_matrix` must be equal to the device count in a pipeline stage.
- When the layout function is invoked to constructs a sharding strategy, each alias name is only allowed to be
used once to shard a tensor.
Args:
device_matrix (tuple): Describe the shape of devices arrangement, its element type is int.
alias_name (tuple): The alias name for each axis of device_matrix, its length shoits element type is string.
When using "interleaved_parallel" as an alias name, the tensor would be split into multiple
copies on the corresponding partition dimension on a single card.
Raises:
TypeError: `device_matrix` is not a tuple type.
TypeError: `alias_name` is not a tuple type.
ValueError: `device_matrix` length is not equal to `alias_name` length.
TypeError: The element of `device_matrix` is not int type.
TypeError: The element of `alias_name` is not a str type.
ValueError: The element of `alias_name` is an empty str.
ValueError: The element of `alias_name` is "None".
ValueError: `alias_name` contains repeated element.
Examples:
>>> from mindspore import Layout
>>> layout = Layout((2, 2, 2), ("dp", "sp", "mp"))
>>> layout0 = layout("dp", "mp")
>>> print(layout0.to_dict())
{"device_matrix": (2, 2, 2), "tensor_map": (2, 0), "interleaved_parallel": False}
>>> # Total device num is 4, but split the tensor in local device into two copies.
>>> layout = Layout((2, 2, 2), ("dp", "sp", "interleaved_parallel"))
>>> layout1 = layout(("dp", "interleaved_parallel"), "sp")
"""
def __init__(self, device_matrix, alias_name):
if not isinstance(device_matrix, tuple):
raise TypeError(f'device_matrix must be tuple type, but got:{type(device_matrix)}')
if not isinstance(alias_name, tuple):
raise TypeError(f'alias_name must be tuple type, but got:{type(alias_name)}')
if len(device_matrix) != len(alias_name):
raise ValueError(f'device_matrix length should be equal to alias_name length')
for in_ele in device_matrix:
if not isinstance(in_ele, int):
raise TypeError(f'The element of device_matrix must be int type, but got:{type(in_ele)}')
for in_ele in alias_name:
if not isinstance(in_ele, str):
raise TypeError(f'The element of alias_name must be str type, but got:{type(in_ele)}')
if not in_ele:
raise ValueError(f"The element of alias_name can not be empty.")
if in_ele == "None":
raise ValueError(f"The element of alias_name can not set 'None', because 'None' means no sharding.")
if len(set(alias_name)) != len(alias_name):
raise ValueError(f'Each element of alias_name {alias_name} should be different')
inter_key = "interleaved_parallel"
if inter_key in alias_name and alias_name.index(inter_key) != len(alias_name) - 1:
raise ValueError(f"When alias_name {alias_name} contains keyword 'interleaved_parallel',"
f" it should be at the last dim of alias_name, which means the virtual sharding.")
self._device_shape = device_matrix
self._alias_name = alias_name
self._tensor_map = None
def __call__(self, *tensor_map):
self._tensor_map = ()
writed_map = ()
for ele in tensor_map:
if isinstance(ele, tuple):
ele_map = ()
for item in ele:
if item == "None":
ele_map += (-1,)
continue
if item not in self._alias_name:
raise ValueError(f'The axis {item} is not found in {self._alias_name}')
if item in writed_map:
raise ValueError(f'The axis {item} has been set more than one in {self._alias_name}')
ele_map += (len(self._alias_name) - 1 - self._alias_name.index(item),)
writed_map += (item,)
self._tensor_map += (ele_map,)
continue
if ele == "None":
self._tensor_map += (-1,)
continue
if ele not in self._alias_name:
raise ValueError(f'The axis {ele} is not found in {self._alias_name}')
if ele in writed_map:
raise ValueError(f'The axis {ele} has been set more than one in {self._alias_name}')
self._tensor_map += (len(self._alias_name) - 1 - self._alias_name.index(ele),)
writed_map += (ele,)
return copy.deepcopy(self)
[docs] def to_dict(self):
"""
Transform layout to a dictionary.
"""
if self._device_shape is None:
raise ValueError("The device_shape of layout is None")
if self._tensor_map is None:
raise ValueError("The tensor_map of layout is None")
interleaved_parallel = "interleaved_parallel" in self._alias_name
return {"device_matrix": self._device_shape, "tensor_map": self._tensor_map,
"interleaved_parallel": interleaved_parallel, "alias_name": self._alias_name}
class Shard(Shard_):
"""Shard operation"""
def __init__(self):
"""Initialize Shard."""
super().__init__('Shard')
self.shard_fn = None
self.fn = None
self.in_strategy = None
self.out_strategy = None
self.parameter_plan = None
self.device = None
self.level = None
def __call__(self, fn, in_strategy, out_strategy=None, parameter_plan=None, device="Ascend", level=0):
parallel_mode = ms.context.get_auto_parallel_context("parallel_mode")
if parallel_mode not in ("auto_parallel", "semi_auto_parallel"):
raise AssertionError(
f"Cell shard only supports auto parallel and semi auto parallel.")
if ms.context.get_context("device_target") not in ("Ascend", "GPU"):
raise AssertionError(
f"'Shard' now only supports 'Ascend' and 'GPU'")
if parallel_mode == "auto_parallel" and \
ms.context.get_auto_parallel_context("search_mode") != "sharding_propagation":
raise AssertionError(f"'search_mode' must be 'sharding_propagation' for 'Shard' when the "
f"'parallel_mode' is 'auto_parallel.'")
if not isinstance(in_strategy, tuple):
raise TypeError(
f"For 'Shard', the 'in_strategy' should be a tuple, but got {type(in_strategy).__name__}.")
inner_type = self._check_layout_inner_type(in_strategy, "in_strategy")
if inner_type == "layout":
in_strategy = self._extract_layout_value(in_strategy, "in_strategy")
if not isinstance(out_strategy, (type(None), tuple)):
raise TypeError(f"For 'Shard', the 'out_strategy' should be None or tuple, "
f"but got {type(out_strategy).__name__}.")
if not isinstance(out_strategy, type(None)):
logger.warning("Out_strategy is not in use currently, will be ignored in the following procedures.")
inner_type = self._check_layout_inner_type(out_strategy, "out_strategy")
if inner_type == "layout":
out_strategy = self._extract_layout_value(out_strategy, "out_strategy")
if not isinstance(device, str):
raise TypeError(f"For 'Shard', the 'device' should be a string, "
f"but got {type(device).__name__}")
if not isinstance(level, int):
raise TypeError(f"For 'Shard', the 'level' should be an integer, "
f"but got {type(level).__name__}")
if ms.get_algo_parameters("fully_use_devices") is True:
logger.warning("After calling 'shard', the environment variable 'fully_use_devices' "
"will be overwritten as False.")
ms.set_algo_parameters(fully_use_devices=False)
if ms.context.get_auto_parallel_context("full_batch_is_set") is False:
logger.warning("When calling the shard interface, "
"'dataset_strategy' or 'full_batch' is not manually set by the user, "
"and the 'dataset_strategy' will be set to 'full_batch'.")
ms.context.set_auto_parallel_context(dataset_strategy="full_batch")
if self._is_attrs_has_been_set(fn, in_strategy, out_strategy, device, level):
return self.shard_fn
shard_ = Shard()
if isinstance(fn, ms.nn.Cell):
for param in fn.trainable_params():
param.is_in_shard = True
# Set parameter layout to corresponding parameter
self._set_param_layout_into_parameter(fn, parameter_plan)
def shard_fn(*args):
@ms.common.jit(hash_args=fn)
def after_shard(*args):
return shard_(fn, in_strategy, out_strategy, device, level)(*args)
return after_shard(*args)
self.shard_fn = shard_fn
self.fn = fn
self.in_strategy = in_strategy
self.out_strategy = out_strategy
self.device = device
self.level = level
return self.shard_fn
@staticmethod
def _search_parameter_by_name(param_name: str, net):
param_name = param_name.replace("self.", "")
for param in net.trainable_params():
if param.name == param_name:
return param
return None
@staticmethod
def _check_layout_is_valid(param_name, param_shape, param_strategy):
if len(param_strategy) != len(param_shape):
raise ValueError(f"For {param_name}, the length of param_strategy: {len(param_strategy)}, "
f"is not equal to param_shape len: {len(param_shape)}.")
for i, _ in enumerate(param_strategy):
if param_shape[i] % param_strategy[i] != 0:
raise ValueError(f"For '{param_name}', the param_shape is {param_shape} and "
f"the setting param_strategy is {param_strategy}. "
f"The param_shape[{i}]: {param_shape[i]} cannot be divisible by "
f"param_strategy[{i}]: {param_strategy[i]}.")
def _set_param_layout_into_parameter(self, fn, parameter_plan):
""" Set param_strategy into parameter if fn is a Cell and parameter_plan is a dict."""
if parameter_plan is None:
return
if isinstance(parameter_plan, dict):
if not isinstance(fn, ms.nn.Cell):
raise TypeError(
f"If parameter_plan is set, type of fn must be mindspore.nn.Cell, but got {type(fn)}")
for k in parameter_plan.keys():
v = parameter_plan[k]
if not isinstance(k, str) or not isinstance(v, (tuple, Layout)):
raise TypeError(f"For 'Shard', the type of each key and value in 'parameter_plan' must be str and "
f"tuple/Layout, but got {type(k).__name__} and {type(v).__name__}")
else:
raise TypeError(f"For 'Shard', the 'parameter_plan' should be a dict or None, "
f"but got {type(parameter_plan).__name__}")
for param_name in parameter_plan.keys():
param_strategy = parameter_plan[param_name]
param = self._search_parameter_by_name(param_name, fn)
if param is None:
logger.warning(
f"{param_name} is not exist, ignored its setting.")
continue
has_set = None
if param.param_info.param_strategy:
has_set = "strategy"
if param.param_info.device_matrix:
has_set = "layout"
if has_set == "strategy":
logger.warning(f"The layout of parameter '{param_name}' has been set to "
f"{param.param_info.param_strategy}, current setting will be ignored.")
elif has_set == "layout":
logger.warning(f"The layout of parameter '{param_name}' has been set, "
f"current setting will be ignored.")
else:
if isinstance(param_strategy, tuple):
self._check_layout_is_valid(param_name, param.shape, param_strategy)
param.param_info.param_strategy = param_strategy
if isinstance(param_strategy, Layout):
param_layout = self._extract_layout_value((param_strategy,), "in_strategy")[0]
param.param_info.device_matrix = param_layout["device_matrix"]
param.param_info.tensor_map = param_layout["tensor_map"]
param.param_info.interleaved_parallel = param_layout["interleaved_parallel"]
param.param_info.alias_name = param_layout["alias_name"]
def _is_attrs_has_been_set(self, fn, in_strategy, out_strategy, device, level):
return self.shard_fn is not None and self.fn == fn and self.in_strategy == in_strategy and \
self.out_strategy == out_strategy and self.device == device and self.level == level
def _check_layout_inner_type(self, strategy, log_info):
"""Check inner item type of layout, should be int or ms.Layout."""
strategy_set = set()
for stra in strategy:
if not isinstance(stra, (tuple, Layout)):
raise TypeError(
f"The '{log_info}' should be a tuple(tuple(int)) or tuple(mindspore.Layout), "
f"but got {type(stra).__name__}")
if isinstance(stra, Layout):
strategy_set.add("layout")
elif isinstance(stra, tuple):
strategy_set.add("tuple")
self._check_tuple_strategy(stra)
if len(strategy_set) != 1:
raise TypeError(
f"For 'Shard', the strategy can only pass in consistent type for all dimensions.")
return strategy_set.pop()
def _extract_layout_value(self, layout, log_info):
"""Extract parallel layout value"""
layout_value = None
if layout is not None:
if not isinstance(layout, tuple):
raise TypeError(f'{log_info} must be tuple type, but got:{type(layout)}')
layout_value = ()
for in_ele in layout:
if not isinstance(in_ele, Layout):
raise TypeError(f"The {log_info} item should be a object of class Layout.")
layout_value += (in_ele.to_dict(),)
return layout_value
def _check_tuple_strategy(self, dim_strategy):
if not all(isinstance(x, int) for x in dim_strategy):
raise TypeError(
f"The tuple strategy for each dimension should be tuple(int).")
[docs]def shard(fn, in_strategy, out_strategy=None, parameter_plan=None, device="Ascend", level=0):
"""
Defining the input and output layouts of this cell and the parallel strategies of remaining ops will be
generated by sharding propagation. In PyNative mode, use this method to specify a Cell for distributed
execution in graph mode. In Graph mode, use this method to specify distribution strategy for a Cell,
strategy for others will be set by sharding propagation.
in_strategy and out_strategy define the input and output layout respectively.
in_strategy/out_strategy should be a tuple, each element of which corresponds to the desired layout of
this input/output, and None represents data_parallel,
which can refer to the description of :func:`mindspore.ops.Primitive.shard`.
The parallel strategies of remaining operators are derived from the strategy specified by the input and output.
Note:
If ms.shard is called, the parallel mode in `set_auto_parallel_context` (parallel_mode) will be set to
"auto_parallel" and the search mode (search_mode) to "sharding_propagation".
If the input contain Parameter, its strategy should be set in `in_strategy`.
Args:
fn (Union[Cell, Function]): Function to be executed in parallel.
Its arguments and return value must be Tensor or Parameter.
If `fn` is a Cell with parameters, `fn` needs to be an instantiated object,
otherwise its arguments cannot be accessed.
in_strategy (tuple): Define the layout of inputs, each element of the tuple should be a tuple(int) or
tuple(mindspore.Layout).
Tuple defines the layout of the corresponding input.
out_strategy (Union[tuple, None]): Define the layout of outputs similar with `in_strategy`.
It is not in use right now. Default: ``None`` .
parameter_plan (Union[dict, None]): Define the layout for the specified parameters. Each element in dict
defines the layout of the parameter like "param_name: layout".
The key is a parameter name of type 'str'.
The value is a 1-D integer tuple or a 1-D mindspore.Layout tuple,
indicating the corresponding layout.
If the parameter name is incorrect or the corresponding parameter
has been set, the parameter setting will be ignored.
Default: ``None`` .
device (string): Select a certain `device` target. It is not in use right now.
Support ["CPU", "GPU", "Ascend"]. Default: ``"Ascend"`` .
level (int): Option for parallel strategy infer algorithm, namely the object function, maximize computation
over communication ratio, maximize speed performance, minimize memory usage etc. It is not in
use right now. Support [0, 1, 2]. Default: ``0`` .
Returns:
Function, return the function that will be executed under auto parallel process.
Raises:
AssertionError: If parallel mode is not "auto_parallel" nor "semi_auto_parallel".
AssertionError: If device_target it not "Ascend" or "GPU".
TypeError: If `in_strategy` is not a tuple.
TypeError: If `out_strategy` is not a tuple or None.
TypeError: If any element in `in_strategy` is not a tuple(int) or tuple(mindspore.Layout).
TypeError: If any element in `out_strategy` is not a tuple(int) or tuple(mindspore.Layout).
TypeError: If `parameter_plan` is not a dict or None.
TypeError: If any key in `parameter_plan` is not a str.
TypeError: If any value in `parameter_plan` is not a tuple(int) or a tuple(mindspore.Layout).
TypeError: If `device` is not a str.
TypeError: If `level` is not an integer.
Supported Platforms:
``Ascend`` ``GPU``
Examples:
>>> import numpy as np
>>> import mindspore as ms
>>> from mindspore import Tensor, nn
>>> from mindspore.communication import init
>>> ms.set_context(mode=ms.GRAPH_MODE)
>>> init()
>>> ms.set_auto_parallel_context(parallel_mode="auto_parallel", search_mode="sharding_propagation",
... device_num=8)
>>>
>>> # Case 1: cell uses functional
>>> class BasicBlock(nn.Cell):
>>> def __init__(self):
>>> super(BasicBlock, self).__init__()
>>> self.dense1 = nn.Dense(64, 64)
>>> self.gelu = nn.GELU()
>>> def my_add(x, y):
>>> x = ops.abs(x)
>>> return x + y
>>> # shard a function with tuple(int) strategies
>>> self.shard_my_add = ms.shard(my_add, in_strategy=((2, 2), (1, 4)), out_strategy=((4, 1),))
>>>
>>> def construct(self, x, u):
>>> x = self.gelu(x)
>>> y = self.gelu(u)
>>> y = x * y
>>> x = self.dense1(x)
>>> x = self.shard_my_add(x, y)
>>> return x
>>>
>>> class NetForward(nn.Cell):
>>> def __init__(self):
>>> super(NetForward, self).__init__()
>>> self.block1 = BasicBlock()
>>> self.block2 = BasicBlock()
>>> self.matmul = ops.MatMul()
>>>
>>> def construct(self, x, y):
>>> x = self.matmul(x, y)
>>> x = self.block1(x, x)
>>> x = self.block2(x, x)
>>> return x
>>>
>>> class Net(nn.Cell):
>>> def __init__(self):
>>> super(Net, self).__init__()
>>> # setting cell sharding strategy and parameter_plan by tuple(int)
>>> self.layer_net1 = NetForward()
>>> self.layer_net1_shard = ms.shard(self.layer_net1, in_strategy=((4, 2), (2, 1)),
... parameter_plan={"self.layer_net1.block1.weight": (4, 1)})
>>>
>>> # setting cell sharding strategy and parameter_plan by tuple(ms.Layout)
>>> self.layer_net2 = NetForward()
>>> layout = Layout((4, 2, 1), ("dp", "mp", "sp"))
>>> in_layout = (layout("dp", "mp"), layout("mp", "sp"))
>>> param_layout = layout("dp", "sp")
>>> self.layer_net2_shard = ms.shard(self.layer_net2, in_strategy=in_layout,
... parameter_plan={"self.layer_net2.block2.weight": param_layout})
>>> self.flatten = nn.Flatten()
>>> self.layer1 = nn.Dense(64, 64)
>>> self.layer2 = nn.Dense(64, 32)
>>> self.add = ops.Add()
>>> self.matmul = ops.MatMul()
>>>
>>> def construct(self, x, y):
>>> x = self.flatten(x)
>>> y = self.flatten(y)
>>> x = self.layer1(x)
>>> x = self.layer_net1_shard(x, y)
>>> x = self.layer_net2_shard(x, y)
>>> x = self.layer2(x)
>>> x = self.matmul(x, Tensor(np.ones(shape=(32, 32)), dtype=ms.float32))
>>> return x
>>>
>>> net = Net()
>>> x = Tensor(np.ones(shape=(64, 1, 8, 8)), dtype=ms.float32)
>>> y = Tensor(np.ones(shape=(64, 1, 8, 8)), dtype=ms.float32)
>>> net(x, y)
>>>
>>> # Case 2: function uses functional sharding
>>> def test_shard(x, y):
... return x + y
>>> x = Tensor(np.ones(shape=(32, 10)), dtype=ms.float32)
>>> y = Tensor(np.ones(shape=(32, 10)), dtype=ms.float32)
>>> output = ms.shard(test_shard, in_strategy=((4, 2), (4, 2)))(x, y)
>>> print(output.shape)
(32, 10)
Tutorial Examples:
- `Functional Operator Sharding
<https://www.mindspore.cn/docs/en/r2.4.0/model_train/parallel/shard_function_parallel.html>`_
- `mindspore.Layout
<https://www.mindspore.cn/docs/en/r2.4.0/api_python/mindspore/mindspore.Layout.html>`_
"""
if not isinstance(fn, (ms.nn.Cell)):
logger.warning("'fn' is not a mindspore.nn.Cell, and its definition cannot involve Parameter; "
"otherwise, the result may be incorrect.")
return Shard()(fn, in_strategy, out_strategy, parameter_plan, device, level)