# Copyright 2020 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.
# ============================================================================
"""comm_helper"""
import os
import glob
import ctypes
import sys
from sys import excepthook
from mindspore import context
from mindspore.parallel._ps_context import _is_role_worker, _is_role_pserver, _is_role_sched, _is_ps_mode,\
_get_ps_context
from mindspore import log as logger
from mindspore._c_expression import CollectiveManager, set_cluster_exit_with_exception, MSContext
from mindspore.common._utils import load_lib
HCCL_LIB = 'libhccl_plugin.so'
def hccl_load_lib():
"""load hccl lib"""
try:
base_dir = os.path.dirname(os.path.realpath(__file__))
lib_path = os.path.join(base_dir, "../lib/plugin/ascend", HCCL_LIB)
ctypes.CDLL(lib_path)
except Exception as exc:
raise RuntimeError('Get hccl lib error.') from exc
_HCCL_TEST_AVAILABLE = False
try:
if MSContext.get_instance().is_ascend_plugin_loaded():
hccl_load_lib()
except RuntimeError:
_HCCL_TEST_AVAILABLE = True
if _HCCL_TEST_AVAILABLE:
try:
import hccl_test.manage.api as hccl
except ImportError:
_HCCL_TEST_AVAILABLE = False
HCCL_WORLD_COMM_GROUP = "hccl_world_group"
NCCL_WORLD_COMM_GROUP = "nccl_world_group"
MCCL_WORLD_COMM_GROUP = "mccl_world_group"
DEVICE_TO_BACKEND = {
"Ascend": "hccl",
"GPU": "nccl",
"CPU": "mccl"
}
class Backend:
"""
Class for available backends.
Note:
The backends' value should be string, e.g., "hccl".
If backend is set to Backend.UNDEFINED, it will be seen as invaliad.
Args:
name (str): The name of backend.
Raises:
TypeError: If name is not a string.
ValueError: If backend is invalid.
Examples:
>>> Backend("abc")
>>> hccl = Backend("hccl")
"""
UNDEFINED = "undefined"
HCCL = "hccl"
NCCL = "nccl"
MCCL = "mccl"
@staticmethod
def __new__(cls, name):
"""Create instance object of Backend."""
if not isinstance(name, str):
raise TypeError("For 'Backend', the class variable 'name' must be a string, "
"but got the type : {}".format(type(name)))
value = getattr(Backend, name.upper(), Backend.UNDEFINED)
if value == Backend.UNDEFINED:
raise ValueError("For 'Backend', the class variable 'name' {} is not supported, "
"please use hccl or nccl.".format(name))
return value
DEFAULT_BACKEND = Backend("hccl")
[文档]class GlobalComm:
"""
World communication information. The GlobalComm is a global class. The members contain:
- ``BACKEND`` : The communication library used, using ``"hccl"`` / ``"nccl"`` / ``"mccl"`` .
``"hccl"`` means Huawei Collective Communication Library(HCCL),
``"nccl"`` means NVIDIA Collective Communication Library(NCCL),
``"mccl"`` means MindSpore Collective Communication Library(MCCL).
- ``WORLD_COMM_GROUP`` : Global communication domain,
using ``"hccl_world_group"`` / ``"nccl_world_group"`` / ``"mccl_world_group"`` .
"""
BACKEND = DEFAULT_BACKEND
WORLD_COMM_GROUP = HCCL_WORLD_COMM_GROUP
INITED = False
CHECK_ENVS = True
class _ExistingGroup:
"""
The communication groups which exist in the progress.
"""
ITEMS = {}
def _hccl_test():
return _HCCL_TEST_AVAILABLE and GlobalComm.BACKEND == Backend.HCCL
def _check_mpi_envs():
"""
Check whether mpi environment variables have been exported or not.
return True if mpi environment variables have been exported, False otherwise.
"""
ompi_command_env = os.getenv("OMPI_COMMAND")
pmix_rank_env = os.getenv("PMIX_RANK")
if ompi_command_env and pmix_rank_env:
return True
return False
def _check_bypass_rank_id_and_size():
'''
Whether bypass calling c++ API to get rank id and size, instead, use fake rank id 0 and rank size 1.
This returns True when this process is Scheduler node or is Server node in old Parameter Server training mode.
'''
if _is_role_sched():
return True
device_target = context.get_context("device_target")
if _is_ps_mode() and _get_ps_context("worker_num") == 1 and device_target == "Ascend":
return True
return False
def _set_elegant_exit_handle():
if _is_role_worker() or _is_role_pserver() or _is_role_sched():
sys.excepthook = lambda *args: (set_cluster_exit_with_exception(), excepthook(*args))
def check_parameter_available(func):
"""
Check parameter is available. If not available, raise Error.
Args:
func (Function): The function to be run.
Raises:
RuntimeError.
Returns:
Wrapper. If not available, raise Error.
"""
def wrapper(*args, **kargs):
if not GlobalComm.INITED:
raise RuntimeError("Distributed Communication has not been inited")
group = None
if "group" in kargs.keys():
group = kargs.get("group")
if group is not None and not isinstance(group, str):
raise TypeError("The parameter 'group' should be str or None, "
"but got the type : {}".format(type(group)))
if group is None:
group = GlobalComm.WORLD_COMM_GROUP
return func(*args, **kargs)
return wrapper
def _is_available():
"""
Returns `True` if distributed module is available.
Note:
Always returns `True` because MindSpore always has distributed ability on all platforms.
"""
return True
def _is_initialized():
"""
Checks if distributed module is successfully initialized.
"""
return CollectiveManager.get_instance().initialized()
def _get_backend():
"""
Returns the backend of communication process groups.
Note:
Only one communication backend is supported by MindSpore for each process.
It should be one of `hccl`/`nccl`/`mccl`.
"""
return GlobalComm.BACKEND
def _is_hccl_available():
"""
Checks if `hccl` backend is available.
"""
return _HCCL_TEST_AVAILABLE
def _is_nccl_available():
"""
Checks if `nccl` backend is available.
"""
base_dir = os.path.dirname(os.path.realpath(__file__))
lib_path = os.path.join(base_dir, "../lib/plugin/gpu*/libnvidia_collective.so")
file_paths = glob.glob(lib_path)
return all(list(load_lib(f) for f in file_paths))
def _is_mpi_available():
"""
Checks if OpenMPI's library is available.
"""
base_dir = os.path.dirname(os.path.realpath(__file__))
lib_path = os.path.join(base_dir, "../lib/libmpi_collective.so")
return load_lib(lib_path)
@check_parameter_available
def _get_rank_helper(group):
"""
The Helper to do get_rank_id.
Args:
group (str): The communication group.
backend (str): The backend, like "hccl".
Raises:
ValueError: If backend is invalid.
Returns:
Integer. The local rank id of the calling process.
"""
if _check_bypass_rank_id_and_size():
rank_id = 0
return rank_id
if _hccl_test():
return hccl.get_rank_id(group)
rank_id = CollectiveManager.get_instance().get_rank_id(group)
return rank_id
@check_parameter_available
def _get_local_rank_helper(group):
"""
The Helper to do get_local_rank_id.
Args:
group (str): The communication group.
backend (str): The backend, like "hccl".
Raises:
ValueError: If backend is invalid.
Returns:
Integer. The local rank id of the calling process.
"""
if _check_bypass_rank_id_and_size():
local_rank_id = 0
return local_rank_id
if _hccl_test():
return hccl.get_local_rank_id(group)
rank_id = CollectiveManager.get_instance().get_local_rank_id(group)
return rank_id
@check_parameter_available
def _get_size_helper(group):
"""
The Helper to do get_rank_size.
Args:
group (str): The communication group.
backend (str): The backend, like "hccl".
Raises:
ValueError: If backend is invalid.
Returns:
Integer. The rank size of specified group.
"""
if _check_bypass_rank_id_and_size():
size = 1
return size
if _hccl_test():
return hccl.get_rank_size(group)
size = CollectiveManager.get_instance().get_group_size(group)
return size
@check_parameter_available
def _get_local_size_helper(group):
"""
The Helper to do get_local_rank_size.
Args:
group (str): The communication group.
backend (str): The backend, like "hccl".
Raises:
ValueError: If backend is invalid.
Returns:
Integer. The local rank size where the calling process is being within specified group.
"""
size = CollectiveManager.get_instance().get_local_group_size(group)
return size
@check_parameter_available
def _get_world_rank_from_group_rank_helper(group, group_rank_id):
"""
The Helper to do get_world_rank_from_group_rank.
Args:
group (str): The user communication group.
group_rank_id (int): A rank id in user communication group.
backend (str): The backend, like "hccl".
Raises:
TypeError: If group_rank_id is not int.
ValueError: If group is "hccl_world_group" or backend is invalid.
Returns:
Integer. A rank id in world communication group.
"""
if not isinstance(group_rank_id, int):
raise TypeError("For 'get_world_rank_from_group_rank', the argument 'group_rank_id' must be"
" type of int, but got 'group_rank_id' type : {}.".format(type(group_rank_id)))
if _hccl_test():
return hccl.get_world_rank_from_group_rank(group, group_rank_id)
world_rank_id = CollectiveManager.get_instance().get_world_rank_from_group_rank(group, group_rank_id)
return world_rank_id
@check_parameter_available
def _get_group_rank_from_world_rank_helper(world_rank_id, group):
"""
The Helper to do get_group_rank_from_world_rank.
Args:
world_rank_id (int): A rank id in world communication group.
group (str): The user communication group.
backend (str): The backend, like "hccl".
Raises:
TypeError: If world_rank_id is not int.
ValueError: If group is 'hccl_world_group' or backend is invalid.
Returns:
Integer. A rank id in user communication group.
"""
group_rank_id = None
if not isinstance(world_rank_id, int):
raise TypeError("For 'get_group_rank_from_world_rank', the argument 'world_rank_id' must be type of int, "
"but got 'world_rank_id' type : {}.".format(type(world_rank_id)))
if _hccl_test():
return hccl.get_group_rank_from_world_rank(world_rank_id, group)
group_rank_id = CollectiveManager.get_instance().get_group_rank_from_world_rank(world_rank_id, group)
return group_rank_id
@check_parameter_available
def _get_group_ranks(group):
"""
The Helper to do get_group_ranks.
Args:
group (str): The communication group.
Returns:
List. The ranks of specified group.
"""
return CollectiveManager.get_instance().get_group_ranks(group)
@check_parameter_available
def _create_group_helper(group, rank_ids):
"""
The Helper to do create_group.
Args:
group (str): The communication group.
rank_ids (list): Rank ids in the group.
backend (str): The backend, like "hccl".
Raises:
TypeError: If rank_ids is not a list.
ValueError: If rank_ids size is not larger than 1 or rank_ids has duplicate data or backend is invalid.
"""
if group in _ExistingGroup.ITEMS.keys():
if rank_ids != _ExistingGroup.ITEMS.get(group):
raise ValueError("The group {} has been created, the rank_list is {}, "
"but current rank_list for the group is {}".
format(group, _ExistingGroup.ITEMS[group], rank_ids))
logger.warning("%r group has existed.", group)
return
if not isinstance(rank_ids, list):
raise TypeError("For 'create_group', the argument 'rank_ids' must be type of list, "
"but got 'rank_ids' type : {}.".format(type(rank_ids)))
rank_size = len(rank_ids)
if rank_size < 1:
raise ValueError("For 'create_group', the argument 'rank_ids' size should be greater than 1, "
"but got 'rank_ids' size : {}.".format(len(rank_ids)))
if len(rank_ids) - len(list(set(rank_ids))) > 0:
raise ValueError("List rank_ids in Group {} has duplicate data!".format(group))
if _hccl_test():
hccl.create_group(group, rank_size, rank_ids)
else:
result = CollectiveManager.get_instance().create_group(group, rank_ids)
if not result:
raise RuntimeError("Failed to create communication group for {} with rank ids {}. "
"If NCCL is used, 'export NCCL_DEBUG=INFO' "
"is suggested before launching jobs.".format(group, rank_ids))
_ExistingGroup.ITEMS[group] = rank_ids
@check_parameter_available
def _destroy_group_helper(group):
"""
The Helper to do destroy_group.
Args:
group (str): The user communication group.
backend (str): The backend, like "hccl".
Raises:
ValueError: If group is "hccl_world_group" or backend is invalid.
"""
if group == GlobalComm.WORLD_COMM_GROUP:
raise ValueError("The world_group does not support destruction.")
if _hccl_test():
hccl.create_group(group)
else:
CollectiveManager.get_instance().destroy_group(group)