Making large AI models cheaper, faster and more accessible
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
 
 
 

507 lines
22 KiB

import operator
from copy import deepcopy
from typing import Dict, List, Union
import torch
from torch.fx import symbolic_trace
from torch.fx.node import Node
from colossalai._analyzer.fx.node_util import MetaInfo
from colossalai.auto_parallel.tensor_shard.constants import RESHAPE_FUNC_OP
from colossalai.auto_parallel.tensor_shard.sharding_strategy import (
CommAction,
CommType,
OperationDataType,
ShardingStrategy,
)
from colossalai.auto_parallel.tensor_shard.solver.strategies_constructor import StrategiesConstructor
from colossalai.device.device_mesh import DeviceMesh
from colossalai.tensor.comm_spec import _all_reduce
from colossalai.tensor.shape_consistency import ShapeConsistencyManager
from colossalai.tensor.sharding_spec import ShardingSpec
from .constants import SHAPE_ARGUMENT_OPS
shape_consistency_manager = ShapeConsistencyManager()
def size_processing(size: Union[int, torch.Size],
dim_partition_dict: Dict[int, List[int]],
device_mesh_info: Dict[int, int],
target_dim: int = None,
node_name: str = None):
"""
This method will be invoked during runtime to convert size node value depending on distributed information.
"""
if target_dim is not None:
assert isinstance(size, int)
if target_dim in dim_partition_dict:
total_shard_size = 1
for shard_dim in dim_partition_dict[target_dim]:
total_shard_size *= device_mesh_info[shard_dim]
size = size * total_shard_size
else:
size = list(size)
for dim, dim_size in enumerate(size):
if dim in dim_partition_dict:
total_shard_size = 1
for shard_dim in dim_partition_dict[dim]:
total_shard_size *= device_mesh_info[shard_dim]
size[dim] = dim_size * total_shard_size
size = torch.Size(size)
return size
def solution_annotation_pass(gm: torch.fx.GraphModule, solution: List[int],
strategies_constructor: StrategiesConstructor):
"""
This method is used to stick the solution strategy to the nodes and add the information
required in runtime into graph as placeholder nodes.
"""
mod_graph = gm.graph
nodes = [strategies_vector.node for strategies_vector in strategies_constructor.leaf_strategies]
no_strategy_nodes = strategies_constructor.no_strategy_nodes
# the dict to get origin sharding spec of node
origin_node_sharding_spec_dict = {}
for node_index, (node, strategy_index) in enumerate(zip(nodes, solution)):
strategies_vector = node.strategies_vector
# stick the solution strategy to the corresponding node
setattr(node, 'best_strategy', strategies_vector[strategy_index])
setattr(node, 'sharding_spec', strategies_vector[strategy_index].get_sharding_spec_by_name(str(node)))
origin_node_sharding_spec_dict[node_index] = strategies_vector[strategy_index].get_sharding_spec_by_name(
str(node))
# attach the corresponding metainfo if node has the attribute `strategies_info`
if hasattr(node, 'strategies_info'):
setattr(node, 'best_strategy_info', node.strategies_info[strategy_index])
# the dict to get input sharding specs of user node
sharding_spec_convert_dict = {}
# the dict to record comm actions of nodes
comm_actions_dict = {}
for index, node in enumerate(nodes):
target_sharding_specs = []
for user_node in node.strategies_vector.successor_nodes:
if user_node in no_strategy_nodes:
target_sharding_spec = node.best_strategy.get_sharding_spec_by_name(str(node.name))
else:
target_sharding_spec = user_node.best_strategy.get_sharding_spec_by_name(str(node.name))
target_sharding_specs.append(target_sharding_spec)
sharding_spec_convert_dict[index] = target_sharding_specs
setattr(node, 'target_sharding_specs', target_sharding_specs)
# the get_attr node strategy is kind of pending strategy, which means we will change it
# to the same strategy of the user node.
if node.op == 'get_attr':
assert len(target_sharding_specs) == 1, f'sharing weight is not supported in current version.'
target_node = node.strategies_vector.successor_nodes[0]
node_name = str(node)
if target_node.op == 'call_function' and target_node.target in RESHAPE_FUNC_OP:
node_name = str(target_node)
target_node = target_node.strategies_vector.successor_nodes[0]
user_strategy = target_node.best_strategy
op_data_in_user = user_strategy.get_op_data_by_name(node_name)
origin_pending_strategy = node.best_strategy
origin_op_data = origin_pending_strategy.get_op_data_by_name(str(node))
new_communication_actions = {}
if op_data_in_user in user_strategy.communication_actions:
new_communication_action = user_strategy.communication_actions.pop(op_data_in_user)
new_communication_action.arg_index = 0
new_communication_actions[origin_op_data] = new_communication_action
node.best_strategy.communication_actions = new_communication_actions
comm_action_dict = {}
for op_data, comm_action in node.best_strategy.communication_actions.items():
comm_action_dict[op_data.name] = comm_action
comm_actions_dict[index] = comm_action_dict
# add above dicts into graph
for node in nodes:
if node.op != 'placeholder':
with mod_graph.inserting_before(node):
input_specs_node = mod_graph.create_node('placeholder', target='sharding_spec_convert_dict')
origin_specs_node = mod_graph.create_node('placeholder', target='origin_node_sharding_spec_dict')
comm_actions_dict_node = mod_graph.create_node('placeholder', target='comm_actions_dict')
break
return gm, sharding_spec_convert_dict, origin_node_sharding_spec_dict, comm_actions_dict
def size_value_converting_pass(gm: torch.fx.GraphModule, device_mesh: DeviceMesh):
"""
In the auto parallel system, tensors may get shard on different devices, so the size of tensors
need to be converted to the size of original tensor and managed by the users, such as torch.view,
torch.reshape, etc. These nodes have enough information like input sharding_spec and
output sharding_spec to decide how to convert the size value.
"""
mod_graph = gm.graph
nodes = tuple(mod_graph.nodes)
node_pairs = {}
# DeviceMesh information instructs the scaling of the size value
device_mesh_info = {}
for dim, dim_size in enumerate(device_mesh.shape):
device_mesh_info[dim] = dim_size
def _extract_target_dim(node):
'''
A helper function to extract the target dimension from size node.
There are two usages of torch.Tensor.size:
1. tensor.size()
2. tensor.size(dim)
If a target_dim is assigned, then the output will be in type of int, instead of torch.Size.
Otherwise, the output will be in type of torch.Size and this function will return None.
'''
target_dim = None
if len(node.args) > 1:
target_dim = node.args[1]
if target_dim < 0:
target_dim += node.args[0]._meta_data.dim()
return target_dim
def _post_processing(node, size_processing_node):
'''
This function is used to process the dependency between the size node and its users after
inserting the size_process_node.
'''
# store original node and processing node pair in node_pairs dictionary
# It will be used to replace the original node with processing node in slice object
node_pairs[node] = size_processing_node
size_processing_node._meta_data = node._meta_data
if hasattr(node.meta['info'], 'activation_checkpoint'):
MetaInfo(size_processing_node,
mod_dir=node.meta['info'].mod_dir,
activation_checkpoint=tuple(node.meta['info'].activation_checkpoint))
user_list = list(node.users.keys())
for user in user_list:
if user == size_processing_node:
continue
new_args = list(user.args)
new_kwargs = dict(user.kwargs)
# the origin node may be a positional argument or key word argument of user node
if node in new_args:
# substitute the origin node with size_processing_node
new_args[new_args.index(node)] = size_processing_node
user.args = tuple(new_args)
elif str(node) in new_kwargs:
# substitute the origin node with size_processing_node
new_kwargs[str(node)] = size_processing_node
user.kwargs = new_kwargs
def _update_slice_object_args(slice_object):
'''
This function is used to update the slice object argument list.
If the slice object contains the Node argument, then the size node will be replaced with
'''
if isinstance(slice_object, slice):
start = slice_object.start
stop = slice_object.stop
step = slice_object.step
if start in node_pairs:
start = node_pairs[start]
if stop in node_pairs:
stop = node_pairs[stop]
if step in node_pairs:
step = node_pairs[step]
return slice(start, stop, step)
elif isinstance(slice_object, int):
if slice_object in node_pairs:
return node_pairs[slice_object]
else:
return slice_object
else:
raise RuntimeError(f"Unsupported slice object type: {type(slice_object)}")
for node in nodes:
if node.op == 'call_method' and node.target == 'size':
# extract useful information from size node
# dim_partition_dict will instruct the size value on which
# dimension should be enlarged.
sharding_spec = node.args[0].sharding_spec
dim_partition_dict = sharding_spec.dim_partition_dict
target_dim = _extract_target_dim(node)
# insert size_processing node
with mod_graph.inserting_after(node):
size_processing_node = mod_graph.create_node('call_function',
size_processing,
args=(node, dim_partition_dict, device_mesh_info,
target_dim, node.name))
_post_processing(node, size_processing_node)
if node.op == 'call_function' and node.target == operator.getitem:
getitem_index = node.args[1]
# slice object is quite special in torch.fx graph,
# On one side, we treat slice object same as type of int,
# so we do not create a node for slice object. On the other side,
# slice object could take fx.Node as its argument. And the user
# relationship cannot be tracked in fx graph.
# Therefore, I record the node_pairs in this pass, and use the it
# to replace the original node argument inside the slice object if
# it has been processed in above pass.
# There are three main usages of operator.getitem:
# getitem(input, int)
# getitem(input, slice)
# getitem(input, Tuple[slice])
# In this pass, we need process the last two cases because
# node arguments may potentially appear in these cases.
if isinstance(getitem_index, slice):
new_slice_item = _update_slice_object_args(getitem_index)
new_args = (node.args[0], new_slice_item)
node.args = new_args
elif isinstance(getitem_index, (tuple, list)):
if not isinstance(getitem_index[0], slice):
continue
new_slice_items = []
for slice_item in getitem_index:
if slice_item is None:
new_slice_items.append(None)
continue
new_slice_item = _update_slice_object_args(slice_item)
new_slice_items.append(new_slice_item)
new_args = (node.args[0], tuple(new_slice_items))
node.args = new_args
return gm
def node_args_converting_pass(gm: torch.fx.GraphModule, device_mesh: DeviceMesh):
"""
This pass will process node args to adapt the distributed tensor layout.
"""
mod_graph = gm.graph
nodes = tuple(mod_graph.nodes)
def _extract_info_from_sharding_spec(sharding_spec):
'''
This function is used to extract the dim_partition_dict and device_mesh from
sharding spec instance or a list of sharding spec.
'''
if isinstance(sharding_spec, ShardingSpec):
dim_partition_dict = sharding_spec.dim_partition_dict
device_mesh = sharding_spec.device_mesh
return dim_partition_dict, device_mesh
if sharding_spec is None:
return None, None
assert isinstance(sharding_spec,
(tuple, list)), 'sharding_spec should be type of ShardingSpec, tuple, list or None'
device_mesh = sharding_spec[0].device_mesh
dim_partition_dict = []
for element in sharding_spec:
dim_partition_dict.append(_extract_info_from_sharding_spec(element))
return dim_partition_dict, sharding_spec
def _process_node_arguments(node):
new_args = []
for arg in node.args:
# There are two args style:
# 1. (input, *shape)
# 2. (input, shape)
# We will extract the elements from shape and add them into the new_args
# Finally, the args style of new_args will be unified to (input, *shape)
if isinstance(arg, Node):
if isinstance(arg._meta_data, (tuple, list)):
new_args.extend(arg._meta_data)
elif isinstance(arg._meta_data, int):
new_args.append(arg._meta_data)
else:
new_args.append(arg)
else:
assert isinstance(arg,
(int, tuple, list)), 'The argument in view node should be either type of Node or int.'
if isinstance(arg, (tuple, list)):
new_args.extend(arg)
else:
new_args.append(arg)
return new_args
def _scale_args_adapt_sharding_spec(dim_partition_dict, device_mesh, node):
new_args = _process_node_arguments(node)
if node.op == 'call_method':
args_to_process = list(new_args[1:])
else:
args_to_process = list(new_args)
for dim, shard_dims in dim_partition_dict.items():
total_shard_size = 1
for shard_dim in shard_dims:
total_shard_size *= device_mesh.shape[shard_dim]
# we will skip the dim with -1 value
if args_to_process[dim] == -1:
continue
else:
# TODO: add assertion here to make sure the dim size is divisible by total_shard_size
args_to_process[dim] //= total_shard_size
args_to_process = tuple(args_to_process)
if node.op == 'call_method':
new_args = (new_args[0],) + args_to_process
else:
new_args = args_to_process
node.args = new_args
def _filter_node_with_shape_args(node):
if node.op == 'call_method':
target = getattr(node.args[0]._meta_data.__class__, node.target)
elif node.op == 'call_function':
target = node.target
else:
target = None
if target in SHAPE_ARGUMENT_OPS:
return True
return False
for node in nodes:
# skip the placeholder node added in _solution_annotation pass
if not hasattr(node, 'sharding_spec'):
continue
output_dim_partition_dict, device_mesh = _extract_info_from_sharding_spec(node.sharding_spec)
if _filter_node_with_shape_args(node):
_scale_args_adapt_sharding_spec(output_dim_partition_dict, device_mesh, node)
return gm
def module_params_sharding_pass(gm: torch.fx.GraphModule, device_mesh: DeviceMesh, overlap=False):
"""
Apply the sharding action to the module parameters and buffers following the
instructions of solver solution.
"""
mod_graph = gm.graph
nodes = tuple(mod_graph.nodes)
# This stream is created for overlapping the communication and computation.
reduction_stream = torch.cuda.Stream()
def _add_hook_for_grad_communication(node, param, name=None):
comm_actions = node.best_strategy.communication_actions
def _filter_param_to_hook(node, op_data, comm_action, name):
if node.op == 'call_module' and op_data.type == OperationDataType.PARAM and op_data.name == name and comm_action.comm_type == CommType.HOOK:
return True
if node.op == 'get_attr' and isinstance(
node._meta_data, torch.nn.parameter.Parameter) and comm_action.comm_type == CommType.HOOK:
return True
return False
for operation_data, comm_action in comm_actions.items():
comm_spec_to_use = comm_action.comm_spec
# register hook to the parameters
if _filter_param_to_hook(node, operation_data, comm_action, name=name):
def wrapper(param, comm_spec, stream, overlap):
def hook_fn(grad):
if overlap:
with torch.cuda.stream(stream):
_all_reduce(grad, comm_spec, async_op=True)
else:
_all_reduce(grad, comm_spec, async_op=False)
param.register_hook(hook_fn)
wrapper(param, comm_spec_to_use, reduction_stream, overlap=overlap)
def _shard_param(param, target_sharding_spec):
# apply the sharding spec of parameters
if target_sharding_spec.dim_partition_dict != {}:
origin_sharding_spec = ShardingSpec(device_mesh, param.shape, {})
setattr(param, 'sharding_spec', origin_sharding_spec)
# TODO: build a ColoParameter class to manager the distributed parameters
# we could use .data here, because all the operations just happen before the real training
# loop, so we don't need to track these operations in the autograd graph.
param = torch.nn.Parameter(
shape_consistency_manager.apply_for_autoparallel_runtime(param.data, param.sharding_spec,
target_sharding_spec).detach().clone())
return param
for node in nodes:
if node.op == 'call_module':
target_module = node.graph.owning_module.get_submodule(node.target)
# TODO: we need to do more actions to take care of the shared parameters.
if hasattr(target_module, 'processed') and target_module.processed:
continue
setattr(target_module, 'processed', True)
for name, param in target_module.named_parameters():
target_sharding_spec = node.best_strategy.get_sharding_spec_by_name(name)
param = _shard_param(param, target_sharding_spec)
setattr(target_module, name, param)
_add_hook_for_grad_communication(node, param, name)
sharded_buffer_dict = {}
# apply the sharding spec of buffers
for name, buffer in target_module.named_buffers():
origin_sharding_spec = ShardingSpec(device_mesh, buffer.shape, {})
setattr(buffer, 'sharding_spec', origin_sharding_spec)
target_sharding_spec = node.best_strategy.get_sharding_spec_by_name(name)
buffer_sharded = shape_consistency_manager.apply(buffer, target_sharding_spec)
sharded_buffer_dict[name] = buffer_sharded
for name, buffer_sharded in sharded_buffer_dict.items():
setattr(target_module, name, buffer_sharded.detach().clone())
if node.op == 'get_attr':
root = node.graph.owning_module
atoms = node.target.split(".")
attr_len = len(atoms)
if attr_len == 1:
target_module = root
target = getattr(root, atoms[0])
else:
target_module = root
for atom in atoms[:-1]:
target_module = getattr(target_module, atom)
target = getattr(target_module, atoms[-1])
target_sharding_spec = node.sharding_spec
target = _shard_param(target, target_sharding_spec)
assert hasattr(target_module, atoms[-1])
setattr(target_module, atoms[-1], target)
_add_hook_for_grad_communication(node, target)
return gm
def implicit_comm_action_apply(gm: torch.fx.GraphModule):
"""
replace the origin kernel into kernel with implicit communication inside.
"""
pass
def runtime_preparation_pass(gm: torch.fx.GraphModule,
solution: List[int],
device_mesh: DeviceMesh,
strategies_constructor: StrategiesConstructor,
overlap=False):
gm, sharding_spec_convert_dict, origin_node_sharding_spec_dict, comm_actions_dict = solution_annotation_pass(
gm, solution, strategies_constructor)
gm = size_value_converting_pass(gm, device_mesh)
gm = node_args_converting_pass(gm, device_mesh)
# TODO: the pass below should be uncommented after the implementation of implicit_comm_action_apply_pass completed.
# gm = implicit_comm_action_apply(gm)
gm = module_params_sharding_pass(gm, device_mesh, overlap=overlap)
return gm, sharding_spec_convert_dict, origin_node_sharding_spec_dict, comm_actions_dict