ColossalAI/colossalai/fx/passes/shard_1d_pass.py

121 lines
4.8 KiB
Python

import torch
from torch.fx.node import map_arg
from torch.fx.node import Node
from torch.fx.passes.split_module import split_module
import colossalai
from colossalai.context import ParallelMode
from colossalai.core import global_context as gpc
def all_gather_function(input_):
world_size = gpc.get_world_size(ParallelMode.PARALLEL_1D)
rank = gpc.get_local_rank(ParallelMode.PARALLEL_1D)
tensor_list = [torch.empty_like(input_) for _ in range(world_size)]
tensor_list[rank] = input_
group = gpc.get_group(ParallelMode.PARALLEL_1D)
torch.distributed.all_gather(tensor_list, input_, group=group)
output = torch.cat(tensor_list, dim=-1).contiguous()
return output
def all_reduce_function(input_):
if gpc.get_world_size(ParallelMode.PARALLEL_1D) == 1:
return input_
torch.distributed.all_reduce(input_, group=gpc.get_group(ParallelMode.PARALLEL_1D))
return input_
def weight_split(weight, dim):
#TODO: this function will be refactored by using ColoTensor dist_spec when a stable reshaper feature is ready to use.
num_partition = gpc.get_world_size(ParallelMode.TENSOR)
shape = weight.shape
length = shape[dim] // num_partition
sharded_weight_list = []
for i in range(num_partition):
sharded_weight_list.append(weight.narrow(dim, i * length, length))
return sharded_weight_list[gpc.get_local_rank(ParallelMode.PARALLEL_1D)]
def replace_all_uses_except_replaced(node, replace_node):
"""
Replace all uses of ``node`` in the Graph with the Node ``replace_node``,
except the user of ``node`` is ``replace_node``.
Args:
replace_node (Node): The node to replace all uses of ``node`` with.
Returns:
The list of Nodes on which this change was made.
"""
to_process = list(node.users)
for use_node in to_process:
if use_node == replace_node:
continue
def may_replace_node(n):
if n == node:
return replace_node
else:
return n
new_args = map_arg(use_node.args, may_replace_node)
new_kwargs = map_arg(use_node.kwargs, may_replace_node)
use_node._args = new_args
use_node._kwargs = new_kwargs
for old_use in use_node._input_nodes.keys():
old_use.users.pop(use_node)
use_node._input_nodes = {}
map_arg(use_node._args, lambda n: use_node._input_nodes.setdefault(n))
map_arg(use_node._kwargs, lambda n: use_node._input_nodes.setdefault(n))
for new_use in use_node._input_nodes.keys():
new_use.users.setdefault(use_node)
return to_process
def column_shard_linear_pass(gm: torch.fx.GraphModule):
mod_graph = gm.graph
for node in mod_graph.nodes:
if node.op == "call_module":
target_module = node.graph.owning_module.get_submodule(node.target)
if isinstance(target_module, torch.nn.Linear):
target_module.weight.data = weight_split(target_module.weight.data, dim=0)
if target_module.bias is not None:
target_module.bias.data = weight_split(target_module.bias.data, dim=0)
# inserting communication node after the sharded linear node
with mod_graph.inserting_after(node):
new_node = mod_graph.create_node('call_function', all_gather_function, args=(node,))
replace_all_uses_except_replaced(node, new_node)
gm.recompile()
return gm
def row_shard_linear_pass(gm: torch.fx.GraphModule):
mod_graph = gm.graph
for node in mod_graph.nodes:
if node.op == "call_module":
target_module = node.graph.owning_module.get_submodule(node.target)
if isinstance(target_module, torch.nn.Linear):
target_module.weight.data = weight_split(target_module.weight.data, dim=-1)
# insert input sharding node before the sharded linear node
with mod_graph.inserting_before(node):
input_node_list = list(node._input_nodes.keys())
assert len(input_node_list) == 1, 'linear forward must have and only have one input tensor.'
input_node = input_node_list[0]
new_input_node = mod_graph.create_node('call_function', weight_split, args=(input_node, -1))
replace_all_uses_except_replaced(input_node, new_input_node)
# inserting communication node after the sharded linear node
with mod_graph.inserting_after(node):
new_node = mod_graph.create_node('call_function', all_reduce_function, args=(node,))
replace_all_uses_except_replaced(node, new_node)
gm.recompile()
return gm
#TODO: add elementwise op process pass, then we can try to use column and row mixed strategy.