mirror of
https://github.com/hpcaitech/ColossalAI.git
synced 2025-09-09 04:50:17 +00:00
[autoparallel] refactor and add rotorc. (#1789)
* [autoparallel] refactor and add rotorc. * [autoparallel] refactor and add rotorc.
This commit is contained in:
@@ -1,6 +1,6 @@
|
||||
import math
|
||||
from abc import ABC
|
||||
from typing import List
|
||||
from typing import Any, Iterable, List
|
||||
|
||||
from torch.utils._pytree import tree_map
|
||||
|
||||
@@ -33,23 +33,25 @@ class Chain:
|
||||
self.xbar = xbar
|
||||
self.ftmp = ftmp
|
||||
self.btmp = btmp
|
||||
self.length = len(ftime)
|
||||
if check_consistency and not self.check_lengths():
|
||||
raise AttributeError("In Chain, input lists do not have consistent lengths")
|
||||
|
||||
def check_lengths(self):
|
||||
return ((len(self.ftime) == self.length) and (len(self.btime) == self.length + 1)
|
||||
and (len(self.x) == self.length + 1) and (len(self.ftmp) == self.length)
|
||||
and (len(self.btmp) == self.length + 1) and (len(self.xbar) == self.length + 1))
|
||||
return ((len(self.ftime) == len(self)) and (len(self.btime) == len(self) + 1) and (len(self.x) == len(self) + 1)
|
||||
and (len(self.ftmp) == len(self)) and (len(self.btmp) == len(self) + 1)
|
||||
and (len(self.xbar) == len(self) + 1))
|
||||
|
||||
def __repr__(self):
|
||||
chain_list = []
|
||||
for i in range(self.length):
|
||||
for i in range(len(self)):
|
||||
chain_list.append((self.ftime[i], self.btime[i], self.x[i], self.xbar[i], self.ftmp[i], self.btmp[i]))
|
||||
i = self.length
|
||||
i = len(self)
|
||||
chain_list.append((None, self.btime[i], self.x[i], self.xbar[i], None, self.btmp[i]))
|
||||
return chain_list.__repr__()
|
||||
|
||||
def __len__(self):
|
||||
return len(self.ftime)
|
||||
|
||||
def discretize_all(self, unit: int):
|
||||
"""Discretize the chain into a list of chains according to unit size."""
|
||||
discretizer = lambda val: math.ceil(val / unit)
|
||||
@@ -163,79 +165,20 @@ class DiscardMemory(MemoryAccess):
|
||||
name = "DM"
|
||||
|
||||
|
||||
class Function:
|
||||
class Sequence(list):
|
||||
|
||||
def __init__(self, name, *args):
|
||||
self.name = name
|
||||
self.args = args
|
||||
self.str_args = ','.join(str(v) for v in self.args)
|
||||
|
||||
def __repr__(self):
|
||||
return "{n}({args})".format(n=self.name, args=self.str_args)
|
||||
|
||||
|
||||
class Sequence:
|
||||
|
||||
def __init__(self, function):
|
||||
self.sequence = [] #List of Operation and Sequence
|
||||
self.function = function #Description the function (name and parameters)
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
|
||||
def __repr__(self):
|
||||
return repr(self.list_operations())
|
||||
|
||||
def list_operations(self):
|
||||
op_list = []
|
||||
for x in self.sequence:
|
||||
for x in self:
|
||||
if isinstance(x, Operation):
|
||||
op_list.append(x)
|
||||
else:
|
||||
assert isinstance(x, Sequence)
|
||||
op_list += x.list_operations()
|
||||
return op_list
|
||||
|
||||
def insert(self, operation):
|
||||
self.sequence.append(operation)
|
||||
|
||||
def remove(self, operation_index):
|
||||
del self.sequence[operation_index]
|
||||
|
||||
def insert_sequence(self, sequence):
|
||||
self.sequence.append(sequence)
|
||||
|
||||
def shift(self, value):
|
||||
for x in self.sequence:
|
||||
x.shift(value)
|
||||
return self
|
||||
|
||||
def remove_useless_write(self):
|
||||
if self.sequence:
|
||||
if isinstance(self.sequence[0], WriteMemory):
|
||||
self.remove(0)
|
||||
return self
|
||||
|
||||
def get_makespan(self, chain):
|
||||
return sum(op.cost(chain) for op in self.list_operations())
|
||||
|
||||
def without_suffix(self):
|
||||
ops = self.list_operations()
|
||||
end_of_first_phase = [i for i in range(len(ops)) if type(ops[i]) is Loss][0]
|
||||
try:
|
||||
last_idx = max(i for i in range(end_of_first_phase) if not type(ops[i]) is ForwardEnable)
|
||||
except ValueError:
|
||||
last_idx = -1
|
||||
if last_idx == end_of_first_phase - 1:
|
||||
return (self, None)
|
||||
chain_length = ops[end_of_first_phase -
|
||||
1].index ## Some assumption here about the sequence (finishes with Forward_L
|
||||
start_of_fwd_enable_chain = ops[last_idx + 1].index ## And starts with B_L), but should be fine in practice
|
||||
result = Sequence(Function("Strip", self.function.name, *self.function.args, start_of_fwd_enable_chain))
|
||||
for i in range(last_idx + 1):
|
||||
result.insert(ops[i])
|
||||
result.insert(Loss())
|
||||
for i in range(chain_length, start_of_fwd_enable_chain - 1, -1):
|
||||
position = end_of_first_phase + 1 + (chain_length - i)
|
||||
assert type(ops[position]) is Backward
|
||||
assert ops[position].index == i
|
||||
for i in range(end_of_first_phase + 1 + 1 + chain_length - start_of_fwd_enable_chain, len(ops)):
|
||||
result.insert(ops[i])
|
||||
return (result, start_of_fwd_enable_chain)
|
||||
|
Reference in New Issue
Block a user