ColossalAI/colossalai/fx/passes/algorithms/ckpt_solver_rotor.py

411 lines
14 KiB
Python

from typing import List, Tuple
import torch
from torch.fx import GraphModule, Node
from colossalai.fx.graph_module import ColoGraphModule
from colossalai.fx.profiler import parameter_size
import math
from .linearize import linearize
from .utils import *
from colossalai.fx.passes.meta_info_prop import MetaInfoProp
from colossalai.fx.codegen.activation_checkpoint_codegen import _find_nested_ckpt_regions
# this is the python compute table code from rotor
# https://gitlab.inria.fr/hiepacs/rotor
# paper link: https://hal.inria.fr/hal-02352969
def _compute_table(chain: Chain, mmax) -> Tuple:
"""Returns the optimal table: a tuple containing:
Opt[m][lmin][lmax] with lmin = 0...chain.length
and lmax = lmin...chain.length (lmax is not included) and m = 0...mmax
what[m][lmin][lmax] is (True,) if the optimal choice is a chain checkpoint
(False, j) if the optimal choice is a leaf checkpoint of length j
The computation uses dynamic programming"""
fw = chain.fweight + [0] ## forward time
bw = chain.bweight ## backward time, not used
cw = chain.cweight + [0] ## size of x (and of y)
cbw = chain.cbweight + [0] ## size of xbar
fwd_mem_tmp = chain.fwd_mem_tmp + [0]
bwd_mem_tmp = chain.bwd_mem_tmp + [0]
# Build table
opt = [[{} for _ in range(chain.length + 1)] for _ in range(mmax + 1)]
what = [[{} for _ in range(chain.length + 1)] for _ in range(mmax + 1)]
## Last one is a dict because its indices go from i to l. Renumbering will wait for C implementation
# Initialize borders of the tables for lmax-lmin = 0
for m in range(mmax + 1):
for i in range(chain.length + 1):
#lmax-lmin = 0
limit = max(cw[i + 1] + cbw[i + 1] + fwd_mem_tmp[i], cw[i + 1] + cbw[i + 1] + bwd_mem_tmp[i])
if m >= limit: ## Equation (1)
opt[m][i][i] = fw[i] + bw[i]
else:
opt[m][i][i] = float("inf")
# Compute everything
for m in range(mmax + 1):
for d in range(1, chain.length + 1):
for i in range(chain.length + 1 - d):
# for idx in range(i+1, chain.length + 1):
idx = i + d
mmin = cw[idx + 1] + cw[i + 1] + fwd_mem_tmp[i]
if idx > i + 1:
mmin = max(mmin, cw[idx + 1] + max(cw[j] + cw[j + 1] + fwd_mem_tmp[j] for j in range(i + 1, idx)))
if m < mmin:
opt[m][i][idx] = float("inf")
else:
leaf_checkpoints = [(j, sum(fw[i:j]) + opt[m - cw[j]][j][idx] + opt[m][i][j - 1])
for j in range(i + 1, idx + 1)
if m >= cw[j]]
if leaf_checkpoints:
best_leaf = min(leaf_checkpoints, key=lambda t: t[1])
else:
best_leaf = None
if m >= cbw[i + 1]:
chain_checkpoint = opt[m][i][i] + opt[m - cbw[i + 1]][i + 1][idx]
else:
chain_checkpoint = float("inf")
if best_leaf and best_leaf[1] <= chain_checkpoint:
opt[m][i][idx] = best_leaf[1]
what[m][i][idx] = (False, best_leaf[0])
else:
opt[m][i][idx] = chain_checkpoint
what[m][i][idx] = (True,)
return (opt, what)
def _rec(chain: Chain, lmin, lmax, cmem, opt_table):
""" chain : the class describing the AC graph
lmin : index of the first forward to execute
lmax : upper bound index of the last forward to execute (not included)
cmem : number of available memory slots
Return the optimal sequence of makespan Opt_hete[cmem][lmin][lmax-lmin]"""
if cmem <= 0:
raise ValueError("Can not process a chain with negative memory {cmem}".format(cmem=cmem))
opt, what = opt_table
sequence = Sequence(Function("Persistent", lmax - lmin, cmem))
if opt[cmem][lmin][lmax] == float("inf"):
raise ValueError("Can not process this chain from index {lmin} to {lmax} with memory {cmem}".format(lmin=lmin,
lmax=lmax,
cmem=cmem))
if lmin == lmax:
if lmin == chain.length:
sequence.insert(Loss())
else:
sequence.insert(ForwardEnable(lmin))
sequence.insert(Backward(lmin))
return sequence
if what[cmem][lmin][lmax][0]:
sequence.insert(ForwardEnable(lmin))
sequence.insert_sequence(_rec(chain, lmin + 1, lmax, cmem - chain.cbweight[lmin + 1], opt_table))
sequence.insert(Backward(lmin))
else:
j = what[cmem][lmin][lmax][1]
sequence.insert(ForwardCheck(lmin))
for k in range(lmin + 1, j):
sequence.insert(ForwardNograd(k))
sequence.insert_sequence(_rec(chain, j, lmax, cmem - chain.cweight[j], opt_table))
sequence.insert_sequence(_rec(chain, lmin, j - 1, cmem, opt_table))
return sequence
def _discretize(mem_unit, values):
return [math.ceil(value / mem_unit) for value in values]
def _compute_size(obj: torch.Tensor) -> int:
return obj.numel() * obj.element_size()
def _compute_output_size(node: List[Node]) -> int:
"""Compute the output size of a node
Args:
node (List[Node]): node, list of torch.fx.Node
Returns:
int: output size
"""
return node[-1].meta['tensor_meta'].numel * torch.tensor([],
dtype=node[-1].meta['tensor_meta'].dtype).element_size()
def _get_inplace(node: Node) -> bool:
"""Get the inplace argument from torch.fx.Node
Args:
node (Node): torch.fx.Node
Returns:
bool: indicates whether this op is inplace
"""
is_inplace = False
if node.op == "call_function":
is_inplace = node.kwargs.get("inplace", False)
elif node.op == "call_module":
is_inplace = getattr(node.graph.owning_module.get_submodule(node.target), "inplace", False)
return is_inplace
def _fwd_xbar(node: List[Node]) -> int:
"""Get the forward xbar of a node
Args:
node (List[Node]): List of torch.fx Node,
indicates a node in linearized graph
Returns:
int: xbar size, unit Byte
"""
xbar = 0
for n in node:
xbar += n.meta['fwd_mem_tmp'] + n.meta['fwd_mem_out']
return xbar
def _fwd_time(node: List[Node]) -> int:
"""Get the foward time of a node
Args:
node (List[Node]): List of torch.fx Node,
indicates a node in linearized graph
Returns:
int: foward time, extimated by flops count
"""
fwd_time = 0
for n in node:
# minimum flop count is needed
fwd_time += max(n.meta['fwd_flop'], 1)
return fwd_time
def _bwd_time(node: List[Node]) -> int:
"""Get the backward time of a node
Args:
node (List[Node]): List of torch.fx Node,
indicates a node in linearized graph
Returns:
int: backward time, extimated by flops count
"""
bwd_time = 0
for n in node:
# minimum flop count is needed
bwd_time += max(n.meta['bwd_flop'], 1)
return bwd_time
def _get_bwd_mem_tmp(node: List[Node]) -> int:
"""Get the backward temp memory of a node
Args:
node (List[Node]): List of torch.fx Node,
indicates a node in linearized graph
Returns:
int: backward temp memory, unit Byte
"""
def _get_deps_size():
deps_size = 0
for k, v in deps.items():
if v > 0:
deps_size += k.meta['bwd_mem_out']
return deps_size
bwd_mem_tmp = 0
deps = {}
# add all the users for last node into deps,
# as those nodes' gradient out will be stored in memory
for child in node[-1].users:
deps[child] = 1
for n in reversed(node):
bwd_mem_tmp = max(bwd_mem_tmp, _get_deps_size() + n.meta['bwd_mem_tmp'])
deps[n] = len(n.all_input_nodes)
for child in n.users:
if child in deps:
deps[child] -= 1
for key in list(deps.keys()):
if deps[key] == 0:
del deps[key]
return bwd_mem_tmp
def _construct_chain(node_list: List[List[Node]], data, mem_unit: int) -> Chain:
fwd_time = []
bwd_time = []
if isinstance(data, torch.Tensor):
xbar_sizes = [_compute_size(data)]
x_sizes = [_compute_size(data)]
elif isinstance(data, list) or isinstance(data, tuple):
xbar_sizes = [sum([_compute_size(obj) for obj in data])]
x_sizes = [sum([_compute_size(obj) for obj in data])]
elif isinstance(data, dict):
xbar_sizes = [sum([_compute_size(obj) for obj in data.values()])]
x_sizes = [sum([_compute_size(obj) for obj in data.values()])]
# currently we can't get the temp memory needed in fwd
tmp_fwd = [0] * len(node_list)
tmp_bwd = []
for idx, node in enumerate(node_list):
fwd_time.append(_fwd_time(node))
bwd_time.append(_bwd_time(node))
x_sizes.append(_compute_output_size(node))
xbar_sizes.append(max(x_sizes[-1], _fwd_xbar(node)))
tmp_bwd.append(_get_bwd_mem_tmp(node))
# if a node with only one inplace op, we need to let x_bar = 0
if len(node) == 1 and _get_inplace(node[0]):
xbar_sizes[-1] = 0
bwd_time.append(0)
# currently we view loss backward temp as zero
tmp_bwd.append(0)
xbar_sizes = _discretize(mem_unit, xbar_sizes)
x_sizes = _discretize(mem_unit, x_sizes)
tmp_fwd = _discretize(mem_unit, tmp_fwd)
tmp_bwd = _discretize(mem_unit, tmp_bwd)
return Chain(fwd_time, bwd_time, x_sizes, xbar_sizes, tmp_fwd, tmp_bwd)
def _annotate_from_sequence(sequence: Sequence, node_list: List[List[Node]]):
op_list = sequence.list_operations()
loss_op = next(op for op in op_list if isinstance(op, Loss))
fwd_list = op_list[:op_list.index(loss_op)]
bwd_list = op_list[op_list.index(loss_op) + 1:]
ckpt_idx = 0
in_ckpt = False
ckpt_region = []
# forward annotation
for idx, op in enumerate(fwd_list, 0):
if in_ckpt:
if isinstance(op, ForwardNograd):
ckpt_region.append(idx)
elif isinstance(op, ForwardEnable):
in_ckpt = False
for node_idx in ckpt_region:
for n in node_list[node_idx]:
setattr(n, "activation_checkpoint", [ckpt_idx])
ckpt_idx += 1
ckpt_region = []
elif isinstance(op, ForwardCheck):
for node_idx in ckpt_region:
for n in node_list[node_idx]:
setattr(n, "activation_checkpoint", [ckpt_idx])
ckpt_idx += 1
ckpt_region = [idx]
else:
if isinstance(op, ForwardCheck):
in_ckpt = True
ckpt_region.append(idx)
# annotate the backward if there is any nested activation checkpoint
in_recompute = False
for op in bwd_list:
if in_recompute:
if isinstance(op, ForwardNograd):
ckpt_region.append(op.index)
elif isinstance(op, ForwardEnable):
for node_idx in ckpt_region:
for n in node_list[node_idx]:
n.activation_checkpoint.append(ckpt_idx)
ckpt_idx += 1
ckpt_region = []
elif isinstance(op, ForwardCheck):
for node_idx in ckpt_region:
for n in node_list[node_idx]:
n.activation_checkpoint.append(ckpt_idx)
ckpt_idx += 1
ckpt_region = [op.index]
elif isinstance(op, Backward):
for node_idx in ckpt_region:
for n in node_list[node_idx]:
n.activation_checkpoint.append(ckpt_idx)
in_recompute = False
else:
if not isinstance(op, Backward):
in_recompute = True
ckpt_idx = 0
ckpt_region = []
if isinstance(op, ForwardCheck):
ckpt_region.append(op.index)
# postprocess, make sure every activation checkpoint label in the
# same activation checkpoint region (level = 0) has the same length
op_list = []
for node in node_list:
op_list += node
ckpt_regions = _find_nested_ckpt_regions(op_list)
for (start_idx, end_idx) in ckpt_regions:
nested_length = max(len(op_list[idx].activation_checkpoint) for idx in range(start_idx, end_idx + 1))
for idx in range(start_idx, end_idx + 1):
op_list[idx].activation_checkpoint += [None] * (nested_length - len(op_list[idx].activation_checkpoint))
def solver_rotor(gm: ColoGraphModule,
data,
mem_limit: int,
mem_slots: int = 500,
cnode: List[str] = None,
eps: float = 0.02) -> ColoGraphModule:
"""solver that automatically find activation checkpoint in rotor's manner
Args:
gm (ColoGraphModule): ColoGraphModule generated by tracing model.
data (torch.Tensor): input data.
mem_limit (int): memory budget in Byte.
mem_slots (int, optional): number of slots for discretizing memory budget. Defaults to 500.
cnode (List[Node], optional): common node list for linearize. Defaults to None.
eps (float): epsilon for memory decay. Defaults to 0.02
Returns:
ColoGraphModule: annotated ColoGraphModuled with __sequence__ attribute
"""
node_list = linearize(gm, cnode)
mem_limit -= parameter_size(gm)
mem_unit = mem_limit * (1.0 - eps) // mem_slots
MetaInfoProp(gm).run(data)
chain: Chain = _construct_chain(node_list, data, mem_unit)
opt_table = _compute_table(chain, mem_slots)
sequence = _rec(chain, 0, chain.length, mem_slots - chain.cweight[0], opt_table)
_annotate_from_sequence(sequence, node_list)
# set __sequence__ attribute to GraphModule
setattr(gm, "__sequence__", sequence)
return gm