2023-03-07 02:34:31 +00:00
|
|
|
import numpy as np
|
2022-06-15 08:36:46 +00:00
|
|
|
import torch
|
2023-03-07 02:34:31 +00:00
|
|
|
import tqdm
|
2022-06-15 08:36:46 +00:00
|
|
|
from torch.fx import symbolic_trace
|
|
|
|
from torch.fx.node import Node
|
2022-11-01 14:53:51 +00:00
|
|
|
|
2022-07-12 05:45:01 +00:00
|
|
|
from colossalai.fx.passes.split_module import split_module
|
2022-06-15 08:36:46 +00:00
|
|
|
|
|
|
|
|
|
|
|
def pipe_split():
|
|
|
|
pass
|
|
|
|
|
|
|
|
|
2023-03-07 02:34:31 +00:00
|
|
|
def block_split():
|
|
|
|
pass
|
|
|
|
|
|
|
|
|
|
|
|
# Construct blocks with the condition that (block_flops / total_flops) >= limit.
|
|
|
|
def construct_blocks(gm: torch.fx.GraphModule, limit=0.01):
|
|
|
|
total_fwd_flop = 0
|
|
|
|
total_bwd_flop = 0
|
|
|
|
for node in gm.graph.nodes:
|
|
|
|
total_fwd_flop += node.fwd_flop
|
|
|
|
total_bwd_flop += node.bwd_flop
|
|
|
|
|
|
|
|
total_flop = total_fwd_flop + total_bwd_flop
|
|
|
|
per_block_flop = total_flop * limit
|
|
|
|
accumulate_fwd_flop = 0
|
|
|
|
accumulate_bwd_flop = 0
|
|
|
|
block_nodes = []
|
|
|
|
for node in gm.graph.nodes:
|
|
|
|
if 'block_split' in node.name:
|
|
|
|
continue
|
|
|
|
accumulate_fwd_flop += node.fwd_flop
|
|
|
|
accumulate_bwd_flop += node.bwd_flop
|
|
|
|
if accumulate_fwd_flop + accumulate_bwd_flop >= per_block_flop:
|
|
|
|
with gm.graph.inserting_after(node):
|
|
|
|
block_node = gm.graph.create_node('call_function', block_split)
|
|
|
|
setattr(block_node, 'fwd_flop', accumulate_fwd_flop)
|
|
|
|
setattr(block_node, 'bwd_flop', accumulate_bwd_flop)
|
|
|
|
accumulate_fwd_flop = 0
|
|
|
|
accumulate_bwd_flop = 0
|
|
|
|
block_nodes.append(block_node)
|
|
|
|
|
|
|
|
return block_nodes
|
|
|
|
|
|
|
|
|
|
|
|
def remove_blocks(gm: torch.fx.GraphModule):
|
|
|
|
for node in gm.graph.nodes:
|
|
|
|
if (node.op, node.target) == ('call_function', block_split):
|
|
|
|
gm.graph.erase_node(node)
|
|
|
|
|
|
|
|
|
|
|
|
def get_compute_costs(node_list):
|
|
|
|
num_nodes = len(node_list)
|
|
|
|
all_compute_cost = np.full((num_nodes, num_nodes), np.inf, dtype=np.float64)
|
|
|
|
|
|
|
|
for start in tqdm.tqdm(range(num_nodes), desc='start pos', position=0):
|
|
|
|
for end in tqdm.tqdm(range(start, num_nodes), desc='end pos', position=1, leave=False):
|
|
|
|
selected_flops = [(node_list[i].fwd_flop + node_list[i].bwd_flop) for i in range(start, end + 1)]
|
|
|
|
all_compute_cost[start, end] = sum(selected_flops)
|
|
|
|
|
|
|
|
return all_compute_cost
|
|
|
|
|
|
|
|
|
|
|
|
def do_dp_split_gpipe_impl(num_nodes, num_stages, num_microbatches, compute_costs, max_compute_cost):
|
|
|
|
"""The core implementation of the DP algorithm."""
|
|
|
|
# Adapted from Alpa DP Formulation.
|
|
|
|
# For f, node ID start from 0
|
|
|
|
# f[number of stages,
|
|
|
|
# node id that is currently being considered]
|
|
|
|
|
|
|
|
# record time cost(assess by fwd+bwd flop now)
|
|
|
|
f = np.full((num_stages + 1, num_nodes + 1), np.inf, dtype=np.float32)
|
|
|
|
|
|
|
|
# record max stage compute cost among all stages in this partition.
|
|
|
|
f_stage_max = np.full((num_stages + 1, num_nodes + 1), 0.0, dtype=np.float32)
|
|
|
|
# record start node index for next stage in this partition
|
|
|
|
f_argmin = np.full((num_stages + 1, num_nodes + 1), -1, dtype=np.int32)
|
|
|
|
f[0, num_nodes] = 0
|
|
|
|
for s in tqdm.tqdm(range(1, num_stages + 1), desc='stage', position=2, leave=False): # pylint: disable=too-many-nested-blocks
|
|
|
|
for i in tqdm.tqdm(range(num_nodes - 1, -1, -1), desc='start node', position=3, leave=False):
|
|
|
|
for k in tqdm.tqdm(range(num_nodes, i, -1), desc='mid node', position=4, leave=False):
|
|
|
|
stage_cost = compute_costs[i, k - 1]
|
|
|
|
new_cost = f[s - 1, k] + stage_cost
|
|
|
|
if (stage_cost <= max_compute_cost and new_cost < f[s, i]):
|
|
|
|
f[s, i] = new_cost
|
|
|
|
f_stage_max[s, i] = max(f_stage_max[s - 1, k], stage_cost)
|
|
|
|
f_argmin[s, i] = k
|
|
|
|
|
|
|
|
best_total_cost = f[num_stages, 0]
|
|
|
|
if np.isinf(best_total_cost):
|
|
|
|
return np.inf, None
|
|
|
|
|
|
|
|
total_cost = f[num_stages, 0] + (num_microbatches - 1) * f_stage_max[num_stages, 0]
|
|
|
|
|
|
|
|
current_s = num_stages
|
|
|
|
current_node = 0
|
|
|
|
|
|
|
|
res = []
|
|
|
|
while current_s > 0 and current_node < num_nodes:
|
|
|
|
next_start_node = f_argmin[current_s, current_node]
|
|
|
|
res.append((current_node, next_start_node))
|
|
|
|
current_s -= 1
|
|
|
|
current_node = next_start_node
|
|
|
|
|
|
|
|
return total_cost, res
|
|
|
|
|
|
|
|
|
|
|
|
def do_dp_split_gpipe(node_list, compute_costs, num_stages: int, num_microbatches: int):
|
|
|
|
# Ignore the memory cost profiling in Alpa's design for convenience.
|
|
|
|
max_compute_costs = np.sort(np.unique(compute_costs))
|
|
|
|
best_cost = np.inf
|
|
|
|
best_solution = None
|
|
|
|
last_max_compute_cost = 0.0
|
|
|
|
gap = 1e6 # temporary magic number, unit: flops
|
|
|
|
|
|
|
|
for max_compute_cost in tqdm.tqdm(max_compute_costs):
|
|
|
|
# Pruning to reduce search space.
|
|
|
|
if max_compute_cost * num_microbatches >= best_cost:
|
|
|
|
break
|
|
|
|
if max_compute_cost - last_max_compute_cost < gap:
|
|
|
|
continue
|
|
|
|
|
|
|
|
cost, solution = do_dp_split_gpipe_impl(len(node_list), num_stages, num_microbatches, compute_costs,
|
|
|
|
max_compute_cost)
|
|
|
|
|
|
|
|
if cost < best_cost:
|
|
|
|
best_cost = cost
|
|
|
|
best_solution = solution
|
|
|
|
last_max_compute_cost = max_compute_cost
|
|
|
|
return best_cost, best_solution
|
|
|
|
|
|
|
|
|
|
|
|
# Auto DP partition based on Alpa.
|
|
|
|
# Adapted to Gpipe Scheduler
|
|
|
|
# split_mode:
|
|
|
|
# 'node': fx_node
|
|
|
|
# 'block': many fx_nodes construct a block
|
|
|
|
def gpipe_dp_split_pass(gm: torch.fx.GraphModule, pp_size: int, num_microbatches: int, mode='block', block_limit=0.01):
|
|
|
|
assert mode in ['node', 'block']
|
|
|
|
|
|
|
|
# nodes or blocks will be used in partition.
|
|
|
|
node_list = []
|
|
|
|
if mode == 'node':
|
|
|
|
for node in gm.graph.nodes:
|
|
|
|
node_list.append(node)
|
|
|
|
elif mode == 'block':
|
|
|
|
node_list = construct_blocks(gm, limit=block_limit)
|
|
|
|
else:
|
|
|
|
pass
|
|
|
|
|
|
|
|
compute_costs = get_compute_costs(node_list)
|
|
|
|
|
|
|
|
best_cost, best_solution = do_dp_split_gpipe(node_list, compute_costs, pp_size, num_microbatches)
|
|
|
|
|
|
|
|
for (_, next_start_node) in best_solution:
|
|
|
|
if pp_size <= 1:
|
|
|
|
break
|
|
|
|
node = node_list[next_start_node]
|
|
|
|
with gm.graph.inserting_before(node):
|
|
|
|
split_node = gm.graph.create_node('call_function', pipe_split)
|
|
|
|
pp_size -= 1
|
|
|
|
|
|
|
|
# remove block node if possible
|
|
|
|
if mode == 'block':
|
|
|
|
remove_blocks(gm)
|
|
|
|
|
|
|
|
gm.recompile()
|
|
|
|
return gm
|
|
|
|
|
|
|
|
|
2023-01-19 05:54:50 +00:00
|
|
|
def avgcompute_split_pass(gm: torch.fx.GraphModule, pp_size: int):
|
|
|
|
"""
|
|
|
|
In avgcompute_split_pass, we split module by the fwd flops.
|
|
|
|
"""
|
|
|
|
mod_graph = gm.graph
|
|
|
|
# To use avgcompute_split_pass, we need run meta_info_prop interpreter first.
|
|
|
|
# If nodes don't have meta info, this pass will fall back to normal balanced split pass.
|
|
|
|
check_node = list(mod_graph.nodes)[0]
|
|
|
|
if 'tensor_meta' not in check_node.meta:
|
|
|
|
return balanced_split_pass(gm, pp_size)
|
|
|
|
|
|
|
|
total_fwd_flop = 0
|
|
|
|
for node in mod_graph.nodes:
|
|
|
|
total_fwd_flop += node.fwd_flop
|
|
|
|
|
|
|
|
partition_flop = total_fwd_flop // pp_size
|
|
|
|
accumulate_fwd_flop = 0
|
|
|
|
for node in mod_graph.nodes:
|
|
|
|
if pp_size <= 1:
|
|
|
|
break
|
|
|
|
if 'pipe_split' in node.name:
|
|
|
|
continue
|
|
|
|
accumulate_fwd_flop += node.fwd_flop
|
|
|
|
if accumulate_fwd_flop >= partition_flop:
|
|
|
|
total_fwd_flop = total_fwd_flop - accumulate_fwd_flop
|
|
|
|
accumulate_fwd_flop = 0
|
|
|
|
pp_size -= 1
|
|
|
|
partition_flop = total_fwd_flop // pp_size
|
|
|
|
with mod_graph.inserting_after(node):
|
|
|
|
split_node = mod_graph.create_node('call_function', pipe_split)
|
|
|
|
gm.recompile()
|
|
|
|
return gm
|
|
|
|
|
|
|
|
|
2022-12-23 03:38:43 +00:00
|
|
|
def avgnode_split_pass(gm: torch.fx.GraphModule, pp_size: int):
|
|
|
|
"""
|
2023-05-24 01:01:50 +00:00
|
|
|
In avgnode_split_pass, simply split graph by node number.
|
2022-12-23 03:38:43 +00:00
|
|
|
"""
|
|
|
|
mod_graph = gm.graph
|
|
|
|
avg_num_node = len(mod_graph.nodes) // pp_size
|
|
|
|
accumulate_num_node = 0
|
|
|
|
for node in mod_graph.nodes:
|
|
|
|
if pp_size <= 1:
|
|
|
|
break
|
|
|
|
accumulate_num_node += 1
|
|
|
|
if accumulate_num_node >= avg_num_node:
|
|
|
|
accumulate_num_node = 0
|
|
|
|
pp_size -= 1
|
|
|
|
if node.next.op == 'output':
|
|
|
|
with mod_graph.inserting_before(node):
|
|
|
|
split_node = mod_graph.create_node('call_function', pipe_split)
|
|
|
|
else:
|
|
|
|
with mod_graph.inserting_after(node):
|
|
|
|
split_node = mod_graph.create_node('call_function', pipe_split)
|
|
|
|
gm.recompile()
|
|
|
|
return gm
|
|
|
|
|
|
|
|
|
2022-06-15 08:36:46 +00:00
|
|
|
def balanced_split_pass(gm: torch.fx.GraphModule, pp_size: int):
|
2022-07-15 06:54:26 +00:00
|
|
|
"""
|
|
|
|
In balanced_split_pass, we split module by the size of parameters(weights+bias).
|
|
|
|
"""
|
2022-06-15 08:36:46 +00:00
|
|
|
mod_graph = gm.graph
|
|
|
|
total_param_amount = 0
|
|
|
|
for param in mod_graph.owning_module.parameters():
|
|
|
|
total_param_amount += param.numel()
|
|
|
|
params_per_partition = total_param_amount // pp_size
|
|
|
|
accumulate_param_amount = 0
|
|
|
|
for node in mod_graph.nodes:
|
|
|
|
if pp_size <= 1:
|
|
|
|
break
|
|
|
|
if node.op == "call_module":
|
|
|
|
target_module = node.graph.owning_module.get_submodule(node.target)
|
|
|
|
for param in target_module.parameters():
|
|
|
|
accumulate_param_amount += param.numel()
|
|
|
|
if accumulate_param_amount >= params_per_partition:
|
|
|
|
accumulate_param_amount = 0
|
|
|
|
pp_size -= 1
|
2022-07-12 05:45:01 +00:00
|
|
|
# If the next node is output node, we will insert split annotation before
|
|
|
|
# node to make sure there is at least one node in last partition.
|
|
|
|
if node.next.op == 'output':
|
|
|
|
with mod_graph.inserting_before(node):
|
|
|
|
split_node = mod_graph.create_node('call_function', pipe_split)
|
|
|
|
else:
|
|
|
|
with mod_graph.inserting_after(node):
|
|
|
|
split_node = mod_graph.create_node('call_function', pipe_split)
|
2022-11-01 14:53:51 +00:00
|
|
|
if pp_size > 1:
|
|
|
|
node_counter = 0
|
|
|
|
for node in mod_graph.nodes:
|
|
|
|
if pp_size <= 1:
|
|
|
|
break
|
|
|
|
if node.op == 'placeholder':
|
|
|
|
continue
|
|
|
|
elif node_counter == 0:
|
|
|
|
node_counter += 1
|
|
|
|
else:
|
|
|
|
pp_size -= 1
|
|
|
|
node_counter = 0
|
|
|
|
with mod_graph.inserting_before(node):
|
|
|
|
split_node = mod_graph.create_node('call_function', pipe_split)
|
|
|
|
|
2022-06-15 08:36:46 +00:00
|
|
|
gm.recompile()
|
|
|
|
return gm
|
|
|
|
|
|
|
|
|
2022-07-15 06:54:26 +00:00
|
|
|
def balanced_split_pass_v2(gm: torch.fx.GraphModule, pp_size: int):
|
|
|
|
"""
|
|
|
|
In balanced_split_pass_v12, we split module by the size of nodes(weights+bias+outputs).
|
|
|
|
"""
|
|
|
|
mod_graph = gm.graph
|
|
|
|
# To use balanced_split_pass_v2, we need run meta_info_prop interpreter first.
|
|
|
|
# If nodes don't have meta info, this pass will fall back to normal balanced split pass.
|
|
|
|
check_node = list(mod_graph.nodes)[0]
|
|
|
|
if 'tensor_meta' not in check_node.meta:
|
|
|
|
return balanced_split_pass(gm, pp_size)
|
|
|
|
|
|
|
|
total_element_size = 0
|
|
|
|
for node in mod_graph.nodes:
|
|
|
|
total_element_size += node.node_size
|
|
|
|
|
|
|
|
partition_size = total_element_size // pp_size
|
|
|
|
accumulate_node_size = 0
|
|
|
|
for node in mod_graph.nodes:
|
|
|
|
if pp_size <= 1:
|
|
|
|
break
|
2022-07-27 05:40:54 +00:00
|
|
|
if 'pipe_split' in node.name:
|
|
|
|
continue
|
2022-07-15 06:54:26 +00:00
|
|
|
accumulate_node_size += node.node_size
|
|
|
|
if accumulate_node_size >= partition_size:
|
2023-01-19 05:54:50 +00:00
|
|
|
total_element_size = total_element_size - accumulate_node_size
|
2022-07-15 06:54:26 +00:00
|
|
|
accumulate_node_size = 0
|
|
|
|
pp_size -= 1
|
2023-01-19 05:54:50 +00:00
|
|
|
partition_size = total_element_size // pp_size
|
2022-07-15 06:54:26 +00:00
|
|
|
with mod_graph.inserting_after(node):
|
|
|
|
split_node = mod_graph.create_node('call_function', pipe_split)
|
|
|
|
gm.recompile()
|
|
|
|
return gm
|
|
|
|
|
|
|
|
|
2022-07-06 05:48:11 +00:00
|
|
|
def uniform_split_pass(gm: torch.fx.GraphModule, pp_size: int):
|
|
|
|
mod_graph = gm.graph
|
|
|
|
valid_children_size = 0
|
|
|
|
valid_children = []
|
|
|
|
for module in mod_graph.owning_module.children():
|
|
|
|
valid_children_size += 1
|
|
|
|
valid_children.append(module)
|
|
|
|
|
|
|
|
if valid_children_size < pp_size:
|
|
|
|
# If valid children is not enough to shard, we will use balanced policy instead of uniform policy.
|
|
|
|
return balanced_split_pass(gm, pp_size)
|
|
|
|
layers_per_partition = valid_children_size // pp_size
|
|
|
|
accumulate_layer_amount = 0
|
|
|
|
for node in mod_graph.nodes:
|
|
|
|
if pp_size <= 1:
|
|
|
|
break
|
|
|
|
if node.op == "call_module":
|
|
|
|
target_module = node.graph.owning_module.get_submodule(node.target)
|
|
|
|
if target_module in valid_children:
|
|
|
|
accumulate_layer_amount += 1
|
|
|
|
if accumulate_layer_amount == layers_per_partition:
|
|
|
|
accumulate_layer_amount = 0
|
|
|
|
pp_size -= 1
|
|
|
|
with mod_graph.inserting_after(node):
|
|
|
|
split_node = mod_graph.create_node('call_function', pipe_split)
|
|
|
|
gm.recompile()
|
|
|
|
return gm
|
|
|
|
|
|
|
|
|
2022-11-29 03:36:28 +00:00
|
|
|
def split_with_split_nodes_pass(annotated_gm: torch.fx.GraphModule, merge_output=False):
|
2022-07-12 06:10:37 +00:00
|
|
|
# TODO(lyl): use partition IR to assign partition ID to each node.
|
|
|
|
# Currently: analyzing graph -> annotate graph by inserting split node -> use split module pass to split graph
|
|
|
|
# In future: graph to partitions -> analyzing partition IR -> recombining partitions to get best performance -> assign partition ID to each node
|
2022-06-15 08:36:46 +00:00
|
|
|
part_idx = 0
|
|
|
|
|
|
|
|
def split_callback(n: torch.fx.Node):
|
|
|
|
nonlocal part_idx
|
|
|
|
if (n.op, n.target) == ('call_function', pipe_split):
|
|
|
|
part_idx += 1
|
|
|
|
return part_idx
|
|
|
|
|
2022-11-29 03:36:28 +00:00
|
|
|
split_mod = split_module(annotated_gm, None, split_callback, merge_output)
|
2022-06-15 08:36:46 +00:00
|
|
|
split_submodules = []
|
|
|
|
for name, submodule in split_mod.named_modules():
|
|
|
|
if isinstance(submodule, torch.fx.GraphModule):
|
|
|
|
for node in submodule.graph.nodes:
|
|
|
|
if (node.op, node.target) == ('call_function', pipe_split):
|
|
|
|
submodule.graph.erase_node(node)
|
|
|
|
submodule.recompile()
|
|
|
|
split_submodules.append(submodule)
|
|
|
|
|
|
|
|
return split_mod, split_submodules
|