github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-03 01:55:12 +00:00
Code Issues Packages Projects Releases Wiki Activity

[autoparallel] adapt solver and CostGraph with new handler (#1695)

Browse Source 
* [autoparallel] adapt solver and CostGraph with new handler

* fix test issue
This commit is contained in:
YuliangLiu0306
2022-10-13 14:04:15 +08:00
committed by GitHub
parent 42b882ef06
commit 81f7530ee7
14 changed files with 834 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tests/test_auto_parallel/test_node_handler/test_norm_pooling_handler.py
View File

@@ -5,8 +5,10 @@ from colossalai.fx import ColoTracer, ColoGraphModule
from colossalai.auto_parallel.solver.op_handler.normal_pooling_handler import NormPoolingHandler
from colossalai.auto_parallel.solver.sharding_strategy import OperationData, OperationDataType, StrategiesVector
from colossalai.device.device_mesh import DeviceMesh
import pytest
@pytest.mark.skip("for higher testing speed")
def test_norm_pool_handler():
model = nn.Sequential(nn.MaxPool2d(4, padding=1).to('meta'))
tracer = ColoTracer()

8
tests/test_auto_parallel/test_node_handler/test_reshape_handler_v2.py
View File

@@ -2,7 +2,7 @@ import torch
import torch.nn as nn
from colossalai.fx import ColoTracer, ColoGraphModule
from colossalai.auto_parallel.solver.op_handler.conv_handler_v2 import ConvFunctionHandler
from colossalai.auto_parallel.solver.op_handler.reshape_handler_v2 import ReshapeHandler
from colossalai.auto_parallel.solver.op_handler.reshape_handler_v2 import ReshapeHandler_V2
from colossalai.auto_parallel.solver.sharding_strategy import OperationData, OperationDataType, StrategiesVector
from colossalai.device.device_mesh import DeviceMesh
@@ -48,9 +48,9 @@ def test_reshape_handler():
strategies_vector=conv_strategies_vector)
conv_handler.register_strategy(compute_resharding_cost=False)
setattr(conv_mod_node, 'strategies_vector', conv_strategies_vector)
reshape_handler = ReshapeHandler(node=reshape_node,
device_mesh=device_mesh,
strategies_vector=reshape_strategies_vector)
reshape_handler = ReshapeHandler_V2(node=reshape_node,
device_mesh=device_mesh,
strategies_vector=reshape_strategies_vector)
reshape_handler.register_strategy(compute_resharding_cost=False)

8
tests/test_auto_parallel/test_node_handler/test_unary_element_wise_handler_v2.py
View File

@@ -2,7 +2,7 @@ from colossalai.fx.tracer.meta_patch.patched_module import linear
import torch
import torch.nn as nn
from colossalai.fx import ColoTracer, ColoGraphModule
from colossalai.auto_parallel.solver.op_handler.unary_elementwise_handler_v2 import UnaryElementwiseHandler
from colossalai.auto_parallel.solver.op_handler.unary_elementwise_handler_v2 import UnaryElementwiseHandler_V2
from colossalai.auto_parallel.solver.op_handler.conv_handler_v2 import ConvFunctionHandler
from colossalai.auto_parallel.solver.sharding_strategy import OperationData, OperationDataType, StrategiesVector
from colossalai.device.device_mesh import DeviceMesh
@@ -50,9 +50,9 @@ def test_elementwise_handler():
strategies_vector=conv_strategies_vector)
conv_handler.register_strategy(compute_resharding_cost=False)
setattr(conv_mod_node, 'strategies_vector', conv_strategies_vector)
relu_handler = UnaryElementwiseHandler(node=relu_mod_node,
device_mesh=device_mesh,
strategies_vector=relu_strategies_vector)
relu_handler = UnaryElementwiseHandler_V2(node=relu_mod_node,
device_mesh=device_mesh,
strategies_vector=relu_strategies_vector)
relu_handler.register_strategy(compute_resharding_cost=False)

99
tests/test_auto_parallel/test_solver_with_resnet_v2.py Normal file
View File

@@ -0,0 +1,99 @@
import torch
from torch.fx import GraphModule
import torch.nn as nn
import pytest
from colossalai.fx.tracer.tracer import ColoTracer
from colossalai.auto_parallel.solver.sharding_strategy import ShardingStrategy, StrategiesVector
from colossalai.tensor.shape_consistency import ShapeConsistencyManager
from colossalai.device.device_mesh import DeviceMesh
from colossalai.auto_parallel.solver.strategies_constructor import StrategiesConstructor_V2
from colossalai.auto_parallel.solver.cost_graph import CostGraph_V2
from copy import deepcopy
from colossalai.auto_parallel.solver.solver import Solver_V2
from torchvision.models import resnet34, resnet50
from colossalai.auto_parallel.solver.constants import *
from colossalai.auto_parallel.solver.graph_analysis import GraphAnalyser
from colossalai.auto_parallel.solver.options import SolverOptions
@pytest.mark.skip("for higher testing speed")
def test_cost_graph():
physical_mesh_id = torch.arange(0, 8)
mesh_shape = (2, 4)
# [[0, 1]
# [2, 3]]
device_mesh = DeviceMesh(physical_mesh_id, mesh_shape)
shape_consistency_manager = ShapeConsistencyManager()
tracer = ColoTracer()
model = resnet50(num_classes=100000)
input_sample = {'x': torch.rand(128, 3, 224, 224).to('meta')}
graph = tracer.trace(root=model, meta_args=input_sample)
# graph():
# %x : torch.Tensor [#users=1] = placeholder[target=x]
# %conv1 : [#users=1] = call_module[target=conv1](args = (%x,), kwargs = {})
# %bn1 : [#users=1] = call_module[target=bn1](args = (%conv1,), kwargs = {})
# %relu : [#users=1] = call_module[target=relu](args = (%bn1,), kwargs = {})
# %maxpool : [#users=2] = call_module[target=maxpool](args = (%relu,), kwargs = {})
# %layer1_0_conv1 : [#users=1] = call_module[target=layer1.0.conv1](args = (%maxpool,), kwargs = {})
# %layer1_0_bn1 : [#users=1] = call_module[target=layer1.0.bn1](args = (%layer1_0_conv1,), kwargs = {})
# %layer1_0_relu : [#users=1] = call_module[target=layer1.0.relu](args = (%layer1_0_bn1,), kwargs = {})
# %layer1_0_conv2 : [#users=1] = call_module[target=layer1.0.conv2](args = (%layer1_0_relu,), kwargs = {})
# %layer1_0_bn2 : [#users=1] = call_module[target=layer1.0.bn2](args = (%layer1_0_conv2,), kwargs = {})
# %add : [#users=1] = call_function[target=operator.add](args = (%layer1_0_bn2, %maxpool), kwargs = {})
# %layer1_0_relu_1 : [#users=2] = call_module[target=layer1.0.relu](args = (%add,), kwargs = {})
# %layer1_1_conv1 : [#users=1] = call_module[target=layer1.1.conv1](args = (%layer1_0_relu_1,), kwargs = {})
# %layer1_1_bn1 : [#users=1] = call_module[target=layer1.1.bn1](args = (%layer1_1_conv1,), kwargs = {})
# %layer1_1_relu : [#users=1] = call_module[target=layer1.1.relu](args = (%layer1_1_bn1,), kwargs = {})
# %layer1_1_conv2 : [#users=1] = call_module[target=layer1.1.conv2](args = (%layer1_1_relu,), kwargs = {})
# %layer1_1_bn2 : [#users=1] = call_module[target=layer1.1.bn2](args = (%layer1_1_conv2,), kwargs = {})
# %add_1 : [#users=1] = call_function[target=operator.add](args = (%layer1_1_bn2, %layer1_0_relu_1), kwargs = {})
# ...
# %avgpool : [#users=1] = call_module[target=avgpool](args = (%layer4_2_relu_1,), kwargs = {})
# %flatten : [#users=1] = call_function[target=torch.flatten](args = (%avgpool, 1), kwargs = {})
# %fc : [#users=1] = call_module[target=fc](args = (%flatten,), kwargs = {})
# return fc
gm = GraphModule(model, graph, model.__class__.__name__)
gm.recompile()
graph_analyser = GraphAnalyser(gm)
liveness_list = graph_analyser.liveness_analysis()
solver_options = SolverOptions(fast=True)
strategies_constructor = StrategiesConstructor_V2(graph, device_mesh, solver_options)
strategies_constructor.build_strategies_and_cost()
cost_graph = CostGraph_V2(strategies_constructor.leaf_strategies)
cost_graph.simplify_graph()
solver = Solver_V2(gm.graph, strategies_constructor, cost_graph, graph_analyser)
ret = solver.call_solver_serialized_args()
print(ret[0])
print(solver.last_s_val)
strategies_list = solver.last_s_val
computation_cost = 0
communication_cost = 0
communication_cost_bn = 0
memory_cost = 0
for index, node in enumerate(graph.nodes):
if node.op == 'call_module':
submod = node.graph.owning_module.get_submodule(node.target)
if type(submod) in BATCHNORM_MODULE_OP:
communication_cost_bn += node.strategies_vector[strategies_list[index]].communication_cost.total
print(node.name, node.strategies_vector[strategies_list[index]].name)
computation_cost += node.strategies_vector[strategies_list[index]].compute_cost.total
communication_cost += node.strategies_vector[strategies_list[index]].communication_cost.total
node_memory_cost = node.strategies_vector[strategies_list[index]].memory_cost.total
if isinstance(node_memory_cost, tuple):
node_memory_cost = node_memory_cost[0]
memory_cost += node_memory_cost.activation + node_memory_cost.parameter
print(f'computation cost is {computation_cost}')
print(f'communication cost is {communication_cost}')
print(f'memory cost is {memory_cost}')
print(f'bn communication cost is {communication_cost_bn}')
if __name__ == '__main__':
test_cost_graph()