github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-04 18:40:28 +00:00
Code Issues Packages Projects Releases Wiki Activity

[misc] update pre-commit and run all files (#4752)

Browse Source 
* [misc] update pre-commit

* [misc] run pre-commit

* [misc] remove useless configuration files

* [misc] ignore cuda for clang-format
This commit is contained in:
Hongxin Liu
2023-09-19 14:20:26 +08:00
committed by GitHub
parent 3c6b831c26
commit 079bf3cb26
1268 changed files with 50037 additions and 38444 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
tests/test_zero/test_low_level/test_grad_acc.py
View File

@@ -14,7 +14,6 @@ from colossalai.zero import LowLevelZeroOptimizer
class MlpModel(nn.Module):
def __init__(self):
super(MlpModel, self).__init__()
self.linear1 = nn.Linear(128, 256)
@@ -36,16 +35,12 @@ def exam_zero_1_2_grad_acc():
# create optimizer
zero1_optimizer = torch.optim.Adam(zero1_model.parameters(), lr=1)
zero2_optimizer = torch.optim.Adam(zero2_model.parameters(), lr=1)
zero1_optimizer = LowLevelZeroOptimizer(zero1_optimizer,
overlap_communication=True,
initial_scale=32,
clip_grad_norm=1.0,
verbose=True)
zero2_optimizer = LowLevelZeroOptimizer(zero2_optimizer,
overlap_communication=True,
partition_grad=True,
initial_scale=32,
clip_grad_norm=1.0)
zero1_optimizer = LowLevelZeroOptimizer(
zero1_optimizer, overlap_communication=True, initial_scale=32, clip_grad_norm=1.0, verbose=True
)
zero2_optimizer = LowLevelZeroOptimizer(
zero2_optimizer, overlap_communication=True, partition_grad=True, initial_scale=32, clip_grad_norm=1.0
)
# create data
seed_all(2021 + local_rank)
input_data1 = torch.randn(32, 128).cuda()
@@ -91,10 +86,9 @@ def exam_zero_1_grad_acc(sync):
# we only test stage 1 here
# in `check_sharded_param_consistency.py`, we will test whether
# level 1 and 2 will produce exactly the same results
zero_optimizer = LowLevelZeroOptimizer(zero_optimizer,
overlap_communication=False,
reduce_bucket_size=262144,
clip_grad_norm=1.0)
zero_optimizer = LowLevelZeroOptimizer(
zero_optimizer, overlap_communication=False, reduce_bucket_size=262144, clip_grad_norm=1.0
)
torch_optimizer = torch.optim.Adam(torch_model.parameters(), lr=1)
@@ -104,7 +98,6 @@ def exam_zero_1_grad_acc(sync):
input_data2 = torch.randn(32, 128).cuda()
def fwd_bwd_func(no_sync, cur_data, check_flag):
# zero1 fwd and bwd
with conditional_context(zero_optimizer.no_sync(), no_sync):
zero_output = zero_model(cur_data)
@@ -135,7 +128,7 @@ def exam_zero_1_grad_acc(sync):
def run_dist(rank, world_size, port):
colossalai.launch(config=dict(), rank=rank, world_size=world_size, port=port, host='localhost')
colossalai.launch(config=dict(), rank=rank, world_size=world_size, port=port, host="localhost")
exam_zero_1_grad_acc(sync=True)
exam_zero_1_grad_acc(sync=False)
@@ -147,5 +140,5 @@ def test_grad_accumulation():
spawn(run_dist, 2)
if __name__ == '__main__':
if __name__ == "__main__":
test_grad_accumulation()

29
tests/test_zero/test_low_level/test_zero1_2.py
View File

@@ -2,7 +2,6 @@ import copy
import pytest
import torch
import torch.distributed as dist
import torch.nn as nn
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.testing import assert_close
@@ -14,7 +13,6 @@ from colossalai.zero import LowLevelZeroOptimizer
class MlpModel(nn.Module):
def __init__(self):
super(MlpModel, self).__init__()
self.linear1 = nn.Linear(123, 253)
@@ -74,14 +72,12 @@ def exam_zero_1_2():
# create optimizer
zero1_optimizer = torch.optim.Adam(zero1_model.parameters(), lr=1)
zero2_optimizer = torch.optim.Adam(zero2_model.parameters(), lr=1)
zero1_optimizer = LowLevelZeroOptimizer(zero1_optimizer,
overlap_communication=True,
initial_scale=128,
verbose=True)
zero2_optimizer = LowLevelZeroOptimizer(zero2_optimizer,
overlap_communication=True,
partition_grad=True,
initial_scale=128)
zero1_optimizer = LowLevelZeroOptimizer(
zero1_optimizer, overlap_communication=True, initial_scale=128, verbose=True
)
zero2_optimizer = LowLevelZeroOptimizer(
zero2_optimizer, overlap_communication=True, partition_grad=True, initial_scale=128
)
# create data
seed_all(2001 + local_rank)
input_data = torch.randn(32, 123).cuda()
@@ -109,7 +105,7 @@ def exam_zero_1_2():
assert torch.equal(z1p.data, z2p.data)
@parameterize('dtype', [torch.float16, torch.bfloat16])
@parameterize("dtype", [torch.float16, torch.bfloat16])
def exam_zero_1_torch_ddp(world_size, dtype: torch.dtype):
"""
In this test, two pairs of model and optimizers are created.
@@ -134,10 +130,9 @@ def exam_zero_1_torch_ddp(world_size, dtype: torch.dtype):
# we only test stage 1 here
# in `check_sharded_param_consistency.py`, we will test whether
# level 1 and 2 will produce exactly the same results
zero_optimizer = LowLevelZeroOptimizer(zero_optimizer,
overlap_communication=True,
initial_scale=1,
reduce_bucket_size=1024 * 1024)
zero_optimizer = LowLevelZeroOptimizer(
zero_optimizer, overlap_communication=True, initial_scale=1, reduce_bucket_size=1024 * 1024
)
torch_optimizer = torch.optim.SGD(torch_model.parameters(), lr=1)
@@ -178,7 +173,7 @@ def exam_zero_1_torch_ddp(world_size, dtype: torch.dtype):
def run_dist(rank, world_size, port):
colossalai.launch(config=dict(), rank=rank, world_size=world_size, port=port, host='localhost')
colossalai.launch(config=dict(), rank=rank, world_size=world_size, port=port, host="localhost")
exam_zero_1_torch_ddp(world_size=world_size)
exam_zero_1_2()
@@ -190,5 +185,5 @@ def test_zero_1_2():
spawn(run_dist, 2)
if __name__ == '__main__':
if __name__ == "__main__":
test_zero_1_2()

19
tests/test_zero/test_low_level/test_zero_ckpt.py
View File

@@ -2,19 +2,17 @@ import copy
import pytest
import torch
import torch.distributed as dist
import torch.nn as nn
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.testing import assert_close
import colossalai
from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
from colossalai.testing import rerun_if_address_is_in_use, spawn
from colossalai.testing.random import seed_all
from colossalai.zero import LowLevelZeroOptimizer
class MlpModel(nn.Module):
def __init__(self):
super(MlpModel, self).__init__()
self.linear1 = nn.Linear(12, 24)
@@ -61,10 +59,9 @@ def exam_zero_1_torch_ddp_ckpt():
# we only test stage 1 here
# the state dicts of stage 1 and stage 2 are the same
zero_optimizer = LowLevelZeroOptimizer(zero_optimizer,
overlap_communication=True,
initial_scale=1,
reduce_bucket_size=262144)
zero_optimizer = LowLevelZeroOptimizer(
zero_optimizer, overlap_communication=True, initial_scale=1, reduce_bucket_size=262144
)
torch_optimizer = torch.optim.Adam(torch_model.parameters(), lr=1)
@@ -88,7 +85,7 @@ def exam_zero_1_torch_ddp_ckpt():
zero_state_dict = zero_optimizer.state_dict()
# examine the original state dict
for torch_state, zero_state in zip(torch_state_dict['state'].values(), zero_state_dict['state'].values()):
for torch_state, zero_state in zip(torch_state_dict["state"].values(), zero_state_dict["state"].values()):
for t_v, z_v in zip(torch_state.values(), zero_state.values()):
loose_close(t_v, z_v)
@@ -100,13 +97,13 @@ def exam_zero_1_torch_ddp_ckpt():
zero_state_dict = zero_optimizer.state_dict()
# examine the loaded state dict
for torch_state, zero_state in zip(torch_state_dict['state'].values(), zero_state_dict['state'].values()):
for torch_state, zero_state in zip(torch_state_dict["state"].values(), zero_state_dict["state"].values()):
for t_v, z_v in zip(torch_state.values(), zero_state.values()):
loose_close(t_v, z_v)
def run_dist(rank, world_size, port):
colossalai.launch(config=dict(), rank=rank, world_size=world_size, port=port, host='localhost')
colossalai.launch(config=dict(), rank=rank, world_size=world_size, port=port, host="localhost")
exam_zero_1_torch_ddp_ckpt()
@@ -117,5 +114,5 @@ def test_zero_ckpt():
spawn(run_dist, 2)
if __name__ == '__main__':
if __name__ == "__main__":
test_zero_ckpt()