mirror of
https://github.com/hpcaitech/ColossalAI.git
synced 2025-09-09 21:09:18 +00:00
[hotfix] fix zero's incompatibility with checkpoint in torch-1.12 (#1786)
* [hotfix] fix zero's incompatibility with checkpoint in torch-1.12 * [zero] add cpu shard init * [zero] add tiny example test * [colo_tensor] fix bugs for torch-1.11
This commit is contained in:
HELSON
GitHub
@@ -1,121 +1,124 @@
|
||||
import torch
|
||||
import colossalai
|
||||
import pytest
|
||||
import torch.multiprocessing as mp
|
||||
import torch.distributed as dist
|
||||
from functools import partial
|
||||
from colossalai.testing import rerun_if_address_is_in_use, parameterize
|
||||
from colossalai.utils import free_port, get_current_device
|
||||
from colossalai.tensor import ProcessGroup as ColoProcessGroup
|
||||
from colossalai.tensor import ColoParameter
|
||||
from colossalai.gemini import TensorState
|
||||
from colossalai.gemini.chunk import Chunk
|
||||
|
||||
|
||||
def dist_sum(x):
|
||||
temp = torch.tensor([x], device=get_current_device())
|
||||
dist.all_reduce(temp)
|
||||
return temp.item()
|
||||
|
||||
|
||||
def add_param(param_list, param_cp_list, *args, **kwargs):
|
||||
param = ColoParameter(torch.randn(*args, **kwargs))
|
||||
param_list.append(param)
|
||||
param_cp_list.append(param.clone())
|
||||
|
||||
|
||||
def check_euqal(param, param_cp):
|
||||
if param.device != param_cp.device:
|
||||
temp = param.data.to(param_cp.device)
|
||||
else:
|
||||
temp = param.data
|
||||
return torch.equal(temp, param_cp.data)
|
||||
|
||||
|
||||
@parameterize('init_device', [None, torch.device('cpu')])
|
||||
@parameterize('keep_gathered', [True, False])
|
||||
@parameterize('pin_memory', [True, False])
|
||||
def exam_chunk_basic(init_device, keep_gathered, pin_memory):
|
||||
world_size = torch.distributed.get_world_size()
|
||||
pg = ColoProcessGroup()
|
||||
my_chunk = Chunk(chunk_size=1024,
|
||||
process_group=pg,
|
||||
dtype=torch.float32,
|
||||
init_device=init_device,
|
||||
keep_gathered=keep_gathered,
|
||||
pin_memory=pin_memory)
|
||||
|
||||
param_list = []
|
||||
param_cp_list = []
|
||||
|
||||
add_param(param_list, param_cp_list, 8, 8, 8, device='cuda')
|
||||
add_param(param_list, param_cp_list, 4, 4)
|
||||
add_param(param_list, param_cp_list, 4, 8, 2, device='cuda')
|
||||
add_param(param_list, param_cp_list, 1, 1, 5)
|
||||
|
||||
for param in param_list:
|
||||
my_chunk.append_tensor(param)
|
||||
assert my_chunk.utilized_size == 597
|
||||
for param, param_cp in zip(param_list, param_cp_list):
|
||||
check_euqal(param, param_cp)
|
||||
my_chunk.close_chunk()
|
||||
|
||||
if keep_gathered is False:
|
||||
assert my_chunk.cpu_shard.size(0) == 1024 // world_size
|
||||
assert my_chunk.device_type == 'cpu'
|
||||
assert my_chunk.can_move
|
||||
my_chunk.shard_move(get_current_device())
|
||||
else:
|
||||
assert my_chunk.chunk_total.size(0) == 1024
|
||||
assert my_chunk.device_type == 'cuda'
|
||||
assert not my_chunk.can_move
|
||||
|
||||
assert dist_sum(my_chunk.valid_end) == my_chunk.utilized_size
|
||||
flag = my_chunk.has_inf_or_nan
|
||||
assert not flag, "has_inf_or_nan is {}".format(flag)
|
||||
|
||||
my_chunk.access_chunk()
|
||||
assert my_chunk.device_type == 'cuda'
|
||||
for param, param_cp in zip(param_list, param_cp_list):
|
||||
check_euqal(param, param_cp)
|
||||
|
||||
assert my_chunk.tensors_state_monitor[TensorState.HOLD] == 4
|
||||
my_chunk.tensor_trans_state(param_list[0], TensorState.COMPUTE)
|
||||
assert my_chunk.tensors_state_monitor[TensorState.HOLD] == 3
|
||||
assert my_chunk.tensors_state_monitor[TensorState.COMPUTE] == 1
|
||||
assert not my_chunk.can_release
|
||||
|
||||
for param in param_list:
|
||||
my_chunk.tensor_trans_state(param, TensorState.COMPUTE)
|
||||
my_chunk.tensor_trans_state(param, TensorState.READY_FOR_REDUCE)
|
||||
|
||||
assert my_chunk.tensors_state_monitor[TensorState.READY_FOR_REDUCE] == 4
|
||||
assert my_chunk.can_reduce
|
||||
my_chunk.reduce()
|
||||
assert my_chunk.tensors_state_monitor[TensorState.HOLD] == 4
|
||||
|
||||
if keep_gathered is False:
|
||||
assert my_chunk.cuda_shard.size(0) == 1024 // world_size
|
||||
assert my_chunk.device_type == 'cuda'
|
||||
assert my_chunk.can_move
|
||||
else:
|
||||
assert my_chunk.chunk_total.size(0) == 1024
|
||||
assert my_chunk.device_type == 'cuda'
|
||||
assert not my_chunk.can_move
|
||||
|
||||
|
||||
def run_dist(rank, world_size, port):
|
||||
colossalai.launch(config={}, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
|
||||
exam_chunk_basic()
|
||||
|
||||
|
||||
@pytest.mark.dist
|
||||
@pytest.mark.parametrize('world_size', [1, 2, 4])
|
||||
@rerun_if_address_is_in_use()
|
||||
def test_chunk_function(world_size):
|
||||
run_func = partial(run_dist, world_size=world_size, port=free_port())
|
||||
mp.spawn(run_func, nprocs=world_size)
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
test_chunk_function(4)
|
||||
from functools import partial
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
import torch.distributed as dist
|
||||
import torch.multiprocessing as mp
|
||||
|
||||
import colossalai
|
||||
from colossalai.gemini import TensorState
|
||||
from colossalai.gemini.chunk import Chunk
|
||||
from colossalai.tensor import ColoParameter
|
||||
from colossalai.tensor import ProcessGroup as ColoProcessGroup
|
||||
from colossalai.testing import parameterize, rerun_if_address_is_in_use
|
||||
from colossalai.utils import free_port, get_current_device
|
||||
|
||||
|
||||
def dist_sum(x):
|
||||
temp = torch.tensor([x], device=get_current_device())
|
||||
dist.all_reduce(temp)
|
||||
return temp.item()
|
||||
|
||||
|
||||
def add_param(param_list, param_cp_list, *args, **kwargs):
|
||||
param = ColoParameter(torch.randn(*args, **kwargs))
|
||||
param_list.append(param)
|
||||
param_cp_list.append(param.clone())
|
||||
|
||||
|
||||
def check_euqal(param, param_cp):
|
||||
if param.device != param_cp.device:
|
||||
temp = param.data.to(param_cp.device)
|
||||
else:
|
||||
temp = param.data
|
||||
return torch.equal(temp, param_cp.data)
|
||||
|
||||
|
||||
@parameterize('init_device', [None, torch.device('cpu')])
|
||||
@parameterize('keep_gathered', [True, False])
|
||||
@parameterize('pin_memory', [True, False])
|
||||
def exam_chunk_basic(init_device, keep_gathered, pin_memory):
|
||||
world_size = torch.distributed.get_world_size()
|
||||
pg = ColoProcessGroup()
|
||||
my_chunk = Chunk(chunk_size=1024,
|
||||
process_group=pg,
|
||||
dtype=torch.float32,
|
||||
init_device=init_device,
|
||||
cpu_shard_init=True,
|
||||
keep_gathered=keep_gathered,
|
||||
pin_memory=pin_memory)
|
||||
|
||||
param_list = []
|
||||
param_cp_list = []
|
||||
|
||||
add_param(param_list, param_cp_list, 8, 8, 8, device='cuda')
|
||||
add_param(param_list, param_cp_list, 4, 4)
|
||||
add_param(param_list, param_cp_list, 4, 8, 2, device='cuda')
|
||||
add_param(param_list, param_cp_list, 1, 1, 5)
|
||||
|
||||
for param in param_list:
|
||||
my_chunk.append_tensor(param)
|
||||
assert my_chunk.utilized_size == 597
|
||||
for param, param_cp in zip(param_list, param_cp_list):
|
||||
check_euqal(param, param_cp)
|
||||
my_chunk.close_chunk()
|
||||
|
||||
if keep_gathered is False:
|
||||
assert my_chunk.cpu_shard.size(0) == 1024 // world_size
|
||||
assert my_chunk.device_type == 'cpu'
|
||||
assert my_chunk.can_move
|
||||
my_chunk.shard_move(get_current_device())
|
||||
else:
|
||||
assert my_chunk.chunk_total.size(0) == 1024
|
||||
assert my_chunk.device_type == 'cuda'
|
||||
assert not my_chunk.can_move
|
||||
|
||||
assert dist_sum(my_chunk.valid_end) == my_chunk.utilized_size
|
||||
flag = my_chunk.has_inf_or_nan
|
||||
assert not flag, "has_inf_or_nan is {}".format(flag)
|
||||
|
||||
my_chunk.access_chunk()
|
||||
assert my_chunk.device_type == 'cuda'
|
||||
for param, param_cp in zip(param_list, param_cp_list):
|
||||
check_euqal(param, param_cp)
|
||||
|
||||
assert my_chunk.tensors_state_monitor[TensorState.HOLD] == 4
|
||||
my_chunk.tensor_trans_state(param_list[0], TensorState.COMPUTE)
|
||||
assert my_chunk.tensors_state_monitor[TensorState.HOLD] == 3
|
||||
assert my_chunk.tensors_state_monitor[TensorState.COMPUTE] == 1
|
||||
assert not my_chunk.can_release
|
||||
|
||||
for param in param_list:
|
||||
my_chunk.tensor_trans_state(param, TensorState.COMPUTE)
|
||||
my_chunk.tensor_trans_state(param, TensorState.READY_FOR_REDUCE)
|
||||
|
||||
assert my_chunk.tensors_state_monitor[TensorState.READY_FOR_REDUCE] == 4
|
||||
assert my_chunk.can_reduce
|
||||
my_chunk.reduce()
|
||||
assert my_chunk.tensors_state_monitor[TensorState.HOLD] == 4
|
||||
|
||||
if keep_gathered is False:
|
||||
assert my_chunk.cuda_shard.size(0) == 1024 // world_size
|
||||
assert my_chunk.device_type == 'cuda'
|
||||
assert my_chunk.can_move
|
||||
else:
|
||||
assert my_chunk.chunk_total.size(0) == 1024
|
||||
assert my_chunk.device_type == 'cuda'
|
||||
assert not my_chunk.can_move
|
||||
|
||||
|
||||
def run_dist(rank, world_size, port):
|
||||
colossalai.launch(config={}, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
|
||||
exam_chunk_basic()
|
||||
|
||||
|
||||
@pytest.mark.dist
|
||||
@pytest.mark.parametrize('world_size', [1, 2, 4])
|
||||
@rerun_if_address_is_in_use()
|
||||
def test_chunk_function(world_size):
|
||||
run_func = partial(run_dist, world_size=world_size, port=free_port())
|
||||
mp.spawn(run_func, nprocs=world_size)
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
test_chunk_function(4)
|
||||
|
||||
@@ -40,7 +40,8 @@ def run_fwd_bwd(model, criterion, optimizer, input_ids, attn_mask):
|
||||
|
||||
|
||||
@parameterize('placement_policy', ['cuda', 'cpu', 'auto', 'const'])
|
||||
def exam_gpt_fwd_bwd(placement_policy):
|
||||
@parameterize('keep_gather', [False, True])
|
||||
def exam_gpt_fwd_bwd(placement_policy, keep_gather):
|
||||
set_seed(42)
|
||||
get_components_func = non_distributed_component_funcs.get_callable('gpt2')
|
||||
model_builder, train_dataloader, test_dataloader, optimizer_class, criterion = get_components_func()
|
||||
@@ -55,7 +56,7 @@ def exam_gpt_fwd_bwd(placement_policy):
|
||||
world_size = torch.distributed.get_world_size()
|
||||
config_dict, _ = search_chunk_configuration(model, search_range_mb=1, search_interval_byte=100)
|
||||
config_dict[world_size]['chunk_size'] = 5000
|
||||
config_dict[world_size]['keep_gathered'] = False
|
||||
config_dict[world_size]['keep_gathered'] = keep_gather
|
||||
chunk_manager = ChunkManager(config_dict)
|
||||
gemini_manager = GeminiManager(placement_policy, chunk_manager)
|
||||
model = ZeroDDP(model, gemini_manager, pin_memory=True)
|
||||
@@ -101,4 +102,4 @@ def test_gpt(world_size):
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
test_gpt(1)
|
||||
test_gpt(4)
|
||||
|
||||
@@ -9,7 +9,7 @@ from torch.nn.parallel import DistributedDataParallel as DDP
|
||||
|
||||
import colossalai
|
||||
from colossalai.amp import convert_to_apex_amp
|
||||
from colossalai.gemini.chunk import ChunkManager, search_chunk_configuration
|
||||
from colossalai.gemini.chunk import ChunkManager, init_chunk_manager, search_chunk_configuration
|
||||
from colossalai.gemini.gemini_mgr import GeminiManager
|
||||
from colossalai.nn.optimizer import HybridAdam
|
||||
from colossalai.nn.parallel import ZeroDDP
|
||||
@@ -98,10 +98,55 @@ def exam_gpt_fwd_bwd(placement_policy):
|
||||
check_param(model, torch_model)
|
||||
|
||||
|
||||
@parameterize('placement_policy', ['cuda', 'cpu'])
|
||||
def exam_tiny_example(placement_policy):
|
||||
set_seed(42)
|
||||
get_components_func = non_distributed_component_funcs.get_callable('gpt2')
|
||||
model_builder, train_dataloader, test_dataloader, optimizer_class, criterion = get_components_func()
|
||||
|
||||
with ColoInitContext(device=get_current_device()):
|
||||
model = model_builder()
|
||||
|
||||
torch_model = model_builder().cuda()
|
||||
for torch_p, p in zip(torch_model.parameters(), model.parameters()):
|
||||
torch_p.data.copy_(p.data)
|
||||
|
||||
chunk_manager = init_chunk_manager(model=model, init_device=get_current_device(), search_range_mb=1)
|
||||
gemini_manager = GeminiManager(placement_policy, chunk_manager)
|
||||
model = ZeroDDP(model, gemini_manager, pin_memory=True)
|
||||
|
||||
optimizer = HybridAdam(model.parameters(), lr=1e-3)
|
||||
zero_optim = ZeroOptimizer(optimizer, model, initial_scale=2)
|
||||
|
||||
amp_config = dict(opt_level='O2', keep_batchnorm_fp32=False, loss_scale=1)
|
||||
torch_optim = torch.optim.Adam(torch_model.parameters(), lr=1e-3)
|
||||
torch_model, torch_optim = convert_to_apex_amp(torch_model, torch_optim, amp_config)
|
||||
torch_model = DDP(torch_model, device_ids=[dist.get_rank()])
|
||||
|
||||
model.eval()
|
||||
torch_model.eval()
|
||||
|
||||
set_seed(dist.get_rank() * 3 + 128)
|
||||
for i, (input_ids, attn_mask) in enumerate(train_dataloader):
|
||||
if i > 2:
|
||||
break
|
||||
|
||||
zero_logits = run_fwd_bwd(model, criterion, zero_optim, input_ids, attn_mask)
|
||||
torch_logits = run_fwd_bwd(torch_model, criterion, torch_optim, input_ids, attn_mask)
|
||||
assert torch.allclose(zero_logits, torch_logits, rtol=1e-3, atol=1e-2)
|
||||
# debug_print([0], zero_logits, torch_logits)
|
||||
|
||||
zero_optim.step()
|
||||
torch_optim.step()
|
||||
|
||||
check_param(model, torch_model)
|
||||
|
||||
|
||||
def run_dist(rank, world_size, port):
|
||||
config = {}
|
||||
colossalai.launch(config=config, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
|
||||
exam_gpt_fwd_bwd()
|
||||
exam_tiny_example()
|
||||
|
||||
|
||||
@pytest.mark.dist
|
||||
@@ -113,4 +158,4 @@ def test_gpt(world_size):
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
test_gpt(1)
|
||||
test_gpt(2)
|
||||
|
||||
Reference in New Issue
Block a user