github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-17 07:00:37 +00:00
Code Issues Packages Projects Releases Wiki Activity

[gemini] improve compatibility and add static placement policy (#4479)

Browse Source 
* [gemini] remove distributed-related part from colotensor (#4379)

* [gemini] remove process group dependency

* [gemini] remove tp part from colo tensor

* [gemini] patch inplace op

* [gemini] fix param op hook and update tests

* [test] remove useless tests

* [test] remove useless tests

* [misc] fix requirements

* [test] fix model zoo

* [test] fix model zoo

* [test] fix model zoo

* [test] fix model zoo

* [test] fix model zoo

* [misc] update requirements

* [gemini] refactor gemini optimizer and gemini ddp (#4398)

* [gemini] update optimizer interface

* [gemini] renaming gemini optimizer

* [gemini] refactor gemini ddp class

* [example] update gemini related example

* [example] update gemini related example

* [plugin] fix gemini plugin args

* [test] update gemini ckpt tests

* [gemini] fix checkpoint io

* [example] fix opt example requirements

* [example] fix opt example

* [example] fix opt example

* [example] fix opt example

* [gemini] add static placement policy (#4443)

* [gemini] add static placement policy

* [gemini] fix param offload

* [test] update gemini tests

* [plugin] update gemini plugin

* [plugin] update gemini plugin docstr

* [misc] fix flash attn requirement

* [test] fix gemini checkpoint io test

* [example] update resnet example result (#4457)

* [example] update bert example result (#4458)

* [doc] update gemini doc (#4468)

* [example] update gemini related examples (#4473)

* [example] update gpt example

* [example] update dreambooth example

* [example] update vit

* [example] update opt

* [example] update palm

* [example] update vit and opt benchmark

* [hotfix] fix bert in model zoo (#4480)

* [hotfix] fix bert in model zoo

* [test] remove chatglm gemini test

* [test] remove sam gemini test

* [test] remove vit gemini test

* [hotfix] fix opt tutorial example (#4497)

* [hotfix] fix opt tutorial example

* [hotfix] fix opt tutorial example
This commit is contained in:
Hongxin Liu
2023-08-24 09:29:25 +08:00
committed by GitHub
parent 285fe7ba71
commit 27061426f7
82 changed files with 1008 additions and 4036 deletions
Download Patch File Download Diff File Expand all files Collapse all files

80
tests/test_zero/test_gemini/test_zeroddp_state_dict.py
View File

@@ -4,31 +4,46 @@ from torch.testing import assert_close
import colossalai
from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
from colossalai.utils.cuda import get_current_device
from colossalai.zero import ColoInitContext, ZeroDDP
from colossalai.zero.gemini.chunk import ChunkManager, search_chunk_configuration
from colossalai.zero.gemini.gemini_mgr import GeminiManager
from colossalai.zero import GeminiDDP
from colossalai.zero.gemini.chunk import search_chunk_configuration
from tests.components_to_test.registry import non_distributed_component_funcs
from tests.test_tensor.common_utils import debug_print, set_seed
from tests.test_tensor.common_utils import set_seed
PLACEMENT_CONFIGS = [
{
'placement_policy': 'static',
'shard_param_frac': 0.0
}, # zero2
{
'placement_policy': 'static',
'shard_param_frac': 1.0
}, # zero3
{
'placement_policy': 'static',
'shard_param_frac': 0.5
}, # zero3-half
{
'placement_policy': 'auto'
}
]
def ignore_the_first_parameter(model: torch.nn.Module):
for name, param in model.named_parameters():
print(f"parameter `{name}` is set ignored")
ZeroDDP.set_params_to_ignore([param])
GeminiDDP.set_params_to_ignore([param])
return
@parameterize('placement_policy', ['cuda', 'cpu', 'auto'])
@parameterize('placement_config', PLACEMENT_CONFIGS)
@parameterize('keep_gathered', [True, False])
@parameterize('model_name', ['gpt2', 'bert'])
def exam_state_dict(placement_policy, keep_gathered, model_name: str):
def exam_state_dict(placement_config, keep_gathered, model_name: str):
set_seed(431)
get_components_func = non_distributed_component_funcs.get_callable(model_name)
model_builder, train_dataloader, test_dataloader, optimizer_class, criterion = get_components_func()
with ColoInitContext(device=get_current_device()):
model = model_builder()
model = model_builder()
torch_model = model_builder()
for torch_p, p in zip(torch_model.parameters(), model.parameters()):
@@ -38,9 +53,7 @@ def exam_state_dict(placement_policy, keep_gathered, model_name: str):
config_dict, *_ = search_chunk_configuration(model, search_range_m=1, search_interval=100)
config_dict[world_size]['chunk_size'] = 5000
config_dict[world_size]['keep_gathered'] = keep_gathered
chunk_manager = ChunkManager(config_dict)
gemini_manager = GeminiManager(placement_policy, chunk_manager)
model = ZeroDDP(model, gemini_manager, pin_memory=True)
model = GeminiDDP(model, config_dict, **placement_config, pin_memory=True)
model.train()
zero_dict = model.state_dict(only_rank_0=False)
@@ -52,16 +65,15 @@ def exam_state_dict(placement_policy, keep_gathered, model_name: str):
assert_close(value, temp_zero_value, rtol=1e-3, atol=1e-5)
@parameterize('placement_policy', ['cuda', 'cpu', 'auto'])
@parameterize('placement_config', PLACEMENT_CONFIGS)
@parameterize('keep_gathered', [True, False])
@parameterize('model_name', ['gpt2', 'bert'])
def exam_load_state_dict(placement_policy, keep_gathered, model_name: str):
def exam_load_state_dict(placement_config, keep_gathered, model_name: str):
set_seed(431)
get_components_func = non_distributed_component_funcs.get_callable(model_name)
model_builder, train_dataloader, test_dataloader, optimizer_class, criterion = get_components_func()
with ColoInitContext(device=get_current_device()):
model = model_builder()
model = model_builder()
set_seed(451)
torch_model = model_builder() # get a different model
@@ -71,13 +83,7 @@ def exam_load_state_dict(placement_policy, keep_gathered, model_name: str):
config_dict[world_size]['chunk_size'] = 5000
config_dict[world_size]['keep_gathered'] = keep_gathered
if placement_policy != 'cuda':
init_device = torch.device('cpu')
else:
init_device = None
chunk_manager = ChunkManager(config_dict, init_device=init_device)
gemini_manager = GeminiManager(placement_policy, chunk_manager)
model = ZeroDDP(model, gemini_manager, pin_memory=True)
model = GeminiDDP(model, config_dict, **placement_config, pin_memory=True)
torch_dict = torch_model.state_dict()
model.load_state_dict(torch_dict, strict=False)
@@ -89,11 +95,37 @@ def exam_load_state_dict(placement_policy, keep_gathered, model_name: str):
assert_close(value, temp_zero_value, rtol=1e-3, atol=1e-5)
@parameterize('placement_config', PLACEMENT_CONFIGS)
@parameterize('model_name', ['gpt2', 'bert'])
def exam_state_dict_shard(placement_config, model_name: str):
get_components_func = non_distributed_component_funcs.get_callable(model_name)
model_builder, train_dataloader, test_dataloader, optimizer_class, criterion = get_components_func()
model = model_builder()
model_size = sum(p.numel() * p.element_size() for p in model.parameters()) / 1024**2
config_dict, *_ = search_chunk_configuration(model, search_range_m=1, search_interval=100)
model = GeminiDDP(model, config_dict, **placement_config)
model.train()
zero_dict = model.state_dict(only_rank_0=False)
accumulated_keys = set()
# ensure number of shards > 1
for shard, _ in model.state_dict_shard(max_shard_size=(model_size / 3), only_rank_0=False):
for key, value in shard.items():
assert key not in accumulated_keys, f"key `{key}` is duplicated."
accumulated_keys.add(key)
assert key in zero_dict, f"{key} not in ZeRO dictionary."
assert torch.equal(value, zero_dict[key]), f"{key} not equal."
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_state_dict()
exam_load_state_dict()
exam_state_dict_shard()
@pytest.mark.dist