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

[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

106
tests/test_zero/test_gemini/test_optim.py
View File

@@ -16,50 +16,30 @@ from tests.components_to_test import run_fwd_bwd
from tests.components_to_test.registry import non_distributed_component_funcs
PLACEMENT_CONFIGS = [
{"placement_policy": "static", "shard_param_frac": 0.0, "offload_optim_frac": 0.0}, # zero2
{"placement_policy": "static", "shard_param_frac": 0.0, "offload_optim_frac": 1.0}, # zero2-offload
{"placement_policy": "static", "shard_param_frac": 0.0, "offload_optim_frac": 0.5}, # zero2-offload-half
{"placement_policy": "static", "shard_param_frac": 1.0}, # zero3
{"placement_policy": "static", "shard_param_frac": 0.5}, # zero3-half
{
'placement_policy': 'static',
'shard_param_frac': 0.0,
'offload_optim_frac': 0.0
}, # zero2
{
'placement_policy': 'static',
'shard_param_frac': 0.0,
'offload_optim_frac': 1.0
}, # zero2-offload
{
'placement_policy': 'static',
'shard_param_frac': 0.0,
'offload_optim_frac': 0.5
}, # zero2-offload-half
{
'placement_policy': 'static',
'shard_param_frac': 1.0
}, # zero3
{
'placement_policy': 'static',
'shard_param_frac': 0.5
}, # zero3-half
{
'placement_policy': 'static',
'shard_param_frac': 1.0,
'offload_optim_frac': 1.0,
'offload_param_frac': 1.0
}, # zero3-offload-all
{
'placement_policy': 'auto'
}
"placement_policy": "static",
"shard_param_frac": 1.0,
"offload_optim_frac": 1.0,
"offload_param_frac": 1.0,
}, # zero3-offload-all
{"placement_policy": "auto"},
]
# this model is large enough to slice to chunks
TEST_MODELS = ['gpt2']
TEST_MODELS = ["gpt2"]
# these models are too small, all parameters in these models are compacted into one chunk
EXAMPLE_MODELS = ['albert', 'beit', 'bert', 'hanging_param_model', 'nested_model', 'repeated_computed_layers']
EXAMPLE_MODELS = ["albert", "beit", "bert", "hanging_param_model", "nested_model", "repeated_computed_layers"]
# bfloat16 cannot represent them exactly
BF16_IGNORED_KEYS = [
'albert.embeddings.word_embeddings.weight',
'albert.embeddings.position_embeddings.weight',
'masked_bias',
"albert.embeddings.word_embeddings.weight",
"albert.embeddings.position_embeddings.weight",
"masked_bias",
]
@@ -78,23 +58,25 @@ def check_param(model: GeminiDDP, torch_model: torch.nn.Module, dtype: torch.dty
if dtype is torch.bfloat16:
rtol, atol = 4e-3, 8e-3
# debug_print([0], "max range: ", key, torch.max(torch.abs(value - temp_zero_value)))
assert_close(value.float(),
temp_zero_value.float(),
rtol=rtol,
atol=atol,
msg=lambda s: s + f'\n{key}\n{temp_zero_value.dtype}')
assert_close(
value.float(),
temp_zero_value.float(),
rtol=rtol,
atol=atol,
msg=lambda s: s + f"\n{key}\n{temp_zero_value.dtype}",
)
@parameterize('placement_config', PLACEMENT_CONFIGS)
@parameterize('model_name', TEST_MODELS)
@parameterize('mixed_precision', [torch.half, torch.bfloat16])
@parameterize("placement_config", PLACEMENT_CONFIGS)
@parameterize("model_name", TEST_MODELS)
@parameterize("mixed_precision", [torch.half, torch.bfloat16])
def exam_model_step(placement_config, model_name: str, mixed_precision: torch.dtype):
set_seed(42)
get_components_func = non_distributed_component_funcs.get_callable(model_name)
model_builder, train_dataloader, test_dataloader, optimizer_class, criterion = get_components_func()
torch_model = model_builder().cuda()
amp_config = dict(opt_level='O2', keep_batchnorm_fp32=False, loss_scale=128)
amp_config = dict(opt_level="O2", keep_batchnorm_fp32=False, loss_scale=128)
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()])
@@ -106,8 +88,8 @@ def exam_model_step(placement_config, model_name: str, mixed_precision: torch.dt
world_size = torch.distributed.get_world_size()
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'] = False
config_dict[world_size]["chunk_size"] = 5000
config_dict[world_size]["keep_gathered"] = False
model = GeminiDDP(model, config_dict, **placement_config, mixed_precision=mixed_precision)
optimizer = HybridAdam(model.parameters(), lr=1e-3)
@@ -135,16 +117,16 @@ def exam_model_step(placement_config, model_name: str, mixed_precision: torch.dt
check_param(model, torch_model, mixed_precision)
@parameterize('placement_config', PLACEMENT_CONFIGS)
@parameterize('model_name', EXAMPLE_MODELS)
@parameterize('mixed_precision', [torch.half, torch.bfloat16])
@parameterize("placement_config", PLACEMENT_CONFIGS)
@parameterize("model_name", EXAMPLE_MODELS)
@parameterize("mixed_precision", [torch.half, torch.bfloat16])
def exam_tiny_example(placement_config, model_name: str, mixed_precision: torch.dtype):
set_seed(2008)
get_components_func = non_distributed_component_funcs.get_callable(model_name)
model_builder, train_dataloader, test_dataloader, optimizer_class, criterion = get_components_func()
torch_model = model_builder().cuda()
amp_config = dict(opt_level='O2', keep_batchnorm_fp32=False, loss_scale=2)
amp_config = dict(opt_level="O2", keep_batchnorm_fp32=False, loss_scale=2)
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()])
@@ -154,12 +136,14 @@ def exam_tiny_example(placement_config, model_name: str, mixed_precision: torch.
for torch_p, p in zip(torch_model.parameters(), model.parameters()):
p.data.copy_(torch_p.data)
model = GeminiDDP(model,
chunk_init_device=get_current_device(),
search_range_m=1,
pin_memory=True,
mixed_precision=mixed_precision,
**placement_config)
model = GeminiDDP(
model,
chunk_init_device=get_current_device(),
search_range_m=1,
pin_memory=True,
mixed_precision=mixed_precision,
**placement_config,
)
optimizer = HybridAdam(model.parameters(), lr=1e-3)
zero_optim = GeminiOptimizer(optimizer, model, initial_scale=2)
@@ -182,7 +166,7 @@ def exam_tiny_example(placement_config, model_name: str, mixed_precision: torch.
torch_loss = run_fwd_bwd(torch_model, input_ids, label, criterion, torch_optim)
loss = run_fwd_bwd(model, input_ids, label, criterion, zero_optim)
assert_close(torch_loss, loss, rtol=rtol, atol=atol) # atol should be 2e-5 for torch lower than 1.12
assert_close(torch_loss, loss, rtol=rtol, atol=atol) # atol should be 2e-5 for torch lower than 1.12
zero_optim.step()
torch_optim.step()
@@ -192,17 +176,17 @@ def exam_tiny_example(placement_config, model_name: str, mixed_precision: torch.
def run_dist(rank, world_size, port):
config = {}
colossalai.launch(config=config, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
colossalai.launch(config=config, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl")
exam_model_step()
exam_tiny_example()
@pytest.mark.dist
@pytest.mark.parametrize('world_size', [1, 4])
@pytest.mark.parametrize("world_size", [1, 4])
@rerun_if_address_is_in_use()
def test_optim(world_size):
spawn(run_dist, world_size)
if __name__ == '__main__':
if __name__ == "__main__":
test_optim(1)