github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-09 04:50:17 +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

63
tests/test_zero/test_gemini/test_grad_clip.py
View File

@@ -16,26 +16,24 @@ 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,
'offload_param_frac': 0.0
}, # zero2
"placement_policy": "static",
"shard_param_frac": 0.0,
"offload_optim_frac": 0.0,
"offload_param_frac": 0.0,
}, # zero2
{
'placement_policy': 'static',
'shard_param_frac': 0.0,
'offload_optim_frac': 1.0,
'offload_param_frac': 0.0
}, # zero2-offload
"placement_policy": "static",
"shard_param_frac": 0.0,
"offload_optim_frac": 1.0,
"offload_param_frac": 0.0,
}, # zero2-offload
{
'placement_policy': 'static',
'shard_param_frac': 0.0,
'offload_optim_frac': 0.5,
'offload_param_frac': 0.0
}, # zero2-offload-half
{
'placement_policy': 'auto'
}
"placement_policy": "static",
"shard_param_frac": 0.0,
"offload_optim_frac": 0.5,
"offload_param_frac": 0.0,
}, # zero2-offload-half
{"placement_policy": "auto"},
]
@@ -52,15 +50,15 @@ def check_param(model: GeminiDDP, torch_model: torch.nn.Module):
assert_close(value, temp_zero_value, rtol=1e-3, atol=4e-3)
@parameterize('placement_config', PLACEMENT_CONFIGS)
@parameterize('model_name', ['gpt2'])
@parameterize("placement_config", PLACEMENT_CONFIGS)
@parameterize("model_name", ["gpt2"])
def exam_grad_clipping(placement_config, model_name: str):
set_seed(1912)
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=32)
amp_config = dict(opt_level="O2", keep_batchnorm_fp32=False, loss_scale=32)
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()])
@@ -72,18 +70,16 @@ def exam_grad_clipping(placement_config, model_name: str):
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
if placement_config['placement_policy'] != 'cuda':
init_device = torch.device('cpu')
config_dict[world_size]["chunk_size"] = 5000
config_dict[world_size]["keep_gathered"] = False
if placement_config["placement_policy"] != "cuda":
init_device = torch.device("cpu")
else:
init_device = None
model = GeminiDDP(model,
chunk_config_dict=config_dict,
chunk_init_device=init_device,
pin_memory=True,
**placement_config)
model = GeminiDDP(
model, chunk_config_dict=config_dict, chunk_init_device=init_device, pin_memory=True, **placement_config
)
optimizer = HybridAdam(model.parameters(), lr=1e-3)
zero_optim = GeminiOptimizer(optimizer, model, initial_scale=32, clipping_norm=1.0)
@@ -106,6 +102,7 @@ def exam_grad_clipping(placement_config, model_name: str):
assert_close(torch_loss, loss)
import apex.amp as apex_amp
torch.nn.utils.clip_grad_norm_(apex_amp.master_params(torch_optim), 1.0)
torch_optim.step()
zero_optim.step()
@@ -115,16 +112,16 @@ def exam_grad_clipping(placement_config, model_name: str):
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_grad_clipping()
@pytest.mark.dist
@pytest.mark.parametrize('world_size', [1, 2])
@pytest.mark.parametrize("world_size", [1, 2])
@rerun_if_address_is_in_use()
def test_grad_clip(world_size):
spawn(run_dist, world_size)
if __name__ == '__main__':
if __name__ == "__main__":
test_grad_clip(2)