2023-06-09 06:36:54 +00:00
|
|
|
import pytest
|
|
|
|
import torch
|
|
|
|
|
|
|
|
import colossalai
|
|
|
|
from colossalai.logging import disable_existing_loggers
|
2023-06-22 02:33:06 +00:00
|
|
|
from colossalai.testing import assert_hf_output_close, clear_cache_before_run, rerun_if_address_is_in_use, spawn
|
|
|
|
from tests.kit.model_zoo import model_zoo
|
|
|
|
from tests.test_shardformer.test_model._utils import build_model, run_forward
|
2023-06-09 06:36:54 +00:00
|
|
|
|
|
|
|
|
2023-06-22 02:33:06 +00:00
|
|
|
def check_forward_backward(org_model, sharded_model, data_gen_fn, output_transform_fn, loss_fn):
|
|
|
|
# check forward
|
|
|
|
org_output, org_loss, shard_output, shard_loss = run_forward(org_model, sharded_model, data_gen_fn,
|
|
|
|
output_transform_fn, loss_fn)
|
|
|
|
assert_hf_output_close(org_output, shard_output)
|
2023-06-09 06:36:54 +00:00
|
|
|
|
2023-06-22 02:33:06 +00:00
|
|
|
# do backward
|
2023-06-09 06:36:54 +00:00
|
|
|
org_loss.backward()
|
|
|
|
shard_loss.backward()
|
2023-06-22 02:33:06 +00:00
|
|
|
|
|
|
|
# check grad equality
|
|
|
|
if org_model.__class__.__name__ == 'BertModel':
|
|
|
|
org_grad = org_model.encoder.layer[0].attention.self.query.weight.grad
|
|
|
|
shard_grad = sharded_model.encoder.layer[0].attention.self.query.weight.grad
|
|
|
|
else:
|
|
|
|
org_grad = org_model.bert.encoder.layer[0].attention.self.query.weight.grad
|
|
|
|
shard_grad = sharded_model.bert.encoder.layer[0].attention.self.query.weight.grad
|
2023-06-09 06:36:54 +00:00
|
|
|
|
|
|
|
shard_grad_list = [torch.zeros([*shard_grad.shape]).to('cuda') for _ in range(2)]
|
|
|
|
shard_grad = torch.distributed.all_gather(shard_grad_list, shard_grad)
|
|
|
|
all_shard_grad = torch.cat(shard_grad_list, dim=0)
|
|
|
|
|
|
|
|
assert torch.allclose(org_loss, shard_loss,
|
|
|
|
atol=1e-5), f"shard model loss is not equal to orgin model loss\n{org_loss}\n{shard_loss}"
|
|
|
|
assert torch.allclose(org_grad, all_shard_grad,
|
2023-06-22 02:33:06 +00:00
|
|
|
atol=1e-5), f"shard model grad is not equal to orgin model grad\n{org_grad}\n{all_shard_grad}"
|
2023-06-09 06:36:54 +00:00
|
|
|
|
|
|
|
|
|
|
|
def check_bert(rank, world_size, port):
|
|
|
|
disable_existing_loggers()
|
2023-06-22 02:33:06 +00:00
|
|
|
colossalai.launch(config={}, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
|
2023-06-19 05:53:17 +00:00
|
|
|
|
2023-06-22 02:33:06 +00:00
|
|
|
sub_model_zoo = model_zoo.get_sub_registry('transformers_bert')
|
|
|
|
for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items():
|
2023-06-30 01:58:08 +00:00
|
|
|
org_model, sharded_model = build_model(model_fn)
|
2023-06-22 02:33:06 +00:00
|
|
|
check_forward_backward(org_model, sharded_model, data_gen_fn, output_transform_fn, loss_fn)
|
2023-06-15 09:56:51 +00:00
|
|
|
|
2023-06-22 02:33:06 +00:00
|
|
|
torch.cuda.empty_cache()
|
2023-06-09 06:36:54 +00:00
|
|
|
|
|
|
|
|
|
|
|
@pytest.mark.dist
|
|
|
|
@rerun_if_address_is_in_use()
|
2023-06-22 02:33:06 +00:00
|
|
|
@clear_cache_before_run()
|
2023-06-09 06:36:54 +00:00
|
|
|
def test_bert():
|
|
|
|
spawn(check_bert, 2)
|
|
|
|
|
|
|
|
|
|
|
|
if __name__ == "__main__":
|
|
|
|
test_bert()
|