2022-03-08 10:18:06 +00:00
|
|
|
#!/usr/bin/env python
|
|
|
|
# -*- encoding: utf-8 -*-
|
|
|
|
|
|
|
|
from copy import deepcopy
|
|
|
|
from functools import partial
|
|
|
|
|
|
|
|
import colossalai
|
|
|
|
import pytest
|
|
|
|
import torch
|
|
|
|
import torch.multiprocessing as mp
|
2022-03-29 01:09:04 +00:00
|
|
|
from colossalai.testing import parameterize, rerun_on_exception
|
2022-03-08 10:18:06 +00:00
|
|
|
from colossalai.utils import free_port
|
2022-03-18 07:44:47 +00:00
|
|
|
from colossalai.zero.init_ctx import ZeroInitContext
|
2022-03-14 07:06:02 +00:00
|
|
|
from colossalai.zero.shard_utils import (BucketTensorShardStrategy, TensorShardStrategy)
|
2022-03-08 10:18:06 +00:00
|
|
|
from colossalai.zero.sharded_model import ShardedModelV2
|
2022-03-18 07:44:47 +00:00
|
|
|
from colossalai.zero.sharded_model.utils import col_model_deepcopy
|
2022-03-08 10:18:06 +00:00
|
|
|
from tests.components_to_test.registry import non_distributed_component_funcs
|
2022-03-18 07:44:47 +00:00
|
|
|
|
2022-03-08 10:18:06 +00:00
|
|
|
from common import CONFIG
|
|
|
|
|
|
|
|
|
2022-03-18 07:44:47 +00:00
|
|
|
@parameterize("shard_strategy_class", [TensorShardStrategy, BucketTensorShardStrategy])
|
|
|
|
def run_zero_state_dict(shard_strategy_class):
|
2022-03-08 10:18:06 +00:00
|
|
|
test_models = ['repeated_computed_layers', 'resnet18']
|
2022-03-18 07:44:47 +00:00
|
|
|
shard_strategy = shard_strategy_class()
|
2022-03-08 10:18:06 +00:00
|
|
|
for model_name in test_models:
|
|
|
|
get_components_func = non_distributed_component_funcs.get_callable(model_name)
|
2022-03-11 06:09:09 +00:00
|
|
|
model_builder, train_dataloader, test_dataloader, optimizer, criterion = get_components_func()
|
2022-03-18 07:44:47 +00:00
|
|
|
|
2022-03-29 01:09:04 +00:00
|
|
|
with ZeroInitContext(target_device=torch.cuda.current_device(),
|
2022-03-18 07:44:47 +00:00
|
|
|
shard_strategy=shard_strategy,
|
|
|
|
shard_param=True,
|
|
|
|
rm_torch_payload_on_the_fly=False):
|
|
|
|
zero_model = model_builder(checkpoint=True)
|
|
|
|
zero_model = ShardedModelV2(zero_model, shard_strategy)
|
|
|
|
|
|
|
|
model = model_builder(checkpoint=True).half()
|
|
|
|
col_model_deepcopy(zero_model, model)
|
|
|
|
model = model.cuda()
|
|
|
|
|
2022-03-08 10:18:06 +00:00
|
|
|
zero_state_dict = zero_model.state_dict()
|
|
|
|
for key, val in model.state_dict().items():
|
|
|
|
assert torch.equal(val, zero_state_dict[key])
|
|
|
|
|
|
|
|
|
2022-03-18 03:35:54 +00:00
|
|
|
def run_dist(rank, world_size, port):
|
|
|
|
colossalai.launch(config=CONFIG, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
|
|
|
|
run_zero_state_dict()
|
|
|
|
|
|
|
|
|
2022-03-08 10:18:06 +00:00
|
|
|
@pytest.mark.dist
|
2022-03-14 07:06:02 +00:00
|
|
|
@pytest.mark.parametrize("world_size", [1, 2])
|
2022-03-25 09:25:12 +00:00
|
|
|
@rerun_on_exception(exception_type=mp.ProcessRaisedException, pattern=".*Address already in use.*")
|
2022-03-18 03:35:54 +00:00
|
|
|
def test_zero_state_dict(world_size):
|
|
|
|
run_func = partial(run_dist, world_size=world_size, port=free_port())
|
2022-03-08 10:18:06 +00:00
|
|
|
mp.spawn(run_func, nprocs=world_size)
|
|
|
|
|
|
|
|
|
|
|
|
if __name__ == '__main__':
|
2022-03-14 07:06:02 +00:00
|
|
|
test_zero_state_dict(2, TensorShardStrategy)
|