2022-03-01 10:17:01 +00:00
|
|
|
#!/usr/bin/env python
|
|
|
|
# -*- encoding: utf-8 -*-
|
|
|
|
|
|
|
|
import copy
|
|
|
|
from functools import partial
|
|
|
|
|
|
|
|
import colossalai
|
|
|
|
import pytest
|
|
|
|
import torch
|
2022-03-02 10:28:29 +00:00
|
|
|
import torch.distributed as dist
|
2022-03-01 10:17:01 +00:00
|
|
|
import torch.multiprocessing as mp
|
|
|
|
from colossalai.utils import free_port
|
2022-03-08 10:18:06 +00:00
|
|
|
from colossalai.zero.shard_utils.tensor_shard_strategy import \
|
|
|
|
TensorShardStrategy
|
2022-03-01 10:17:01 +00:00
|
|
|
from colossalai.zero.sharded_model import ShardedModelV2
|
2022-03-08 10:18:06 +00:00
|
|
|
from tests.components_to_test.registry import non_distributed_component_funcs
|
|
|
|
from torch.nn.parallel import DistributedDataParallel as DDP
|
2022-03-02 10:28:29 +00:00
|
|
|
|
2022-03-08 10:18:06 +00:00
|
|
|
from common import CONFIG, check_grads, check_grads_padding
|
2022-03-01 10:17:01 +00:00
|
|
|
|
|
|
|
|
2022-03-08 10:18:06 +00:00
|
|
|
def run_fwd_bwd(model, data, label, criterion, enable_autocast=False):
|
2022-03-01 10:17:01 +00:00
|
|
|
model.train()
|
|
|
|
with torch.cuda.amp.autocast(enabled=enable_autocast):
|
2022-03-08 10:18:06 +00:00
|
|
|
y = model(data)
|
|
|
|
loss = criterion(y, label)
|
2022-03-01 10:17:01 +00:00
|
|
|
loss = loss.float()
|
2022-03-02 10:28:29 +00:00
|
|
|
if isinstance(model, ShardedModelV2):
|
|
|
|
model.backward(loss)
|
|
|
|
else:
|
|
|
|
loss.backward()
|
|
|
|
|
2022-03-01 10:17:01 +00:00
|
|
|
|
2022-03-09 02:39:02 +00:00
|
|
|
def run_bert_fwd_bwd(model, data, label, enable_autocast=False):
|
|
|
|
model.train()
|
|
|
|
with torch.cuda.amp.autocast(enabled=enable_autocast):
|
|
|
|
output = model(input_ids=data, labels=label)
|
|
|
|
loss = output[0]
|
|
|
|
if isinstance(model, ShardedModelV2):
|
|
|
|
model.backward(loss)
|
|
|
|
else:
|
|
|
|
loss.backward()
|
|
|
|
|
|
|
|
|
2022-03-01 10:17:01 +00:00
|
|
|
def run_dist(rank, world_size, port):
|
2022-03-08 04:03:35 +00:00
|
|
|
colossalai.launch(config=CONFIG, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
|
2022-03-09 02:39:02 +00:00
|
|
|
test_models = ['repeated_computed_layers', 'resnet18', 'bert']
|
|
|
|
shard_strategy = TensorShardStrategy()
|
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)
|
|
|
|
model, train_dataloader, test_dataloader, optimizer, criterion = get_components_func()
|
2022-03-09 02:39:02 +00:00
|
|
|
model = model(checkpoint=True).half().cuda()
|
2022-03-08 10:18:06 +00:00
|
|
|
zero_model = ShardedModelV2(copy.deepcopy(model), shard_strategy)
|
2022-03-02 10:28:29 +00:00
|
|
|
if dist.get_world_size() > 1:
|
2022-03-08 10:18:06 +00:00
|
|
|
model = DDP(model)
|
|
|
|
|
|
|
|
for i, (data, label) in enumerate(train_dataloader):
|
|
|
|
if i > 2:
|
|
|
|
break
|
2022-03-09 02:39:02 +00:00
|
|
|
|
|
|
|
if model_name == 'bert':
|
|
|
|
data, label = data.cuda(), label.cuda()
|
|
|
|
run_bert_fwd_bwd(model, data, label, False)
|
|
|
|
run_bert_fwd_bwd(zero_model, data, label, False)
|
|
|
|
else:
|
|
|
|
data, label = data.half().cuda(), label.cuda()
|
|
|
|
run_fwd_bwd(model, data, label, criterion, False)
|
|
|
|
run_fwd_bwd(zero_model, data, label, criterion, False)
|
|
|
|
|
2022-03-08 10:18:06 +00:00
|
|
|
if dist.get_world_size() > 1:
|
|
|
|
check_grads_padding(model, zero_model, loose=True)
|
|
|
|
else:
|
|
|
|
check_grads(model, zero_model, loose=True)
|
2022-03-01 10:17:01 +00:00
|
|
|
|
|
|
|
|
|
|
|
@pytest.mark.dist
|
2022-03-08 04:03:35 +00:00
|
|
|
@pytest.mark.parametrize("world_size", [1, 2, 4])
|
|
|
|
def test_shard_model_v2(world_size):
|
2022-03-01 10:17:01 +00:00
|
|
|
run_func = partial(run_dist, world_size=world_size, port=free_port())
|
|
|
|
mp.spawn(run_func, nprocs=world_size)
|
|
|
|
|
|
|
|
|
|
|
|
if __name__ == '__main__':
|
2022-03-08 04:03:35 +00:00
|
|
|
test_shard_model_v2(world_size=2)
|