[BUG] Unreasonable memory comsumption

[ZhiYuanZeng]

🐛 Describe the bug

Creating an TransformerEncoder causes memory overflow, but the same config works with the huggingface transformers module.

# config.py
from colossalai.amp import AMP_TYPE

fp16=dict(
    mode=AMP_TYPE.TORCH
)
NUM_MICRO_BATCHES=8
parallel = dict(
    tensor=dict(size=4, mode='2d')
)

# launch command: python -m torch.distributed.launch --nproc_per_node 4 --master_port 12345 xxx.py
# Memory overflow on Nvidia 2080 Ti
from titans.layer.block import TransformerEncoderLayer,TransformerEncoder
colossalai.launch_from_torch(config='/home/zyzeng/fastnlp/examples/config.py')
backbone=TransformerEncoder(
    TransformerEncoderLayer(hidden_size=768, nhead=12, dim_feedforward=768*4), 
    num_layers=12
)

# No memory overflow on Nvidia 2080 Ti
from transformers import BertModel, AutoConfig
config=AutoConfig.from_pretrained('bert-base-uncased')
model=BertModel(config)
model.cuda()

Environment

No response

[ZhiYuanZeng]

I also have a question: is the Tensor parallel module must be created after colossalai.launch?