distributed training error

Hi, I try the distributed training with 2 machines. There are 4 GPUs in each machine.
in the master machine, I run:
python -u tools/run_net.py \
--cfg configs/Kinetics/SLOWFAST_8x8_R50.yaml \
--shard_id 0 \
--num_shards 2 \
--init_method tcp://127.0.0.1:40000 \
DATA.PATH_TO_DATA_DIR data/k400 \
DATA.PATH_PREFIX /root/data/datasets/k400_non_local_wangxiaolong/compress \
NUM_GPUS 4 \
TRAIN.BATCH_SIZE 32 \
TRAIN.EVAL_PERIOD 1

in the second machine, I run:
python -u tools/run_net.py \
--cfg configs/Kinetics/SLOWFAST_8x8_R50.yaml \
--shard_id 1 \
--num_shards 2 \
--init_method tcp://10.110.17.113:50933 \
DATA.PATH_TO_DATA_DIR data/k400 DATA.PATH_PREFIX \
/root/data/datasets/k400_non_local_wangxiaolong/compress NUM_GPUS 4 \
TRAIN.BATCH_SIZE 32 \
TRAIN.EVAL_PERIOD 1

the log on the master is :
[INFO: kinetics.py: 78]: Constructing Kinetics train...
[INFO: kinetics.py: 113]: Constructing kinetics dataloader (size: 234584) from data/k400/train.csv
[INFO: kinetics.py: 78]: Constructing Kinetics val...
[INFO: kinetics.py: 113]: Constructing kinetics dataloader (size: 19760) from data/k400/val.csv
[INFO: train_net.py: 307]: Start epoch: 1
then it hangs

the code and the data on both machines are totally the same.
the world size is 8.
I find the trainnig on both machines get pending at the the same code line :
optimizer.step()

Does anyone know how does that happen? What should I do ?