多卡运行报错 TypeError: can't pickle _thread.RLock objects

您好，这边用了最新的代码之后，使用多卡进行预训练就会报错，主要是出现在mp.spawn那一步，错误信息如下：
Traceback (most recent call last):
File "pretrain.py", line 133, in
main()
File "pretrain.py", line 129, in main
trainer.train_and_validate(args)
File "/data/leo/Projects/uer-py-1/uer/trainer.py", line 56, in train_and_validate
mp.spawn(worker, nprocs=args.ranks_num, args=(args.gpu_ranks, args, model), daemon=False)
File "/home/leo/anaconda3/envs/py36/lib/python3.6/site-packages/torch/multiprocessing/spawn.py", line 200, in spawn
return start_processes(fn, args, nprocs, join, daemon, start_method='spawn')
File "/home/leo/anaconda3/envs/py36/lib/python3.6/site-packages/torch/multiprocessing/spawn.py", line 149, in start_processes
process.start()
File "/home/leo/anaconda3/envs/py36/lib/python3.6/multiprocessing/process.py", line 105, in start
self._popen = self._Popen(self)
File "/home/leo/anaconda3/envs/py36/lib/python3.6/multiprocessing/context.py", line 284, in _Popen
return Popen(process_obj)
File "/home/leo/anaconda3/envs/py36/lib/python3.6/multiprocessing/popen_spawn_posix.py", line 32, in __init__
super().__init__(process_obj)
File "/home/leo/anaconda3/envs/py36/lib/python3.6/multiprocessing/popen_fork.py", line 19, in __init__
self._launch(process_obj)
File "/home/leo/anaconda3/envs/py36/lib/python3.6/multiprocessing/popen_spawn_posix.py", line 47, in _launch
reduction.dump(process_obj, fp)
File "/home/leo/anaconda3/envs/py36/lib/python3.6/multiprocessing/reduction.py", line 60, in dump
ForkingPickler(file, protocol).dump(obj)
TypeError: can't pickle _thread.RLock objects

如果用单卡训练 wordsize=1时，就没有问题。

以下是我的训练脚本：

python pretrain.py \
--dataset_path /data/leo/Projects/uer-py-1/corpora/medical_zh_albert_512.pt \
--vocab_path /data/leo/Projects/UER-py/models/google_zh_vocab.txt \
--pretrained_model_path /data/leo/Projects/UER-py/output_pre/r_512_419/r_512_mlm_from_base_100gpus_110w.bin \
--output_model_path outputs/pretrin/pretrain_r_512_medical_zh_albert_512.bin \
--config_path /data/leo/Projects/uer-py-1/models/bert_base_config.json \
--total_steps 5000000 \
--save_checkpoint_steps 1000 \
--report_steps 100 \
--accumulation_steps 1 \
--batch_size 25 \
--tokenizer bert \
--embedding word_pos_seg \
--encoder transformer \
--whole_word_masking \
--target albert \
--learning_rate 2e-5 \
--warmup 0.1 \
--world_size 3 \
--gpu_ranks 0 1 2 \
--fp16

请帮助解答一下，谢谢！