tczhangzhi / pytorch-distributed

A quickstart and benchmark for pytorch distributed training.

tczhangzhi/pytorch-distributed Issues

Connections time out
Updated 7 months ago
RuntimeError: Timed out initializing process group in store based barrier on rank: 2, for key: store_based_barrier_key:1 (world_size=2, worker_count=4, timeout=0:30:00)
Updated a year ago2
Can't pickle <function main_worker
Updated 2 years ago
Windows 提示Distributed package doesn't have NCCL "Distributed package doesn't have NCCL built in
Updated 2 years ago1
how can i modify it ?
Updated 3 years ago
how can i modify it
Updated 3 years ago
多卡显存占用不均衡
Updated 3 years ago1
No need to divide with nprocs again
Closed 3 years ago1
关于损失backward问题
Updated 3 years ago1
您好，我想请问一下Pytorch 在多进程分布式训练的时候，一开始加载数据集dataloader的时候，由于是多进程的，就会导致多个进程一起加载dataloader，这些dataloder全部加载到内存中使得内存爆炸，请问有没有什么办法解决呢
Updated 3 years ago
apex 并行加速中的data_prefetcher normalize为什么做两次？
Closed 3 years ago1
大佬，请问如何指定gpu训练
Updated 3 years ago
请问使用DistributedSampler，各个GPU的数据是如何分配的？是连续(互不相同)的还是相同的？
Updated 3 years ago1
使用mp.spawn启动比torch.distributed.launch慢很多
Closed 4 years ago2
distributed.py seed
Closed 4 years ago6
About saving model in the multi-GPU training.
Closed 4 years ago3
DDP掉精度
Closed 4 years ago3
[Discussion] mp: duplicate of torch.cuda.set_device(local_rank) and images = images.cuda(local_rank, non_blocking=True)
Updated 4 years ago4
2080Ti多卡训练，GPU利用率忽高忽低，这个怎么回事
Closed 4 years ago2
Does the test accuracy need to be synchronized in distributed.py?
Closed 4 years ago7