distributed-training

name: distributed-training description: Multi-GPU and distributed training patterns with PyTorch DDP. Use when scaling training across GPUs. metadata: category: tooling trigger-keywords: "distributed,multi-gpu,parallel,ddp,scale" applicable-stages: "10,12" priority: "7" version: "1.0" author: researchclaw references: "PyTorch DDP Tutorial, pytorch.org; Goyal et al., Accurate Large Minibatch SGD, 2017"

Distributed Training Best Practice

1. Use DistributedDataParallel (DDP) over DataParallel for multi-GPU
2. Initialize process group: dist.init_process_group(backend='nccl')
3. Use DistributedSampler for data sharding
4. Synchronize batch norm: nn.SyncBatchNorm.convert_sync_batchnorm()
5. Only save checkpoint on rank 0
6. Scale learning rate linearly with world size
7. Use gradient accumulation for effectively larger batch sizes