1，DataParallel

#指定设备
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# cuda
if cfg.cuda:
	model = model.to(device)  # 模型 -> 单卡
    loss = loss.to(device)  # loss -> 单卡
    if torch.cuda.device_count() > 1:
    	device_ids = range(torch.cuda.device_count()) # [0,1,2,3]
        model = torch.nn.DataParallel(model, device_ids=device_ids)
        model = model.to(device)  # 模型 -> 多卡
#  模型加载
model.load_state_dict({k.replace('module.',''):v for k,v in torch.load(checkout).items()})
2，DistributedDataParallel， DDP，仅支持Linux系统
 
from torch.utils.data.distributed import DistributedSampler
# 1) 初始化
torch.distributed.init_process_group(backend="nccl")
# 2） 配置每个进程的gpu
torch.cuda.set_device(local_rank)     # local_rank默认为0
device = torch.device("cuda", local_rank)
# 3）使用DistributedSampler
train_loader = DataLoader(dataset=dataset,
                         batch_size=batch_size,
                         sampler=DistributedSampler(dataset))
# 4) 封装之前要把模型移到对应的gpu
model.to(device)   # cpu -> gpu
if torch.cuda.device_count() > 1:
    print("Let's use", torch.cuda.device_count(), "GPUs!")
    model = torch.nn.parallel.DistributedDataParallel(model, 
                                                      device_ids=[local_rank], 
                                                      output_device=local_rank, 
                                                      broadcast_buffers=False,
                                                      find_unused_parameters=True)
python -m torch.distributed.launch --nproc_per_node=4 train.py  # 执行                                         
3，模型推理
 3.1 设置显卡
 
# -*- coding: utf-8 -*-
import torch
from torchvision import models
def main():
    model = models.resnet50(pretrained=True).eval()
    model = model.cuda(3)
    x = torch.randn(1,3,224,224)
    x = x.cuda(3)
    out = model(x)
    print(out.size())
    print(out.get_device())
if __name__ == '__main__':
    main()
设置显卡二
 
# -*- coding: utf-8 -*-
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "0,1,2,3"
import torch
from torchvision import models
device = torch.device('cuda:3' if torch.cuda.is_available() else 'cpu')
if __name__ == '__main__':
    resnet18 = models.resnet18(pretrained=True)
    resnet18.to(device).eval()
    x = torch.randn(1, 3, 224, 224).to(device)
    print(x.device)
    out = resnet18(x)
    print(out.size())
    print(out.device)
数据在多张卡切换
 
# -*- coding: utf-8 -*-
import os; os.environ["CUDA_VISIBLE_DEVICES"] = "0,1,2,3"
import torch
device0 = torch.device("cuda:0")
device1 = torch.device("cuda:1")
if __name__ == '__main__':
    x = torch.randn(1,3,224,224).to(device0)
    print(x.get_device())
    x = x.cuda(device1)
    print(x.get_device())
3.2 多线程+显卡
				
1. 一机多卡（one matchine multi-GPU）
1.1 DataParallel
DataParallel（DP）：Parameter Server模式，一张卡位reducer，实现也超级简单，一行代码。 有个不能接受的缺陷是：DataParallel是基于Parameter server的算法，所有的loss都在主卡上计算，负载不均衡的问题比较严重，有时在模型较大的时候（比如bert-large），主卡占满了，其他的卡一半都占不到，及其浪费资源。
值得注意的是，模型和数据都需要先 load