PyTorch多GPU训练时同步梯度是mean还是sum?

PyTorch 通过两种方式可以进行多GPU训练: DataParallel, DistributedDataParallel. 当使用DataParallel的时候, 梯度的计算结果和在单卡上跑是一样的, 对每个数据计算出来的梯度进行累加. 当使用DistributedDataParallel的时候, 每个卡单独计算梯度, 然后多卡的梯度再进行平均.

下面是实验验证:

DataParallel

python 复制代码
import torch
import os
import torch.nn as nn

def main():
    model = nn.Linear(2, 3).cuda()
    model = torch.nn.DataParallel(model, device_ids=[0, 1])
    input = torch.rand(2, 2)
    labels = torch.tensor([[1, 0, 0], [0, 1, 0]]).cuda()
    (model(input) * labels).sum().backward()
    print('input', input)
    print([p.grad for p in model.parameters()])


if __name__=="__main__":
    main()

执行CUDA_VISIBLE_DEVICES=0,1 python t.py可以看到输出, 代码中对两个样本分别求梯度, 梯度等于样本的值, DataParallel把两个样本的梯度累加起来在不同GPU中同步.

bash 复制代码
input tensor([[0.4362, 0.4574],
        [0.2052, 0.2362]])
[tensor([[0.4363, 0.4573],
        [0.2052, 0.2362],
        [0.0000, 0.0000]], device='cuda:0'), tensor([1., 1., 0.], device='cuda:0')]

DistributedDataParallel

python 复制代码
import torch
import os
import torch.distributed as dist
import torch.multiprocessing as mp
import torch.nn as nn
import torch.optim as optim
from torch.nn.parallel import DistributedDataParallel as DDP


def example(rank, world_size):
    # create default process group
    dist.init_process_group("gloo", rank=rank, world_size=world_size)
    # create local model
    model = nn.Linear(2, 3).to(rank)
    print('model param', 'rank', rank, [p for p in model.parameters()])
    # construct DDP model
    ddp_model = DDP(model, device_ids=[rank])
    print('ddp model param', 'rank', rank, [p for p in ddp_model.parameters()])
    # forward pass
    input = torch.randn(1, 2).to(rank)
    outputs = ddp_model(input)
    labels = torch.randn(1, 3).to(rank) * 0
    labels[0, rank] = 1
    # backward pass
    (outputs * labels).sum().backward()
    print('rank', rank, 'grad', [p.grad for p in ddp_model.parameters()])
    print('rank', rank, 'input', input, 'outputs', outputs)
    print('rank', rank, 'labels', labels)
    # update parameters
    optimizer.step()

def main():
    world_size = 2
    mp.spawn(example,
        args=(world_size,),
        nprocs=world_size,
        join=True)

if __name__=="__main__":
    # Environment variables which need to be
    # set when using c10d's default "env"
    # initialization mode.
    os.environ["MASTER_ADDR"] = "localhost"
    os.environ["MASTER_PORT"] = "29504"
    main()

执行CUDA_VISIBLE_DEVICES=0,1 python t1.py可以看到输出, 代码中对两个样本分别求梯度, 梯度等于样本的值, 最终的梯度是各个GPU的梯度的平均.

bash 复制代码
model param rank 0 [Parameter containing:
tensor([[-0.4819,  0.0253],
        [ 0.0858,  0.2256],
        [ 0.5614,  0.2702]], device='cuda:0', requires_grad=True), Parameter containing:
tensor([-0.0090,  0.4461, -0.3493], device='cuda:0', requires_grad=True)]
model param rank 1 [Parameter containing:
tensor([[-0.3737,  0.3062],
        [ 0.6450,  0.2930],
        [-0.2422,  0.2089]], device='cuda:1', requires_grad=True), Parameter containing:
tensor([-0.5868,  0.2106, -0.4461], device='cuda:1', requires_grad=True)]
ddp model param rank 1 [Parameter containing:
tensor([[-0.4819,  0.0253],
        [ 0.0858,  0.2256],
        [ 0.5614,  0.2702]], device='cuda:1', requires_grad=True), Parameter containing:
tensor([-0.0090,  0.4461, -0.3493], device='cuda:1', requires_grad=True)]
ddp model param rank 0 [Parameter containing:
tensor([[-0.4819,  0.0253],
        [ 0.0858,  0.2256],
        [ 0.5614,  0.2702]], device='cuda:0', requires_grad=True), Parameter containing:
tensor([-0.0090,  0.4461, -0.3493], device='cuda:0', requires_grad=True)]
rank 1 grad [tensor([[ 0.2605,  0.1631],
        [-0.0934, -0.5308],
        [ 0.0000,  0.0000]], device='cuda:1'), tensor([0.5000, 0.5000, 0.0000], device='cuda:1')]
rank 0 grad [tensor([[ 0.2605,  0.1631],
        [-0.0934, -0.5308],
        [ 0.0000,  0.0000]], device='cuda:0'), tensor([0.5000, 0.5000, 0.0000], device='cuda:0')]
rank 1 input tensor([[-0.1868, -1.0617]], device='cuda:1') outputs tensor([[ 0.0542,  0.1906, -0.7411]], device='cuda:1',
       grad_fn=<AddmmBackward0>)
rank 0 input tensor([[0.5209, 0.3261]], device='cuda:0') outputs tensor([[-0.2518,  0.5644,  0.0314]], device='cuda:0',
       grad_fn=<AddmmBackward0>)
rank 1 labels tensor([[-0., 1., -0.]], device='cuda:1')
rank 0 labels tensor([[1., 0., -0.]], device='cuda:0')
相关推荐
chh5634 小时前
从 BERT 到大模型:让模型从“会接龙”到“会回答
人工智能·深度学习·bert
TomEval4 小时前
【测AI】第26篇:线上模型监控与告警体系 — 测试不是终点
人工智能·python·功能测试·ai
ZhengEnCi5 小时前
AI 总模仿代码库里的过时写法怎么办:先清坏样例,再用类型与测试让 AI 无法犯错
人工智能·ai编程
一只爱撸猫的程序猿5 小时前
五周、两个开源仓库、九次拦截之后:AI 时代程序员缺的到底是什么
人工智能·后端·开源
GitCode官方6 小时前
开源鸿蒙跨平台直播|AI 时代,用 cj-tauri 开发鸿蒙 PC 应用
人工智能·华为·harmonyos·atomgit
老金带你玩AI6 小时前
GPT-6.2更强更便宜了,为什么还有人骂?
人工智能
CyrusYiVPP6 小时前
实测 uv 0.13.7:同一套包,pip 装 94 秒,它只用 19 秒
人工智能·python
泯泷7 小时前
OpenClaw 退潮背后,是 AI 圈“技术自嗨”的终结
人工智能·chatgpt·llm
罗西的思考7 小时前
【轻量化复刻大模型 / KIMI K3】Muon 优化器及 训练系统
人工智能·算法·机器学习
屿华AI7 小时前
AUTOSAR 实战-诊断工具 | 用 Claude Code 开发诊断测试工具:AI 协作开发实践
人工智能·测试工具