intro：

        想要看一下在DDP中pytorch是怎么调用torch.distributed的通信后端的，所以写了一个猴子补丁。

训练脚本：

from torchvision.datasets import MNIST
from torchvision.transforms import ToTensor
from torch import nn
import torch
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.utils.data.distributed import DistributedSampler
from torch.utils.data import DataLoader
import torch.nn.functional as F
import os
import distributed_patch

# 设置 NCCL 日志环境变量
'''
os.environ["NCCL_DEBUG"] = "INFO"
os.environ["NCCL_DEBUG_SUBSYS"] = "ALL"  # 或者 COLL
os.environ["NCCL_LOG_FILE"] = "nccl_log.txt"

# 运行 PyTorch 分布式代码
'''




class Net(nn.Module):  # 模型定义
    def __init__(self):
        super(Net, self).__init__()
        self.flatten = nn.Flatten()
        self.seq = nn.Sequential(
            nn.Linear(28 * 28, 128),
            nn.ReLU(),
            nn.Linear(128, 64),
            nn.ReLU(),
            nn.Linear(64, 10)
        )

    def forward(self, x):
        x = self.flatten(x)
        return self.seq(x)


def main():
    dist.init_process_group(backend='nccl')  # 【集合通讯】其他进程连master，大家互认

    rank = dist.get_rank()
    world_size = dist.get_world_size()
    device_name = f'cuda:{rank}'

    checkpoint = None  # 各自加载checkpoint
    try:
        checkpoint = torch.load('checkpoint.pth', map_location='cpu')  # checkpoint是cuda:0保存的，加载默认会读到cuda:0，所以明确指定给cpu
    except:
        pass

    model = Net().to(device_name)
    if checkpoint and rank == 0:  # rank0恢复模型参数
        model.load_state_dict(checkpoint['model'])

    model = DDP(model)  # 【集合通讯】rank0广播参数给其他进程

    optimizer = torch.optim.Adam(model.parameters(), lr=0.001)  # model参数一致，则optim会保证其初始状态一致
    if checkpoint:
        optimizer.load_state_dict(checkpoint['optimizer'])  # 各自加载checkpoint

    train_dataset = MNIST(root='./data', download=True, transform=ToTensor(), train=True)  # 各自加载dataset
    sampler = DistributedSampler(train_dataset)  # 指派子集给各进程
    train_dataloader = DataLoader(train_dataset, batch_size=32, sampler=sampler, persistent_workers=True, num_workers=2)

    val_dataset = MNIST(root='./data', download=True, transform=ToTensor(), train=False)
    val_dataloader = DataLoader(val_dataset, batch_size=32, shuffle=True, persistent_workers=True, num_workers=2)

    for epoch in range(20):
        sampler.set_epoch(epoch)  # 【集合通讯】生成随机种子，rank0广播给其他进程

        model.train()
        for x, y in train_dataloader:
            x, y = x.to(device_name), y.to(device_name)
            pred_y = model(x)  # 【集合通讯】rank0广播model buffer给其他进程
            loss = F.cross_entropy(pred_y, y)
            optimizer.zero_grad()
            loss.backward()  # 【集合通讯】每个参数的梯度做all reduce（每个进程会收到其他进程的梯度，并求平均）
            optimizer.step()

        dist.reduce(loss, dst=0)  # 【集合通讯】rank0汇总其他进程的loss

        if rank == 0:
            train_avg_loss = loss.item() / world_size

            # evaluate
            raw_model = model.module
            val_loss = 0
            with torch.no_grad():
                for x, y in val_dataloader:
                    x, y = x.to(device_name), y.to(device_name)
                    pred_y = raw_model(x)
                    loss = F.cross_entropy(pred_y, y)
                    val_loss += loss.item()
            val_avg_loss = val_loss / len(val_dataloader)
            print(f'train_loss:{train_avg_loss} val_loss:{val_avg_loss}')

            # checkpoint
            torch.save({'model': model.module.state_dict(), 'optimizer': optimizer.state_dict()}, '.checkpoint.pth')
            os.replace('.checkpoint.pth', 'checkpoint.pth')

        dist.barrier()  # 【集合通讯】等待rank0跑完eval



if __name__ == '__main__':
    main()

# torchrun --nproc_per_node 1 pytorch_dis_gpu.py

插桩脚本：

import torch.distributed as dist

# 保存原始函数引用
original_functions = {
    "init_process_group": dist.init_process_group,
    "all_reduce": dist.all_reduce,
    "reduce": dist.reduce,
    "broadcast": dist.broadcast,
    "barrier": dist.barrier,
    "get_rank": dist.get_rank,
    "get_world_size": dist.get_world_size
}

# 插桩函数
def patched_init_process_group(*args, **kwargs):
    print("[distributed] init_process_group called")
    return original_functions["init_process_group"](*args, **kwargs)

def patched_all_reduce(tensor, op=dist.ReduceOp.SUM, group=None, async_op=False):
    print("[distributed] all_reduce called")
    return original_functions["all_reduce"](tensor, op, group, async_op)

def patched_reduce(tensor, dst, op=dist.ReduceOp.SUM, group=None, async_op=False):
    print("[distributed] reduce called")
    return original_functions["reduce"](tensor, dst, op, group, async_op)

def patched_broadcast(tensor, src, group=None, async_op=False):
    print("[distributed] broadcast called")
    return original_functions["broadcast"](tensor, src, group, async_op)

def patched_barrier(*args, **kwargs):
    print("[distributed] barrier called")
    return original_functions["barrier"](*args, **kwargs)

def patched_get_rank(*args, **kwargs):
    print("[distributed] get_rank called")
    return original_functions["get_rank"](*args, **kwargs)

def patched_get_world_size(*args, **kwargs):
    print("[distributed] get_world_size called")
    return original_functions["get_world_size"](*args, **kwargs)

# 替换分布式接口函数为插桩版本
dist.init_process_group = patched_init_process_group
dist.all_reduce = patched_all_reduce
dist.reduce = patched_reduce
dist.broadcast = patched_broadcast
dist.barrier = patched_barrier
dist.get_rank = patched_get_rank
dist.get_world_size = patched_get_world_size

问题：

发现实际上并没有像想象中那样调用Broadcast！这是一直以来理解中的误区！