PyTorch——利用Accelerate轻松控制多个CPU/GPU/TPU加速计算

前言

• CPU？GPU？TPU？
  • 计算设备太多，很混乱？
  • 切换环境，代码大量改来改去？
  • 不懂怎么调用多个CPU/GPU/TPU？或者想轻松调用？
• OK！OK！OK！
  • 来自HuggingFace的Accelerate库帮你轻松解决这些问题，只需几行代码改动就可以快速完成计算设备的自动调整。
    
• 相关地址
  • 官方文档：https://huggingface.co/docs/accelerate/index
  • GitHub：https://github.com/huggingface/accelerate
  • 安装（推荐用>=0.14的版本） $ pip install accelerate
• 下面就来说说怎么用
  • 你也可以直接看我在Kaggle上做好的完整的Notebook示例

官方示例

• 先大致看个样
• 移除掉以前.to(device)部分的代码，引入Accelerator对model、optimizer、data、loss.backward()做下处理即可

import torch
import torch.nn.functional as F
from datasets import load_dataset
from accelerate import Accelerator# device = 'cpu'
accelerator = Accelerator()# model = torch.nn.Transformer().to(device)
model = torch.nn.Transformer()
optimizer = torch.optim.Adam(model.parameters())dataset = load_dataset('my_dataset')
data = torch.utils.data.DataLoader(dataset, shuffle=True)model, optimizer, data = accelerator.prepare(model, optimizer, data)model.train()
for epoch in range(10):for source, targets in data:# source = source.to(device)# targets = targets.to(device)optimizer.zero_grad()output = model(source)loss = F.cross_entropy(output, targets)# loss.backward()accelerator.backward(loss)optimizer.step()

单个程序内控制多个CPU/GPU/TPU

• 详细内容请参考官方Example

简单说一下

• 对于单个计算设备，像前面那个简单示例改下代码即可
• 多个计算设备（例如GPU）的情况下，有一点特殊的要处理，下面做个完整的PyTorch训练示例
  • 你可以拿这个和我之前发的示例做个对比 CNN图像分类-FashionMNIST
  • 也可以直接看我在Kaggle上做好的完整的Notebook示例

设备环境

• 看看当前的显卡设备（2颗Tesla T4），命令 $ nvidia-smi

Thu Apr 27 10:53:26 2023       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 470.161.03   Driver Version: 470.161.03   CUDA Version: 11.4     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla T4            Off  | 00000000:00:04.0 Off |                    0 |
| N/A   43C    P8     9W /  70W |      0MiB / 15109MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   1  Tesla T4            Off  | 00000000:00:05.0 Off |                    0 |
| N/A   41C    P8     9W /  70W |      0MiB / 15109MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------++-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

• 安装或更新Accelerate，命令 $ !pip install --upgrade accelerate

导包

import torch
from torch import nn
from torch.utils.data import DataLoader
from torchvision.transforms import ToTensor, Compose
import torchvision.datasets as datasets
from accelerate import Accelerator
from accelerate import notebook_launcher

加载数据 FashionMNIST

train_data = datasets.FashionMNIST(root="./data",train=True,download=True,transform=Compose([ToTensor()])
)test_data = datasets.FashionMNIST(root="./data",train=False,download=True,transform=Compose([ToTensor()])
)print(train_data.data.shape)
print(test_data.data.shape)

创建一个简单的CNN模型

class CNNModel(nn.Module):def __init__(self):super(CNNModel, self).__init__()self.module1 = nn.Sequential(nn.Conv2d(1, 32, kernel_size=5, stride=1, padding=2),nn.BatchNorm2d(32),nn.ReLU(),nn.MaxPool2d(kernel_size=2, stride=2))  self.module2 = nn.Sequential(nn.Conv2d(32, 64, kernel_size=5, stride=1, padding=2),nn.BatchNorm2d(64),nn.ReLU(),nn.MaxPool2d(kernel_size=2, stride=2))self.flatten = nn.Flatten()self.linear1 = nn.Linear(7 * 7 * 64, 64)self.linear2 = nn.Linear(64, 10)self.relu = nn.ReLU()def forward(self, x):out = self.module1(x)out = self.module2(out)out = self.flatten(out)out = self.linear1(out)out = self.relu(out)out = self.linear2(out)return out

训练函数-只包含训练

• 注意看accelerator相关代码
• 若要实现多设备控制训练，for epoch in range(epoch_num):中末尾处的代码必不可少

def training_function():# 参数配置epoch_num = 4batch_size = 64learning_rate = 0.005# device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')# 数据train_loader = DataLoader(dataset=train_data, batch_size=batch_size, shuffle=True)val_loader = DataLoader(test_data, batch_size=batch_size, shuffle=True)# 模型/损失函数/优化器# model = CNNModel().to(device)model = CNNModel()criterion = nn.CrossEntropyLoss()optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)accelerator = Accelerator()model, optimizer, train_loader, val_loader = accelerator.prepare(model, optimizer, train_loader, val_loader)# 开始训练for epoch in range(epoch_num):# 训练model.train()for i, (X_train, y_train) in enumerate(train_loader):# X_train = X_train.to(device)# y_train = y_train.to(device)out = model(X_train)loss = criterion(out, y_train)optimizer.zero_grad()# loss.backward()accelerator.backward(loss)optimizer.step()if (i + 1) % 100 == 0:print(f"{accelerator.device} Train... [epoch {epoch + 1}/{epoch_num}, step {i + 1}/{len(train_loader)}]\t[loss {loss.item()}]")# 等待每个GPU上的模型执行完当前的epoch，并进行合并同步accelerator.wait_for_everyone() model = accelerator.unwrap_model(model)# 现在所有GPU上都一样了，可以保存modelaccelerator.save(model, "model.pth") 

训练函数-包含训练和验证

• 相比前面的代码，多了“验证”相关的代码
• 验证时，因为使用多个设备进行训练，所以会比较特殊，会涉及到多个设备的验证结果合并的问题

def training_function():# 参数配置epoch_num = 4batch_size = 64learning_rate = 0.005# 数据train_loader = DataLoader(dataset=train_data, batch_size=batch_size, shuffle=True)val_loader = DataLoader(test_data, batch_size=batch_size, shuffle=True)# 模型/损失函数/优化器model = CNNModel()criterion = nn.CrossEntropyLoss()optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)accelerator = Accelerator()model, optimizer, train_loader, val_loader = accelerator.prepare(model, optimizer, train_loader, val_loader)# 开始训练for epoch in range(epoch_num):# 训练model.train()for i, (X_train, y_train) in enumerate(train_loader):out = model(X_train)loss = criterion(out, y_train)optimizer.zero_grad()accelerator.backward(loss)optimizer.step()if (i + 1) % 100 == 0:print(f"{accelerator.device} Train... [epoch {epoch + 1}/{epoch_num}, step {i + 1}/{len(train_loader)}]\t[loss {loss.item()}]")# 验证model.eval()correct, total = 0, 0for X_val, y_val in val_loader:with torch.no_grad():output = model(X_val)_, pred = torch.max(output, 1)# 合并每个GPU的验证数据pred, y_val = accelerator.gather_for_metrics((pred, y_val))total += y_val.size(0)correct += (pred == y_val).sum()# 用main process打印accuracyaccelerator.print(f'epoch {epoch + 1}/{epoch_num}, accuracy = {100 * (correct.item() / total):.2f}')# 等待每个GPU上的模型执行完当前的epoch，并进行合并同步accelerator.wait_for_everyone() model = accelerator.unwrap_model(model)# 现在所有GPU上都一样了，可以保存modelaccelerator.save(model, "model.pth") 

训练

• 如果你在本地训练的话，直接调用前面定义的函数training_function即可。最后在命令行启动训练脚本 $ accelerate launch example.py。

training_function()

• 如果你在Kaggle/Colab上面，则需要利用notebook_launcher进行训练

# num_processes=2 指定使用2个GPU，因为当前我申请了2颗 Nvidia T4
notebook_launcher(training_function, num_processes=2)

• 下面是2个GPU训练时的控制台输出样例

Launching training on 2 GPUs.
cuda:0 Train... [epoch 1/4, step 100/469]	[loss 0.43843933939933777]
cuda:1 Train... [epoch 1/4, step 100/469]	[loss 0.5267877578735352]
cuda:0 Train... [epoch 1/4, step 200/469]	[loss 0.39918822050094604]cuda:1 Train... [epoch 1/4, step 200/469]	[loss 0.2748252749443054]cuda:1 Train... [epoch 1/4, step 300/469]	[loss 0.54105544090271]cuda:0 Train... [epoch 1/4, step 300/469]	[loss 0.34716445207595825]cuda:1 Train... [epoch 1/4, step 400/469]	[loss 0.2694844901561737]
cuda:0 Train... [epoch 1/4, step 400/469]	[loss 0.4343942701816559]
epoch 1/4, accuracy = 88.49
cuda:0 Train... [epoch 2/4, step 100/469]	[loss 0.19695354998111725]
cuda:1 Train... [epoch 2/4, step 100/469]	[loss 0.2911057770252228]
cuda:0 Train... [epoch 2/4, step 200/469]	[loss 0.2948791980743408]
cuda:1 Train... [epoch 2/4, step 200/469]	[loss 0.292676717042923]
cuda:0 Train... [epoch 2/4, step 300/469]	[loss 0.222089946269989]
cuda:1 Train... [epoch 2/4, step 300/469]	[loss 0.28814008831977844]
cuda:0 Train... [epoch 2/4, step 400/469]	[loss 0.3431250751018524]
cuda:1 Train... [epoch 2/4, step 400/469]	[loss 0.2546379864215851]
epoch 2/4, accuracy = 87.31
cuda:1 Train... [epoch 3/4, step 100/469]	[loss 0.24118559062480927]cuda:0 Train... [epoch 3/4, step 100/469]	[loss 0.363821804523468]cuda:0 Train... [epoch 3/4, step 200/469]	[loss 0.36783623695373535]
cuda:1 Train... [epoch 3/4, step 200/469]	[loss 0.18346744775772095]
cuda:0 Train... [epoch 3/4, step 300/469]	[loss 0.23459288477897644]
cuda:1 Train... [epoch 3/4, step 300/469]	[loss 0.2887689769268036]
cuda:0 Train... [epoch 3/4, step 400/469]	[loss 0.3079166114330292]
cuda:1 Train... [epoch 3/4, step 400/469]	[loss 0.18255220353603363]
epoch 3/4, accuracy = 88.46
cuda:1 Train... [epoch 4/4, step 100/469]	[loss 0.27428603172302246]
cuda:0 Train... [epoch 4/4, step 100/469]	[loss 0.17705145478248596]
cuda:1 Train... [epoch 4/4, step 200/469]	[loss 0.2811894416809082]
cuda:0 Train... [epoch 4/4, step 200/469]	[loss 0.22682836651802063]
cuda:0 Train... [epoch 4/4, step 300/469]	[loss 0.2291710525751114]
cuda:1 Train... [epoch 4/4, step 300/469]	[loss 0.32024848461151123]
cuda:0 Train... [epoch 4/4, step 400/469]	[loss 0.24648766219615936]
cuda:1 Train... [epoch 4/4, step 400/469]	[loss 0.0805584192276001]
epoch 4/4, accuracy = 89.38

• 下面是1个TPU训练时的控制台输出样例

Launching training on CPU.
xla:0 Train... [epoch 1/4, step 100/938]	[loss 0.6051161289215088]
xla:0 Train... [epoch 1/4, step 200/938]	[loss 0.27442359924316406]
xla:0 Train... [epoch 1/4, step 300/938]	[loss 0.557417631149292]
xla:0 Train... [epoch 1/4, step 400/938]	[loss 0.1840067058801651]
xla:0 Train... [epoch 1/4, step 500/938]	[loss 0.5252436399459839]
xla:0 Train... [epoch 1/4, step 600/938]	[loss 0.2718536853790283]
xla:0 Train... [epoch 1/4, step 700/938]	[loss 0.2763175368309021]
xla:0 Train... [epoch 1/4, step 800/938]	[loss 0.39897507429122925]
xla:0 Train... [epoch 1/4, step 900/938]	[loss 0.28720396757125854]
epoch = 0, accuracy = 86.36
xla:0 Train... [epoch 2/4, step 100/938]	[loss 0.24496735632419586]
xla:0 Train... [epoch 2/4, step 200/938]	[loss 0.37713131308555603]
xla:0 Train... [epoch 2/4, step 300/938]	[loss 0.3106330633163452]
xla:0 Train... [epoch 2/4, step 400/938]	[loss 0.40438592433929443]
xla:0 Train... [epoch 2/4, step 500/938]	[loss 0.38303741812705994]
xla:0 Train... [epoch 2/4, step 600/938]	[loss 0.39199298620224]
xla:0 Train... [epoch 2/4, step 700/938]	[loss 0.38932573795318604]
xla:0 Train... [epoch 2/4, step 800/938]	[loss 0.26298171281814575]
xla:0 Train... [epoch 2/4, step 900/938]	[loss 0.21517205238342285]
epoch = 1, accuracy = 90.07
xla:0 Train... [epoch 3/4, step 100/938]	[loss 0.366019606590271]
xla:0 Train... [epoch 3/4, step 200/938]	[loss 0.27360212802886963]
xla:0 Train... [epoch 3/4, step 300/938]	[loss 0.2014923095703125]
xla:0 Train... [epoch 3/4, step 400/938]	[loss 0.21998485922813416]
xla:0 Train... [epoch 3/4, step 500/938]	[loss 0.28129786252975464]
xla:0 Train... [epoch 3/4, step 600/938]	[loss 0.42534705996513367]
xla:0 Train... [epoch 3/4, step 700/938]	[loss 0.22158119082450867]
xla:0 Train... [epoch 3/4, step 800/938]	[loss 0.359947144985199]
xla:0 Train... [epoch 3/4, step 900/938]	[loss 0.3221997022628784]
epoch = 2, accuracy = 90.36
xla:0 Train... [epoch 4/4, step 100/938]	[loss 0.2814193069934845]
xla:0 Train... [epoch 4/4, step 200/938]	[loss 0.16465164721012115]
xla:0 Train... [epoch 4/4, step 300/938]	[loss 0.2897304892539978]
xla:0 Train... [epoch 4/4, step 400/938]	[loss 0.13403896987438202]
xla:0 Train... [epoch 4/4, step 500/938]	[loss 0.1135573536157608]
xla:0 Train... [epoch 4/4, step 600/938]	[loss 0.14964193105697632]
xla:0 Train... [epoch 4/4, step 700/938]	[loss 0.20239461958408356]
xla:0 Train... [epoch 4/4, step 800/938]	[loss 0.23625142872333527]
xla:0 Train... [epoch 4/4, step 900/938]	[loss 0.3418393135070801]
epoch = 3, accuracy = 90.11

多个服务器、多个程序间控制多个CPU/GPU/TPU

• 详细内容请参考官方Example
• 包括
  • 单服务器内，多个程序控制多个计算设备
  • 多个服务器间，多个程序控制多个计算设备
• 写好代码后，请先在每个服务器下执行$ accelerate config生成对应的配置文件，下面是个样例

(huggingface) PS C:\Users\alion\temp> accelerate config
------------------------------------------------------------------------------------------------------------------------In which compute environment are you running?
This machine
------------------------------------------------------------------------------------------------------------------------Which type of machine are you using?
multi-GPU
How many different machines will you use (use more than 1 for multi-node training)? [1]: 2
------------------------------------------------------------------------------------------------------------------------What is the rank of this machine?
0
What is the IP address of the machine that will host the main process? 192.168.101
What is the port you will use to communicate with the main process? 12345
Are all the machines on the same local network? Answer `no` if nodes are on the cloud and/or on different network hosts [YES/no]: yes
Do you wish to optimize your script with torch dynamo?[yes/NO]:no
Do you want to use DeepSpeed? [yes/NO]: no
Do you want to use FullyShardedDataParallel? [yes/NO]: no
Do you want to use Megatron-LM ? [yes/NO]: no
How many GPU(s) should be used for distributed training? [1]:2
What GPU(s) (by id) should be used for training on this machine as a comma-seperated list? [all]:0
------------------------------------------------------------------------------------------------------------------------Do you wish to use FP16 or BF16 (mixed precision)?
fp16
accelerate configuration saved at C:\Users\alion/.cache\huggingface\accelerate\default_config.yaml

• 最后在每个服务器启动训练脚本 $ accelerate launch example.py（如果你是单台服务器多个程序，那就只启动一台的脚本就完了）

参考链接

• https://github.com/huggingface/accelerate
• https://www.kaggle.com/code/muellerzr/multi-gpu-and-accelerate
• https://github.com/huggingface/notebooks/blob/main/examples/accelerate_examples/simple_nlp_example.ipynb
• https://github.com/huggingface/accelerate/tree/main/examples