本文旨在帮助Pytorch使用者快速上手使用寒武纪MLU。以代码块为主，文字尽可能简洁，许多部分对标NVIDIA CUDA。不正确的地方请留言更正。本文不定期更新。

---

前言

大背景：信创改造、信创国产化、GPU国产化。

为使PyTorch支持寒武纪MLU，寒武纪对机器学习框架PyTorch进行了部分定制。若要在寒武纪MLU上运行PyTorch，需要安装并使用寒武纪定制的 Cambricon PyTorch。

---

Cambricon PyTorch的Python包torch_mlu导入

Cambricon CATCH是寒武纪发布的一款Python包（包名torch_mlu），提供了在MLU设备上进行张量计算的能力。安装好Cambricon CATCH后，便可使用torch_mlu模块：

import torch # 需安装Cambricon PyTorch
import torch_mlu # 动态扩展MLU后端

附 Cambricon PyTorch源码编译安装

导入 torch 和 torch_mlu 后可以测试在MLU上完成加法运算：

t0 = torch.randn(2, 2, device='mlu') # 在MLU设备上生成Tensor
t1 = torch.randn(2, 2, device='mlu')
result = t0 + t1 # 在MLU设备上完成加法运算

将模型加载到MLU上model.to(‘mlu’)

以ResNet18为例，将模型加载到MLU上用 model.to('mlu')，对标cuda的 model.to(device) ：

# 定义模型
model = models.__dict__["resnet50"]()
# 将模型加载到MLU上。
mlu_model = model.to('mlu')

定义损失函数，然后将其拷贝至MLU

# 构造损失函数
criterion = nn.CrossEntropyLoss()
# 将损失函数拷贝到MLU上
criterion.to('mlu')

将数据从CPU拷贝到MLU设备

x = torch.randn(1000000, dtype=torch.float)
x_mlu = x.to(torch.device('mlu'), non_blocking=True)

以mnist.py为例的训练代码demo

import torch # 导入原生 PyTorch
import torch_mlu # 导入 Cambricon PyTorch
from torch.utils.data import DataLoader
from torchvision.datasets import mnist
from torch import nn
from torch import optim
from torchvision import transforms
from torch.optim.lr_scheduler import StepLR
import torch.nn.functional as F# 定义模型
class Net(nn.Module):def __init__(self):super(Net, self).__init__()self.conv1 = nn.Conv2d(1, 32, 3, 1)self.conv2 = nn.Conv2d(32, 64, 3, 1)self.dropout1 = nn.Dropout2d(0.25)self.dropout2 = nn.Dropout2d(0.5)self.fc1 = nn.Linear(9216, 128)self.fc2 = nn.Linear(128, 10)# 定义前向计算def forward(self, x):x = self.conv1(x)x = F.relu(x)x = self.conv2(x)x = F.relu(x)x = F.max_pool2d(x, 2)x = self.dropout1(x)x = torch.flatten(x, 1)x = self.fc1(x)x = F.relu(x)x = self.dropout2(x)x = self.fc2(x)output = F.log_softmax(x, dim=1)return output# 模型训练
def train(model, train_data, optimizer, epoch):model = model.train()for batch_idx, (img, label) in enumerate(train_data):img = img.mlu()label = label.mlu()optimizer.zero_grad()out = model(img)loss = F.nll_loss(out, label)# 反向计算loss.backward()# 梯度更新optimizer.step()if batch_idx % 100 == 0:print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(epoch, batch_idx * len(img), len(train_data.dataset),100. * batch_idx / len(train_data), loss.item()))# 模型推理
def validate(val_loader, model):test_loss = 0correct = 0model.eval()with torch.no_grad():for images, target in val_loader:images = images.mlu()target = target.mlu()output = model(images)test_loss += F.nll_loss(output, target, reduction='sum').item()pred = output.argmax(dim=1, keepdim=True)correct += pred.eq(target.view_as(pred)).sum().item()test_loss /= len(val_loader.dataset)# 打印精度结果print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(test_loss, correct, len(val_loader.dataset),100. * correct / len(val_loader.dataset)))# 主函数
def main():# 定义预处理函数data_tf = transforms.Compose([transforms.ToTensor(),transforms.Normalize([0.1307],[0.3081])])# 获取 MNIST 数据集train_set = mnist.MNIST('./data', train=True, transform=data_tf, download=True)test_set = mnist.MNIST('./data', train=False, transform=data_tf, download=True)train_data = DataLoader(train_set, batch_size=64, shuffle=True)test_data = DataLoader(test_set, batch_size=1000, shuffle=False)net_orig = Net()# 模型拷贝到MLU设备net = net_orig.mlu()optimizer = optim.Adadelta(net.parameters(), 1)# 训练10个epochnums_epoch = 10# 训练完成后保存模型save_model = True# 学习率调整策略scheduler = StepLR(optimizer, step_size=1, gamma=0.7)for epoch in range(nums_epoch):train(net, train_data, optimizer, epoch)validate(test_data, net)scheduler.step()if save_model:  # 将训练好的模型保存为model.pthif epoch == nums_epoch-1:checkpoint = {"state_dict":net.state_dict(), "optimizer":optimizer.state_dict(), "epoch": epoch}torch.save(checkpoint, 'model.pth')if __name__ == '__main__':main()

---

参考引用

寒武纪PyTorch v1.13.1用户手册