PaddlePaddle

https://www.paddlepaddle.org.cn/documentation/docs/zh/
Admin
PaddlePaddle is an open-source deep learning framework designed to be easy-to-use, efficient,...
Tokens:1,376,406
Snippets:8,060
Trust Score:9.6
Update:7 months ago
Show doc for...
Tokens:
Raw
### Launch Distributed Training Job with Fleetrun

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/06_distributed_training/cluster_quick_start_ps_cn

This command-line instruction shows how to launch a distributed training job using the 'fleetrun' utility provided by PaddlePaddle. It specifies the number of server and trainer nodes required for the distributed setup. The example command initiates a training script named 'train.py' with one server node and two trainer nodes.

```bash
fleetrun --server_num=1 --trainer_num=2 train.py

```

--------------------------------

### Initialize Distributed Training Environment (Parameter Server Mode)

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/06_distributed_training/cluster_quick_start_ps_cn

Initializes the distributed training environment for parameter server mode. This requires enabling static graph mode and ensuring the environment variables for distributed setup (node count, IPs, etc.) are pre-configured.

```python
# Current parameter server mode only supports static graph mode, so specify ``paddle.enable_static()`` before training
paddle.enable_static()
fleet.init(is_collective=False)
```

--------------------------------

### Install PaddlePaddle XPU2 KP Package

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/dev_guides/op_optimization/kernel_primitive_api/model_example_cn

Installs the PaddlePaddle XPU2 KP installation package. This command is specific to Python 3.7 and the xpu2 hardware. Ensure you have the correct Python version and hardware environment.

```bash
pip install https://paddle-wheel.bj.bcebos.com/2.3.0/xpu2/kp/paddlepaddle_xpu-2.3.0-cp37-cp37m-linux_x86_64.whl
```

--------------------------------

### PaddlePaddle NLP Training Example

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/model_convert/convert_from_pytorch/nlp_fast_explore_cn

This code snippet demonstrates a typical NLP training workflow using PaddlePaddle. It includes dataset loading, model and tokenizer definition, data preprocessing, batch sampling, optimizer setup, and the training loop. Dependencies include paddle, paddlenlp, and numpy.

```python
from functools import partial
import numpy as np

import paddle
from paddlenlp.datasets import load_dataset
from paddlenlp.transformers import BertForSequenceClassification, AutoTokenizer
# 定义数据集
train_ds = load_dataset("chnsenticorp", splits=["train"])
# 定义模型
model = BertForSequenceClassification.from_pretrained("bert-wwm-chinese", num_classes=len(train_ds.label_list))
# 定义 Tokenizer，实现文本到 ID 的转化
tokenizer = AutoTokenizer.from_pretrained("bert-wwm-chinese")
# 数据处理，从文本到处理好的数据，input_ids，token_type_ids
def convert_example(example, tokenizer):
    encoded_inputs = tokenizer(text=example["text"], max_seq_len=512, pad_to_max_seq_len=True)
    return tuple([np.array(x, dtype="int64") for x in [
            encoded_inputs["input_ids"], encoded_inputs["token_type_ids"], [example["label"]]]])

train_ds = train_ds.map(partial(convert_example, tokenizer=tokenizer))
# 定义 BatchSampler，组合 batch，shuffle 数据
batch_sampler = paddle.io.BatchSampler(dataset=train_ds, batch_size=8, shuffle=True)
train_data_loader = paddle.io.DataLoader(dataset=train_ds, batch_sampler=batch_sampler, return_list=True)
# 定义 optimizer 优化器
optimizer = paddle.optimizer.AdamW(learning_rate=0.001, parameters=model.parameters())
# 定义 loss
criterion = paddle.nn.loss.CrossEntropyLoss()
# 训练
for input_ids, token_type_ids, labels in train_data_loader():
    logits = model(input_ids, token_type_ids)
    loss = criterion(logits, labels)
    loss.backward()
    optimizer.step()
    optimizer.clear_grad()

```

--------------------------------

### Download and Install CMake from Source

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/install/compile/arm-compile

This snippet downloads the CMake source code, extracts it, and then builds and installs it. This is necessary if the system's default CMake version is below 3.15.

```bash
wget https://github.com/Kitware/CMake/releases/download/v3.16.8/cmake-3.16.8.tar.gz
tar -xzf cmake-3.16.8.tar.gz && cd cmake-3.16.8
./bootstrap && make && sudo make install
```

--------------------------------

### Build and Install Paddle iluvatar-gpu Plugin from Source

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/hardware_support/iluvatar_gpu/install_cn

Compiles the Paddle iluvatar-gpu plugin from source code and installs it using the provided build and install scripts. This method is used when pre-built wheel packages are not available or a custom build is required.

```bash
# 进入硬件后端(天数 iluvatar_gpu)目录
cd backends/iluvatar_gpu

# 先安装飞桨 CPU 安装包
python -m pip install  --pre paddlepaddle -i https://www.paddlepaddle.org.cn/packages/nightly/cpu/

# 安装编译所需依赖
cd /tmp
wget https://github.com/protocolbuffers/protobuf/releases/download/v21.12/protoc-21.12-linux-x86_64.zip
unzip protoc-21.12-linux-x86_64.zip
mv bin/protoc /usr/local/bin/
rm -rf protoc-21.12-linux-x86_64.zip include bin
cd - 

pip install --upgrade setuptools wheel

# 执行编译脚本
bash build_paddle.sh

# 编译产出在 build_pip 路径下，使用安装脚本进行安装
bash install_paddle.sh

```

--------------------------------

### Clone PaddleClas and Install Requirements

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/dev_guides/op_optimization/kernel_primitive_api/model_example_cn

Clones the PaddleClas repository from GitHub and installs the necessary Python dependencies. This step is crucial for accessing the model configurations and training scripts.

```bash
git clone -b develop https://github.com/PaddlePaddle/PaddleClas.git
cd PaddleClas
python -m pip install -r requirements.txt
```

--------------------------------

### Python Package Setup Script (setup.py)

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/custom_op/cross_ecosystem_custom_op/design_and_migration_cn

This Python script uses setuptools to build a Python package. It leverages torch.utils.cpp_extension to compile C++ source files into Python-extendable modules, defining the package name, included packages, and custom extension modules.

```python
from setuptools import setup, find_packages
from torch.utils import cpp_extension

setup(
    name="extension",
    packages=find_packages(include=['extension']),
    ext_modules=[
        cpp_extension.CUDAExtension(
            name="extension_cpp",
            sources=["csrc/muladd.cc"],
        )
    ],
    cmdclass={'build_ext': cpp_extension.BuildExtension},
)

```

--------------------------------

### Start and Enter TianShu IXUCA Docker Container

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/hardware_support/iluvatar_gpu/install_cn

Starts a Docker container named 'paddle-ixuca-dev' with necessary volume mounts for development and then enters the container's bash shell. This container provides the isolated environment for PaddlePaddle development with TianShu GPGPU.

```bash
# 启动容器
docker run -itd --name paddle-ixuca-dev -v /usr/src:/usr/src -v /lib/modules:/lib/modules \
    -v /dev:/dev -v /home:/home --privileged --cap-add=ALL --pid=host \
    ccr-2vdh3abv-pub.cnc.bj.baidubce.com/device/paddle-ixuca:latest
docker exec -it paddle-ixuca-dev bash

```

--------------------------------

### Full Example Imports (Python)

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/06_distributed_training/cluster_quick_start_collective_cn

Provides a comprehensive set of imports for a complete distributed training example in PaddlePaddle, including core libraries, neural network modules, and distributed training specific APIs.

```python
# -*- coding: UTF-8 -*-
import numpy as np
import matplotlib.pyplot as plt
import paddle
import paddle.nn.functional as F
from paddle.vision.transforms import ToTensor
# 一、导入分布式专用 Fleet API
from paddle.distributed import fleet
# 构建分布式数据加载器所需 API
from paddle.io import DataLoader, DistributedBatchSampler
```

--------------------------------

### Example PaddleJob Launch Log Output (Log)

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/06_distributed_training/deployment_cn

A typical log output from a PaddleJob distributed training launch. This shows configuration details such as devices, run mode, and pod status, which are useful for verifying correct distributed setup.

```log
LAUNCH INFO 2022-05-18 11:53:09,773 -----------  Configuration  ----------------------
LAUNCH INFO 2022-05-18 11:53:09,773 devices: 6,7
LAUNCH INFO 2022-05-18 11:53:09,773 elastic_level: -1
LAUNCH INFO 2022-05-18 11:53:09,773 elastic_timeout: 30
LAUNCH INFO 2022-05-18 11:53:09,773 gloo_port: 6767
LAUNCH INFO 2022-05-18 11:53:09,773 host: None
LAUNCH INFO 2022-05-18 11:53:09,773 job_id: default
LAUNCH INFO 2022-05-18 11:53:09,773 legacy: False
LAUNCH INFO 2022-05-18 11:53:09,773 log_dir: log
LAUNCH INFO 2022-05-18 11:53:09,773 log_level: INFO
LAUNCH INFO 2022-05-18 11:53:09,774 master: None
LAUNCH INFO 2022-05-18 11:53:09,774 max_restart: 3
LAUNCH INFO 2022-05-18 11:53:09,774 nnodes: 1
LAUNCH INFO 2022-05-18 11:53:09,774 nproc_per_node: None
LAUNCH INFO 2022-05-18 11:53:09,774 rank: -1
LAUNCH INFO 2022-05-18 11:53:09,774 run_mode: collective
LAUNCH INFO 2022-05-18 11:53:09,774 server_num: None
LAUNCH INFO 2022-05-18 11:53:09,774 servers:
LAUNCH INFO 2022-05-18 11:53:09,774 trainer_num: None
LAUNCH INFO 2022-05-18 11:53:09,774 trainers:
LAUNCH INFO 2022-05-18 11:53:09,774 training_script: demo.py
LAUNCH INFO 2022-05-18 11:53:09,774 training_script_args: []
LAUNCH INFO 2022-05-18 11:53:09,774 with_gloo: 0
LAUNCH INFO 2022-05-18 11:53:09,774 --------------------------------------------------
LAUNCH INFO 2022-05-18 11:53:09,783 Job: default, mode collective, replicas 1[1:1], elastic False
LAUNCH INFO 2022-05-18 11:53:09,784 Run Pod: gistdo, replicas 2, status ready
LAUNCH INFO 2022-05-18 11:53:09,806 Watching Pod: gistdo, replicas 2, status running

```

--------------------------------

### Download and Prepare ImageNet Dataset

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/dev_guides/op_optimization/kernel_primitive_api/model_example_cn

Downloads the ImageNet dataset (or a subset like CIFAR100 for this example) and prepares it for model training. It involves downloading a tar archive, extracting it, and creating symbolic links for the training and validation lists.

```bash
cd dataset
rm -rf ILSVRC2012
wget -nc https://paddle-imagenet-models-name.bj.bcebos.com/data/whole_chain/whole_chain_CIFAR100.tar
tar xf whole_chain_CIFAR100.tar
ln -s whole_chain_CIFAR100 ILSVRC2012
cd ILSVRC2012
mv train.txt train_list.txt
mv test.txt val_list.txt
```

--------------------------------

### Define Distributed Training Strategy (Asynchronous)

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/06_distributed_training/cluster_quick_start_ps_cn

Defines the distributed training strategy using `fleet.DistributedStrategy`. This example specifically sets `a_sync` to `True` to enable asynchronous training, which is a common strategy in parameter server mode.

```python
# Define asynchronous training
dist_strategy = fleet.DistributedStrategy()
dist_strategy.a_sync = True
```

--------------------------------

### Implementing setuptools.setup in Paddle for PyTorch Compatibility

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/model_convert/convert_from_pytorch/api_difference/composite_implement/setuptools.setup

Shows how to adapt the setuptools.setup function for use with Paddle, specifically when integrating with PyTorch and custom C++ extensions. It checks for and handles the 'cmdclass' attribute related to Paddle's C++ extension setup.

```python
# PyTorch 写法
setuptools.setup(*args, **attrs)

# Paddle 写法
if "cmdclass" in attrs:
    if "paddle.utils.cpp_extension.BuildExtension" in attrs["cmdclass"]:
        paddle.utils.cpp_extension.setup(attrs)
```

--------------------------------

### Run Paddle Dynamic Graph Pipeline Parallelism Example

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/06_distributed_training/pipeline_parallel_cn

Provides the command to launch the dynamic graph pipeline parallelism script using Paddle's distributed launch utility. It requires setting the visible CUDA devices and executing the Python file containing the training logic.

```bash
export CUDA_VISIBLE_DEVICES=0,1
python -m paddle.distributed.launch alexnet_dygraph_pipeline.py # alexnet_dygraph_pipeline.py is the python file for users to run dynamic graph pipeline parallelism
```

--------------------------------

### Install GPU Version of PaddlePaddle (CUDA 13.0)

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/install/pip/linux-pip

Installs the GPU version of PaddlePaddle 3.3.0 compatible with CUDA 13.0. TensorRT 10.5.0.18 can be optionally installed.

```bash
python3 -m pip install paddlepaddle-gpu==3.3.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu130/
```

--------------------------------

### Install GPU Version of PaddlePaddle (CUDA 12.9)

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/install/pip/linux-pip

Installs the GPU version of PaddlePaddle 3.3.0 compatible with CUDA 12.9. TensorRT 10.5.0.18 can be optionally installed.

```bash
python3 -m pip install paddlepaddle-gpu==3.3.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu129/
```

--------------------------------

### 使用 nvidia-docker 执行 GPU 训练

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/install/Tables

使用 nvidia-docker 命令来运行包含 GPU 支持的 Docker 镜像。此方法要求在物理机上安装 GPU 最新驱动，并已安装 nvidia-docker。它将当前工作目录挂载到容器的 /work 目录。

```bash
nvidia-docker run -it -v $PWD:/work ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddle:3.3.0-gpu-cuda12.9-cudnn9.9 /bin/bash
```

--------------------------------

### Install GPU Version of PaddlePaddle (CUDA 12.6)

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/install/pip/linux-pip

Installs the GPU version of PaddlePaddle 3.3.0 compatible with CUDA 12.6. TensorRT 10.5.0.18 can be optionally installed.

```bash
python3 -m pip install paddlepaddle-gpu==3.3.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
```

--------------------------------

### Install GPU Version of PaddlePaddle (CUDA 11.8)

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/install/pip/linux-pip

Installs the GPU version of PaddlePaddle 3.3.0 compatible with CUDA 11.8. TensorRT 8.5.3.1 can be optionally installed.

```bash
python3 -m pip install paddlepaddle-gpu==3.3.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu118/
```

--------------------------------

### Adapting setup.py for PaddlePaddle Compatibility

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/custom_op/cross_ecosystem_custom_op/design_and_migration_cn

This modified `setup.py` script enables compatibility with PaddlePaddle's custom operator compilation. By adding `import paddle` and `paddle.enable_compat()`, it allows the use of PyTorch's `cpp_extension` module, which is proxied by PaddlePaddle.

```python
+import paddle
+paddle.enable_compat()  # Enable torch proxy globally

from setuptools import setup, find_packages
# 如下的 torch extension 已经被 PaddlePaddle 的同等功能替代（即 paddle.utils.cpp_extension）
# 下面的代码完全不需要修改即可运行
from torch.utils import cpp_extension

setup(
    name="extension",
    packages=find_packages(include=['extension']),
    ext_modules=[
        cpp_extension.CUDAExtension(
            name="extension_cpp",
            sources=["csrc/muladd.cc"],
        )
    ],
    cmdclass={'build_ext': cpp_extension.BuildExtension},
)

```

--------------------------------

### Pull TianShu IXUCA Docker Image

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/hardware_support/iluvatar_gpu/install_cn

Pulls the latest TianShu IXUCA development Docker image. This image contains the necessary pre-installed libraries for the TianShu GPGPU environment. It's the recommended starting point for setting up the development environment.

```bash
# 拉取镜像
docker pull ccr-2vdh3abv-pub.cnc.bj.baidubce.com/device/paddle-ixuca:latest

```

--------------------------------

### Verify PaddlePaddle Installation

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/install/pip/linux-pip

After installation, this command sequence enters the Python interpreter, imports PaddlePaddle, and runs a check to confirm successful installation. A success message indicates the installation was completed correctly.

```python
python3
import paddle
paddle.utils.run_check()
```

--------------------------------

### Install GPU Version of PaddlePaddle (CUDA 11.8)

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/install/pip/windows-pip

Installs the GPU version of PaddlePaddle (3.3.0) compatible with CUDA 11.8. It also suggests installing TensorRT 8.5.1.7 if TensorRT functionality is needed.

```shell
python -m pip install paddlepaddle-gpu==3.3.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu118/
```

--------------------------------

### Check System Architecture and Bitness

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/install/index_cn

This command confirms that your Python and pip installations are 64-bit and that your processor architecture is x86_64 (or compatible). This is a prerequisite for installing PaddlePaddle.

```shell
python -c "import platform;print(platform.architecture()[0]);print(platform.machine())"
```

--------------------------------

### Launch Multi-Card Training Script (Bash)

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/model_convert/convert_from_pytorch/cv_experience_cn

Example bash script for launching distributed training across multiple GPUs (e.g., 8 cards). It uses paddle.distributed.launch and specifies the GPUs to be used.

```bash
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

python3 -m paddle.distributed.launch \
    --gpus="0,1,2,3,4,5,6,7" \
    train.py \
    --data-path /paddle/data/ILSVRC2012_torch \
    --lr 0.01 \
    --batch-size 32 \
    --output-dir "./output/"
```

--------------------------------

### Run Model Parallelism Example using PaddlePaddle Distributed Launch

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/06_distributed_training/model_parallel_cn

This command demonstrates how to execute the model parallelism dynamic graph example. It requires setting the CUDA_VISIBLE_DEVICES environment variable to specify the GPUs to use and then launching the Python script using `paddle.distributed.launch`. This setup is essential for distributed training on multiple GPUs.

```bash
export CUDA_VISIBLE_DEVICES=0,1
python -m paddle.distributed.launch mp_dygraph.py
```

--------------------------------

### Check Pip Version

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/install/index_cn

Ensures that your pip package installer meets the minimum version requirement of 20.2.2. An outdated pip can lead to installation issues.

```shell
python -m ensurepip
python -m pip --version
```

--------------------------------

### PaddlePaddle Project Structure Example

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/model_convert/convert_from_pytorch/cv_experience_cn

This example demonstrates a recommended directory structure for a PaddlePaddle project, designed to be similar to PyTorch projects for easier comparison and migration. It includes common directories for datasets, models, transforms, and scripts for training and evaluation.

```tree
paddlevision
    |--datasets                             # 数据集加载与处理相关代码
    |--models                               # 模型组网相关代码
         |--__init__.py
         |--_utils.py                       # 工具类及函数
         |--misc_paddle.py                  # 组成网络的子模块
         |--mobilenet_v3_paddle.py          # 模型网络结构定义
    |--transforms                           # 数据增强实现
    |--__init__.py
__init__.py
metric.py                                   # 评价指标实现
presets.py                                  # 数据增强设定
train.py                                    # 模型训练代码
utils.py                                    # 工具类及函数

```

--------------------------------

### Install PaddlePaddle CPU Version

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/install/index_cn

Installs the CPU-only version of PaddlePaddle using pip. This command specifies the version and the package index URL for CPU builds.

```shell
python -m pip install paddlepaddle==3.3.0 -i https://www.paddlepaddle.org.cn/packages/stable/cpu/
```

--------------------------------

### Configure Build Environment and Compile PaddlePaddle

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/install/compile/arm-compile

This section details the steps to prepare the build directory, set the maximum number of open files, configure the build using CMake with specific options for ARMv8 architecture and Python 3.10, and finally compile the PaddlePaddle library using make.

```bash
mkdir build && cd build
ulimit -n 4096
cmake .. -DPY_VERSION=3.10 -DPYTHON_EXECUTABLE=`which python3` -DWITH_ARM=ON -DWITH_GPU=OFF -DWITH_TESTING=OFF -DCMAKE_BUILD_TYPE=Release -DON_INFER=ON -DWITH_XBYAK=OFF
make TARGET=ARMV8 -j$(nproc)
```

--------------------------------

### Install PaddlePaddle with Pip

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/06_distributed_training/deployment_cn

Installs a specific version of the PaddlePaddle GPU package using pip. Ensure you have a compatible CUDA and cuDNN version installed on your system.

```bash
$ python -m pip install paddlepaddle-gpu==2.2.2.post112 -f https://www.paddlepaddle.org.cn/whl/linux/mkl/avx/stable.html
```

--------------------------------

### Conversion Examples

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/model_convert/convert_from_pytorch/api_difference/torch_more_args/torch.scalar_tensor

Illustrates how to convert PyTorch tensor creation syntax to PaddlePaddle syntax for device placement, gradient handling, and pinned memory.

```APIDOC
## Conversion Examples

### Description
Illustrates how to convert PyTorch tensor creation syntax to PaddlePaddle syntax for device placement, gradient handling, and pinned memory.

### Method
N/A (Code examples)

### Endpoint
N/A

### Parameters
#### Path Parameters
None

#### Query Parameters
None

#### Request Body
None

### Request Example
#### device: Tensor 的设备
```python
# PyTorch 写法
torch.tensor(3, device=torch.device('cpu'))

# Paddle 写法
y = paddle.to_tensor(3, dtype=paddle.float32, place=paddle.CPUPlace())
```

#### requires_grad：是否不阻断梯度传导
```python
# PyTorch 写法
x = torch.tensor(3, requires_grad=True)

# Paddle 写法
x = paddle.to_tensor(3, dtype=paddle.float32, stop_gradient=False)
```

#### pin_memory：是否分配到固定内存上
```python
# PyTorch 写法
x = torch.tensor(3, pin_memory=True)

# Paddle 写法
x = paddle.to_tensor(3, dtype=paddle.float32).pin_memory()
```

### Response
#### Success Response (200)
Provides code snippets demonstrating conversions.

#### Response Example
```json
{
  "conversion_examples": "See code snippets above"
}
```
```

--------------------------------

### Install PaddlePaddle CPU Package

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/hardware_support/iluvatar_gpu/install_cn

Installs the nightly build of the PaddlePaddle CPU package using pip. This is a prerequisite for installing the TianShu GPGPU plugin, as the GPU version is an add-on to the core PaddlePaddle framework.

```bash
# 先安装飞桨 CPU 安装包
python -m pip install  --pre paddlepaddle -i https://www.paddlepaddle.org.cn/packages/nightly/cpu/ 

```

--------------------------------

### Install and Install Pre-commit Hooks (Python/Git)

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/dev_guides/code_contributing_path_cn

These commands install the 'pre-commit' tool and then install the git hooks for the Paddle repository. Pre-commit automates code formatting and pre-commit checks, ensuring code quality and consistency before committing.

```bash
➜  pip install pre-commit==2.17.0
➜  pre-commit install
```

--------------------------------

### setuptools.setup Function Signature

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/model_convert/convert_from_pytorch/api_difference/composite_implement/setuptools.setup

Displays the function signature for setuptools.setup. This is a standard Python function used for packaging projects.

```python
setuptools.setup(*args, **attrs)
```

--------------------------------

### Download Example Image for Transformations

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/beginner/data_preprocessing_cn

This command downloads a sample image named 'flower_demo.png' which is used in subsequent examples to demonstrate the effects of various image transformations provided by PaddlePaddle.

```bash
# 下载示例图片
! wget https://paddle-imagenet-models-name.bj.bcebos.com/data/demo_images/flower_demo.png
```

--------------------------------

### Verify PaddlePaddle Installation

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/model_convert/convert_from_pytorch/nlp_fast_explore_cn

Checks if PaddlePaddle is installed correctly and prints the installed version. This involves importing the paddle library and running a built-in check function.

```python
import paddle
paddle.utils.run_check()
print(paddle.__version__)
```

--------------------------------

### AMP Training Setup (Partial)

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/06_distributed_training/data_parallel/amp_cn

Illustrates the initial steps for setting up Automatic Mixed Precision (AMP) training in PaddlePaddle. This includes defining the model and optimizer, which are prerequisites before applying AMP context managers and loss scaling.

```python
model = SimpleNet(input_size, output_size)  # 定义模型

optimizer = paddle.optimizer.SGD(learning_rate=0.0001, parameters=model.parameters())  # 定义优化器


```

--------------------------------

### Install Protocol Buffers Compiler

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/hardware_support/iluvatar_gpu/install_cn

Downloads and installs the Protocol Buffers compiler (protoc) version 21.12. This compiler is a build dependency for compiling the Paddle iluvatar-gpu plugin from source.

```bash
# 安装编译所需依赖
cd /tmp
wget https://github.com/protocolbuffers/protobuf/releases/download/v21.12/protoc-21.12-linux-x86_64.zip
unzip protoc-21.12-linux-x86_64.zip
mv bin/protoc /usr/local/bin/
rm -rf protoc-21.12-linux-x86_64.zip include bin
cd - 

```

--------------------------------

### 手动挂载 CUDA 库和设备到 Docker 容器

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/install/Tables

在未安装 nvidia-docker 的情况下，通过手动导出 CUDA 库和 Linux 设备到 Docker 容器内来启用 GPU 支持。此方法需要用户自行查找并导出相关的库和设备文件。

```bash
export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') \
$(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"
export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}')
docker run ${CUDA_SO} \
${DEVICES} -it ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddle:latest-gpu
```

--------------------------------

### Verify PaddlePaddle Installation

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/06_distributed_training/deployment_cn

Executes a Python command to verify the PaddlePaddle installation. This command imports the paddle library and runs a check function, which should output expected results if the installation is successful.

```python
python -c "import paddle; paddle.utils.run_check()"
```

--------------------------------

### Parameter Mapping and Conversion Examples

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/model_convert/convert_from_pytorch/api_difference/torch_more_args/torch.nn.MSELoss

Examples demonstrating how to map PyTorch's MSELoss parameters (size_average, reduce) to PaddlePaddle's MSELoss reduction parameter.

```APIDOC
## Parameter Mapping

PyTorch parameters `size_average` and `reduce` are deprecated and correspond to the `reduction` parameter in PaddlePaddle.

| PyTorch      | PaddlePaddle | Remarks                                     |
|--------------|--------------|---------------------------------------------|
| `size_average` | -            | Deprecated. Needs to be transcribed.        |
| `reduce`       | -            | Deprecated. Needs to be transcribed.        |
| `reduction`    | `reduction`  | Specifies the reduction method for the output. |

## Conversion Examples

### `size_average`/`reduce` to `reduction='sum'`

```python
# PyTorch syntax
torch.nn.MSELoss(size_average=False, reduce=True)
torch.nn.MSELoss(size_average=False)

# PaddlePaddle syntax
paddle.nn.MSELoss(reduction='sum')
```

### `size_average`/`reduce` to `reduction='mean'`

```python
# PyTorch syntax
torch.nn.MSELoss(size_average=True, reduce=True)
torch.nn.MSELoss(reduce=True)
torch.nn.MSELoss(size_average=True)
torch.nn.MSELoss()

# PaddlePaddle syntax
paddle.nn.MSELoss(reduction='mean')
```

### `size_average`/`reduce` to `reduction='none'`

```python
# PyTorch syntax
torch.nn.MSELoss(size_average=True, reduce=False)
torch.nn.MSELoss(size_average=False, reduce=False)
torch.nn.MSELoss(reduce=False)

# PaddlePaddle syntax
paddle.nn.MSELoss(reduction='none')
```
```

--------------------------------

### Check Python Version

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/install/index_cn

Verifies that your installed Python version is compatible with PaddlePaddle. Supported versions are 3.9, 3.10, 3.11, 3.12, and 3.13.

```shell
python --version
```

--------------------------------

### Device Synchronization Examples (PyTorch vs PaddlePaddle)

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/model_convert/convert_from_pytorch/api_difference/input_args_type_diff/torch.cuda.synchronize

Illustrates how to call synchronize functions with different device specifications in PyTorch and PaddlePaddle, showing parameter mapping.

```python
# PyTorch 写法
torch.cuda.synchronize('cuda:0')

# Paddle 写法
paddle.device.synchronize('gpu:0')
```

```python
# PyTorch 写法
torch.cuda.synchronize(2)

# Paddle 写法
paddle.device.synchronize('gpu:2')
```

--------------------------------

### Install PaddlePaddle GPU Version (CUDA 12.6)

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/install/index_cn

Installs the GPU version of PaddlePaddle with CUDA 12.6 support using pip. This command specifies the version and the package index URL for the corresponding CUDA toolkit.

```shell
python -m pip install paddlepaddle-gpu==3.3.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
```

--------------------------------

### Install PaddlePaddle GPU Version (CUDA 12.9)

Source: https://www.paddlepaddle.org.cn/documentation/docs/zh/install/index_cn

Installs the GPU version of PaddlePaddle with CUDA 12.9 support using pip. This command specifies the version and the package index URL for the corresponding CUDA toolkit.

```shell
python -m pip install paddlepaddle-gpu==3.3.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu129/
```
Tokens:
Raw