PaddleOCR-VL NVIDIA Blackwell-Architecture GPUs Usage Tutorial

INFO: Unless otherwise specified, the term "PaddleOCR-VL" in this tutorial refers to the PaddleOCR-VL model series (e.g., PaddleOCR-VL-1.5). References specific to the PaddleOCR-VL v1 version will be explicitly noted.

This tutorial provides guidance on using PaddleOCR-VL on NVIDIA Blackwell-architecture GPUs, covering the complete workflow from environment preparation to service deployment.

NVIDIA Blackwell-architecture GPUs include, but are not limited to:

• RTX 5090
• RTX 5080
• RTX 5070、RTX 5070 Ti
• RTX 5060、RTX 5060 Ti
• RTX 5050

PaddleOCR-VL has been verified for accuracy and speed on the RTX 5070. However, due to hardware diversity, compatibility with other NVIDIA Blackwell-architecture GPUs has not yet been confirmed. We welcome the community to test on different hardware setups and share your results.

Before starting the tutorial, please ensure that your NVIDIA driver supports CUDA 12.9 or higher.

TIP: Before reading this hardware-specific tutorial, we recommend first reading the Process Guide in the main PaddleOCR-VL Usage Tutorial to determine which chapters apply to your goal, and then returning here to read the corresponding sections.

1. Environment Preparation

This section introduces how to set up the PaddleOCR-VL runtime environment using one of the following two methods:

• Method 1: Use the official Docker image.

• Method 2: Manually install PaddlePaddle and PaddleOCR.

We strongly recommend using the Docker image to minimize potential environment-related issues.

1.1 Method 1: Using Docker Image

We recommend using the official Docker image (requires Docker version >= 19.03, GPU-equipped machine with NVIDIA driver supporting CUDA 12.9 or higher):

docker run \
    -it \
    --gpus all \
    --network host \
    --user root \
    ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-vl:latest-nvidia-gpu-sm120 \
    /bin/bash
# Call PaddleOCR CLI or Python API in the container

If you wish to use PaddleOCR-VL in an offline environment, replace ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-vl:latest-nvidia-gpu-sm120 (image size approximately 10 GB) in the above command with the offline version image ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-vl:latest-nvidia-gpu-sm120-offline (image size approximately 12 GB).

TIP: Images with the latest-xxx tag correspond to the latest version of PaddleOCR. If you want to use a specific version of the PaddleOCR image, you can replace latest in the tag with the desired version number: paddleocr<major>.<minor>. For example: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-vl:paddleocr3.3-nvidia-gpu-sm120-offline

1.2 Method 2: Manually Install PaddlePaddle and PaddleOCR

If Docker is not an option, you can manually install PaddlePaddle and PaddleOCR. Python version 3.8–3.13 is required.

We strongly recommend installing PaddleOCR-VL in a virtual environment to avoid dependency conflicts. For example, create a virtual environment using Python's standard venv library:

# Create a virtual environment
python -m venv .venv_paddleocr
# Activate the environment
source .venv_paddleocr/bin/activate

Run the following commands to complete the installation:

# Note that PaddlePaddle for cu129 is being installed here
python -m pip install paddlepaddle-gpu==3.2.1 -i https://www.paddlepaddle.org.cn/packages/stable/cu129/
python -m pip install -U "paddleocr[doc-parser]"

Please ensure that PaddlePaddle framework version 3.2.1 or higher is installed.

2. Quick Start

Please refer to PaddleOCR-VL Usage Tutorial - 2. Quick Start.

3. Improving Inference Performance with VLM Inference Services

The inference performance under default configurations may not be fully optimized and may not meet actual production requirements. This section introduces how to improve PaddleOCR-VL inference performance through a VLM inference service. In this hardware-specific guide, the examples use vLLM and SGLang as the backends for the VLM inference service.

3.1 Starting the VLM Inference Service

There are two methods to start the VLM inference service; choose one:

• Method 1: Start the service using the official Docker image.

• Method 2: Manually install dependencies and start the service via PaddleOCR CLI.

We strongly recommend using the Docker image to minimize potential environment-related issues.

3.1.1 Method 1: Using Docker Image

PaddleOCR provides a Docker image for quickly starting the vLLM inference service. Use the following command to start the service (requires Docker version >= 19.03, GPU-equipped machine with NVIDIA driver supporting CUDA 12.9 or higher):

docker run \
    -it \
    --gpus all \
    --network host \
    ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-genai-vllm-server:latest-nvidia-gpu-sm120 \
    paddleocr genai_server --model_name PaddleOCR-VL-1.5-0.9B --host 0.0.0.0 --port 8118 --backend vllm

If you wish to start the service in an offline environment, replace ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-genai-vllm-server:latest-nvidia-gpu-sm120 (image size approximately 13 GB) in the above command with the offline version image ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-genai-vllm-server:latest-nvidia-gpu-sm120-offline (image size approximately 15 GB).

When launching the vLLM inference service, we provide a set of default parameter settings. If you need to adjust parameters such as GPU memory usage, you can configure additional parameters yourself. Please refer to 3.3.1 Server-side Parameter Adjustment to create a configuration file, then mount the file into the container and specify the configuration file using backend_config in the command to start the service, for example:

docker run \
    -it \
    --rm \
    --gpus all \
    --network host \
    -v vllm_config.yml:/tmp/vllm_config.yml \
    ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-genai-vllm-server:latest-nvidia-gpu-sm120 \
    paddleocr genai_server --model_name PaddleOCR-VL-1.5-0.9B --host 0.0.0.0 --port 8118 --backend vllm --backend_config /tmp/vllm_config.yml

TIP: Images with the latest-xxx tag correspond to the latest version of PaddleOCR. If you want to use a specific version of the PaddleOCR image, you can replace latest in the tag with the desired version number: paddleocr<major>.<minor>. For example: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-genai-vllm-server:paddleocr3.3-nvidia-gpu-sm120-offline

3.1.2 Method 2: Installation and Usage via PaddleOCR CLI

Since inference acceleration frameworks may conflict with packages already installed in the current environment, it is recommended to install them in a virtual environment:

# If a virtual environment is currently activated, deactivate it first using `deactivate`
# Create a virtual environment
python -m venv .venv_vlm
# Activate the environment
source .venv_vlm/bin/activate

vLLM and SGLang depend on FlashAttention, and installing FlashAttention may require CUDA compilation tools such as nvcc. If these tools are not available in your environment (for example, when using the paddleocr-vl image), you can obtain a prebuilt FlashAttention package (version 2.8.3 required) from this repository, install it first, and then proceed with subsequent commands. For example, in the paddleocr-vl image, run python -m pip install https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.3.14/flash_attn-2.8.3+cu128torch2.8-cp310-cp310-linux_x86_64.whl. This step is not required for FastDeploy.

Install PaddleOCR and the dependencies of inference acceleration services, using vLLM as an example:

# Install PaddleOCR
python -m pip install "paddleocr[doc-parser]"
# Install inference acceleration service dependencies
paddleocr install_genai_server_deps vllm

Usage of the paddleocr install_genai_server_deps command:

paddleocr install_genai_server_deps <inference acceleration framework name>

Currently supported framework names are vllm and sglang, corresponding to vLLM and SGLang, respectively.

WARNING: The transformers library versions required by vLLM, SGLang and Transformers engine are currently incompatible, so Transformers engine cannot be installed together with vLLM or SGLang in the same environment. If using Transformers + vLLM or Transformers + SGLang inference, please deploy the layout detection model and VLM service in different environments.

After installation, you can start the service using the paddleocr genai_server command:

paddleocr genai_server --model_name PaddleOCR-VL-1.5-0.9B --backend vllm --port 8118

The parameters supported by this command are as follows:

3.2 Client Usage

For client-side invocation methods, please refer to PaddleOCR-VL Usage Tutorial - 3.2 Client Usage Methods. If you run the client on this hardware, make sure to specify device="gpu".

3.3 Performance Tuning

Please refer to PaddleOCR-VL Usage Tutorial - 3.3 Performance Tuning.

4. Service Deployment

This section mainly introduces how to deploy PaddleOCR-VL as a service and invoke it. There are two methods available; choose one:

• Method 1: Deploy using Docker Compose.

• Method 2: Manually install dependencies for deployment.

Please note that the PaddleOCR-VL service introduced in this section differs from the VLM inference service in the previous section: the latter is responsible for only one part of the complete process (i.e., VLM inference) and is called as an underlying service by the former.

4.1 Method 1: Deploy Using Docker Compose

1. Download the Compose file and the environment variable configuration file separately from here and here to your local machine.

2. Execute the following command in the directory containing the compose.yaml and .env files to start the server, which will listen on port 8080 by default:

   shellCopy

   # Must be executed in the directory containing compose.yaml and .env files
   docker compose up
   

   After startup, you will see output similar to the following:

   textCopy

   paddleocr-vl-api             | INFO:     Started server process [1]
   paddleocr-vl-api             | INFO:     Waiting for application startup.
   paddleocr-vl-api             | INFO:     Application startup complete.
   paddleocr-vl-api             | INFO:     Uvicorn running on http://0.0.0.0:8080 (Press CTRL+C to quit)
   

This method accelerates VLM inference using the vLLM framework and is more suitable for production environment deployment.

Additionally, after starting the server in this manner, no internet connection is required except for image pulling. For deployment in an offline environment, you can first pull the images involved in the Compose file on a connected machine, export them, and transfer them to the offline machine for import to start the service in an offline environment.

Docker Compose starts two containers sequentially by reading configurations from the .env and compose.yaml files, running the underlying VLM inference service and the PaddleOCR-VL service (pipeline service) respectively.

The meanings of each environment variable contained in the .env file are as follows:

- `API_IMAGE_TAG_SUFFIX`: The tag suffix of the image used to launch the pipeline service.
- `VLM_BACKEND`: The VLM inference backend.
- `VLM_IMAGE_TAG_SUFFIX`: The tag suffix of the image used to launch the VLM inference service.

You can modify compose.yaml to meet custom requirements, for example:

Edit <code>paddleocr-vl-api.ports</code> in the <code>compose.yaml</code> file to change the port. For example, if you need to change the service port to 8111, make the following modifications:

  paddleocr-vl-api:
    ...
    ports:
-     - 8080:8080
+     - 8111:8080
    ...

Edit <code>environment</code> in the <code>compose.yaml</code> file to change the GPU used. For example, if you need to use card 1 for deployment, make the following modifications:

  paddleocr-vl-api:
    ...
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
-             device_ids: ["0"]
+             device_ids: ["1"]
              capabilities: [gpu]
    ...
  paddleocr-vlm-server:
    ...
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
-             device_ids: ["0"]
+             device_ids: ["1"]
              capabilities: [gpu]
    ...

If you want to adjust the VLM server configuration, refer to <a href="./PaddleOCR-VL.en.md#331-server-side-parameter-adjustment">3.3.1 Server Parameter Adjustment</a> to generate a configuration file.

After generating the configuration file, add the following <code>paddleocr-vlm-server.volumes</code> and <code>paddleocr-vlm-server.command</code> fields to your <code>compose.yaml</code>. Replace <code>/path/to/your_config.yaml</code> with your actual configuration file path.

  paddleocr-vlm-server:
    ...
    volumes: /path/to/your_config.yaml:/home/paddleocr/vlm_server_config.yaml
    command: paddleocr genai_server --model_name PaddleOCR-VL-1.5-0.9B --host 0.0.0.0 --port 8118 --backend vllm --backend_config /home/paddleocr/vlm_server_config.yaml
    ...

Refer to the <a href="./PaddleOCR-VL.en.md#44-pipeline-configuration-adjustment-instructions">4.4 Pipeline Configuration Adjustment Instructions</a> section.

4.2 Method 2: Manually Deployment

Please refer to PaddleOCR-VL Usage Tutorial - 4.2 Method 2: Manual Deployment.

4.3 Client Invocation Methods

Please refer to PaddleOCR-VL Usage Tutorial - 4.3 Client Invocation Methods.

4.4 Pipeline Configuration Adjustment Instructions

Please refer to PaddleOCR-VL Usage Tutorial - 4.4 Pipeline Configuration Adjustment Instructions.

5. Model Fine-Tuning

Please refer to PaddleOCR-VL Usage Tutorial - 5. Model Fine-Tuning.