PaddleOCR-VL Usage Tutorial

PaddleOCR-VL is an advanced and efficient document parsing model designed specifically for element recognition in documents. Its core component is PaddleOCR-VL-0.9B, a compact yet powerful Vision-Language Model (VLM) composed of a NaViT-style dynamic resolution visual encoder and the ERNIE-4.5-0.3B language model, enabling precise element recognition. The model supports 109 languages and excels in recognizing complex elements (such as text, tables, formulas, and charts) while maintaining extremely low resource consumption. Comprehensive evaluations on widely used public benchmarks and internal benchmarks demonstrate that PaddleOCR-VL achieves SOTA performance in both page-level document parsing and element-level recognition. It significantly outperforms existing Pipeline-based solutions, document parsing multimodal schemes, and advanced general-purpose multimodal large models, while offering faster inference speeds. These advantages make it highly suitable for deployment in real-world scenarios.

Process Guide

Before starting, please refer to the next section for information on the inference device support provided by PaddleOCR-VL to determine if your device meets the operational requirements. If your device meets the requirements, please select the relevant section to read based on your needs.

For some inference hardware, you may need to refer to other environment configuration documents we provide, but the process remains the same and does not affect your reading of the following process guide:

1. Want to quickly experience PaddleOCR-VL:

   If you wish to quickly experience the inference effects of PaddleOCR-VL, please read 1. Environment Preparation and 2. Quick Start.

2. Want to use PaddleOCR-VL in a production environment:

   Although the quick experience allows you to feel the effects of PaddleOCR-VL, it may not be optimal in terms of inference speed and GPU memory usage. If you wish to apply PaddleOCR-VL in a production environment and have higher requirements for inference performance, please read 3. Enhancing VLM Inference Performance Using Inference Acceleration Frameworks.

3. Want to deploy PaddleOCR-VL as an API service:

   If you want to deploy PaddleOCR-VL as a web service (API) so that other devices or applications can access and call it through a specific URL without configuring the environment, we offer two methods:

   • Deployment using Docker Compose (one-click start, recommended): Please read 4.1 Method 1: Deploy Using Docker Compose and 4.3 Client-Side Invocation.
   • Manual deployment: Please read 1. Environment Preparation, 4.2 Method 2: Manual Deployment, and 4.3 Client-Side Invocation.

4. Want to fine-tune PaddleOCR-VL to adapt to specific business needs:

   If you find that the accuracy performance of PaddleOCR-VL in specific business scenarios does not meet expectations, please read 5. Model Fine-tuning.

Inference Device Support for PaddleOCR-VL

Currently, PaddleOCR-VL offers four inference methods, with varying levels of support for different inference devices. Please confirm that your inference device meets the requirements in the table below before proceeding with PaddleOCR-VL deployment:

Inference Method	NVIDIA GPU	KUNLUNXIN XPU	HYGON DCU	MetaX GPU	Iluvatar GPU	x64 CPU
PaddlePaddle	✅	✅	✅	🚧	🚧	✅
vLLM	✅	🚧	✅	🚧	🚧	❌
SGLang	✅	🚧	🚧	🚧	🚧	❌
FastDeploy	✅	✅	🚧	🚧	🚧	❌

Tip

• When using NVIDIA GPU for inference, ensure that the Compute Capability (CC) and CUDA version meet the requirements:

  • PaddlePaddle: CC ≥ 7.0, CUDA ≥ 11.8
  • vLLM: CC ≥ 8.0, CUDA ≥ 12.6
  • SGLang: 8.0 ≤ CC < 12.0, CUDA ≥ 12.6
  • FastDeploy: 8.0 ≤ CC < 12.0, CUDA ≥ 12.6
  • Common GPUs with CC ≥ 8 include RTX 30/40/50 series and A10/A100, etc. For more models, refer to CUDA GPU Compute Capability

• vLLM compatibility note: Although vLLM can be launched on NVIDIA GPUs with CC 7.x such as T4/V100, timeout or OOM issues may occur, and its use is not recommended.
• Currently, PaddleOCR-VL does not support ARM architecture CPUs. More hardware support will be expanded based on actual needs in the future, so stay tuned!
• vLLM, SGLang, and FastDeploy cannot run natively on Windows or macOS. Please use the Docker images we provide.

Since different hardware requires different dependencies, if your hardware meets the requirements in the table above, please refer to the following table for the corresponding tutorial to configure your environment:

Hardware Type	Environment Configuration Tutorial
x64 CPU	This tutorial
NVIDIA GPU	- NVIDIA Blackwell architecture GPU (e.g., RTX 50 series) refer to PaddleOCR-VL NVIDIA Blackwell Architecture GPU Environment Configuration Tutorial - Other NVIDIA GPUs refer to this tutorial
KUNLUNXIN XPU	PaddleOCR-VL XPU Environment Configuration Tutorial
HYGON DCU	PaddleOCR-VL DCU Environment Configuration Tutorial

Tip

For example, if you are using an RTX 50 series GPU that meets the device requirements for both PaddlePaddle and vLLM inference methods, please refer to the PaddleOCR-VL NVIDIA Blackwell Architecture GPU Environment Configuration Tutorial to complete the environment configuration before using PaddleOCR-VL.

1. Environment Preparation

This section explains how to set up the runtime environment for PaddleOCR-VL. Choose one of the following two methods:

• Method 1: Use the official Docker image.

• Method 2: Manually install PaddlePaddle and PaddleOCR.

1.1 Method 1: Using Docker Image

We recommend using the official Docker image (requires Docker version >= 19.03, GPU-equipped machine with NVIDIA drivers supporting CUDA 12.6 or later):

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

If you need to use PaddleOCR-VL in an offline environment, replace ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-vl:latest (image size approximately 8 GB) in the above command with the offline version image ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-vl:latest-offline (image size is approximately 10 GB). You will need to pull the image on an internet-connected machine, import it into the offline machine, and then start the container using this image on the offline machine. For example:

# Execute on an internet-connected machine
docker pull ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-vl:latest-offline
# Save the image to a file
docker save ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-vl:latest-offline -o paddleocr-vl-latest-offline.tar

# Transfer the image file to the offline machine

# Execute on the offline machine
docker load -i paddleocr-vl-latest-offline.tar
# After that, you can use `docker run` to start the container on the offline machine

1.2 Method 2: Manually Install PaddlePaddle and PaddleOCR

If you cannot use Docker, you can manually install PaddlePaddle and PaddleOCR. The required Python version is 3.8–3.12.

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

# 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:

# The following command installs the PaddlePaddle version for CUDA 12.6. For other CUDA versions and the CPU version, please refer to https://www.paddlepaddle.org.cn/install/quick?docurl=/documentation/docs/zh/develop/install/pip/linux-pip.html
python -m pip install paddlepaddle-gpu==3.2.1 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
python -m pip install -U "paddleocr[doc-parser]"
# For Linux systems, please directly copy and execute the following commands without modifying the cuda version in the link:
python -m pip install https://paddle-whl.bj.bcebos.com/nightly/cu126/safetensors/safetensors-0.6.2.dev0-cp38-abi3-linux_x86_64.whl
# For Windows systems, directly copy and execute the following command:
python -m pip install https://xly-devops.cdn.bcebos.com/safetensors-nightly/safetensors-0.6.2.dev0-cp38-abi3-win_amd64.whl

Important

Please ensure that you install PaddlePaddle framework version 3.2.1 or above, along with the special version of safetensors. For macOS users, please use Docker to set up the environment.

2. Quick Start

PaddleOCR-VL supports two usage methods: CLI command line and Python API. The CLI command line method is simpler and suitable for quickly verifying functionality, while the Python API method is more flexible and suitable for integration into existing projects.

Tip

The methods introduced in this section are primarily for rapid validation. Their inference speed, memory usage, and stability may not meet the requirements of a production environment. If deployment to a production environment is needed, we strongly recommend using a dedicated inference acceleration framework. For specific methods, please refer to the next section.

2.1 Command Line Usage

Run a single command to quickly test the PaddleOCR-VL ：

paddleocr doc_parser -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/paddleocr_vl_demo.png

# Use --use_doc_orientation_classify to enable document orientation classification
paddleocr doc_parser -i ./paddleocr_vl_demo.png --use_doc_orientation_classify True

# Use --use_doc_unwarping to enable document unwarping module
paddleocr doc_parser -i ./paddleocr_vl_demo.png --use_doc_unwarping True

# Use --use_layout_detection to enable layout detection
paddleocr doc_parser -i ./paddleocr_vl_demo.png --use_layout_detection False

Command line supports more parameters. Click to expand for detailed parameter descriptions

Parameter	Description	Type
input	Data to be predicted, required. For example, the local path of an image file or PDF file: /root/data/img.jpg;Such as a URL link, for example, the network URL of an image file or PDF file:Example;Such as a local directory, which should contain the images to be predicted, for example, the local path: /root/data/(Currently, prediction for directories containing PDF files is not supported. PDF files need to be specified with a specific file path).	str
save_path	Specify the path where the inference result file will be saved. If not set, the inference results will not be saved locally.	str
layout_detection_model_name	Name of the layout area detection and ranking model. If not set, the default model of the pipeline will be used.	str
layout_detection_model_dir	Directory path of the layout area detection and ranking model. If not set, the official model will be downloaded.	str
layout_threshold	Score threshold for the layout model. Any value between 0-1. If not set, the default value is used, which is 0.5.
layout_nms	Whether to use post-processing NMS for layout detection. If not set, the initialized default value will be used.	bool
layout_unclip_ratio	Expansion coefficient for the detection boxes of the layout area detection model.Any floating-point number greater than 0. If not set, the initialized default value will be used.	float
layout_merge_bboxes_mode	Merging mode for the detection boxes output by the model in layout detection. large when set to large, it means that among the detection boxes output by the model, for overlapping and contained boxes, only the outermost largest box is retained, and the overlapping inner boxes are deleted; small, when set to small, it means that among the detection boxes output by the model, for overlapping and contained boxes, only the innermost contained small box is retained, and the overlapping outer boxes are deleted; union,no filtering is performed on the boxes, and both inner and outer boxes are retained; If not set, the initialized parameter value will be used.	str
vl_rec_model_name	Name of the multimodal recognition model. If not set, the default model will be used.	str
vl_rec_model_dir	Directory path of the multimodal recognition model. If not set, the official model will be downloaded.	str
vl_rec_backend	Inference backend used by the multimodal recognition model.	str
vl_rec_server_url	If the multimodal recognition model uses an inference service, this parameter is used to specify the server URL.	str
vl_rec_max_concurrency	If the multimodal recognition model uses an inference service, this parameter is used to specify the maximum number of concurrent requests.	str
vl_rec_api_key	If the multimodal recognition model uses an inference service, this parameter is used to specify the API key of the service.	str
doc_orientation_classify_model_name	Name of the document orientation classification model. If not set, the initialized default value will be used.	str
doc_orientation_classify_model_dir	Directory path of the document orientation classification model. If not set, the official model will be downloaded.	str
doc_unwarping_model_name	Name of the text image rectification model. If not set, the initialized default value will be used.	str
doc_unwarping_model_dir	Directory path of the text image rectification model. If not set, the official model will be downloaded.	str
use_doc_orientation_classify	Whether to load and use the document orientation classification module. If not set, the initialized default value will be used, which is initialized toFalse.	bool
use_doc_unwarping	Whether to load and use the text image rectification module. If not set, the initialized default value will be used, which is initialized to False.	bool
use_layout_detection	Whether to load and use the layout area detection and ranking module. If not set, the initialized default value will be used, which is initialized to True.	bool
use_chart_recognition	Whether to use the chart parsing function. If not set, the initialized default value will be used, which is initialized to False.	bool
format_block_content	Controls whether to format the block_content content within as Markdown. If not set, the initialized default value will be used, which defaults to initialization asFalse.	bool
use_queues	Used to control whether to enable internal queues. When set to True, data loading (such as rendering PDF pages as images), layout detection model processing, and VLM inference will be executed asynchronously in separate threads, with data passed through queues, thereby improving efficiency. This approach is particularly efficient for PDF documents with a large number of pages or directories containing a large number of images or PDF files.	bool
prompt_label	The prompt type setting for the VL model, which takes effect if and only if use_layout_detection=False.	str
repetition_penalty	The repetition penalty parameter used in VL model sampling.	float
temperature	The temperature parameter used in VL model sampling.	float
top_p	The top-p parameter used in VL model sampling.	float
min_pixels	The minimum number of pixels allowed when the VL model preprocesses images.	int
max_pixels	The maximum number of pixels allowed when the VL model preprocesses images.	int
device	The device used for inference. Supports specifying specific card numbers: CPU: For example,cpu indicates using the CPU for inference; GPU: For example,gpu:0 indicates using the first GPU for inference; NPU: For example,npu:0 indicates using the first NPU for inference; XPU: For example,xpu:0 indicates using the first XPU for inference; MLU: For example,mlu:0 indicates using the first MLU for inference; DCU: For example,dcu:0 indicates using the first DCU for inference; If not set, the initialized default value will be used. During initialization, the local GPU device 0 will be used preferentially. If it is not available, the CPU device will be used.	str
enable_hpi	Whether to enable high-performance inference.	bool
use_tensorrt	Whether to enable the TensorRT subgraph engine of Paddle Inference. If the model does not support acceleration via TensorRT, acceleration will not be used even if this flag is set. For PaddlePaddle version with CUDA 11.8, the compatible TensorRT version is 8.x (x>=6). It is recommended to install TensorRT 8.6.1.6.	bool
precision	Computational precision, such as fp32, fp16.	str
enable_mkldnn	Whether to enable MKL-DNN accelerated inference. If MKL-DNN is not available or the model does not support acceleration via MKL-DNN, acceleration will not be used even if this flag is set.	bool
mkldnn_cache_capacity	MKL-DNN cache capacity.	int
cpu_threads	The number of threads used for inference on the CPU.	int
paddlex_config	The file path for PaddleX pipeline configuration.	str

The inference result will be printed in the terminal. The default output of the PP-StructureV3 pipeline is as follows:

👉Click to expand

 
{'res': {'input_path': 'paddleocr_vl_demo.png', 'page_index': None, 'model_settings': {'use_doc_preprocessor': False, 'use_layout_detection': True, 'use_chart_recognition': False, 'format_block_content': False}, 'layout_det_res': {'input_path': None, 'page_index': None, 'boxes': [{'cls_id': 6, 'label': 'doc_title', 'score': 0.9636914134025574, 'coordinate': [np.float32(131.31366), np.float32(36.450516), np.float32(1384.522), np.float32(127.984665)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9281806349754333, 'coordinate': [np.float32(585.39465), np.float32(158.438), np.float32(930.2184), np.float32(182.57469)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9840355515480042, 'coordinate': [np.float32(9.023666), np.float32(200.86115), np.float32(361.41583), np.float32(343.8828)]}, {'cls_id': 14, 'label': 'image', 'score': 0.9871416091918945, 'coordinate': [np.float32(775.50574), np.float32(200.66502), np.float32(1503.3807), np.float32(684.9304)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9801855087280273, 'coordinate': [np.float32(9.532196), np.float32(344.90594), np.float32(361.4413), np.float32(440.8244)]}, {'cls_id': 17, 'label': 'paragraph_title', 'score': 0.9708921313285828, 'coordinate': [np.float32(28.040405), np.float32(455.87976), np.float32(341.7215), np.float32(520.7117)]}, {'cls_id': 24, 'label': 'vision_footnote', 'score': 0.9002962708473206, 'coordinate': [np.float32(809.0692), np.float32(703.70044), np.float32(1488.3016), np.float32(750.5238)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9825374484062195, 'coordinate': [np.float32(8.896561), np.float32(536.54895), np.float32(361.05237), np.float32(655.8058)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9822263717651367, 'coordinate': [np.float32(8.971573), np.float32(657.4949), np.float32(362.01715), np.float32(774.625)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9767460823059082, 'coordinate': [np.float32(9.407074), np.float32(776.5216), np.float32(361.31067), np.float32(846.82874)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9868153929710388, 'coordinate': [np.float32(8.669495), np.float32(848.2543), np.float32(361.64703), np.float32(1062.8568)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9826608300209045, 'coordinate': [np.float32(8.8025055), np.float32(1063.8615), np.float32(361.46588), np.float32(1182.8524)]}, {'cls_id': 22, 'label': 'text', 'score': 0.982555627822876, 'coordinate': [np.float32(8.820602), np.float32(1184.4663), np.float32(361.66394), np.float32(1302.4507)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9584776759147644, 'coordinate': [np.float32(9.170288), np.float32(1304.2161), np.float32(361.48898), np.float32(1351.7483)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9782056212425232, 'coordinate': [np.float32(389.1618), np.float32(200.38202), np.float32(742.7591), np.float32(295.65146)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9844875931739807, 'coordinate': [np.float32(388.73303), np.float32(297.18463), np.float32(744.00024), np.float32(441.3034)]}, {'cls_id': 17, 'label': 'paragraph_title', 'score': 0.9680547714233398, 'coordinate': [np.float32(409.39468), np.float32(455.89386), np.float32(721.7174), np.float32(520.9387)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9741666913032532, 'coordinate': [np.float32(389.71606), np.float32(536.8138), np.float32(742.7112), np.float32(608.00165)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9840384721755981, 'coordinate': [np.float32(389.30988), np.float32(609.39636), np.float32(743.09247), np.float32(750.3231)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9845995306968689, 'coordinate': [np.float32(389.13272), np.float32(751.7772), np.float32(743.058), np.float32(894.8815)]}, {'cls_id': 22, 'label': 'text', 'score': 0.984852135181427, 'coordinate': [np.float32(388.83267), np.float32(896.0371), np.float32(743.58215), np.float32(1038.7345)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9804865717887878, 'coordinate': [np.float32(389.08478), np.float32(1039.9119), np.float32(742.7585), np.float32(1134.4897)]}, {'cls_id': 22, 'label': 'text', 'score': 0.986461341381073, 'coordinate': [np.float32(388.52643), np.float32(1135.8137), np.float32(743.451), np.float32(1352.0085)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9869391918182373, 'coordinate': [np.float32(769.8341), np.float32(775.66235), np.float32(1124.9813), np.float32(1063.207)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9822869896888733, 'coordinate': [np.float32(770.30383), np.float32(1063.938), np.float32(1124.8295), np.float32(1184.2192)]}, {'cls_id': 17, 'label': 'paragraph_title', 'score': 0.9689218997955322, 'coordinate': [np.float32(791.3042), np.float32(1199.3169), np.float32(1104.4521), np.float32(1264.6985)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9713128209114075, 'coordinate': [np.float32(770.4253), np.float32(1279.6072), np.float32(1124.6917), np.float32(1351.8672)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9236552119255066, 'coordinate': [np.float32(1153.9058), np.float32(775.5814), np.float32(1334.0654), np.float32(798.1581)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9857938885688782, 'coordinate': [np.float32(1151.5197), np.float32(799.28015), np.float32(1506.3619), np.float32(991.1156)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9820687174797058, 'coordinate': [np.float32(1151.5686), np.float32(991.91095), np.float32(1506.6023), np.float32(1110.8875)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9866049885749817, 'coordinate': [np.float32(1151.6919), np.float32(1112.1301), np.float32(1507.1611), np.float32(1351.9504)]}]}}}

For explanation of the result parameters, refer to 2.2 Python Script Integration.

Note: The default model for the pipeline is relatively large, which may result in slower inference speed. It is recommended to use inference acceleration frameworks to enhance VLM inference performance for faster inference.

2.2 Python Script Integration

The command line method is for quick testing and visualization. In actual projects, you usually need to integrate the model via code. You can perform pipeline inference with just a few lines of code as shown below:

from paddleocr import PaddleOCRVL

pipeline = PaddleOCRVL()
# pipeline = PaddleOCRVL(use_doc_orientation_classify=True) # Use use_doc_orientation_classify to enable/disable document orientation classification model
# pipeline = PaddleOCRVL(use_doc_unwarping=True) # Use use_doc_unwarping to enable/disable document unwarping module
# pipeline = PaddleOCRVL(use_layout_detection=False) # Use use_layout_detection to enable/disable layout detection module
output = pipeline.predict("./paddleocr_vl_demo.png")
for res in output:
    res.print() ## Print the structured prediction output
    res.save_to_json(save_path="output") ## Save the current image's structured result in JSON format
    res.save_to_markdown(save_path="output") ## Save the current image's result in Markdown format

For PDF files, each page will be processed individually and generate a separate Markdown file. If you want to convert the entire PDF to a single Markdown file, use the following method:

from pathlib import Path
from paddleocr import PaddleOCRVL

input_file = "./your_pdf_file.pdf"
output_path = Path("./output")

pipeline = PaddleOCRVL()
output = pipeline.predict(input=input_file)

markdown_list = []
markdown_images = []

for res in output:
    md_info = res.markdown
    markdown_list.append(md_info)
    markdown_images.append(md_info.get("markdown_images", {}))

markdown_texts = pipeline.concatenate_markdown_pages(markdown_list)

mkd_file_path = output_path / f"{Path(input_file).stem}.md"
mkd_file_path.parent.mkdir(parents=True, exist_ok=True)

with open(mkd_file_path, "w", encoding="utf-8") as f:
    f.write(markdown_texts)

for item in markdown_images:
    if item:
        for path, image in item.items():
            file_path = output_path / path
            file_path.parent.mkdir(parents=True, exist_ok=True)
            image.save(file_path)

Note:

• In the example code, the parameters use_doc_orientation_classify and use_doc_unwarping are all set to False by default. These indicate that document orientation classification and document image unwarping are disabled. You can manually set them to True if needed.

The above Python script performs the following steps:

(1) Instantiate the pipeline object. Specific parameter descriptions are as follows:

Parameter	Parameter Description	Parameter Type	Default Value
layout_detection_model_name	Name of the layout area detection and ranking model. If set to None, the default model of the pipeline will be used.	str|None	None
layout_detection_model_dir	Directory path of the layout area detection and ranking model. If set to None, the official model will be downloaded.	str|None	None
layout_threshold	Score threshold for the layout model. float: Any floating-point number between 0-1; dict: {0:0.1} The key is the class ID, and the value is the threshold for that class; None: If set to None, the parameter value initialized by the pipeline will be used.	float|dict|None	None
layout_nms	Whether to use post-processing NMS for layout detection. If set to None, the parameter value initialized by the pipeline will be used.	bool|None	None
layout_unclip_ratio	Expansion coefficient for the detection box of the layout area detection model. float: Any floating-point number greater than 0 Tuple[float,float]: The respective expansion coefficients in the horizontal and vertical directions; dict: where the key of the dict is of int type, representing cls_id, and the value is oftuple type, such as{0: (1.1, 2.0)}, indicating that the center of the detection box for class 0 output by the model remains unchanged, with the width expanded by 1.1 times and the height expanded by 2.0 times; None: If set to None, the parameter value initialized by the pipeline will be used.	float|Tuple[float,float]|dict|None	None
layout_merge_bboxes_mode	Merging mode for the detection boxes output by the model in layout detection. large when set to large, it means that among the detection boxes output by the model, for overlapping and contained boxes, only the outermost largest box is retained, and the overlapping inner boxes are deleted; small, when set to small, it means that among the detection boxes output by the model, for overlapping and contained boxes, only the innermost contained small box is retained, and the overlapping outer boxes are deleted; union,no filtering is performed on the boxes, and both inner and outer boxes are retained; If not set, the initialized parameter value will be used.	str|dict|None	None
vl_rec_model_name	Name of the multimodal recognition model. If not set, the default model will be used.	str|None	None
vl_rec_model_dir	Directory path of the multimodal recognition model. If not set, the official model will be downloaded.	str|None	None
vl_rec_backend	Inference backend used by the multimodal recognition model.	str|None	None
vl_rec_server_url	If the multimodal recognition model uses an inference service, this parameter is used to specify the server URL.	str|None	None
vl_rec_max_concurrency	If the multimodal recognition model uses an inference service, this parameter is used to specify the maximum number of concurrent requests.	str|None	None
vl_rec_api_key	If the multimodal recognition model uses an inference service, this parameter is used to specify the API key of the service.	str|None	None
doc_orientation_classify_model_name	Name of the document orientation classification model. If not set, the initialized default value will be used.	str|None	None
doc_orientation_classify_model_dir	Directory path of the document orientation classification model. If not set, the official model will be downloaded.	str|None	None
doc_unwarping_model_name	Name of the text image rectification model. If not set, the initialized default value will be used.	str|None	None
doc_unwarping_model_dir	Directory path of the text image rectification model. If not set, the official model will be downloaded.	str|None	None
use_doc_orientation_classify	Whether to load and use the document orientation classification module. If not set, the initialized default value will be used, which is initialized toFalse.	bool|None	None
use_doc_unwarping	Whether to load and use the text image rectification module. If not set, the initialized default value will be used, which is initialized to False.	bool|None	None
use_layout_detection	Whether to load and use the layout area detection and ranking module. If not set, the initialized default value will be used, which is initialized to True.	bool|None	None
use_chart_recognition	Whether to use the chart parsing function. If not set, the initialized default value will be used, which is initialized to False.	bool|None	None
format_block_content	Controls whether to format the block_content content within as Markdown. If not set, the initialized default value will be used, which defaults to initialization asFalse.	bool|None	None
device	The device used for inference. Supports specifying specific card numbers: CPU: For example,cpu indicates using the CPU for inference; GPU: For example,gpu:0 indicates using the first GPU for inference; NPU: For example,npu:0 indicates using the first NPU for inference; XPU: For example,xpu:0 indicates using the first XPU for inference; MLU: For example,mlu:0 indicates using the first MLU for inference; DCU: For example,dcu:0 indicates using the first DCU for inference; If not set, the initialized default value will be used. During initialization, the local GPU device 0 will be used preferentially. If it is not available, the CPU device will be used.	str|None	None
enable_hpi	Whether to enable high-performance inference.	bool	False
use_tensorrt	Whether to enable the TensorRT subgraph engine of Paddle Inference. If the model does not support acceleration via TensorRT, acceleration will not be used even if this flag is set. For PaddlePaddle version with CUDA 11.8, the compatible TensorRT version is 8.x (x>=6). It is recommended to install TensorRT 8.6.1.6.	bool	False
precision	Computational precision, such as fp32, fp16.	str	"fp32"
enable_mkldnn	Whether to enable MKL-DNN accelerated inference. If MKL-DNN is not available or the model does not support acceleration via MKL-DNN, acceleration will not be used even if this flag is set.	bool	True
mkldnn_cache_capacity	MKL-DNN cache capacity.	int	10
cpu_threads	The number of threads used for inference on the CPU.	int	8
paddlex_config	The file path for PaddleX pipeline configuration.	str	None

(2) Call the predict()method of the PaddleOCR-VL pipeline object for inference prediction. This method will return a list of results. Additionally, the pipeline also provides the predict_iter()Method. The two are completely consistent in terms of parameter acceptance and result return. The difference lies in that predict_iter()returns a generator, which can process and obtain prediction results step by step. It is suitable for scenarios involving large datasets or where memory conservation is desired. You can choose either of these two methods based on actual needs. Below are the parameters of the predict()method and their descriptions:

Parameter	Parameter Description	Parameter Type	Default Value
input	Data to be predicted, supporting multiple input types. Required. Python Var: such as numpy.ndarray representing image data str: such as the local path of an image file or PDF file: /root/data/img.jpg;such as a URL link, such as the network URL of an image file or PDF file:Example;such as a local directory, which should contain the images to be predicted, such as the local path: /root/data/(Currently, prediction for directories containing PDF files is not supported. PDF files need to be specified with a specific file path) list: List elements should be of the aforementioned data types, such as [numpy.ndarray, numpy.ndarray], ["/root/data/img1.jpg", "/root/data/img2.jpg"], ["/root/data1", "/root/data2"].	Python Var|str|list
use_doc_orientation_classify	Whether to use the document orientation classification module during inference. Setting it to None means using the instantiation parameter; otherwise, this parameter takes precedence.	bool|None	None
use_doc_unwarping	Whether to use the text image rectification module during inference. Setting it to None means using the instantiation parameter; otherwise, this parameter takes precedence.	bool|None	None
use_layout_detection	Whether to use the layout region detection and sorting module during inference. Setting it to None means using the instantiation parameter; otherwise, this parameter takes precedence.	bool|None	None
use_chart_recognition	Whether to use the chart parsing module during inference. Setting it to None means using the instantiation parameter; otherwise, this parameter takes precedence.	bool|None	None
layout_threshold	The parameter meaning is basically the same as the instantiation parameter. Setting it to None means using the instantiation parameter; otherwise, this parameter takes precedence.	float|dict|None	None
layout_nms	The parameter meaning is basically the same as the instantiation parameter. Setting it to None means using the instantiation parameter; otherwise, this parameter takes precedence.	bool|None	None
layout_unclip_ratio	The parameter meaning is basically the same as the instantiation parameter. Setting it to None means using the instantiation parameter; otherwise, this parameter takes precedence.	float|Tuple[float,float]|dict|None	None
layout_merge_bboxes_mode	The parameter meaning is basically the same as the instantiation parameter. Setting it to None means using the instantiation parameter; otherwise, this parameter takes precedence.	str|dict|None	None
use_queues	Used to control whether to enable internal queues. When set to True, data loading (such as rendering PDF pages as images), layout detection model processing, and VLM inference will be executed asynchronously in separate threads, with data passed through queues, thereby improving efficiency. This approach is particularly efficient for PDF documents with many pages or directories containing a large number of images or PDF files.	bool|None	None
prompt_label	The prompt type setting for the VL model, which takes effect only when use_layout_detection=False. The fillable parameters are ocr、formula、table and chart.	str|None	None
format_block_content	The parameter meaning is basically the same as the instantiation parameter. Setting it to None means using the instantiation parameter; otherwise, this parameter takes precedence.	bool|None	None
repetition_penalty	The repetition penalty parameter used for VL model sampling.	float|None	None
temperature	Temperature parameter used for VL model sampling.	float|None	None
top_p	Top-p parameter used for VL model sampling.	float|None	None
min_pixels	The minimum number of pixels allowed when the VL model preprocesses images.	int|None	None
max_pixels	The maximum number of pixels allowed when the VL model preprocesses images.	int|None	None

(3) Process the prediction results: The prediction result for each sample is a corresponding Result object, supporting operations such as printing, saving as an image, and saving as a json file:

Method	Method Description	Parameter	Parameter Type	Parameter Description	Default Value
print()	Print results to the terminal	format_json	bool	Whether to format the output content using JSON indentation.	True
		indent	int	Specify the indentation level to beautify the output JSON data, making it more readable. Only valid when format_json is True.	4
		ensure_ascii	bool	Control whether non- ASCII characters are escaped as Unicode. When set to True, all non- ASCII characters will be escaped; False retains the original characters. Only valid when format_json is True.	False
save_to_json()	Save the results as a json format file	save_path	str	The file path for saving. When it is a directory, the saved file name will be consistent with the input file type naming.	None
		indent	int	Specify the indentation level to beautify the output JSONdata, making it more readable. Only valid when format_jsonis True.	4
		ensure_ascii	bool	Control whether non- ASCII characters are escaped as Unicode. When set to True, all non- ASCII characters will be escaped; False retains the original characters. Only valid when format_json is True.	False
save_to_img()	Save the visualized images of each intermediate module in png format	save_path	str	The file path for saving, supporting directory or file paths.	None
save_to_markdown()	Save each page in an image or PDF file as a markdown format file separately	save_path	str	The file path for saving. When it is a directory, the saved file name will be consistent with the input file type naming	None
		pretty	bool	Whether to beautify the markdown output results, centering charts, etc., to make the markdown rendering more aesthetically pleasing.	True
		show_formula_number	bool	Control whether to retain formula numbers in markdown. When set to True, all formula numbers are retained; False retains only the formulas	False
save_to_html()	Save the tables in the file as html format files	save_path	str	The file path for saving, supporting directory or file paths.	None
save_to_xlsx()	Save the tables in the file as xlsx format files	save_path	str	The file path for saving, supporting directory or file paths.	None

- Calling the `print()` method will print the results to the terminal. The content printed to the terminal is explained as follows: - `input_path`: `(str)` The input path of the image or PDF to be predicted. - `page_index`: `(Union[int, None])` If the input is a PDF file, it indicates the current page number of the PDF; otherwise, it is `None`. - `model_settings`: `(Dict[str, bool])` Model parameters required for configuring PaddleOCR-VL. - `use_doc_preprocessor`: `(bool)` Controls whether to enable the document preprocessing sub-pipeline. - `use_layout_detection`: `(bool)` Controls whether to enable the layout detection module. - `use_chart_recognition`: `(bool)` Controls whether to enable the chart recognition function. - `format_block_content`: `(bool)` Controls whether to save the formatted markdown content in `JSON`. - `doc_preprocessor_res`: `(Dict[str, Union[List[float], str]])` A dictionary of document preprocessing results, which exists only when `use_doc_preprocessor=True`. - `input_path`: `(str)` The image path accepted by the document preprocessing sub-pipeline. When the input is a `numpy.ndarray`, it is saved as `None`; here, it is `None`. - `page_index`: `None`. Since the input here is a `numpy.ndarray`, the value is `None`. - `model_settings`: `(Dict[str, bool])` Model configuration parameters for the document preprocessing sub-pipeline. - `use_doc_orientation_classify`: `(bool)` Controls whether to enable the document image orientation classification sub-module. - `use_doc_unwarping`: `(bool)` Controls whether to enable the text image distortion correction sub-module. - `angle`: `(int)` The prediction result of the document image orientation classification sub-module. When enabled, it returns the actual angle value. - `parsing_res_list`: `(List[Dict])` A list of parsing results, where each element is a dictionary. The list order is the reading order after parsing. - `block_bbox`: `(np.ndarray)` The bounding box of the layout area. - `block_label`: `(str)` The label of the layout area, such as `text`, `table`, etc. - `block_content`: `(str)` The content within the layout area. - `block_id`: `(int)` The index of the layout area, used to display the layout sorting results. - `block_order` `(int)` The order of the layout area, used to display the layout reading order. For non-sorted parts, the default value is `None`. - Calling the `save_to_json()` method will save the above content to the specified `save_path`. If a directory is specified, the saved path will be `save_path/{your_img_basename}_res.json`. If a file is specified, it will be saved directly to that file. Since json files do not support saving numpy arrays, the `numpy.array` types within will be converted to list form. - `input_path`: `(str)` The input path of the image or PDF to be predicted. - `page_index`: `(Union[int, None])` If the input is a PDF file, it indicates the current page number of the PDF; otherwise, it is `None`. - `model_settings`: `(Dict[str, bool])` Model parameters required for configuring PaddleOCR-VL. - `use_doc_preprocessor`: `(bool)` Controls whether to enable the document preprocessing sub-pipeline. - `use_layout_detection`: `(bool)` Controls whether to enable the layout detection module. - `use_chart_recognition`: `(bool)` Controls whether to enable the chart recognition function. - `format_block_content`: `(bool)` Controls whether to save the formatted markdown content in `JSON`. - `doc_preprocessor_res`: `(Dict[str, Union[List[float], str]])` A dictionary of document preprocessing results, which exists only when `use_doc_preprocessor=True`. - `input_path`: `(str)` The image path accepted by the document preprocessing sub-pipeline. When the input is a `numpy.ndarray`, it is saved as `None`; here, it is `None`. - `page_index`: `None`. Since the input here is a `numpy.ndarray`, the value is `None`. - `model_settings`: `(Dict[str, bool])` Model configuration parameters for the document preprocessing sub-pipeline. - `use_doc_orientation_classify`: `(bool)` Controls whether to enable the document image orientation classification sub-module. - `use_doc_unwarping`: `(bool)` Controls whether to enable the text image distortion correction sub-module. - `angle`: `(int)` The prediction result of the document image orientation classification sub-module. When enabled, it returns the actual angle value. - `parsing_res_list`: `(List[Dict])` A list of parsing results, where each element is a dictionary. The list order represents the reading order after parsing. - `block_bbox`: `(np.ndarray)` The bounding box of the layout region. - `block_label`: `(str)` The label of the layout region, such as `text`, `table`, etc. - `block_content`: `(str)` The content within the layout region. - `block_id`: `(int)` The index of the layout region, used to display the layout sorting results. - `block_order` `(int)` The order of the layout region, used to display the layout reading order. For non-sorted parts, the default value is `None`. - Calling the `save_to_img()` method will save the visualization results to the specified `save_path`. If a directory is specified, visualized images for layout region detection, global OCR, layout reading order, etc., will be saved. If a file is specified, it will be saved directly to that file. (Pipelines typically contain many result images, so it is not recommended to directly specify a specific file path, as multiple images will be overwritten, retaining only the last one.) - Calling the `save_to_markdown()` method will save the converted Markdown file to the specified `save_path`. The saved file path will be `save_path/{your_img_basename}.md`. If the input is a PDF file, it is recommended to directly specify a directory; otherwise, multiple markdown files will be overwritten. Additionally, it also supports obtaining visualized images and prediction results with results through attributes, as follows:

Attribute	Attribute Description
json	Obtain the prediction jsonresult in the format
img	obtain in the format of dictvisualized image
markdown	obtain in the format of dictmarkdown result

- The prediction result obtained through the `json` attribute is data of dict type, with relevant content consistent with that saved by calling the `save_to_json()` method. - The prediction result returned by the `img` attribute is data of dict type. The keys are `layout_det_res`, `overall_ocr_res`, `text_paragraphs_ocr_res`, `formula_res_region1`, `table_cell_img`, and `seal_res_region1`, with corresponding values being `Image.Image` objects: used to display visualized images of layout region detection, OCR, OCR text paragraphs, formulas, tables, and seal results, respectively. If optional modules are not used, the dict only contains `layout_det_res`. - The prediction result returned by the `markdown` attribute is data of dict type. The keys are `markdown_texts`, `markdown_images`, and `page_continuation_flags`, with corresponding values being markdown text, images displayed in Markdown (`Image.Image` objects), and a bool tuple used to identify whether the first element on the current page is the start of a paragraph and whether the last element is the end of a paragraph, respectively.

3. Enhancing VLM Inference Performance Using Inference Acceleration Frameworks

The inference performance under default configurations is not fully optimized and may not meet actual production requirements. This step primarily introduces how to use the vLLM, SGLang and FastDeploy inference acceleration frameworks to enhance the inference performance of PaddleOCR-VL.

3.1 Launching the VLM Inference Service

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

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

• Method 2: Launch the service by manually installing dependencies via the PaddleOCR CLI.

3.1.1 Method 1: Using Docker Image

PaddleOCR provides Docker images for quickly launching vLLM or FastDeploy inference services. You can use the following commands to start the services (requires Docker version >= 19.03, a machine equipped with a GPU, and NVIDIA drivers supporting CUDA 12.6 or later):

Launch vLLM ServiceLaunch FastDeploy Service

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

If you wish to start the service in an environment without internet access, replace ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-genai-vllm-server:latest (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-offline (image size approximately 15 GB).

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

If you wish to start the service in an environment without internet access, replace ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-genai-fastdeploy-server:latest (image size approximately 43 GB) in the above command with the offline version image ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-genai-fastdeploy-server:latest-offline (image size approximately 45 GB).

When starting the vLLM or FastDeploy inference service, we provide a set of default parameter settings. If you have additional requirements for adjusting parameters such as GPU memory usage, you can configure more parameters yourself. Please refer to 3.3.1 Server-side Parameter Adjustment to create a configuration file, then mount this file into the container, and specify the configuration file using backend_config in the command to start the service. Taking vLLM as an 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 \
    paddleocr genai_server --model_name PaddleOCR-VL-0.9B --host 0.0.0.0 --port 8118 --backend vllm --backend_config /tmp/vllm_config.yml

3.1.2 Method 2: Installation and Usage via PaddleOCR CLI

Since inference acceleration frameworks may have dependency conflicts with the PaddlePaddle framework, it is recommended to install them in a virtual environment. Taking vLLM as an example:

# If there is an active virtual environment currently, deactivate it first using `deactivate`
# Create a virtual environment
python -m venv .venv_vlm
# Activate the environment
source .venv_vlm/bin/activate
# Install PaddleOCR
python -m pip install "paddleocr[doc-parser]"
# Install dependencies for the inference acceleration service
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, sglang and fastdeploy, corresponding to vLLM, SGLang and FastDeploy, respectively.

The vLLM and SGLang installed via paddleocr install_genai_server_deps are both CUDA 12.6 versions; ensure that your local NVIDIA drivers are consistent with or higher than this version.

The paddleocr install_genai_server_deps command may require CUDA compilation tools such as nvcc during execution. If these tools are not available in your environment (e.g., when using the paddleocr-vl image), you can obtain a precompiled version of FlashAttention from this repository. Install the precompiled package before executing subsequent commands. For example, if you are in the paddleocr-vl image, execute python -m pip install https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.3.14/flash_attn-2.8.2+cu128torch2.8-cp310-cp310-linux_x86_64.whl.

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

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

The parameters supported by this command are as follows:

Parameter	Description
--model_name	Model name
--model_dir	Model directory
--host	Server hostname
--port	Server port number
--backend	Backend name, i.e., the name of the inference acceleration framework used; options are vllm or sglang
--backend_config	Can specify a YAML file containing backend configurations

3.2 Client Usage Methods

After launching the VLM inference service, the client can call the service through PaddleOCR. Please note that because the client needs to invoke the sequential model for layout detection, it is still recommended to run the client on GPU or other acceleration devices to achieve more stable and efficient performance.

3.2.1 CLI Invocation

Specify the backend type (vllm-server or sglang-server) using --vl_rec_backend and the service address using --vl_rec_server_url, for example:

paddleocr doc_parser --input paddleocr_vl_demo.png --vl_rec_backend vllm-server --vl_rec_server_url http://127.0.0.1:8118/v1

3.2.2 Python API Invocation

Pass the vl_rec_backend and vl_rec_server_url parameters when creating a PaddleOCRVL object:

pipeline = PaddleOCRVL(vl_rec_backend="vllm-server", vl_rec_server_url="http://127.0.0.1:8118/v1")

3.3 Performance Tuning

The default configurations are optimized for single NVIDIA A100 GPUs with exclusive client access and may not be suitable for other environments. If users encounter performance issues in actual use, the following optimization methods can be attempted.

3.3.1 Server-Side Parameter Adjustment

Different inference acceleration frameworks support different parameters. Refer to their official documentation for available parameters and adjustment timing:

• vLLM Official Parameter Tuning Guide
• SGLang Hyperparameter Tuning Documentation
• FastDeploy Best Practices

The PaddleOCR VLM inference service supports parameter tuning through configuration files. The following example shows how to adjust the gpu-memory-utilization and max-num-seqs parameters for the vLLM server:

1. Create a YAML file vllm_config.yaml with the following content:

gpu-memory-utilization: 0.3
max-num-seqs: 128

1. Specify the configuration file path when starting the service, for example, using the paddleocr genai_server command:

paddleocr genai_server --model_name PaddleOCR-VL-0.9B --backend vllm --backend_config vllm_config.yaml

If using a shell that supports process substitution (like Bash), you can also pass configuration items directly without creating a configuration file:

paddleocr genai_server --model_name PaddleOCR-VL-0.9B --backend vllm --backend_config <(echo -e 'gpu-memory-utilization: 0.3\nmax-num-seqs: 128')

3.3.2 Client-Side Parameter Adjustment

PaddleOCR groups sub-images from single or multiple input images and sends concurrent requests to the server, so the number of concurrent requests significantly impacts performance.

• For CLI and Python API, adjust the maximum number of concurrent requests using the vl_rec_max_concurrency parameter;
• For service deployment, modify the VLRecognition.genai_config.max_concurrency field in the configuration file.

When there is a 1:1 client-to-VLM inference service ratio and sufficient server resources, increasing concurrency can improve performance. If the server needs to support multiple clients or has limited computing resources, reduce concurrency to avoid resource overload and service abnormalities.

3.3.3 Common Hardware Performance Tuning Recommendations

The following configurations are for scenarios with a 1:1 client-to-VLM inference service ratio.

NVIDIA RTX 3060

• Server-Side
  • vLLM: gpu-memory-utilization: 0.8
  • FastDeploy:
    • gpu-memory-utilization: 0.8
    • max-concurrency: 2048

4. Service Deployment

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

• Method 1: Deploy using Docker Compose (recommended).

• Method 2: Manual Deployment.

Note that the PaddleOCR-VL service described 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 (Recommended)

You can obtain the Compose file and the environment variables configuration file from here and here, respectively, and download them to your local machine. Then, in the directory where the files were just downloaded, execute the following command to start the server, which will listen on port 8080 by default:

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

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

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 solution accelerates VLM inference based on frameworks like vLLM, making it more suitable for production environment deployment. However, it requires the machine to be equipped with a GPU and the NVIDIA driver to support CUDA 12.6 or higher.

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

Docker Compose starts two containers in sequence by reading the configurations in the .env and compose.yaml files, running the underlying VLM inference service and the PaddleOCR-VL service (Pipeline) 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 start the pipeline service. The default is latest-offline, indicating the use of an offline GPU image.
• VLM_BACKEND: The VLM inference backend, currently supporting vllm and fastdeploy. The default is vllm.
• VLM_IMAGE_TAG_SUFFIX: The tag suffix of the image used to start the VLM inference service. The default is latest-offline, indicating the use of an offline GPU image.

You can meet custom requirements by modifying .env and compose.yaml, for example:

1. Change the port of the PaddleOCR-VL service

Edit paddleocr-vl-api.ports in the compose.yaml 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
    ...

2. Specify the GPU used by the PaddleOCR-VL service

Edit device_ids in the compose.yaml file to change the GPU used. For example, if you need to use GPU 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]
    ...

3. Adjust VLM server-side configuration

If you want to adjust the VLM server-side configuration, please refer to 3.3.1 Server-side Parameter Adjustment to generate a configuration file. After generating the configuration file, add the following paddleocr-vlm-server.volumes and paddleocr-vlm-server.command fields to your compose.yaml. Please replace /path/to/your_config.yaml 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-0.9B --host 0.0.0.0 --port 8118 --backend vllm --backend_config /home/paddleocr/vlm_server_config.yaml
    ...

4. Change the VLM inference backend

Modify VLM_BACKEND in the .env file, for example, to change the VLM inference backend to fastdeploy:

  API_IMAGE_TAG_SUFFIX=latest-offline
- VLM_BACKEND=vllm
+ VLM_BACKEND=fastdeploy
  VLM_IMAGE_TAG_SUFFIX=latest-offline

5. Adjust pipeline configurations (such as model path, batch size, deployment device, etc.)

Refer to section 4.4 Pipeline Configuration Adjustment Instructions in this document.

4.2 Method 2: Manual Deployment

Execute the following command to install the service deployment plugin via the PaddleX CLI:

paddlex --install serving

Then, start the server using the PaddleX CLI:

paddlex --serve --pipeline PaddleOCR-VL

After startup, you will see output similar to the following, with the server listening on port 8080 by default:

INFO:     Started server process [63108]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://0.0.0.0:8080 (Press CTRL+C to quit)

The command-line options related to serving are as follows:

Name	Description
--pipeline	PaddleX pipeline registration name or pipeline configuration file path.
--device	Deployment device for the pipeline. By default, a GPU will be used if available; otherwise, a CPU will be used."
--host	Hostname or IP address to which the server is bound. Defaults to 0.0.0.0.
--port	Port number on which the server listens. Defaults to 8080.
--use_hpip	If specified, uses high-performance inference. Refer to the High-Performance Inference documentation for more information.
--hpi_config	High-performance inference configuration. Refer to the High-Performance Inference documentation for more information.

If you need to adjust pipeline configurations (such as model path, batch size, deployment device, etc.), you can specify the --pipeline parameter as a custom configuration file path. For the correspondence between PaddleOCR pipelines and PaddleX pipeline registration names, as well as how to obtain and modify PaddleX pipeline configuration files, please refer to PaddleOCR and PaddleX. Furthermore, section 4.1.3 will introduce how to adjust the pipeline configuration based on common requirements.

4.3 Client-Side Invocation

Below are the API reference and examples of multi-language service invocation:

API Reference

Main operations provided by the service:

• The HTTP request method is POST.
• Both the request body and response body are JSON data (JSON objects).
• When the request is processed successfully, the response status code is200, and the properties of the response body are as follows:

Name	Type	Meaning
logId	string	The UUID of the request.
errorCode	integer	Error code. Fixed as 0.
errorMsg	string	Error description. Fixed as "Success".
result	object	Operation result.

• When the request is not processed successfully, the properties of the response body are as follows:

Name	Type	Meaning
logId	string	The UUID of the request.
errorCode	integer	Error code. Same as the response status code.
errorMsg	string	Error description.

The main operations provided by the service are as follows:

• infer

Perform layout parsing.

POST /layout-parsing

• The properties of the request body are as follows:

Name	Type	Meaning	Required
file	string	The URL of an image file or PDF file accessible to the server, or the Base64-encoded result of the content of the aforementioned file types. By default, for PDF files with more than 10 pages, only the first 10 pages will be processed. To remove the page limit, add the following configuration to the pipeline configuration file: Serving: extra: max_num_input_imgs: null	Yes
fileType	integer|null	File type.0 represents a PDF file,1 represents an image file. If this property is not present in the request body, the file type will be inferred from the URL.	No
useDocOrientationClassify	boolean | null	Please refer to the description of the use_doc_orientation_classify parameter in the predict method of the PaddleOCR-VL object.	No
useDocUnwarping	boolean|null	Please refer to the description of the use_doc_unwarping parameter in the predict method of the PaddleOCR-VL object.	No
useLayoutDetection	boolean|null	Please refer to the description of the use_layout_detection parameter in the predict method of the PaddleOCR-VL object.	No
useChartRecognition	boolean|null	Please refer to the description of the use_chart_recognition parameter in the predict method of the PaddleOCR-VL object.	No
layoutThreshold	number|object|null	Please refer to the description of the layout_threshold parameter in the predict method of the PaddleOCR-VL object.	No
layoutNms	boolean|null	Please refer to the description of the layout_nms parameter in the predict method of the PaddleOCR-VL object.	No
layoutUnclipRatio	number|array|object|null	Please refer to the description of the layout_unclip_ratio parameter in the predict method of the PaddleOCR-VL object.	No
layoutMergeBboxesMode	string|object|null	Please refer to the description of the layout_merge_bboxes_mode parameter in the predict method of the PaddleOCR-VL object.	No
promptLabel	string|null	Please refer to the description of the prompt_label parameter in the predict method of the PaddleOCR-VL object.	No
formatBlockContent	boolean|null	Please refer to the description of the format_block_content parameter in the predict method of the PaddleOCR-VL object.	No
repetitionPenalty	number|null	Please refer to the description of the repetition_penalty parameter in the predict method of the PaddleOCR-VL object.	No
temperature	number|null	Please refer to the description of the temperature parameter in the predict method of the PaddleOCR-VL object.	No
topP	number|null	Please refer to the description of the top_p parameter in the predict method of the PaddleOCR-VL object.	No
minPixels	number|null	Please refer to the description of the min_pixels parameter in the predict method of the PaddleOCR-VL object.	No
maxPixels	number|null	Please refer to the description of the max_pixels parameter in the predict method of the PaddleOCR-VL object.	No
prettifyMarkdown	boolean	Whether to output beautified Markdown text. The default is true.	No
showFormulaNumber	boolean	Whether to include formula numbers in the output Markdown text. The default is false.	No
visualize	boolean|null	Whether to return visualization result images and intermediate images during the processing. Pass true: Return images. Pass false: Do not return images. If this parameter is not provided in the request body or null is passed: Follow the setting in the configuration file Serving.visualize. For example, add the following field in the configuration file: Serving: visualize: False Images will not be returned by default, and the default behavior can be overridden by the visualize parameter in the request body. If this parameter is not set in either the request body or the configuration file (or null is passed in the request body and the configuration file is not set), images will be returned by default.	No

• When the request is processed successfully, the result in the response body has the following attributes:

Name	Type	Meaning
layoutParsingResults	array	Layout parsing results. The array length is 1 (for image input) or the actual number of document pages processed (for PDF input). For PDF input, each element in the array represents the result of each actual page processed in the PDF file.
dataInfo	object	Input data information.

Each element inlayoutParsingResults is an object with the following attributes:

Meaning	Name	Type
prunedResult	object	A simplified version of the res field in the JSON representation of the results generated by the predict method of the object, with the input_path and page_index fields removed.
markdown	object	Markdown results.
outputImages	object|null	Refer to the img property description of the prediction results. The image is in JPEG format and encoded using Base64.
inputImage	string|null	Input image. The image is in JPEG format and encoded using Base64.

markdownis an objectwith the following properties:

Name	Type	Meaning
text	string	Markdown text.
images	object	Key-value pairs of relative paths to Markdown images and Base64-encoded images.
isStart	boolean	Whether the first element on the current page is the start of a paragraph.
isEnd	boolean	Whether the last element on the current page is the end of a paragraph.

Multi-Language Service Invocation Examples

Python

import base64
import requests
import pathlib

API_URL = "http://localhost:8080/layout-parsing" # Service URL

image_path = "./demo.jpg"

# Encode the local image in Base64
with open(image_path, "rb") as file:
    image_bytes = file.read()
    image_data = base64.b64encode(image_bytes).decode("ascii")

payload = {
    "file": image_data, # Base64-encoded file content or file URL
    "fileType": 1, # File type, 1 indicates an image file
}

# Call the API
response = requests.post(API_URL, json=payload)

# Process the returned data from the interface
assert response.status_code == 200
result = response.json()["result"]
for i, res in enumerate(result["layoutParsingResults"]):
    print(res["prunedResult"])
    md_dir = pathlib.Path(f"markdown_{i}")
    md_dir.mkdir(exist_ok=True)
    (md_dir / "doc.md").write_text(res["markdown"]["text"])
    for img_path, img in res["markdown"]["images"].items():
        img_path = md_dir / img_path
        img_path.parent.mkdir(parents=True, exist_ok=True)
        img_path.write_bytes(base64.b64decode(img))
    print(f"Markdown document saved at {md_dir / 'doc.md'}")
    for img_name, img in res["outputImages"].items():
        img_path = f"{img_name}_{i}.jpg"
        pathlib.Path(img_path).parent.mkdir(exist_ok=True)
        with open(img_path, "wb") as f:
            f.write(base64.b64decode(img))
        print(f"Output image saved at {img_path}")

C++

#include <iostream>
#include <filesystem>
#include <fstream>
#include <vector>
#include <string>
#include "cpp-httplib/httplib.h" // https://github.com/Huiyicc/cpp-httplib
#include "nlohmann/json.hpp" // https://github.com/nlohmann/json
#include "base64.hpp" // https://github.com/tobiaslocker/base64

namespace fs = std::filesystem;

int main() {
    httplib::Client client("localhost", 8080);

    const std::string filePath = "./demo.jpg";

    std::ifstream file(filePath, std::ios::binary | std::ios::ate);
    if (!file) {
        std::cerr << "Error opening file: " << filePath << std::endl;
        return 1;
    }

    std::streamsize size = file.tellg();
    file.seekg(0, std::ios::beg);
    std::vector buffer(size);
    if (!file.read(buffer.data(), size)) {
        std::cerr << "Error reading file." << std::endl;
        return 1;
    }

    std::string bufferStr(buffer.data(), static_cast(size));
    std::string encodedFile = base64::to_base64(bufferStr);

    nlohmann::json jsonObj;
    jsonObj["file"] = encodedFile;
    jsonObj["fileType"] = 1;

    auto response = client.Post("/layout-parsing", jsonObj.dump(), "application/json");

    if (response && response->status == 200) {
        nlohmann::json jsonResponse = nlohmann::json::parse(response->body);
        auto result = jsonResponse["result"];

        if (!result.is_object() || !result.contains("layoutParsingResults")) {
            std::cerr << "Unexpected response format." << std::endl;
            return 1;
        }

        const auto& results = result["layoutParsingResults"];
        for (size_t i = 0; i < results.size(); ++i) {
            const auto& res = results[i];

            if (res.contains("prunedResult")) {
                std::cout << "Layout result [" << i << "]: " << res["prunedResult"].dump() << std::endl;
            }

            if (res.contains("outputImages") && res["outputImages"].is_object()) {
                for (auto& [imgName, imgBase64] : res["outputImages"].items()) {
                    std::string outputPath = imgName + "_" + std::to_string(i) + ".jpg";
                    fs::path pathObj(outputPath);
                    fs::path parentDir = pathObj.parent_path();
                    if (!parentDir.empty() && !fs::exists(parentDir)) {
                        fs::create_directories(parentDir);
                    }

                    std::string decodedImage = base64::from_base64(imgBase64.get());

                    std::ofstream outFile(outputPath, std::ios::binary);
                    if (outFile.is_open()) {
                        outFile.write(decodedImage.c_str(), decodedImage.size());
                        outFile.close();
                        std::cout << "Saved image: " << outputPath << std::endl;
                    } else {
                        std::cerr << "Failed to save image: " << outputPath << std::endl;
                    }
                }
            }
        }
    } else {
        std::cerr << "Request failed." << std::endl;
        if (response) {
            std::cerr << "HTTP status: " << response->status << std::endl;
            std::cerr << "Response body: " << response->body << std::endl;
        }
        return 1;
    }

    return 0;
}

Java

import okhttp3.*;
import com.fasterxml.jackson.databind.ObjectMapper;
import com.fasterxml.jackson.databind.JsonNode;
import com.fasterxml.jackson.databind.node.ObjectNode;

import java.io.File;
import java.io.FileOutputStream;
import java.io.IOException;
import java.util.Base64;
import java.nio.file.Paths;
import java.nio.file.Files;

public class Main {
    public static void main(String[] args) throws IOException {
        String API_URL = "http://localhost:8080/layout-parsing";
        String imagePath = "./demo.jpg";

        File file = new File(imagePath);
        byte[] fileContent = java.nio.file.Files.readAllBytes(file.toPath());
        String base64Image = Base64.getEncoder().encodeToString(fileContent);

        ObjectMapper objectMapper = new ObjectMapper();
        ObjectNode payload = objectMapper.createObjectNode();
        payload.put("file", base64Image);
        payload.put("fileType", 1);

        OkHttpClient client = new OkHttpClient();
        MediaType JSON = MediaType.get("application/json; charset=utf-8");

        RequestBody body = RequestBody.create(JSON, payload.toString());

        Request request = new Request.Builder()
                .url(API_URL)
                .post(body)
                .build();

        try (Response response = client.newCall(request).execute()) {
            if (response.isSuccessful()) {
                String responseBody = response.body().string();
                JsonNode root = objectMapper.readTree(responseBody);
                JsonNode result = root.get("result");

                JsonNode layoutParsingResults = result.get("layoutParsingResults");
                for (int i = 0; i < layoutParsingResults.size(); i++) {
                    JsonNode item = layoutParsingResults.get(i);
                    int finalI = i;
                    JsonNode prunedResult = item.get("prunedResult");
                    System.out.println("Pruned Result [" + i + "]: " + prunedResult.toString());

                    JsonNode outputImages = item.get("outputImages");
                    outputImages.fieldNames().forEachRemaining(imgName -> {
                        try {
                            String imgBase64 = outputImages.get(imgName).asText();
                            byte[] imgBytes = Base64.getDecoder().decode(imgBase64);
                            String imgPath = imgName + "_" + finalI + ".jpg";

                            File outputFile = new File(imgPath);
                            File parentDir = outputFile.getParentFile();
                            if (parentDir != null && !parentDir.exists()) {
                                parentDir.mkdirs();
                                System.out.println("Created directory: " + parentDir.getAbsolutePath());
                            }

                            try (FileOutputStream fos = new FileOutputStream(outputFile)) {
                                fos.write(imgBytes);
                                System.out.println("Saved image: " + imgPath);
                            }
                        } catch (IOException e) {
                            System.err.println("Failed to save image: " + e.getMessage());
                        }
                    });
                }
            } else {
                System.err.println("Request failed with HTTP code: " + response.code());
            }
        }
    }
}

Go

package main

import (
    "bytes"
    "encoding/base64"
    "encoding/json"
    "fmt"
    "io/ioutil"
    "net/http"
    "os"
    "path/filepath"
)

func main() {
    API_URL := "http://localhost:8080/layout-parsing"
    filePath := "./demo.jpg"

    fileBytes, err := ioutil.ReadFile(filePath)
    if err != nil {
        fmt.Printf("Error reading file: %v\n", err)
        return
    }
    fileData := base64.StdEncoding.EncodeToString(fileBytes)

    payload := map[string]interface{}{
        "file":     fileData,
        "fileType": 1,
    }
    payloadBytes, err := json.Marshal(payload)
    if err != nil {
        fmt.Printf("Error marshaling payload: %v\n", err)
        return
    }

    client := &http.Client{}
    req, err := http.NewRequest("POST", API_URL, bytes.NewBuffer(payloadBytes))
    if err != nil {
        fmt.Printf("Error creating request: %v\n", err)
        return
    }
    req.Header.Set("Content-Type", "application/json")

    res, err := client.Do(req)
    if err != nil {
        fmt.Printf("Error sending request: %v\n", err)
        return
    }
    defer res.Body.Close()

    if res.StatusCode != http.StatusOK {
        fmt.Printf("Unexpected status code: %d\n", res.StatusCode)
        return
    }

    body, err := ioutil.ReadAll(res.Body)
    if err != nil {
        fmt.Printf("Error reading response: %v\n", err)
        return
    }

    type Markdown struct {
        Text   string            `json:"text"`
        Images map[string]string `json:"images"`
    }

    type LayoutResult struct {
        PrunedResult map[string]interface{} `json:"prunedResult"`
        Markdown     Markdown               `json:"markdown"`
        OutputImages map[string]string      `json:"outputImages"`
        InputImage   *string                `json:"inputImage"`
    }

    type Response struct {
        Result struct {
            LayoutParsingResults []LayoutResult `json:"layoutParsingResults"`
            DataInfo             interface{}    `json:"dataInfo"`
        } `json:"result"`
    }

    var respData Response
    if err := json.Unmarshal(body, &respData); err != nil {
        fmt.Printf("Error parsing response: %v\n", err)
        return
    }

    for i, res := range respData.Result.LayoutParsingResults {
        fmt.Printf("Result %d - prunedResult: %+v\n", i, res.PrunedResult)

        mdDir := fmt.Sprintf("markdown_%d", i)
        os.MkdirAll(mdDir, 0755)
        mdFile := filepath.Join(mdDir, "doc.md")
        if err := os.WriteFile(mdFile, []byte(res.Markdown.Text), 0644); err != nil {
            fmt.Printf("Error writing markdown file: %v\n", err)
        } else {
            fmt.Printf("Markdown document saved at %s\n", mdFile)
        }

        for path, imgBase64 := range res.Markdown.Images {
            fullPath := filepath.Join(mdDir, path)
            if err := os.MkdirAll(filepath.Dir(fullPath), 0755); err != nil {
                fmt.Printf("Error creating directory for markdown image: %v\n", err)
                continue
            }
            imgBytes, err := base64.StdEncoding.DecodeString(imgBase64)
            if err != nil {
                fmt.Printf("Error decoding markdown image: %v\n", err)
                continue
            }
            if err := os.WriteFile(fullPath, imgBytes, 0644); err != nil {
                fmt.Printf("Error saving markdown image: %v\n", err)
            }
        }

        for name, imgBase64 := range res.OutputImages {
            imgBytes, err := base64.StdEncoding.DecodeString(imgBase64)
            if err != nil {
                fmt.Printf("Error decoding output image %s: %v\n", name, err)
                continue
            }
            filename := fmt.Sprintf("%s_%d.jpg", name, i)

            if err := os.MkdirAll(filepath.Dir(filename), 0755); err != nil {
                fmt.Printf("Error creating directory for output image: %v\n", err)
                continue
            }

            if err := os.WriteFile(filename, imgBytes, 0644); err != nil {
                fmt.Printf("Error saving output image %s: %v\n", filename, err)
            } else {
                fmt.Printf("Output image saved at %s\n", filename)
            }
        }
    }
}

C#

using System;
using System.IO;
using System.Net.Http;
using System.Text;
using System.Threading.Tasks;
using Newtonsoft.Json.Linq;

class Program
{
    static readonly string API_URL = "http://localhost:8080/layout-parsing";
    static readonly string inputFilePath = "./demo.jpg";

    static async Task Main(string[] args)
    {
        var httpClient = new HttpClient();

        byte[] fileBytes = File.ReadAllBytes(inputFilePath);
        string fileData = Convert.ToBase64String(fileBytes);

        var payload = new JObject
        {
            { "file", fileData },
            { "fileType", 1 }
        };
        var content = new StringContent(payload.ToString(), Encoding.UTF8, "application/json");

        HttpResponseMessage response = await httpClient.PostAsync(API_URL, content);
        response.EnsureSuccessStatusCode();

        string responseBody = await response.Content.ReadAsStringAsync();
        JObject jsonResponse = JObject.Parse(responseBody);

        JArray layoutParsingResults = (JArray)jsonResponse["result"]["layoutParsingResults"];
        for (int i = 0; i < layoutParsingResults.Count; i++)
        {
            var res = layoutParsingResults[i];
            Console.WriteLine($"[{i}] prunedResult:\n{res["prunedResult"]}");

            JObject outputImages = res["outputImages"] as JObject;
            if (outputImages != null)
            {
                foreach (var img in outputImages)
                {
                    string imgName = img.Key;
                    string base64Img = img.Value?.ToString();
                    if (!string.IsNullOrEmpty(base64Img))
                    {
                        string imgPath = $"{imgName}_{i}.jpg";
                        byte[] imageBytes = Convert.FromBase64String(base64Img);

                        string directory = Path.GetDirectoryName(imgPath);
                        if (!string.IsNullOrEmpty(directory) && !Directory.Exists(directory))
                        {
                            Directory.CreateDirectory(directory);
                            Console.WriteLine($"Created directory: {directory}");
                        }

                        File.WriteAllBytes(imgPath, imageBytes);
                        Console.WriteLine($"Output image saved at {imgPath}");
                    }
                }
            }
        }
    }
}

Node.js

const axios = require('axios');
const fs = require('fs');
const path = require('path');

const API_URL = 'http://localhost:8080/layout-parsing';
const imagePath = './demo.jpg';
const fileType = 1;

function encodeImageToBase64(filePath) {
  const bitmap = fs.readFileSync(filePath);
  return Buffer.from(bitmap).toString('base64');
}

const payload = {
  file: encodeImageToBase64(imagePath),
  fileType: fileType
};

axios.post(API_URL, payload)
  .then(response => {
    const results = response.data.result.layoutParsingResults;
    results.forEach((res, index) => {
      console.log(`\n[${index}] prunedResult:`);
      console.log(res.prunedResult);

      const outputImages = res.outputImages;
      if (outputImages) {
        Object.entries(outputImages).forEach(([imgName, base64Img]) => {
          const imgPath = `${imgName}_${index}.jpg`;

          const directory = path.dirname(imgPath);
          if (!fs.existsSync(directory)) {
            fs.mkdirSync(directory, { recursive: true });
            console.log(`Created directory: ${directory}`);
          }

          fs.writeFileSync(imgPath, Buffer.from(base64Img, 'base64'));
          console.log(`Output image saved at ${imgPath}`);
        });
      } else {
        console.log(`[${index}] No outputImages.`);
      }
    });
  })
  .catch(error => {
    console.error('Error during API request:', error.message || error);
  });

PHP

<?php

$API_URL = "http://localhost:8080/layout-parsing";
$image_path = "./demo.jpg";

$image_data = base64_encode(file_get_contents($image_path));
$payload = array("file" => $image_data, "fileType" => 1);

$ch = curl_init($API_URL);
curl_setopt($ch, CURLOPT_POST, true);
curl_setopt($ch, CURLOPT_POSTFIELDS, json_encode($payload));
curl_setopt($ch, CURLOPT_HTTPHEADER, array('Content-Type: application/json'));
curl_setopt($ch, CURLOPT_RETURNTRANSFER, true);
$response = curl_exec($ch);
curl_close($ch);

$result = json_decode($response, true)["result"]["layoutParsingResults"];

foreach ($result as $i => $item) {
    echo "[$i] prunedResult:\n";
    print_r($item["prunedResult"]);

    if (!empty($item["outputImages"])) {
        foreach ($item["outputImages"] as $img_name => $img_base64) {
            $output_image_path = "{$img_name}_{$i}.jpg";

            $directory = dirname($output_image_path);
            if (!is_dir($directory)) {
                mkdir($directory, 0777, true);
                echo "Created directory: $directory\n";
            }

            file_put_contents($output_image_path, base64_decode($img_base64));
            echo "Output image saved at $output_image_path\n";
        }
    } else {
        echo "No outputImages found for item $i\n";
    }
}
?>

4.4 Pipeline Configuration Adjustment Instructions

Note

If you do not need to adjust pipeline configurations, you can ignore this section.

Adjusting the PaddleOCR-VL configuration for service deployment involves only three steps:

1. Obtain the configuration file
2. Modify the configuration file
3. Apply the configuration file

4.4.1 Obtain the Configuration File

If you are deploying using Docker Compose:

Download the corresponding pipeline configuration file based on the backend you are using:

• vLLM: pipeline_config_vllm.yaml
• FastDeploy: pipeline_config_fastdeploy.yaml

If you are deploying by manually installing dependencies:

Execute the following command to generate the pipeline configuration file:

paddlex --get_pipeline_config PaddleOCR-VL

4.4.2 Modify the Configuration File

Enhance VLM Inference Performance Using Acceleration Frameworks

To improve VLM inference performance using acceleration frameworks such as vLLM (refer to Section 2 for detailed instructions on starting the VLM inference service), modify the VLRecognition.genai_config.backend and VLRecognition.genai_config.server_url fields in the pipeline configuration file, as shown below:

VLRecognition:
  ...
  genai_config:
    backend: vllm-server
    server_url: http://127.0.0.1:8118/v1

The Docker Compose solution already uses an acceleration framework by default.

Enable Document Image Preprocessing Functionality

The service started with default configurations does not support document preprocessing. If a client attempts to invoke this functionality, an error message will be returned. To enable document preprocessing, set use_doc_preprocessor to True in the pipeline configuration file and start the service using the modified configuration file.

Disable Result Visualization Functionality

The service returns visualized results by default, which introduces additional overhead. To disable this functionality, add the following configuration to the pipeline configuration file (Serving is a top-level field):

Serving:
  visualize: False

Additionally, you can set the visualize field to false in the request body to disable visualization for a single request.

Configure Return of Image URLs

For visualized result images and images included in Markdown, the service returns them in Base64 encoding by default. To return images as URLs instead, add the following configuration to the pipeline configuration file (Serving is a top-level field):

Serving:
  extra:
    file_storage:
      type: bos
      endpoint: https://bj.bcebos.com
      bucket_name: some-bucket
      ak: xxx
      sk: xxx
      key_prefix: deploy
    return_img_urls: True

Currently, storing generated images in Baidu Intelligent Cloud Object Storage (BOS) and returning URLs is supported. The parameters are described as follows:

• endpoint: Access domain name (required).
• ak: Baidu Intelligent Cloud Access Key (required).
• sk: Baidu Intelligent Cloud Secret Key (required).
• bucket_name: Storage bucket name (required).
• key_prefix: Unified prefix for object keys.
• connection_timeout_in_mills: Request timeout in milliseconds.

For more information on obtaining AK/SK and other details, refer to the Baidu Intelligent Cloud Official Documentation.

Modify PDF Parsing Page Limit

For performance considerations, the service processes only the first 10 pages of received PDF files by default. To adjust the page limit, add the following configuration to the pipeline configuration file (Serving is a top-level field):

Serving:
  extra:
    max_num_input_imgs: <new page limit, e.g., 100>

Set max_num_input_imgs to null to remove the page limit.

4.4.3 Apply the Configuration File

If you deployed using Docker Compose:

If you are deploying using Docker Compose:

Set the services.paddleocr-vl-api.volumes field in the Compose file to mount the pipeline configuration file to the /home/paddleocr directory. For example:

services:
  paddleocr-vl-api:
    ...
    volumes:
      - pipeline_config_vllm.yaml:/home/paddleocr/pipeline_config.yaml
...

In a production environment, you can also build the image yourself and package the configuration file into the image.

If you are deploying by manually installing dependencies:

When starting the service, specify the --pipeline parameter as the path to your custom configuration file.

5. Model Fine-Tuning

If you find that PaddleOCR-VL does not meet accuracy expectations in specific business scenarios, we recommend using the ERNIEKit suite to perform supervised fine-tuning (SFT) on the PaddleOCR-VL-0.9B model. For detailed instructions, refer to the ERNIEKit Official Documentation.

Currently, fine-tuning of layout detection and ranking models is not supported.

Comments