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.
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](#1-environment-preparation) and [2. Quick Start](#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-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](#41-method-1-deploy-using-docker-compose-recommended) and [4.3 Client-Side Invocation](#43-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](#5-model-fine-tuning).
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:
> > - Common GPUs with CC ≥ 8 include RTX 30/40/50 series and A10/A100, etc. For more models, refer to [CUDA GPU Compute Capability](https://developer.nvidia.com/cuda-gpus)
> 2. 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.
> 3. 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!
> 4. 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:
> 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](./PaddleOCR-VL-NVIDIA-Blackwell.en.md) to complete the environment configuration before using PaddleOCR-VL.
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):
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:
**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:
# 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
> **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.
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.
> 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.
For example, the local path of an image file or PDF file: <code>/root/data/img.jpg</code>;<b>Such as a URL link</b>, for example, the network URL of an image file or PDF file:<ahref="https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/demo_paper.png">Example</a>;<b>Such as a local directory</b>, which should contain the images to be predicted, for example, the local path: <code>/root/data/</code>(Currently, prediction for directories containing PDF files is not supported. PDF files need to be specified with a specific file path).</td>
<td>Expansion coefficient for the detection boxes of the layout area detection model.Any floating-point number greater than <code>0</code>. If not set, the initialized default value will be used.</td>
<td>Merging mode for the detection boxes output by the model in layout detection.
<ul>
<li><b>large</b> 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;</li>
<li><b>small</b>, 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;</li>
<li><b>union</b>,no filtering is performed on the boxes, and both inner and outer boxes are retained;</li></ul>
If not set, the initialized parameter value will be used.
<td>Whether to load and use the document orientation classification module. If not set, the initialized default value will be used, which is initialized to<code>False</code>.</td>
<td>Whether to load and use the text image rectification module. If not set, the initialized default value will be used, which is initialized to <code>False.</td>
<td>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 <code>True</code>.</td>
<td>Whether to use the chart parsing function. If not set, the initialized default value will be used, which is initialized to <code>False</code>.</td>
<td>Controls whether to format the <code>block_content</code> content within as Markdown. If not set, the initialized default value will be used, which defaults to initialization as<code>False</code>.</td>
<td>Used to control whether to enable internal queues. When set to <code>True</code>, 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.</td>
<li><b>CPU</b>: For example,<code>cpu</code> indicates using the CPU for inference;</li>
<li><b>GPU</b>: For example,<code>gpu:0</code> indicates using the first GPU for inference;</li>
<li><b>NPU</b>: For example,<code>npu:0</code> indicates using the first NPU for inference;</li>
<li><b>XPU</b>: For example,<code>xpu:0</code> indicates using the first XPU for inference;</li>
<li><b>MLU</b>: For example,<code>mlu:0</code> indicates using the first MLU for inference;</li>
<li><b>DCU</b>: For example,<code>dcu:0</code> indicates using the first DCU for inference;</li>
</ul>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.</td>
<td>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.<br/>For PaddlePaddle version with CUDA 11.8, the compatible TensorRT version is 8.x (x&gt;=6). It is recommended to install TensorRT 8.6.1.6.<br/>
<td>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.</td>
<b>Note: </b> 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](#31-starting-the-vlm-inference-service) for faster inference.
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:
```python
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
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:
- 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:
<td>Whether to use post-processing NMS for layout detection. If set to <code>None</code>, the parameter value initialized by the pipeline will be used.</td>
Expansion coefficient for the detection box of the layout area detection model.
<ul>
<li><b>float</b>: Any floating-point number greater than <code>0</code></li>
<li><b>Tuple[float,float]</b>: The respective expansion coefficients in the horizontal and vertical directions;</li>
<li><b>dict</b>: where the key of the dict is of <b>int</b> type, representing <code>cls_id</code>, and the value is of</code>tuple <code>type, such as</code>{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;</li>
<td>Merging mode for the detection boxes output by the model in layout detection.
<ul>
<li><b>large</b> 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;</li>
<li><b>small</b>, 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;</li>
<li><b>union</b>,no filtering is performed on the boxes, and both inner and outer boxes are retained;</li></ul>
If not set, the initialized parameter value will be used.
<td>Whether to load and use the document orientation classification module. If not set, the initialized default value will be used, which is initialized to<code>False</code>.</td>
<td>Whether to load and use the text image rectification module. If not set, the initialized default value will be used, which is initialized to <code>False.</td>
<td>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 <code>True</code>.</td>
<td>Whether to use the chart parsing function. If not set, the initialized default value will be used, which is initialized to <code>False</code>.</td>
<td>Controls whether to format the <code>block_content</code> content within as Markdown. If not set, the initialized default value will be used, which defaults to initialization as<code>False</code>.</td>
<td>The device used for inference. Supports specifying specific card numbers:<ul>
<li><b>CPU</b>: For example,<code>cpu</code> indicates using the CPU for inference;</li>
<li><b>GPU</b>: For example,<code>gpu:0</code> indicates using the first GPU for inference;</li>
<li><b>NPU</b>: For example,<code>npu:0</code> indicates using the first NPU for inference;</li>
<li><b>XPU</b>: For example,<code>xpu:0</code> indicates using the first XPU for inference;</li>
<li><b>MLU</b>: For example,<code>mlu:0</code> indicates using the first MLU for inference;</li>
<li><b>DCU</b>: For example,<code>dcu:0</code> indicates using the first DCU for inference;</li>
</ul>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.</td>
<td>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.<br/>For PaddlePaddle version with CUDA 11.8, the compatible TensorRT version is 8.x (x&gt;=6). It is recommended to install TensorRT 8.6.1.6.<br/>
<td>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.</td>
<details><summary>(2) Call the <code>predict()</code>method of the PaddleOCR-VL pipeline object for inference prediction. This method will return a list of results. Additionally, the pipeline also provides the <code>predict_iter()</code>Method. The two are completely consistent in terms of parameter acceptance and result return. The difference lies in that <code>predict_iter()</code>returns a <code>generator</code>, 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 <code>predict()</code>method and their descriptions:</summary>
<li><b>Python Var</b>: such as <code>numpy.ndarray</code> representing image data</li>
<li><b>str</b>: such as the local path of an image file or PDF file: <code>/root/data/img.jpg</code>;<b>such as a URL link</b>, such as the network URL of an image file or PDF file:<ahref="https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/demo_paper.png">Example</a>;<b>such as a local directory</b>, which should contain the images to be predicted, such as the local path: <code>/root/data/</code>(Currently, prediction for directories containing PDF files is not supported. PDF files need to be specified with a specific file path)</li>
<li><b>list</b>: List elements should be of the aforementioned data types, such as <code>[numpy.ndarray, numpy.ndarray]</code>, <code>["/root/data/img1.jpg", "/root/data/img2.jpg"]</code>, <code>["/root/data1", "/root/data2"].</code></li>
<td>Whether to use the document orientation classification module during inference. Setting it to <code>None</code> means using the instantiation parameter; otherwise, this parameter takes precedence.</td>
<td>Whether to use the text image rectification module during inference. Setting it to <code>None</code> means using the instantiation parameter; otherwise, this parameter takes precedence.</td>
<td>Whether to use the layout region detection and sorting module during inference. Setting it to <code>None</code> means using the instantiation parameter; otherwise, this parameter takes precedence.</td>
<td>Whether to use the chart parsing module during inference. Setting it to <code>None</code> means using the instantiation parameter; otherwise, this parameter takes precedence.</td>
<td>The parameter meaning is basically the same as the instantiation parameter. Setting it to <code>None</code> means using the instantiation parameter; otherwise, this parameter takes precedence.</td>
<td>The parameter meaning is basically the same as the instantiation parameter. Setting it to <code>None</code> means using the instantiation parameter; otherwise, this parameter takes precedence.</td>
<td>The parameter meaning is basically the same as the instantiation parameter. Setting it to <code>None</code> means using the instantiation parameter; otherwise, this parameter takes precedence.</td>
<td>The parameter meaning is basically the same as the instantiation parameter. Setting it to <code>None</code> means using the instantiation parameter; otherwise, this parameter takes precedence.</td>
<td>Used to control whether to enable internal queues. When set to <code>True</code>, 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.</td>
<td>The prompt type setting for the VL model, which takes effect only when <code>use_layout_detection=False</code>. The fillable parameters are <code>ocr</code>、<code>formula</code>、<code>table</code> and <code>chart</code>.</td>
<td>The parameter meaning is basically the same as the instantiation parameter. Setting it to <code>None</code> means using the instantiation parameter; otherwise, this parameter takes precedence.</td>
<details><summary>(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 <code>json</code> file:</summary>
<td>Specify the indentation level to beautify the output <code>JSON</code> data, making it more readable. Only valid when <code>format_json</code> is <code>True</code>.</td>
<td>Control whether non- <code>ASCII</code> characters are escaped as <code>Unicode</code>. When set to <code>True</code>, all non- <code>ASCII</code> characters will be escaped; <code>False</code> retains the original characters. Only valid when <code>format_json</code> is <code>True</code>.</td>
<td>Specify the indentation level to beautify the output <code>JSON</code>data, making it more readable. Only valid when <code>format_json</code>is <code>True</code>.</td>
<td>Control whether non- <code>ASCII</code> characters are escaped as <code>Unicode</code>. When set to <code>True</code>, all non- <code>ASCII</code> characters will be escaped; <code>False</code> retains the original characters. Only valid when <code>format_json</code> is <code>True</code>.</td>
<td>Whether to beautify the <code>markdown</code> output results, centering charts, etc., to make the <code>markdown</code> rendering more aesthetically pleasing.</td>
<td>Control whether to retain formula numbers in <code>markdown</code>. When set to <code>True</code>, all formula numbers are retained; <code>False</code> retains only the formulas</td>
-`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`.
- 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:<table>
- 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.</details>
## 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.
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):
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).
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](#331-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:
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:
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](https://github.com/mjun0812/flash-attention-prebuild-wheels). 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 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.**
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.
Different inference acceleration frameworks support different parameters. Refer to their official documentation for available parameters and adjustment timing:
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:
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.
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.
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.
You can obtain the Compose file and the environment variables configuration file from [here](https://github.com/PaddlePaddle/PaddleOCR/blob/main/deploy/paddleocr_vl_docker/compose.yaml) and [here](https://github.com/PaddlePaddle/PaddleOCR/blob/main/deploy/paddleocr_vl_docker/.env), 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:
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:
<details>
<summary>1. Change the port of the PaddleOCR-VL service</summary>
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:
<summary>2. Specify the GPU used by the PaddleOCR-VL service</summary>
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:
If you want to adjust the VLM server-side configuration, please refer to [3.3.1 Server-side Parameter Adjustment](#331-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.
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](../paddleocr_and_paddlex.en.md). Furthermore, section 4.1.3 will introduce how to adjust the pipeline configuration based on common requirements.
<li>Both the request body and response body are JSON data (JSON objects).</li>
<li>When the request is processed successfully, the response status code is<code>200</code>, and the properties of the response body are as follows:</li>
</ul>
<table>
<thead>
<tr>
<th>Name</th>
<th>Type</th>
<th>Meaning</th>
</tr>
</thead>
<tbody>
<tr>
<td><code>logId</code></td>
<td><code>string</code></td>
<td>The UUID of the request.</td>
</tr>
<tr>
<td><code>errorCode</code></td>
<td><code>integer</code></td>
<td>Error code. Fixed as <code>0</code>.</td>
</tr>
<tr>
<td><code>errorMsg</code></td>
<td><code>string</code></td>
<td>Error description. Fixed as <code>"Success"</code>.</td>
</tr>
<tr>
<td><code>result</code></td>
<td><code>object</code></td>
<td>Operation result.</td>
</tr>
</tbody>
</table>
<ul>
<li>When the request is not processed successfully, the properties of the response body are as follows:</li>
</ul>
<table>
<thead>
<tr>
<th>Name</th>
<th>Type</th>
<th>Meaning</th>
</tr>
</thead>
<tbody>
<tr>
<td><code>logId</code></td>
<td><code>string</code></td>
<td>The UUID of the request.</td>
</tr>
<tr>
<td><code>errorCode</code></td>
<td><code>integer</code></td>
<td>Error code. Same as the response status code.</td>
</tr>
<tr>
<td><code>errorMsg</code></td>
<td><code>string</code></td>
<td>Error description.</td>
</tr>
</tbody>
</table>
<p>The main operations provided by the service are as follows:</p>
<ul>
<li><b><code>infer</code></b></li>
</ul>
<p>Perform layout parsing.</p>
<p><code>POST /layout-parsing</code></p>
<ul>
<li>The properties of the request body are as follows:</li>
<td>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.<br/>To remove the page limit, add the following configuration to the pipeline configuration file:<pre><code>Serving:
<td>File type.<code>0</code> represents a PDF file,<code>1</code> represents an image file. If this property is not present in the request body, the file type will be inferred from the URL.</td>
<td>Please refer to the description of the <code>use_doc_orientation_classify</code> parameter in the <code>predict</code> method of the PaddleOCR-VL object.</td>
<td>Please refer to the description of the <code>use_doc_unwarping</code> parameter in the <code>predict</code> method of the PaddleOCR-VL object.</td>
<td>No</td>
</tr>
<tr>
<td><code>useLayoutDetection</code></td>
<td><code>boolean</code>|<code>null</code></td>
<td>Please refer to the description of the <code>use_layout_detection</code> parameter in the <code>predict</code> method of the PaddleOCR-VL object.</td>
<td>No</td>
</tr>
<tr>
<td><code>useChartRecognition</code></td>
<td><code>boolean</code>|<code>null</code></td>
<td>Please refer to the description of the <code>use_chart_recognition</code> parameter in the <code>predict</code> method of the PaddleOCR-VL object.</td>
<td>Please refer to the description of the <code>layout_unclip_ratio</code> parameter in the <code>predict</code> method of the PaddleOCR-VL object.</td>
<td>Please refer to the description of the <code>layout_merge_bboxes_mode</code> parameter in the <code>predict</code> method of the PaddleOCR-VL object.</td>
<td>Please refer to the description of the <code>prompt_label</code> parameter in the <code>predict</code> method of the PaddleOCR-VL object.</td>
<td>No</td>
</tr>
<tr>
<td><code>formatBlockContent</code></td>
<td><code>boolean</code>|<code>null</code></td>
<td>Please refer to the description of the <code>format_block_content</code> parameter in the <code>predict</code> method of the PaddleOCR-VL object.</td>
<td>No</td>
</tr>
<tr>
<td><code>repetitionPenalty</code></td>
<td><code>number</code>|<code>null</code></td>
<td>Please refer to the description of the <code>repetition_penalty</code> parameter in the <code>predict</code> method of the PaddleOCR-VL object.</td>
<td>No</td>
</tr>
<tr>
<td><code>temperature</code></td>
<td><code>number</code>|<code>null</code></td>
<td>Please refer to the description of the <code>temperature</code> parameter in the <code>predict</code> method of the PaddleOCR-VL object.</td>
<td>No</td>
</tr>
<tr>
<td><code>topP</code></td>
<td><code>number</code>|<code>null</code></td>
<td>Please refer to the description of the <code>top_p</code> parameter in the <code>predict</code> method of the PaddleOCR-VL object.</td>
<td>No</td>
</tr>
<tr>
<td><code>minPixels</code></td>
<td><code>number</code>|<code>null</code></td>
<td>Please refer to the description of the <code>min_pixels</code> parameter in the <code>predict</code> method of the PaddleOCR-VL object.</td>
<td>No</td>
</tr>
<tr>
<td><code>maxPixels</code></td>
<td><code>number</code>|<code>null</code></td>
<td>Please refer to the description of the <code>max_pixels</code> parameter in the <code>predict</code> method of the PaddleOCR-VL object.</td>
<td>No</td>
</tr>
<tr>
<td><code>prettifyMarkdown</code></td>
<td><code>boolean</code></td>
<td>Whether to output beautified Markdown text. The default is <code>true</code>.</td>
<td>No</td>
</tr>
<tr>
<td><code>showFormulaNumber</code></td>
<td><code>boolean</code></td>
<td>Whether to include formula numbers in the output Markdown text. The default is <code>false</code>.</td>
<td>No</td>
</tr>
<tr>
<td><code>visualize</code></td>
<td><code>boolean</code>|<code>null</code></td>
<td>Whether to return visualization result images and intermediate images during the processing.<ulstyle="margin: 0 0 0 1em; padding-left: 0em;">
<li>Pass <code>true</code>: Return images.</li>
<li>Pass <code>false</code>: Do not return images.</li>
<li>If this parameter is not provided in the request body or <code>null</code> is passed: Follow the setting in the configuration file <code>Serving.visualize</code>.</li>
</ul>
<br/>For example, add the following field in the configuration file:<br/>
<pre><code>Serving:
visualize: False</code></pre>Images will not be returned by default, and the default behavior can be overridden by the <code>visualize</code> parameter in the request body. If this parameter is not set in either the request body or the configuration file (or <code>null</code> is passed in the request body and the configuration file is not set), images will be returned by default.</td>
<td>No</td>
</tr>
</tbody>
</table>
<ul>
<li>When the request is processed successfully, the <code>result</code> in the response body has the following attributes:</li>
</ul>
<table>
<thead>
<tr>
<th>Name</th>
<th>Type</th>
<th>Meaning</th>
</tr>
</thead>
<tbody>
<tr>
<td><code>layoutParsingResults</code></td>
<td><code>array</code></td>
<td>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.</td>
</tr>
<tr>
<td><code>dataInfo</code></td>
<td><code>object</code></td>
<td>Input data information.</td>
</tr>
</tbody>
</table>
<p>Each element in<code>layoutParsingResults</code> is an <code>object</code> with the following attributes:</p>
<table>
<thead>
<tr>
<th>Meaning</th>
<th>Name</th>
<th>Type</th>
</tr>
</thead>
<tbody>
<tr>
<td><code>prunedResult</code></td>
<td><code>object</code></td>
<td>A simplified version of the <code>res</code> field in the JSON representation of the results generated by the <code>predict</code> method of the object, with the <code>input_path</code> and <code>page_index</code> fields removed.</td>
</tr>
<tr>
<td><code>markdown</code></td>
<td><code>object</code></td>
<td>Markdown results.</td>
</tr>
<tr>
<td><code>outputImages</code></td>
<td><code>object</code>|<code>null</code></td>
<td>Refer to the <code>img</code> property description of the prediction results. The image is in JPEG format and encoded using Base64.</td>
</tr>
<tr>
<td><code>inputImage</code></td>
<td><code>string</code>|<code>null</code></td>
<td>Input image. The image is in JPEG format and encoded using Base64.</td>
</tr>
</tbody>
</table>
<p><code>markdown</code>is an <code>object</code>with the following properties:</p>
<table>
<thead>
<tr>
<th>Name</th>
<th>Type</th>
<th>Meaning</th>
</tr>
</thead>
<tbody>
<tr>
<td><code>text</code></td>
<td><code>string</code></td>
<td>Markdown text.</td>
</tr>
<tr>
<td><code>images</code></td>
<td><code>object</code></td>
<td>Key-value pairs of relative paths to Markdown images and Base64-encoded images.</td>
</tr>
<tr>
<td><code>isStart</code></td>
<td><code>boolean</code></td>
<td>Whether the first element on the current page is the start of a paragraph.</td>
</tr>
<tr>
<td><code>isEnd</code></td>
<td><code>boolean</code></td>
<td>Whether the last element on the current page is the end of a paragraph.</td>
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:
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.
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):
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):
Currently, storing generated images in Baidu Intelligent Cloud Object Storage (BOS) and returning URLs is supported. The parameters are described as follows:
-`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](https://cloud.baidu.com/doc/BOS/index.html).
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):
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:
If you find that PaddleOCR-VL does not meet accuracy expectations in specific business scenarios, we recommend using the [ERNIEKit suite](https://github.com/PaddlePaddle/ERNIE/tree/release/v1.4) to perform supervised fine-tuning (SFT) on the PaddleOCR-VL-0.9B model. For detailed instructions, refer to the [ERNIEKit Official Documentation](https://github.com/PaddlePaddle/ERNIE/blob/release/v1.4/docs/paddleocr_vl_sft.md).
