GPU support

Achieving Optimal GPU Performance with Docling

This guide describes how to maximize GPU performance for Docling pipelines. It covers device selection, pipeline differences, and provides example snippets for configuring batch size and concurrency in the VLM pipeline for both Linux and Windows.

!!! note

Improvements and optimizations strategies for maximizing the GPU performance is an
active topic. Regularly check these guidelines for updates.

Standard Pipeline

Enable GPU acceleration by configuring the accelerator device and concurrency options using Docling's API:

from docling.datamodel.accelerator_options import AcceleratorDevice, AcceleratorOptions

# Configure accelerator options for GPU
accelerator_options = AcceleratorOptions(
    device=AcceleratorDevice.CUDA,  # or AcceleratorDevice.AUTO
)

Batch size and concurrency for document processing are controlled for each stage of the pipeline as:

from docling.datamodel.pipeline_options import (
    ThreadedPdfPipelineOptions,
)

pipeline_options = ThreadedPdfPipelineOptions(
    ocr_batch_size=64,  # default 4
    layout_batch_size=64,  # default 4
    table_batch_size=4,  # currently not using GPU batching
)

Setting a higher page_batch_size will run the Docling models (in particular the layout detection stage) with a GPU batch inference mode.

Complete example

For a complete example see gpu_standard_pipeline.py.

OCR engines

The current Docling OCR engines rely on third-party libraries, hence GPU support depends on the availability in the respective engines.

The only setup which is known to work at the moment is RapidOCR with the torch backend, which can be enabled via

pipeline_options = PdfPipelineOptions()
pipeline_options.ocr_options = RapidOcrOptions(
    backend="torch",
)

More details in the GitHub discussion #2451.

VLM Pipeline

For best GPU utilization, use a local inference server. Docling supports inference servers which exposes the OpenAI-compatible chat completion endpoints. For example:

• vllm: http://localhost:8000/v1/chat/completions (available only on Linux)
• LM Studio: http://localhost:1234/v1/chat/completions (available both on Linux and Windows)
• Ollama: http://localhost:11434/v1/chat/completions (available both on Linux and Windows)

Start the inference server

Here is an example on how to start the vllm inference server with optimum parameters for Granite Docling.

vllm serve ibm-granite/granite-docling-258M \
  --host 127.0.0.1 --port 8000 \
  --max-num-seqs 512 \
  --max-num-batched-tokens 8192 \
  --enable-chunked-prefill \
  --gpu-memory-utilization 0.9

Configure Docling

Configure the VLM pipeline using Docling's VLM options:

from docling.datamodel.pipeline_options import VlmPipelineOptions

vlm_options = VlmPipelineOptions(
    enable_remote_services=True,
    vlm_options={
        "url": "http://localhost:8000/v1/chat/completions",  # or any other compatible endpoint
        "params": {
            "model": "ibm-granite/granite-docling-258M",
            "max_tokens": 4096,
        },
        "concurrency": 64,  # default is 1
        "prompt": "Convert this page to docling.",
        "timeout": 90,
    }
)

Additionally to the concurrency, we also have to set the page_batch_size Docling parameter. Make sure to set settings.perf.page_batch_size >= vlm_options.concurrency.

from docling.datamodel.settings import settings

settings.perf.page_batch_size = 64  # default is 4

Complete example

For a complete example see gpu_vlm_pipeline.py.

Available models

Both LM Studio and Ollama rely on llama.cpp as runtime engine. For using this engine, models have to be converted to the gguf format.

Here is a list of known models which are available in gguf format and how to use them.

TBA.

Performance results

Test data

Test infrastructure

Results

* cpu-only timing computed with 16 pytorch threads.