Best performance settings

[piotrkandziora]

Hello,

What best performance settings do you guys recommend for analyzing documents?

I am currently experimenting on a single AWS EC2 instance (m5.2xlarge, 8xCPU, 32GB, dockerized docling with num_threads=8 Acceleration Options - other instance with GPU support is also possible for me). Currently I am parsing pdf file (~140 pages) and it last quite long. In the future I would like to analyze pdfs and confluence pages.

My current experimental pipeline settings:

    settings.debug.profile_pipeline_timings = True
    pipeline_options = PdfPipelineOptions()
    pipeline_options.do_ocr = False
    pipeline_options.do_table_structure = True
    pipeline_options.table_structure_options.do_cell_matching = True
    pipeline_options.do_picture_classification = True
    pipeline_options.do_picture_description = True
    pipeline_options.generate_page_images = True
    pipeline_options.generate_picture_images = True
    pipeline_options.accelerator_options = AcceleratorOptions(num_threads=8, device=AcceleratorDevice.AUTO)
    doc_converter = DocumentConverter(
        format_options={
            InputFormat.PDF: PdfFormatOption(pipeline_options=pipeline_options)
        }
    )

As it's hosted on AWS I am thinking also about leveraging Bedrock for VLM.

[dosubot[bot]]

Hi @piotrkandziora! I'm Dosu and I’m helping the docling team.

For best performance with Docling on AWS, focus on these settings and strategies:

• Use GPU acceleration if available—switch your EC2 instance to one with an NVIDIA GPU and use a CUDA-enabled Docling Docker image. Set device="cuda" in AcceleratorOptions and match num_threads to your vCPU count for parallelism. This can drastically speed up image, OCR, and vision-language model tasks source.
• Disable expensive pipeline features you don't need, like table extraction, image generation, code enrichment, formula enrichment, and picture classification/description. These steps slow down processing, especially for large PDFs source.
• For fastest PDF conversion, use the pypdfium2 backend instead of the default. This is much faster and more stable for large documents source.
• Use batching and parallelism: Threaded pipelines and batch sizes for OCR/layout/table can yield up to 40% speedup. Consider splitting large documents into smaller chunks or processing page ranges in parallel source.
• Keep images_scale at the default (1.0) unless you need high accuracy for small text—higher values increase processing time and memory usage source.
• Profile your pipeline timings (settings.debug.profile_pipeline_timings = True) to identify bottlenecks. This will help you see which steps are slowest and adjust accordingly source.
• Pre-download models and mount them into your container to avoid startup delays and network dependencies source.
• For Confluence pages and structured formats (DOCX, HTML, Markdown), use the SimplePipeline, which skips OCR and enrichment for fastest results source.

If you want to leverage AWS Bedrock for VLM, Docling can route PDFs and images to VLM pipelines, but VLMs do not replace OCR for text recognition—OCR must be explicitly enabled. Timeout and slow inference can result from inefficient batching or model loading, so tune batch size and worker count for best results source.

Let me know if you want a sample config or have specific bottlenecks in your pipeline!

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}