CPU and fast conversion

Markdrop runs on CPU-only machines. No NVIDIA GPU is required.

Recommended: --fast mode on CPU

Default conversion loads Docling layout models (Torch, ~GB download, minutes per document). For quick text extraction on CPU, use fast mode:

markdrop convert report.pdf --output_dir out --fast

Fast mode uses PyMuPDF only — typically seconds per document instead of minutes.

	Default	--fast
Engine	Docling + PyMuPDF reconcile	PyMuPDF
Torch / ML models	Yes	No
Table structure	Yes	No (text layer only)
Scanned PDFs / OCR	Docling handles some cases	Poor — use default mode
Typical CPU time (15-page paper)	~2–10 min	~5–30 sec

Install optional better Markdown output for fast mode:

pip install "markdrop[lite]"

This adds pymupdf4llm for higher-quality Markdown. Without it, Markdrop falls back to block-based text extraction.

Default mode on CPU

Default mode works on CPU but is slow:

• First run downloads Docling models (one-time, large).
• Expect roughly 8–15 seconds per page on a modern laptop CPU for digital PDFs.
• A 15-page paper is often 2–5 minutes after models are cached.

Tips:

• Run once on a small PDF to warm the model cache before batch jobs.
• Use --fast when you only need searchable text from digital PDFs.
• Use default mode when tables, figures, and layout matter.

Google Colab

Colab free tier is CPU or limited GPU. Prefer:

!pip install markdrop
!markdrop convert /content/paper.pdf --output_dir /content/out --fast

For full layout quality on Colab, use a GPU runtime and default mode (still slow on first cell).

Environment variables

No GPU-specific configuration is required. Docling automatically uses CPU when CUDA is unavailable.