Skip to main content

DeepSeek-OCR

Run from the repository root in a CUDA-enabled Eole environment. The examples use BF16, GPU 0, and FlashAttention; install flash-attn --no-build-isolation or select the PyTorch attention backend in the configuration.

Prepare the checkpoint​

With Eole 0.6.4 or newer, choose either preparation option below. Both use the same directory expected by the scripts and YAMLs:

export EOLE_MODEL_DIR=/path/to/models

Download pre-converted weights: the public Eole DeepSeek-OCR package includes converted weights, configuration and tokenizer. No conversion or HF token is required:

hf download eole-nlp/DeepSeek-OCR-eole --local-dir "$EOLE_MODEL_DIR/DeepSeek-OCR"

Or convert the original HF checkpoint: the converter accepts the upstream repository ID or an existing local HF directory:

eole convert HF --model_dir deepseek-ai/DeepSeek-OCR \
  --output "$EOLE_MODEL_DIR/DeepSeek-OCR" --dtype bf16

Use a new destination for either option. Add --token "$HF_TOKEN" to conversion if authentication is required. To pin a pre-converted download, add --revision COMMIT_SHA; validated pins are in the RunPod model catalog. Downloading pre-converted weights does not remove runtime kernel or GPU memory requirements.

After either preparation option:

python recipes/deepseekocr/test_inference.py

The script prints OCR and grounded Markdown outputs for two bundled images in eole/tests/data/images.

Convert a PDF to Markdown​

Install the PDF helper dependencies and supply your own PDF and output directory:

pip install pymupdf img2pdf
python recipes/deepseekocr/pdf_ocr_mmd.py \
  -c recipes/deepseekocr/predict-pdf.yaml \
  --input /path/to/document.pdf --output-dir ./ocr-output

The helper renders pages to images, runs OCR, and writes document.mmd, document_det.mmd (grounding coordinates), and document_layouts.pdf (annotated pages). Batch size defaults to one; increase it in the YAML only if VRAM permits. All pages are currently rendered into host memory, so split very large PDFs.