Samarth0710/nemotron-ocr-v2-pdf-runner

nemotron-ocr-v2 PDF runner

Wrapper script around nvidia/nemotron-ocr-v2 that rasterizes a PDF (via pdf2image / poppler) and runs the V2 pipeline page-by-page, writing one .txt file per page.

Verified working

• Hardware: NVIDIA L4 (sm_89), CUDA toolkit 12.8, PyTorch 2.11+cu128, Python 3.12
• Test input: arXiv 1706.03762 (Attention Is All You Need), page 1
• Result: 25 text regions extracted, full abstract recovered as readable paragraphs

Setup

# CUDA toolkit (must match torch CUDA major version)
# Install cuda-nvcc-12-8 + cuda-cudart-dev-12-8 + cuda-libraries-dev-12-8

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128
pip install hatchling hatch-vcs huggingface_hub pdf2image pillow

huggingface-cli download nvidia/nemotron-ocr-v2 --local-dir nemotron-ocr-v2
cd nemotron-ocr-v2/nemotron-ocr
TORCH_CUDA_ARCH_LIST="8.9" pip install --no-build-isolation -v .

apt-get install -y poppler-utils

Usage

python run_ocr_pdf.py sample.pdf --page 1
python run_ocr_pdf.py sample.pdf --page 0   # all pages
python run_ocr_pdf.py sample.pdf --lang multi --merge-level sentence

Outputs ocr_out/page_NNN.png and ocr_out/page_NNN.txt.