Skip to main content

​
System Requirements

Before installing olmOCR, ensure your system meets these requirements:

​
Hardware Requirements

GPU

Recent NVIDIA GPU with CUDA supportTested GPUs:
  • RTX 4090
  • L40S
  • A100
  • H100

Storage

Minimum 30GB of free disk spaceRequired for:
  • Model weights
  • Temporary processing files
  • Output storage

​
Software Requirements

  • Python: 3.11 or higher (pyproject.toml:20)
  • CUDA: Compatible version for your GPU
  • Operating System: Ubuntu/Debian (other Linux distributions may work)
GPU is required for running inference. CPU-only mode is not supported for the inference pipeline.

​
Installation Steps

1

Install System Dependencies

Install poppler-utils and additional fonts for rendering PDF images:
sudo apt-get update
sudo apt-get install poppler-utils ttf-mscorefonts-installer msttcorefonts fonts-crosextra-caladea fonts-crosextra-carlito gsfonts lcdf-typetools
These fonts are required for properly rendering PDF pages to images.
2

Create Conda Environment

Set up a dedicated conda environment with Python 3.11:
conda create -n olmocr python=3.11
conda activate olmocr
3

Clone and Install olmOCR

Clone the repository and install the package:
git clone https://github.com/allenai/olmocr.git
cd olmocr
pip install -e .
This installs all core dependencies including:
  • pypdf (>=5.2.0) - PDF parsing
  • pypdfium2 - PDF rendering
  • torch (>=2.5.1) - Deep learning framework
  • transformers (>=4.46.2) - Model loading
  • Pillow - Image processing
  • And more (see pyproject.toml:21-42 for full list)
4

Install SGLang for GPU Inference (Recommended)

For GPU-accelerated inference, install sglang with flashinfer support:
pip install sgl-kernel==0.0.3.post1 --force-reinstall --no-deps
pip install "sglang[all]==0.4.2" --find-links https://flashinfer.ai/whl/cu124/torch2.4/flashinfer/
This step is optional but highly recommended for production use. Without sglang, you won’t be able to run GPU inference.
5

Verify Installation

Verify your installation by checking the package version:
python -c "import olmocr; print(olmocr.VERSION)"

​
Optional Dependencies

​
Development Tools

If you plan to contribute to olmOCR, install the development dependencies: 
pip install -e ".[dev]"
This includes:
  • Testing tools (pytest, pytest-cov)
  • Code formatters (black, isort, ruff)
  • Type checking (mypy)
  • Documentation tools (Sphinx)

​
Training Dependencies

For fine-tuning models, install the training extras: 
pip install -e ".[train]"
This adds:
  • accelerate - Distributed training
  • peft - Parameter-efficient fine-tuning
  • wandb - Experiment tracking
  • datasets - Data loading

​
GPU Configuration

​
Memory Requirements

The pipeline automatically adjusts GPU memory usage based on available VRAM:
  • < 60GB VRAM: Uses 80% memory fraction for KV cache
  • >= 60GB VRAM: Uses default memory allocation
The memory fraction is automatically configured in pipeline.py:508 based on your GPU.

​
Checking GPU Availability

Verify that PyTorch can detect your GPU: 
python -c "import torch; print(f'CUDA available: {torch.cuda.is_available()}'); print(f'GPU count: {torch.cuda.device_count()}')"
Expected output: 
CUDA available: True
GPU count: 1

​
Troubleshooting

​
Poppler Not Found

If you get an error about poppler not being found: 
# Verify poppler is installed
pdftoppm -v
If not found, reinstall: 
sudo apt-get install --reinstall poppler-utils

​
CUDA Version Mismatch

If you encounter CUDA version errors, ensure your PyTorch installation matches your CUDA version: 
# Check CUDA version
nvidia-smi

# Install PyTorch with matching CUDA version
# Visit https://pytorch.org for the correct command

​
Font Rendering Issues

If PDFs render with missing characters, ensure all fonts are installed: 
fc-cache -f -v

​
Next Steps

Now that you have olmOCR installed, you’re ready to process your first PDF!

Quickstart Guide

Learn how to convert PDFs in minutes

Build docs developers (and LLMs) love

Get started for freeTalk to us