---
title: PaliGemma Demo
emoji: 🤲
colorFrom: green
colorTo: yellow
sdk: gradio
sdk_version: 4.22.0
app_file: app.py
pinned: false
license: apache-2.0
---

# PaliGemma Demo

See [Blogpost] and [`big_vision README.md`] for details about the model.


[Blogpost]: https://huggingface.co/blog/paligemma

[`big_vision README.md`]: https://github.com/google-research/big_vision/blob/main/big_vision/configs/proj/paligemma/README.md

## Development

Local testing (CPU, Python 3.12):

```bash
pip -m venv env
. env/bin/activate
pip install -qr requirements-cpu.txt
python app.py
```

Environment variables:

- `MOCK_MODEL=yes`: For quick UI testing.
- `RAM_CACHE_GB=18`: Enables caching of 3 bf16 models in memory: a single bf16
  model is about 5860 MB. Use with care on spaces with little RAM. For example,
  on a `A10G large` space you can cache five models in RAM, so you would set
  `RAM_CACHE_GB=30`.
- `HOST_COLOCATION=4`: If host RAM/disk is shared between 4 processes (e.g. the
  Huggingface `A10 large` Spaces).


Loading models:

- The set of models loaded is defined in `./models.py`.
- You must first acknowledge usage conditions to access models.
- When testing locally, you'll have to run `huggingface_cli login`.
- When running in a Huggingface Space, you'll have to set a `HF_TOKEN` secret.