# Android / Termux install notes

Android is not a supported OBLITERATUS runtime. These notes exist so a
Termux (or other Android Python) install does not fight host wheels or
download a second setuptools behind PEP 517 isolation.

`sys.platform` on current Android CPython is `android`, not `linux`. The
locked CPU PyTorch index in `pyproject.toml` is intentionally limited to
`linux` and `win32` manylinux/Windows wheels. Do not point Android at that
index.

## Layout

1. Install the platform's own `python`, `setuptools`, and scientific
   wheels first (`torch`, `numpy`, and whatever Termux already packages).
   Those builds carry the Android patches that PyPI wheels do not.
   `safetensors` and `tokenizers` are Rust extensions; rustup does not
   host `aarch64-linux-android`, so install Termux `rust` before pip
   and export the API level Python was built against (24 on current
   Termux):

   ```bash
   pkg install rust
   export ANDROID_API_LEVEL=24
   ```

2. Isolated pip builds of meson packages (pandas, matplotlib) will try
   to compile `cmake`/`ninja` from source and can hang under proot.
   Install those binaries from Termux first (`pkg install ninja cmake
   python-matplotlib python-scipy python-pillow python-psutil
   python-pyarrow python-greenlet python-brotli python-cryptography
   python-rpds-py python-msgpack libraqm harfbuzz`). PyPI `psutil`
   sdists reject `sys.platform == "android"`; the Termux package is
   patched. Termux matplotlib needs `libraqm`. PyPI `tokenizers` abi3
   wheels fail on CPython 3.14 (`PyBaseObject_Type`); rebuild without
   the `abi3` cargo feature so the extension is `cp314` instead of
   `abi3`. Then install OBLITERATUS without build isolation:

   ```bash
   python -m pip install --no-build-isolation -e ".[spaces]"
   ```

3. If pip still tries to compile `torch` or `numpy` from source, install
   the project over the host packages instead:

   ```bash
   python -m pip install --no-build-isolation --no-deps -e .
   python -m pip install --no-build-isolation "gradio>=6.7,<7.0"
   ```

Do not use `uv sync --locked` on Android. The lock resolves Linux/Windows
CPU PyTorch and does not contain `android_*` wheels for torch, numpy,
pyarrow, tokenizers, or bitsandbytes.

The `quantization` and `qwen-hybrid` extras are CUDA-oriented and are not
an Android install path.