dx-engine

Python bindings for the DEEPX Runtime (dx_rt). dx-engine exposes the libdxrt inference stack to Python via a pybind11 extension so you can load DEEPX-compiled models (.dxnn), run inference on DEEPX NPUs, inspect device status, and collect profiling data from Python.

• Package name (PyPI / pip): dx-engine
• Import name: dx_engine
• Supported Python: 3.8 – 3.14
• Requires: libdxrt runtime installed on the system (typically via the libdxrt-bin Debian package).

Installation

From PyPI

python -m pip install dx-engine

From a prebuilt wheel (recommended)

Wheels are shipped inside the libdxrt-bin Debian package under /usr/share/libdxrt-bin/python/. Install the wheel whose cpXY tag matches your Python version:

# 1. Install the system runtime
sudo dpkg -i libdxrt-bin_<version>_<arch>.deb

# 2. Install the matching wheel into your venv/interpreter
PYTAG=cp$(python -c 'import sys;print(f"{sys.version_info[0]}{sys.version_info[1]}")')
pip install /usr/share/libdxrt-bin/python/dx_engine-*-${PYTAG}-${PYTAG}-*.whl

sudo dpkg -i automatically installs the wheel that matches the system python3 interpreter. Any user-owned venv must repeat the pip install step because dpkg runs as root and cannot safely write into a venv.

From source

Build wheels for one or more Python versions using isolated uv-managed venvs. See BUILD_WHEELS.md for the full procedure.

cd dx_rt/python_package
./build_all_wheels.sh              # 3.8 – 3.14
./build_all_wheels.sh 3.12         # single version

Produced wheels are written to ./wheelhouse/.

Quick Start

import numpy as np
from dx_engine import InferenceEngine, InferenceOption

# 1. Load a compiled model
option = InferenceOption()
engine = InferenceEngine("model.dxnn", option)

# 2. Prepare inputs matching the model's expected shape/dtype
input_tensor = np.zeros(engine.get_input_shape(), dtype=np.float32)

# 3. Run inference
outputs = engine.run([input_tensor])

# 4. Inspect results
for i, out in enumerate(outputs):
    print(f"output[{i}]: shape={out.shape}, dtype={out.dtype}")

Public API

Re-exported from dx_engine (src/dx_engine/__init__.py):

Runtime Requirements

dx_engine links against libdxrt.so at import time. If you see ImportError: libdxrt.so.<major>: cannot open shared object file, verify:

dpkg -l libdxrt-bin | grep ^ii
ldconfig -p | grep libdxrt.so

If the library is installed but not in the loader cache, run sudo ldconfig once.

The libdxrt build linked at wheel build time and the one present at runtime must share the same major version. Reinstall the matching wheel whenever you upgrade libdxrt-bin.

Verify the Installation

python -c "import dx_engine; print(dx_engine.__version__)"

License

Proprietary — Copyright (C) 2018- DEEPX Ltd. All rights reserved.

This software is provided exclusively to customers supplied with a DEEPX NPU. Unauthorized sharing or use is prohibited. See LICENSE for the full terms; each wheel embeds the same file under dx_engine-<ver>.dist-info/licenses/LICENSE.