Skip to content

README.md

Latest commit

 

History

History
86 lines (59 loc) · 2.12 KB

README.md

File metadata and controls

86 lines (59 loc) · 2.12 KB
Raw

UAII Python SDK

Load a model (or UAII IR), run inference, and profile — without reading C++ internals.

Prerequisites

  1. Build the C++ runtime (produces uaii_capi shared library):
cmake -S ../.. -B ../../build -DUAII_BUILD_TESTS=ON
cmake --build ../../build --config Release --parallel
  1. Bundle the native library into the package (recommended for pip install):
python scripts/bundle_native.py --build-dir ../../build
  1. Optional native extension (pybind11):
cmake -S ../.. -B ../../build -DUAII_BUILD_PYTHON=ON
cmake --build ../../build --config Release --parallel
  1. Install this package editable:
pip install -e .

If the shared library is not found under uaii/_native/ or the build tree:

# Windows PowerShell
$env:UAII_CAPI_PATH = "C:\path\to\uaii_capi.dll"

Defaults are fail-closed (weight_init="none"). GPU backend names use host-fallback unless real device kernels exist.

Build a wheel (with natives)

From the repo root, after a CMake build of uaii_capi:

# Linux / macOS
bash bindings/python/scripts/build_wheel.sh

# Windows PowerShell
powershell -File bindings/python/scripts/build_wheel.ps1

These scripts:

  1. Configure and build with CMake (-DUAII_BUILD_PLUGINS=ON; tip: add -DUAII_BUILD_PYTHON=ON for pybind).
  2. Run scripts/bundle_native.py to copy uaii_capi into uaii/_native/.
  3. Run python -m build to produce wheels under bindings/python/dist/.

Manual equivalent:

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release --parallel
python bindings/python/scripts/bundle_native.py --build-dir build
python -m pip install build
python -m build bindings/python

CI builds a CPU wheel on Ubuntu via .github/workflows/python-wheel.yml (and the multi-OS wheels.yml on tags).

Quick start

import uaii

print(uaii.version(), "c_api", uaii.c_api_version())

s = uaii.Session()
s.load("path/to/model.uaii.json", backend="cpu", weight_init="ones")
s.set("x", [1.0, 2.0, 3.0, 4.0])
s.run()
print(s.get("y_prob"))

See examples/python/load_run_profile.py in the repo root.