oapython 0.7.2

OA — Vulkan GPU compute & ML foundation library (SUPER EXPERIMENTAL Python bindings)

oapython — OA Python bindings (super experimental)

oapython exposes OA, a Vulkan 1.4 GPU compute & ML engine, to Python via nanobind. You install oapython; you import it as oa.

⚠️ Super-experimental tech preview. The API, kernels, and formats change without notice. This is not a stable or supported API — expect breakage. The C++ library is the primary surface; these bindings are an early look.

Install

pip install oapython        # then: import oa

Requirements: Linux x86-64, CPython 3.12, glibc ≥ 2.39, and a Vulkan 1.1+ driver on the machine. The wheel loads Vulkan dynamically at runtime — no loader is bundled, so the target needs a working Vulkan ICD (NVIDIA, AMD/RADV, or Intel/ANV).

What you get

The bindings call the same compute kernels as the native engine, so the Python API adds no measurable overhead (nanobind is ~100 ns/call — negligible against GPU-op latency). On identical byte-level NLP models, the Python training loop runs at C++ throughput.

End-to-end examples (MNIST classifier, byte-level RNN/GRU/Transformer/Mamba training) live under Tutorial/Py/ in the repository.

Links

• Source & docs: https://github.com/realminc/oa
• Guides: https://dev.realm.software

License

Business Source License 1.1 — source-available; each version converts to Apache-2.0 four years after release. © 2025–2026 Lukasz Biernat (trading as Realm).