axera

the python api for axengine runtime

Pulsar2 Agent, Build a 10‑stage automated pipeline powered by AI Agent. Convert floating‑point models into out‑of‑the‑box full AX‑chip delivery packages with a single natural‑language instruction to empower algorithm engineers.

Python scripts performing Open Vocabulary Object Detection using the YOLO-World model in ONNX. And Export the ONNX model for AXera's NPU

Try to export the ONNX QDQ model that conforms to the AXERA NPU quantization specification. Currently, only w8a8 is supported.

llama.cpp backend that runs GGUF models directly on the Axera AX8850 NPU — no model conversion, no per-model compile. 24-30 t/s decode on a Raspberry Pi 5 with the CPU idle.

PixArt-Sigma 512 text-to-image on the AXERA AX650/AX8850 NPU (Radxa AICore AX-M1) - 27s per image, Pulsar2 conversion pipeline and runtime