What problem does it solve? Deploying ASR models (encoder, decoder, punctuation) on the Kirin9030 HarmonyOS NPU requires 4bit W4A16 quantization and omg offline model conversion, a process full of platform-specific pitfalls such as UINT4-only weight packing, container-only omg execution, and cryptic conversion errors. This Skill codifies the full pipeline and troubleshooting knowledge so the conversion succeeds on the first attempts. ## Core Features & Use Cases - Five-stage quantization pipeline: dopt PTQ calibration, quant parameter export, ONNX export with grid-aligned weights, graph rewriting, dimension padding, and in-container omg conversion to omc. - Kirin9030 platform rules: Enforces the two hard constraints—omg must run inside a glibc >= 2.35 container with the AscendC toolchain activated, and W4 kernels only accept UINT4 weights on the 4bit grid. - Model-specific strategies and packaging: Covers encoder/decoder/punc mixed-precision configs, encoder splitting into multiple omc segments, and CANNPAK packing of all omc files plus embedding binaries into a single on-device bin. - Use Case: You need to deploy a Conformer encoder on a Kirin9030 phone. Follow the pipeline to calibrate with real data, export grid-aligned weights, rewrite the ONNX graph, run omg in the container, split the encoder into segments, and pack everything into one bin—while routing errors like "CompressInt4ToInt8 Value -1 out of range" to their root causes. ## Quick Start Ask the assistant to quantize your ASR encoder to 4bit W4A16 and convert it with omg for Kirin9030 deployment, and it will walk you through the five-stage pipeline and platform rules.
