atc-model-converter - Convert PyTorch or ONNX models for Ascend NPU
Analyzes model code, exports PyTorch to ONNX, converts ONNX to OM with ATC, adapts NPU inference, and verifies precision.
Tags
Updated: 2026-10-19Capabilities
Typical Inputs
Typical Outputs
What this skill does
- Discover model parameters
- Probe runtime tensor shapes
- Export PyTorch to ONNX
- Inspect ONNX I/O
- Convert ONNX to OM
- Adapt end-to-end NPU inference
- Compare ONNX and OM precision
- Generate reproducible README
Inputs
- Model weights path
- Model code repository
- Task type
- Target input size
- Test images or data
- Target Ascend SoC version
Outputs
- Parameter discovery report
- ONNX model file
- OM model file
- NPU inference pipeline
- Precision comparison results
- Reproducible README
Requirements
- Python 3.10 or lower
- NumPy below 2.0
- CANN 8.1.RC1 or newer
- Ascend ATC environment
- ONNX opset support