Complete guide for converting ONNX models to Ascend AI processor compatible format using ATC (Ascend Tensor Compiler) tool.
| Python | 3.7, 3.8, 3.9, or 3.10 | Python 3.11+ incompatible with CANN 8.1.RC1 | | NumPy | < 2.0 (e.g., 1.26.4) | CANN uses deprecated NumPy API | | ONNX Opset | 11 or 13 (for CANN 8.1.RC1) | Higher opset versions not supported |
SoC version in ATC conversion must exactly match your target device! ```bash # Get exact SoC version from your device npu-smi info | grep Name # Output: Name: 910B3 → Use: --socversion=Ascend910B3 # Output: Name: 310P3 → Use: --socversion=Ascend310P3 ``` Common Error: ``` [ACL ERROR] EE1001: supported socVersion=Ascend910B3,
Complete toolkit for Huawei Ascend NPU model conversion and inference. (1) Convert ONNX models to .om format using ATC tool with multi-CANN version support (8.3.RC1, 8.5.0+). (2) Run Python inference on OM models using ais_bench. (3) Compare precision between CPU ONNX and NPU OM outputs. (4) End-to-end YOLO inference with Ultralytics preprocessing/postprocessing - supports Detection, Pose, Segmentation, OBB tasks. Use when converting, testing, or deploying models on Ascend AI processors. Source: ascend-ai-coding/awesome-ascend-skills.