name: onnx2ms-runtime-operator-fix description: Use when user reports MindSpore Lite runtime errors, .ms model inference failures, or mentions "鸿蒙运行时报错", "ms模型报错", "推理报错", "推理失败", "运行时错误", "benchmark报错", "模型加载失败", "执行失败". Analyzes error logs, provides diagnosis, suggestions, and can fix unsupported operators by replacing them with equivalent subgraphs.

ONNX to MS Runtime Operator Fix Skill

Analyzes and fixes MindSpore Lite (.ms) model runtime errors during inference. Now supports device-specific operator detection and automatic replacement for unsupported operators.

When to Use

Trigger this skill when user mentions:

"鸿蒙运行时报错"
"ms模型报错"
"推理报错" / "推理失败"
"运行时错误"
"benchmark报错"
"模型加载失败"
"执行失败"
"NPU算子不支持" / "CPU算子不支持"
"Split算子报错" / "Mod算子报错"

Supported Error Types

Error Type	Pattern Keywords	Diagnosis	Suggestions
shape_mismatch	shape mismatch, dimension, size mismatch, expected shape	输入tensor维度与模型期望不匹配	检查输入数据shape，使用--inputShape参数
dtype_error	dtype error, data type mismatch, type mismatch, float16, int8	输入/输出数据类型与模型定义不符	使用--inputDataType/--outputDataType重新转换
memory_error	OOM, out of memory, allocate failed, memory error	内存不足导致推理失败	启用INT8量化，减小batch size
unsupported_op_cpu	op not supported CPU, kernel not found CPU, Expand not supported	算子在CPU设备上未实现	切换设备或使用ohos_runtime_op_fix修复
unsupported_op_npu	op not supported NPU, kernel not found NPU, Split not supported	算子在NPU设备上未实现	切换设备或使用ohos_runtime_op_fix修复
unsupported_op_gpu	op not supported GPU, kernel not found GPU, OpenCL error	算子在GPU设备上未实现	切换设备或使用ohos_runtime_op_fix修复
unsupported_op_runtime	op not supported, unsupported operator (generic)	算子在目标设备上未实现	查看算子支持列表，切换设备尝试
device_error	GPU not available, NPU error, OpenCL error, device failed	目标设备不可用或驱动问题	检查设备驱动，切换到CPU运行
input_error	input error, input file not found, invalid input	输入数据文件不存在或格式错误	检查--inDataFile路径
model_load_error	model load failed, model invalid, parse error, file corrupted	.ms文件损坏或格式错误	检查文件完整性，重新转换
quant_accuracy_error	accuracy threshold, cosine distance, precision loss	INT8量化后精度不达标	调整校准数据集
negative_dimension	negative dimension, CheckTensorsInvalid, Resize()	模型包含动态shape维度	使用--inputShape参数指定输入维度

Supported Operator Replacements

Operator	Device Issue	Replacement	Strategy	Description
Split	NPU不支持	Slices	split_to_slices	将Split算子替换为多个Slice算子
Mod	NPU不支持	Sub+Mul+Div	mod_to_sub_mul_div	将Mod算子替换为等价计算子图
Expand	CPU不支持	Tile	expand_to_tile	将Expand算子替换为Tile算子

Workflow

Case A: User Reports Runtime Error (Diagnosis Only)

User provides error log describing the failure
Call ohos_runtime_analyze with error_log parameter
Script parses and matches all error types from log
Detects device (CPU/NPU/GPU) and specific operator if applicable
Return diagnosis and suggestions for each matched type

Case B: NPU/CPU Unsupported Operator (Auto Fix)

User reports: "模型在NPU上推理报错 Split算子不支持"
Call ohos_runtime_analyze with error_log and device="NPU"
Detect: unsupported_op_npu with operator "Split"
Check replacement config: Split has fix strategy (split_to_slices)
Suggest: 使用 ohos_runtime_op_fix 替换 Split 算子
Execute: ohos_runtime_op_fix(onnx_path, device="NPU", operators=["Split"])
Generate fixed ONNX file
Re-convert and re-run benchmark

Case C: Explicit Fix Request

User calls ohos_runtime_op_fix with ONNX file path
System detects unsupported operators for specified device
Execute operator replacement using configured strategy
Generate fixed ONNX file
Suggest user to re-run conversion

Tools

ohos_runtime_analyze

Call: ohos_runtime_analyze
Arguments: {
    "error_log": "<error-log-string>",     // Required
    "ms_path": "<path-to-ms-file>",        // Optional, for context
    "device": "CPU|GPU|NPU"                 // Optional, target device
}

Output:

{
    "matched_errors": [
        {
            "error_type": "unsupported_op_npu",
            "diagnosis": "算子在NPU设备上未实现",
            "suggestions": [
                "使用 ohos_runtime_op_fix 工具替换 Split 算子",
                "替换策略: 将Split算子替换为多个Slice算子",
                ...
            ],
            "matched_patterns": ["op not supported NPU"],
            "confidence": 0.9,
            "operator": "Split",
            "device": "NPU",
            "replacement": {
                "operator": "Split",
                "strategy": "split_to_slices",
                "has_fix": true
            }
        }
    ],
    "total_matches": 1,
    "has_errors": true,
    "detected_device": "NPU",
    "detected_operator": "Split"
}

ohos_runtime_op_fix

Call: ohos_runtime_op_fix
Arguments: {
    "onnx_path": "<path-to-onnx-file>",
    "output_path": "<output-path>",        // Optional, defaults to *_runtime_fixed.onnx
    "device": "CPU|NPU|GPU",               // Required, target device
    "operators": ["Split", ...]            // Optional, auto-detect if not specified
}

Output:

{
    "success": true,
    "device": "NPU",
    "operators_replaced": ["Split"],
    "nodes_per_operator": {"Split": 3},
    "output_path": "model_runtime_fixed.onnx",
    "message": "成功替换 1 种算子"
}

Example Usage

Example 1: NPU Split Operator Error

User: "模型在NPU上推理报错 Split算子不支持"

Call: ohos_runtime_analyze
Arguments: {
    "error_log": "op not supported: Split is not implemented for NPU device",
    "device": "NPU"
}

Result:
- error_type: unsupported_op_npu
- detected_operator: Split
- replacement: {strategy: "split_to_slices", has_fix: true}
- suggestions: 使用 ohos_runtime_op_fix 替换 Split 算子

Call: ohos_runtime_op_fix
Arguments: {
    "onnx_path": "model.onnx",
    "device": "NPU",
    "operators": ["Split"]
}

Result:
- 3 Split nodes replaced with Slices
- Output: model_runtime_fixed.onnx
- Suggest: re-run conversion

Example 2: CPU Expand Operator Error

User: "模型在CPU上报错 Expand算子不支持"

Call: ohos_runtime_analyze
Arguments: {
    "error_log": "kernel not found for op type Expand on CPU",
    "device": "CPU"
}

Result:
- error_type: unsupported_op_cpu
- detected_operator: Expand
- replacement: {strategy: "expand_to_tile", has_fix: true}

Call: ohos_runtime_op_fix
Arguments: {
    "onnx_path": "model.onnx",
    "device": "CPU"
}

Result:
- Expand nodes replaced with Tile
- Suggest: re-run conversion

Example 3: Multiple Errors

User: "运行时报错：GPU not available，而且 out of memory"

Call: ohos_runtime_analyze
Arguments: {
    "error_log": "GPU not available, out of memory, allocate failed"
}

Result:
- matched_errors: [device_error, memory_error]
- total_matches: 2
- 每种错误返回独立的诊断和建议

File Locations

Analyzer Script: .opencode/skills/onnx2ms-runtime-operator-fix/scripts/runtime_error_analyzer.py
Operator Replace Script: .opencode/skills/onnx2ms-runtime-operator-fix/scripts/runtime_op_replace.py
Replacement Config: .opencode/skills/onnx2ms-runtime-operator-fix/config/operator_replacements.json
Error Examples: .opencode/skills/onnx2ms-runtime-operator-fix/examples/*.json
MCP Server: ohos_model_claw/mcp_server.py
Error Parser: ohos_model_claw/error_parser.py
Core Logic: ohos_model_claw/ohos_model_claw.py

Notes

Returns ALL matched error types (not just first match)
Device-specific patterns allow precise diagnosis (CPU/NPU/GPU)
Operator extraction from error logs enables targeted fixes
Replacement strategies are configured in operator_replacements.json
Confidence score indicates pattern match ratio
Independent from onnx-converter-operator-fix (handles runtime, not conversion)
Some operators may not have replacement strategies (check has_fix flag)

Relationship with Other Skills

Skill	Scope	Trigger	Fix Available
onnx-converter-operator-fix	转换阶段算子错误	ONNX转换时ERROR日志	HardSigmoid/HardSwish/Swish/Mish
onnx2ms-runtime-operator-fix	运行时推理错误	benchmark或推理时报错	Split/Mod/Expand (device-specific)

This skill handles errors AFTER conversion with device-specific diagnosis and operator replacement, while onnx-converter-operator-fix handles errors DURING conversion.