菜单
Adds a new operation to OpenVINO Frontend pipelines with translator updates, registration, and tests.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/openvinotoolkit/openvino --skill add-fe-op命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | add-fe-op |
| description | Adds a new operation to OpenVINO Frontend pipelines with translator updates, registration, and tests. |
Enable a new operation in an OpenVINO Frontend (FE) pipeline — translator creation, registration, conversion validation, and tests.
Each frontend has its own detailed workflow. Read the one matching the target framework:
| Frontend | Skill file | What it covers |
|---|---|---|
| PyTorch | pytorch.md | NodeContext API, op_table.cpp registration (TorchScript + FX/Export), translate_1to1_match_* wrappers, Python layer tests, Mark operations |
| ONNX | onnx.md | ov::frontend::onnx::Node API, ONNX_OP macro registration with OPSET_SINCE/OPSET_RANGE, .prototxt test models, C++ GTest cases, normalize-step transformations |
| Frontend | Skill file | When to use |
|---|---|---|
| ONNX | conversion-issues/onnx.md | Conversion failures, accuracy bugs, shape/type mismatches, opset version gaps |
| PyTorch | conversion-issues/pytorch.md | Conversion failures, accuracy bugs, tracing mode issues, normalize-step failures |
The following sections apply to frontends not covered by the framework-specific files above (e.g., TensorFlow), or as a general reference when no framework-specific skill is available.
aten::op).aten.op.default).ONNX_OP macro registration exists in translator file.tests/frontend (not available for PyTorch).tests/layer_tests where supported.Prefer runtime shape computation over conversion-time shape extraction.
ShapeOf-based graph logic for shape-dependent behavior.get_shape() / fully static shape reads during FE conversion.get_partial_shape() is acceptable for compile-time rank-only decisions.Handle data types according to framework semantics.
Avoid constants tied to compile-time element type queries.
Constant values from get_element_type() when type can be dynamic at conversion time.