openvinotoolkit

Write unit tests for OpenVINO ov::Model graph transformations (passes). Use when the user asks to write, add, or refactor tests for a transformation pass (MatcherPass, ModelPass), or to modernize legacy transformation tests.

Troubleshooting all sorts of failures, crashes, exceptions and errors using debug capabilities. Analyze accuracy, performance, model compilation, or memory issues. Dump tensors and intermediate blobs. Serialize and visualize IRs, execution graphs. Enable verbose, logging. Profile execution. Compare layer outputs. Inspect, trace or dump transformations. Identify executed operations, nodes, primitives, kernels.

Upgrade the PyTorch version used by OpenVINO tests (torch / torchvision / torchaudio) and resolve fallout — missing operator translators, new functionalized `*_copy` aten ops, decomposition changes, FX-only tests failing in TorchScript mode, and accuracy regressions caused by stricter typing. Use when the user asks to "bump torch", "update pytorch to X.Y", "upgrade torch tests", or when pytorch_tests / model_hub pytorch tests fail after a torch version change. Do not use for: enabling a single new PyTorch operator unrelated to a version bump, GenAI / Optimum upgrades, or plugin-level numerical bugs unrelated to the frontend.

Adds a core operator to the OpenVINO toolkit. Use when asked to implement a new operation into OpenVINO.

Add a new operation to the OpenVINO CPU plugin — node registration, JIT/oneDNN executors (AVX2/AVX-512/AMX), and functional tests.

Adds a new operation to OpenVINO Frontend pipelines with translator updates, registration, and tests.

Adds a new OpenVINO fusion transformation (subgraph to one or several operations) and corresponding tests.

Analyze a HuggingFace model and attempt OpenVINO conversion — probe properties, run strategy matrix, classify failures, and produce a structured routing report.

Validate a newly supported optimum-intel model with OpenVINO GenAI. Use when: checking new model support, verifying model export to OpenVINO IR, running GenAI inference test with llm_bench, benchmarking model accuracy with who-what-benchmark.

Enables VLM models for GenAI VLM pipelines.

Analyze failed llm_bench execution results for a model. Use when: checking llm_bench log, troubleshooting llm_bench fails.

Analyze results of WWB execution with a specific model with all possible backends. Use when: checking wwb logs, troubleshooting wwb fails.

Open a pull request to openvinotoolkit/openvino.genai. Use when: submitting changes, creating PR, opening pull request, pushing work for review, finalizing a task with a PR.

Update OpenVINO GenAI site documentation for API or feature changes. Use when: new pipelines, models, or use-cases are introduced; site docs need to reflect new capabilities.

Adds or modifies robot hardware integrations under physicalai.robot. Use when implementing the Robot protocol, SO101 or Trossen WidowX drivers, robot connect helpers, verify.py checks, optional extras so101 or trossen, or tests in tests/unit/robot.

Adds or modifies a camera backend under physicalai.capture. Use when implementing a new Camera type, extending create_camera in src/physicalai/capture/factory.py, discovery helpers, optional pip extras for vendor SDKs, SharedCamera transport, or tests under tests/unit/capture with fake devices.

Configures preprocessors, postprocessors, and runners around InferenceModel via manifest specs and ComponentRegistry. Use when editing physicalai.inference.preprocessors or postprocessors, manifest preprocessor/postprocessor lists, instantiate_component, registered type names, or class_path init_args for inference pipeline components.

Loads and validates policies exported from Physical AI Studio for Runtime deployment. Use when working on InferenceModel, InferenceModel.from_pretrained, manifest.json, adapter auto-detection (onnx, openvino), backend/device kwargs, Hugging Face Hub policy packages, or the Runtime side of the export/load contract that Studio produces with physicalai export.

Runs exported policies on hardware with PolicyRuntime, execution modes, and physicalai run. Use when wiring PolicyRuntime, SyncExecution or RTC execution, runtime YAML configs, action queues, runtime callbacks, or docs/how-to/runtime run-policy-on-robot and execution modes.

Standardize GitHub Actions workflows with minimal functional changes.

Increment the version of NNCF in src/nncf/version.py.

Generate release notes for the new NNCF release.

Validate a HuggingFace tokenizer with OpenVINO Tokenizers and OpenVINO GenAI. Use when: checking if a tokenizer converts and works correctly, verifying tokenizer/detokenizer accuracy, testing normalization steps, checking GenAI Tokenizer compatibility.

Diagnose tokenizer conversion issues for OpenVINO Tokenizers. Use when: tokenizer-checker reports failures, need to pinpoint root cause location (Python conversion vs C++ operation), identify which pipeline stage diverges, determine whether to use tokenizer-fix-python or tokenizer-fix-cpp skill.

Fix Python conversion issues in OpenVINO Tokenizers. Use when: tokenizer-diagnostics reports root_cause_location=python, unsupported types need new handlers in hf_parser.py, pipeline step mapping or merging is incorrect, tokenizer_pipeline.py step classes need fixes.