# TensorRT-Model-Connect

> Build TensorRT bundles from Python-first checkpoints and run them through a native C++ task API.

Documentation: https://nvidia.github.io/TensorRT-Model-Connect/
Repository: https://github.com/NVIDIA/TensorRT-Model-Connect

## Start here

- [Overview](https://nvidia.github.io/TensorRT-Model-Connect/)
- [Get Started](https://nvidia.github.io/TensorRT-Model-Connect/getting-started/overview)
- [Supported Models](https://nvidia.github.io/TensorRT-Model-Connect/models-recipes/overview)
- [User Guides](https://nvidia.github.io/TensorRT-Model-Connect/user-guides/overview)
- [Tutorials](https://nvidia.github.io/TensorRT-Model-Connect/learning-path)
- [Reference](https://nvidia.github.io/TensorRT-Model-Connect/api/overview)
- [Developer Guide](https://nvidia.github.io/TensorRT-Model-Connect/developer-guide/overview)
- [Release and Support](https://nvidia.github.io/TensorRT-Model-Connect/release-support/overview)
- [AI and Agent Guide](https://nvidia.github.io/TensorRT-Model-Connect/agent-guide)

## Agent contract

- Read all applicable AGENTS.md files before acting; the closest scoped file wins.
- Preserve user changes and inspect branch, status, and remotes before editing.
- Do not weaken tests or thresholds to make CI pass.
- Do not commit, push, publish, merge, delete, reset, or rebase without authorization.
- Use exact hf_id, revision, manifest profile, task, precision, topology, and runtime path.
- Keep native support and platform-specialized provider claims separate.
- Report source/static, GPU execution, model parity, performance, and publication evidence separately.
- Never expose credentials, gated assets, private URLs, or internal artifacts.

## Sources of truth

- Python family ownership: python/tensorrt_model_connect/families/<family>/MODEL.toml
- Native runtime ownership: src/runtime/models/<owner>/MODEL.toml
- Exact native model contracts: tests/e2e/models/<family>/manifests/*.json
- CLI: website/docs/api/cli-reference.md and trtmc --help
- Repository agent rules: AGENTS.md