Skip to main content

Reference

Reference pages are for exact lookup. Begin with the Quick Start if you have not built a bundle, or use the User Guides for goal-oriented procedures.

TensorRT-Model-Connect exposes three public entry layers:

APIEntry pointBest for
Python build APIpython -m tensorrt_model_connect build and tensorrt_model_connect.build()Resolving a supported checkpoint and building a .bundle.
Native CLItrtmc inspect and task commands such as trtmc runInspecting a bundle or invoking one abstract Task interface.
C++ Task API#include <trtmc/runtime/family_loader.h> and trtmc::load_task()Native applications that need task-specific results.

The build and runtime entry points are intentionally separate. The Python builder resolves exactly one families/<family>/support.py, imports only that family's model.py, and writes a bundle. The native loader reads the bundle's family, task, and backend, then loads exactly one family DSO and one backend DSO from the explicit runtime root.

Hugging Face model ID or local snapshot
-> python -m tensorrt_model_connect build
-> model.bundle
-> trtmc::load_task() or trtmc TASK --runtime-root DIR
-> task-specific output

There is no Python runtime wrapper, central model registry, runtime-strategy switch, backend search path, or fallback runtime discovery in the current architecture.