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 separates build tools from runtime entry points:

API	Entry point	Best for
Python build API	python -m tensorrt_model_connect build and tensorrt_model_connect.build()	Resolving a supported checkpoint and building a .bundle.
Native CLI	trtmc inspect and task commands such as trtmc run	Inspecting a bundle or invoking one abstract Task interface.
C Task SDK	trtmc_get_api() from trtmc/trtmc.h	Typed native calls through the public C boundary for migrated families.
Header-only C++ SDK	#include <trtmc/trtmc.hpp> and trtmc::Model::load()	User-compiled convenience wrappers and RAII over that same C boundary.

The new Task SDK is experimental pending its first stable release. Runtime and family implementations upgrade together; the older internal C++ loader API is not a stable user ABI. See C and C++ Task SDK.

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 directory containing the loaded libtrtmc_runtime, or from an explicit runtime-root override.

Hugging Face model ID or local snapshot
  -> python -m tensorrt_model_connect build
  -> model.bundle
  -> trtmc::Model::load() or trtmc TASK
  -> task-specific output

The installed trtmc Python entrypoint uses the Python builder only for build; other commands replace that process with the packaged native executable. There is no resident Python inference wrapper. The SDK defaults to its installed library directory, with an explicit runtime-root override when needed; it does not retry arbitrary backends or family implementations after a failed call.