Reference

Reference pages are for exact lookup, not progressive learning or task instruction. Begin with your first NLP inference if you have not built a bundle, or use User Guides when you need a goal-oriented procedure.

TensorRT-Model-Connect exposes four public entry layers:

API	Entry point	Best for
Python builder	tensorrt_model_connect.build() and trtmc build	Building .bundle bundles from Hugging Face IDs or local model directories.
Python runtime wrapper	tensorrt_model_connect.Pipeline	Text and vision-language generation through the native trtmc executable from Python.
C++ runtime	#include <trtmc/pipeline.h> and trtmc::load()	Native applications that want task-specific inference results.
C-linkage subset	trtmc_create_pipeline_ex() and trtmc_generate_batch()	C++ shims and experimental FFI integration; the current header/handle is not yet a complete pure-C ownership API.

The command-line interface is a thin adapter over these APIs:

• trtmc build is implemented by src/cli/main.cpp delegating to python/tensorrt_model_connect/build_cli.py.
• Runtime subcommands such as trtmc run are implemented under src/cli/.
• tensorrt_model_connect.Pipeline is a subprocess wrapper over trtmc run and trtmc inspect, not an in-process binding to IPipeline.

The core contract is always the same:

Hugging Face model or local model directory
  -> trtmc build
  -> model.bundle
  -> trtmc::load() or trtmc run
  -> task-specific output