Quick Start

Install the wheel produced by the release or local package stage:

python -m pip install /path/to/tensorrt_model_connect-0.1.0-*.whl

If the selected family owns extra build or reference dependencies, install its plain requirements file. From a checkout or unpacked source release:

python -m pip install -r families/sana_wm/requirements.txt

The wheel carries the same owner file. After installing the wheel, locate it from the installed families package:

FAMILY_REQUIREMENTS="$(python -c 'from pathlib import Path; import families; print(Path(families.__file__).parent / "sana_wm" / "requirements.txt")')"
python -m pip install -r "$FAMILY_REQUIREMENTS"

There is no central family extra or dependency registry. A family without a requirements.txt needs only the pinned base environment and the wheel.

Build a bundle directly from a Hugging Face model ID:

python -m tensorrt_model_connect build openai-community/gpt2 \
  --precision fp16 \
  --output gpt2.bundle

The CLI downloads the snapshot, reads config.json or model_index.json, and asks every dependency-free family support.py. Exactly one family must claim the checkpoint. That family supplies the default task; pass --task only when selecting another task supported by the same family. The build then imports only the selected families.gpt2.model and calls build(request, writer) once. A prepared local snapshot can be passed in place of the model ID.

Run the bundle:

trtmc run gpt2.bundle \
  --prompt "Hello" \
  --max-new-tokens 32

The loader reads the bundle header and loads exactly libtrtmc_backend_trt.so and libtrtmc_model_gpt2.so from the directory containing the active libtrtmc_runtime.so: the package bin directory for a wheel install, or the library directory for a native CMake install. It returns the abstract task interface declared by the bundle.

To load from another directory, pass --runtime-root DIR naming a directory that contains the matching libtrtmc_core.so, libtrtmc_runtime.so, libtrtmc_backend_trt.so, and libtrtmc_model_gpt2.so. The CLI never searches environment variables, the current directory, or a fallback runtime.