Build a Bundle

Start from an exact supported model ID or compatible local snapshot:

python -m tensorrt_model_connect build Qwen/Qwen3-0.6B \
  --revision MODEL_COMMIT \
  --precision fp16 \
  --max-sequence-length 256 \
  --output qwen3-0.6b.bundle

--output is required. Pin an immutable revision for reproducible evidence. The resolver requires exactly one families/*/support.py match and imports only that family's model.py.

Choose explicit typed inputs

Need	Option
Non-default family task	--task TASK
Precision/backend	--precision, --backend
Text/image/video build bounds	--max-sequence-length, --image-height, --image-width, --video-num-frames
Batch/topology	--max-batch-size, --tensor-parallel-size, --context-parallel-size
Family-owned quantization	--quantization NAME, repeatable --fp32-layer INDEX
Compatible dynamic KV build	--dynamic-kv-cache

The selected family implements or explicitly rejects each non-default value. Do not copy an option combination to a different checkpoint and infer support. There is no generic config file/options registry or build fallback.

Retain the build receipt

Record the exact source revision, model ID/revision, family requirements, complete command, output checksum, family/task/backend from inspection, precision/topology, target GPU, and TensorRT/CUDA cohort. A successful build proves artifact construction only; it does not prove Task correctness or parity.

Next, inspect the bundle, then run its declared Task.