Build a Bundle

Start from an exact Hugging Face ID in Supported Models or a compatible local checkpoint directory:

trtmc build Qwen/Qwen3-0.6B \
  --model-revision MODEL_COMMIT \
  -o qwen3-0.6b.bundle

Omit --model-revision for exploration only. A reproducible result pins an immutable revision and records the complete command.

Choose one configuration source

Need	Surface	Example
Common CLI option	Dedicated flag	--precision fp16
Registered feature schema	Config file	--config profile.json
One schema override	Repeatable key/value	--set namespace.field=value
Model-owned parallel build	Topology flag	--tensor-parallel-size 4

trtmc build MODEL_ID \
  --model-revision MODEL_COMMIT \
  --precision fp16 \
  --config build-profile.json \
  --set qwen_vl_vision.dynamic_resolution=true \
  -o model.bundle

Do not copy that combination to an arbitrary family. The selected family owns which schemas, precision modes, quantization formats, graph shapes, and topologies it supports.

Retain the build receipt

Record at least:

• exact model ID and immutable revision;
• output bundle name and checksum;
• complete build command and config file;
• build environment, TensorRT/CUDA cohort, and SM architecture;
• family, runtime strategy, precision, quantization, and topology; and
• whether the resulting bundle is native or platform-specialized.

Run Inspect a Bundle before inference. The CLI Reference is the source for the complete option inventory.