harvestsu/z-image-turbo-jetson-trt-artifacts

Z-Image-Turbo Jetson TensorRT Artifacts

This repository hosts generated artifacts for running Tongyi-MAI/Z-Image-Turbo with TensorRT on NVIDIA Jetson.

Code, Dockerfiles, export scripts, and reproduction docs live in the companion source repository. If you publish the source under a different GitHub organization/name, update this section to point there.

Validated Target

Current published artifacts are intended for this target unless a folder name states otherwise:

Field	Value
Device	Jetson Orin NX 16GB
JetPack	6.x
OS	Ubuntu 22.04
CUDA	12.6 host libraries
TensorRT	10.3
Precision	BF16 TensorRT engines
Runtime	TensorRT Python bindings

TensorRT engines are not portable across arbitrary GPUs, TensorRT versions, or JetPack/CUDA stacks. If your target differs, rebuild engines from ONNX on that target.

Layout

runtime-minimal/
  tokenizer/
  scheduler/
  vae/config.json

engines/
  orin-nx-jp6-trt10.3/
    384-bf16/
      *.engine
    512-bf16/
      *.engine
    text-encoder-split-g4/
      *.engine

manifests/
  engines-orin-nx-jp6-trt10.3-384-bf16.sha256
  engines-orin-nx-jp6-trt10.3-512-bf16.sha256
  text-encoder-split-g4-orin-nx-jp6-trt10.3.sha256

runtime-minimal/ contains only small config/tokenizer files required by the no-PyTorch runtime. It does not contain the full upstream model weights.

Expected Size

Approximate validated artifact sizes:

Artifact	Size
384 BF16 TensorRT engines	~12GB
512 BF16 TensorRT engines	~12GB
Split text encoder engines, group size 4	~15GB
Minimal tokenizer/scheduler/VAE config	Small

The original upstream diffusers snapshot is about 20GB and is not duplicated here. Download it from Tongyi-MAI/Z-Image-Turbo when using export or PyTorch fallback paths.

Use From Runtime

Example host layout after downloading this artifact repo:

/home/harvest/models/z-image-trt-artifacts/
  engines/orin-nx-jp6-trt10.3/384-bf16/
  engines/orin-nx-jp6-trt10.3/text-encoder-split-g4/
  runtime-minimal/

Run the no-PyTorch 384 text-to-image path:

DOCKER_IMAGE=z-image-jetson-no-torch:latest \
MODEL_ROOT_HOST=/home/harvest/models/z-image-trt-artifacts \
ENGINE_DIR_384_HOST=/home/harvest/models/z-image-trt-artifacts/engines/orin-nx-jp6-trt10.3/384-bf16 \
TEXT_ENCODER_ENGINE_DIR_HOST=/home/harvest/models/z-image-trt-artifacts/engines/orin-nx-jp6-trt10.3/text-encoder-split-g4 \
MODEL_DIR=/models/runtime-minimal \
RESOLUTION=384 \
scripts/run/run_3drope_no_torch.sh

Validation Reference

Validated on Jetson Orin NX 16GB:

Runtime	Resolution	Steps	Total	TRT denoise	Result
PyTorch-buffer TRT	384	4	73.2s	37.1s	Correct
PyTorch-buffer TRT	512	4	100.2s	63.1s	Correct
no-PyTorch TRT	384	4	92.8s	56.2s	Correct

The no-PyTorch runtime image is 413MB and imports TensorRT, CUDA Runtime through ctypes, NumPy, Pillow, and tokenizers. It does not import PyTorch, diffusers, or transformers.

License

The source project code is MIT licensed.

These artifacts are derived from Tongyi-MAI/Z-Image-Turbo, whose model license is Apache-2.0. Follow the upstream model terms when redistributing or using generated artifacts.