diantoudefengshan/Jetson-PI-GGUF

Jetson-PI PI0 / PI0.5 GGUF Models

GGUF conversions of PI0 and PI0.5 for deployment with Jetson-PI-Edge, a llama.cpp-based inference engine for real-time vision-language-action model execution on NVIDIA Jetson and other CUDA-capable edge platforms.

These files include the language/action model and the matching SigLIP vision encoder/projector. The PI0.5 vision model uses hidden_act: gelu and exports clip.use_gelu=true.

Model Files

Model	File	Description
PI0	pi0/pi_llm.gguf	PI0 language model and action expert, F16
PI0	pi0/mmproj-model-f16.gguf	Matching PI0 SigLIP vision encoder/projector, F16
PI0.5	pi05/pi_llm.gguf	PI0.5 language model and action expert, F16
PI0.5	pi05/mmproj-model-f16.gguf	Matching PI0.5 SigLIP vision encoder/projector, F16

The main model and mmproj file must come from the same model directory. Do not mix PI0 and PI0.5 files.

Runtime

Clone and build Jetson-PI-Edge with CUDA support:

git clone https://github.com/PKU-SEC-Lab/Jetson-PI-Edge.git
cd Jetson-PI-Edge
cmake -S . -B build -DGGML_CUDA=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --target llama-server -j

Start the persistent foreground server for PI0:

PI_MODEL=pi0 ./build/bin/llama-server \
  -m /path/to/pi0/pi_llm.gguf \
  --mmproj /path/to/pi0/mmproj-model-f16.gguf \
  -ngl 99 \
  --host 0.0.0.0 \
  --port 8080

For PI0.5:

PI_MODEL=pi05 ./build/bin/llama-server \
  -m /path/to/pi05/pi_llm.gguf \
  --mmproj /path/to/pi05/mmproj-model-f16.gguf \
  -ngl 99 \
  --host 0.0.0.0 \
  --port 8080

See the Jetson-PI-Edge README for the foreground HTTP API, Python API, FlashRT integration, robot-state format, image submission, and action output format.

Intended Use

The models are intended for research on local and onboard VLA inference, robot-control systems, deployment optimization, and reproducibility of Jetson-PI experiments. They accept camera images, a natural-language task instruction, and robot state through the Jetson-PI-Edge runtime and produce a continuous action chunk.

These checkpoints are not general-purpose chat models and should not be loaded through an ordinary text-only llama.cpp chat endpoint.

Conversion and Validation

The files were converted with the PI conversion tools included in Jetson-PI-Edge:

• convert_hf_to_gguf.py
• tools/mtmd/legacy-models/convert_image_encoder_to_gguf.py
• tools/mtmd/legacy-models/pi0_surgery.py

The converted PI0 and PI0.5 models were loaded and exercised through the Jetson-PI-Edge foreground server with aligned model inputs, camera images, robot state, and action noise. For reproducible correctness checks, compare the pre-unnormalization action output against the corresponding official OpenPI implementation under identical inputs.

License

The GGUF files are derived from their corresponding upstream PI checkpoints. Use of the converted weights remains subject to the original model license and terms. Jetson-PI-Edge source code is distributed under its repository license. Users should review both sets of terms before redistribution or deployment.

Acknowledgments

This work builds on OpenPI, llama.cpp, and the PI model family from Physical Intelligence.

Citation

If these models or Jetson-PI help your research, please cite:

@article{yang2026jetson,
  title={Jetson-PI: Towards Onboard Real-Time Robot Control via Foresight-Aligned Asynchronous Inference},
  author={Yang, Zebin and Wang, Qi and Wang, Yunhe and Guo, Xiurui and Yu, Bo and Liu, Shaoshan and Xu, Jiafeng and Dong, Hao and Li, Meng},
  journal={arXiv preprint arXiv:2607.12659},
  year={2026}
}