seeed-projects/seeed-jetson-developtool/gpt-oss-live

GPT-OSS-20B on Jetson Orin

Runs OpenAI's GPT-OSS-20B open-weight model on a Jetson Orin device using llama.cpp compiled with CUDA support. The model is converted from HuggingFace format to GGUF and quantized to Q4_K for on-device inference. An optional OpenWebUI frontend is included.

Hardware: reComputer Super J4012 or other Jetson Orin Prerequisites: JetPack 6.x, Miniconda installed, ~40GB free disk space

Execution model

Run one phase at a time. After each phase:

• Relay all output lines to the user.
• If output contains [STOP] → stop immediately, consult the failure decision tree.
• If output ends with [OK] → tell the user "Phase N complete" and proceed to the next phase.

Phase 1 — deps (~5 min)

Install Miniconda (skip if already installed) and system build dependencies.

# Only if Miniconda is not yet installed:
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-aarch64.sh
chmod +x Miniconda3-latest-Linux-aarch64.sh
./Miniconda3-latest-Linux-aarch64.sh
source ~/.bashrc

sudo apt update
sudo apt install -y build-essential cmake git

Expected: cmake --version and gcc --version both return successfully. [OK]

Phase 2 — build (~2 hours)

Clone and build llama.cpp with CUDA enabled. This step is long — expect roughly 2 hours on Jetson Orin. Do not interrupt the build.

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build -DGGML_CUDA=ON
cmake --build build --parallel

Expected: ./build/bin/llama-cli --version prints a version string. [OK]

Phase 3 — model (varies)

Download the GPT-OSS-20B weights from HuggingFace, upload to Jetson, then convert and quantize.

1. Download from: https://huggingface.co/openai/gpt-oss-20b/tree/main
2. Transfer the downloaded model directory to the Jetson (e.g. via scp or USB drive).
3. Create the conda environment and convert to GGUF:

conda create -n gpt-oss python=3.10
conda activate gpt-oss
cd llama.cpp
pip install .
python convert_hf_to_gguf.py --outfile <path_of_output> <path_of_input_model>

1. Quantize to Q4_K:

./build/bin/llama-quantize <f16_gguf> <output_q4_gguf> Q4_K

Replace <path_of_input_model>, <path_of_output>, <f16_gguf>, and <output_q4_gguf> with your actual paths.

Expected: Q4_K .gguf file created, size roughly 10–12 GB. [OK]

Phase 4 — run

Launch inference with the quantized model. -ngl 40 offloads 40 layers to GPU.

cd llama.cpp
./build/bin/llama-cli -m <model_path> -ngl 40

Expected: model loads and accepts prompts interactively. [OK]

Optional — WebUI

Serve a browser-based chat interface on port 8080, backed by llama.cpp on port 8081.

conda create -n open-webui python=3.11
conda activate open-webui
pip install open-webui
open-webui serve

Then open http://<jetson-ip>:8080 in a browser and set the OpenAI connection URL to http://127.0.0.1:8081.

Failure decision tree

Reference files

• references/source.body.md — original tutorial article (background reading, not required for execution)