Deploy DeepSeek on Jetson with MLC

Execution model

Run one phase at a time. After each phase:

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

Prerequisites

Phase 1 — Preflight

Verify JetPack version, available RAM, and disk space before touching Docker.

cat /etc/nv_tegra_release
free -h
df -h /
df -h /ssd 2>/dev/null || true

Expected: L4T R35.x (JP5) or R36.x (JP6), ≥8 GB RAM free, ≥50 GB disk available. [OK] when all three pass. [STOP] if RAM or disk is insufficient.

Phase 2 — Install Docker + nvidia-container

sudo apt update

# JetPack 5.x
sudo apt install -y nvidia-container

# JetPack 6.x — also install curl, then Docker
sudo apt install -y nvidia-container curl
curl https://get.docker.com | sh
sudo systemctl --now enable docker

# Add current user to docker group
sudo usermod -aG docker $USER
newgrp docker

Verify:

docker --version
docker run --rm --runtime nvidia --gpus all ubuntu:22.04 nvidia-smi

Expected: nvidia-smi output shows the Jetson GPU. [OK] when GPU is visible inside the container.

Move Docker storage to SSD (strongly recommended)

Edit /etc/docker/daemon.json:

{
  "data-root": "/ssd/docker",
  "runtimes": {
    "nvidia": {
      "path": "nvidia-container-runtime",
      "runtimeArgs": []
    }
  }
}

sudo systemctl restart docker
docker info | grep "Docker Root Dir"

[OK] when Docker Root Dir points to your SSD path.

Phase 3 — Pull MLC container and download DeepSeek model

# JP5.x:
docker pull dustynv/mlc-llm:r35.4.1

# JP6.x:
docker pull dustynv/mlc-llm:r36.2.0

docker images | grep mlc-llm

Download model weights inside the container:

docker run -it --rm \
  --runtime nvidia \
  --network host \
  -v /ssd/models:/models \
  dustynv/mlc-llm:r36.2.0 \
  bash -c "huggingface-cli download deepseek-ai/DeepSeek-R1-Distill-Qwen-7B --local-dir /models/deepseek-r1-7b"

[OK] when model files are present under /ssd/models/. [STOP] if download fails — see failure decision tree.

Phase 4 — Launch inference

docker run -it --rm \
  --runtime nvidia \
  --network host \
  -v /ssd/models:/models \
  dustynv/mlc-llm:r36.2.0 \
  python3 -m mlc_llm serve /models/deepseek-r1-7b \
    --device cuda \
    --host 0.0.0.0 \
    --port 8080

Test the endpoint:

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-r1-7b","messages":[{"role":"user","content":"Hello"}]}'

[OK] when the API returns a JSON response with a completion.

For full step-by-step commands, screenshots, and model configuration options, read references/source.body.md.

Failure decision tree

Reference files

• references/source.body.md — full original Seeed tutorial with complete MLC configuration, model options, and effect demonstration (reference only)