A GPU Local-AI Lab on an External NTFS SSD (Docker on a Sparse ext4 Image)

Containers and models on an external drive without reformatting it: Docker's data root inside a sparse ext4 image, GPU in containers, measured tokens/sec.

2026-10-12#AI Lab

work: Oct 2026 · Built · AI, Docker, Linux, GPU, Ollama

Goal

Run local models on the workstation's RTX 3070 (8 GB) and keep the internal NVMe clean: models, containers, RAG data and datasets all on an external 1 TB NVMe. That SSD already held data on NTFS, and I didn't want to wipe it.

Layout

/mnt/rev-ssd/ai-lab/
├── models/  agents/  rag/  datasets/  training/  evals/
├── logs/    backups/ config/
└── docker-storage.ext4   ← 100 GB sparse ext4 image

The SSD mounts with ntfs-3g (the in-kernel ntfs3 mount failed on this disk)
docker-storage.ext4 is loop-mounted at /mnt/ai-docker
Docker's data-root and containerd's root both point inside it

Why a sparse image

Docker's overlay storage needs a real Linux filesystem: permissions, xattrs, overlayfs. NTFS can't give it that. A sparse ext4 image only uses the blocks it has written, so "100 GB" costs nothing until it fills, and all of Docker's state lives in one file I can back up or move.

truncate -s 100G /mnt/rev-ssd/ai-lab/docker-storage.ext4
mkfs.ext4 /mnt/rev-ssd/ai-lab/docker-storage.ext4
sudo mount -o loop /mnt/rev-ssd/ai-lab/docker-storage.ext4 /mnt/ai-docker
# /etc/docker/daemon.json → { "data-root": "/mnt/ai-docker/docker" }
# containerd config.toml   → root = "/mnt/ai-docker/containerd"

GPU in containers

With the NVIDIA Container Toolkit:

sudo docker run --rm --gpus all nvidia/cuda:12.9.1-base-ubuntu24.04 nvidia-smi

showed the RTX 3070 on driver 580.x. Then Ollama and Open WebUI through Docker Compose: Ollama on 127.0.0.1:11434 with models stored on the SSD, and Open WebUI on localhost:3000 with its data inside the ext4 image.

Results

Model	Size on disk	Result
Qwen3 8B (abliterated)	~5 GB	fits fully in VRAM
DeepSeek-R1 14B (abliterated)	~9 GB	~10.4 tok/s, ~71% GPU / 29% CPU split

Takeaway: on an 8 GB card, 4B–9B quantized models are the sweet spot. Anything bigger works, but slowly.

nvidia-smi with Ollama loaded — The RTX 3070 with a 14B model loaded: about 6.9 of 8 GB in use by the llama server.

ollama ps CPU/GPU split — ollama ps: the 14B model is split 29% CPU / 71% GPU, which is where the slowdown comes from.

Operational choices

Docker and containerd are disabled at boot. The lab starts when I want it, not on every login.
One command brings the stack up or down.
It survived a reboot: the image remounted and every container and model came back without reformatting.
Open item: make the SSD mount reliable at boot without fighting GNOME's automounter.

Sources: Build Local AI Lab (12 Oct 2026)