Hardware requirements
What a machine needs to run potocolom, and which GPUs work.
potocolom runs on one Linux x86_64 machine with Docker. To generate images on that machine, it needs an NVIDIA or AMD GPU. A machine with no GPU can run the API and use a GPU worker on another machine.
Minimum and comfortable
- GPU VRAM: 8 GB minimum. 12-16 GB is comfortable for SDXL-class models at 1024 px, and it lets you switch models without eviction.
- Disk: 20 GB free minimum. The weights are 2-7 GB per model, plus your images. 50 GB or more is comfortable.
- RAM: 8 GB minimum. 16 GB is comfortable.
Each model manifest sets its own VRAM floor in min_vram_gb. The worker measures free VRAM before it loads a model. Run scripts/preflight.sh to see which shipped models fit on your card.
The worker keeps 1536 MiB of measured free VRAM in reserve. Raise VRAM_RESERVE_MB when the GPU also drives a display. Set it to 0 for a GPU that only runs potocolom.
GPU vendors
| GPU | Profile | What it needs |
|---|---|---|
| NVIDIA | gpu |
The NVIDIA Container Toolkit. A host driver of 525.60.13 or newer. The worker image uses CUDA 12.8. |
| AMD | rocm |
The ROCm kernel driver, so that /dev/kfd and /dev/dri exist. The GPU must be on the supported list: gfx900, gfx906, gfx908, gfx90a, gfx942, gfx1030, gfx1100, gfx1101, gfx1102, gfx1200, gfx1201. |
| None | api |
A GPU worker on another host, with the shared FLEET_SECRET. |
The installer checks for NVIDIA before AMD. Set POTOCOLOM_PROFILE to choose a profile yourself.
For RDNA3 consumer cards, such as the RX 7600 XT, the ROCm build supports the card directly. Do not set HSA_OVERRIDE_GFX_VERSION.
The release does not provide CPU inference, ARM images, or Intel or Apple GPU support.
To check a machine before you install, see Checking a machine.