|
Operating System
|
Linux is recommended for production use
; Windows is supported
|
On Windows using Docker Desktop
or Docker Engine in WSL—see Section
2.3.
|
|
NVIDIA Drivers
|
Driver that reports the card and its
VRAM size
|
nvidia-smi must display the driver version and
graphics memory.
|
|
NVIDIA Container
Toolkit
|
required – pass the GPU to the container using
--gpus all
|
Test: docker run --rm --gpus all
nvidia/cuda:12.8.0-base-ubuntu22.04 nvidia-smi. If the
test does not run, any further configuration
is pointless.
|
|
Docker
|
Version 23 or
later
|
docker --version
|
|
Membership in the group
docker
|
A company user is
a member
|
`docker ps` must work without `sudo`;
administrator privileges are not
required.
|
|
Approval with the
catalog data
|
mounted on the host and
passed into the container
|
The mount target must match the configured
prefix (default: /media/windowsshare) and
be available before the container starts—otherwise, the
container will see an empty directory.
|
|
Certificate Mount
|
Only for
HTTPS operation
|
If the resource is missing, the
service silently falls back to HTTP.
|
|
exps-Mount (Checkpoint
73)
|
Only for the "visual,"
"din_lite," and "visual_norm_blend_optin" presets
|
The model file (1.3 GB) is not included in the
image; six of the nine presets run without this
mount.
|
|
HuggingFace Token /
Model License
|
Not
required for the standard build
|
The default gemma4 build (Apache 2.0,
ungated) can be built without a token or license consent.
An HF token is only required for opt-in builds that
also include gemma3 (rollback/A-B).
|
