Docker deployments

Recipes for serving models with Subconscious SGLang from prebuilt Docker images. Each subfolder targets one GPU type and holds its own guide (README.md) and launch script (launch.sh).

Folder	GPUs	Model	Default image
h100/	H100 / H200 (sm_90)	Qwen3.6-27B, Subconscious x-mode	subconsciouslabs/orangeline-dell:sm_90-v0.2

Quick start

The images are private, so log in to Docker Hub first, once per host. The username is subconsciouslabs; the password is the access token we share with you:

docker login -u subconsciouslabs                                       # paste the token at the password prompt
echo "$DOCKER_TOKEN" | docker login -u subconsciouslabs --password-stdin   # or non-interactively

Then start the server:

cd h100
./launch.sh                      # uses every GPU on the host
DETACH=1 ./launch.sh             # same, in the background (logs: docker logs -f subc)
GPUS=0,1 PORT=8001 NAME=subc-2 ./launch.sh # uses GPUs 0 and 1, serves on port 8001, container named subc-2

The launch script finds the GPUs, picks the tensor-parallel size, sizes the memory pools, and starts an OpenAI-compatible server on port 8000 (or the port specified with PORT). Once the log shows the server is ready:

curl http://localhost:8000/v1/models
curl http://localhost:8000/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"model": "Qwen/Qwen3.6-27B", "messages": [{"role": "user", "content": "Hello"}]}'

Settings such as GPUS, PORT, MODEL and ENABLE_HICACHE are passed as environment variables; the folder's README lists them all.

Host requirements

All deployments need:

• NVIDIA driver and the NVIDIA Container Toolkit. If docker run fails with could not select device driver "" with capabilities: [[gpu]], the toolkit is missing or not registered with Docker: sudo nvidia-ctk runtime configure --runtime=docker && sudo systemctl restart docker.
• A Docker Hub login. The subconsciouslabs repositories are private, so run docker login (see Quick start) before the first launch. Without it, docker run stops with pull access denied ... may require 'docker login'.
• Model weights, either in a Hugging Face cache on the host or in a local directory. Each folder's README gives the details.

Images

Images are built from the SubconsciousCache-SGLang repository and tagged <sm_arch>-v<release>, for example sm_90-v0.1. An image only runs on the GPU architecture in its tag: an sm_90 image will not start on B200 (sm_100) or A100 (sm_80). Use the folder that matches your GPUs.