reference
Configuration and commands
Settings, ports, storage and the commands you will use day to day.
Settings
Everything is set with environment variables, passed with -e NAME=value. The image carries a commented template with every setting. Copy it out, uncomment what you need and pass it with --env-file:
bash
docker run --rm --entrypoint cat ghcr.io/deeplm/dlm-ops-dev:0.2.0-rc.4 /etc/dlm-ops/dlm-ops.env.example > dlm-ops.envThese are the settings you are most likely to need.
| Setting | Default | Meaning |
|---|---|---|
DLM_OPS_SLURM | auto | Where cluster data comes from: auto, local, replay (the demo cluster), off, or ssh (experimental). |
DLM_OPS_ENGINE | auto | Where the model runs: local (this container), dmr (Docker Model Runner, the Mac's GPU) or auto. |
DLM_OPS_DEVICE | auto | auto, gpu or cpu. |
DLM_OPS_GPU_MIN_FREE_MIB | 4608 | Use a GPU only with this much free memory. |
DLM_OPS_SLEEP_IDLE | 900 | Seconds idle before the model is unloaded to free GPU memory. 0 turns it off. |
DLM_OPS_THREADS | min(8, usable CPUs) | CPU threads. |
DLM_OPS_CTX | 32768 | Context tokens per conversation. |
DLM_OPS_PASSWORD_FILE | unset | A file holding the sign-in password, at least 12 characters. Unset means a sign-in link that stays valid until restart. |
DLM_OPS_OFFLINE | unset | Set to 1 to never download the model. |
DLM_OPS_PUBLIC_URL | unset | The base URL printed in the sign-in link. Set it if you publish a different host port. |
DLM_OPS_ALLOWED_HOSTS | empty | Extra accepted Host header names, comma-separated. |
DLM_OPS_AUDIT | stdout | Audit events: stdout, file, both or off. |
Ports and storage
| What | Value |
|---|---|
| Port | 8090, published on 127.0.0.1 only |
| Volume | dlm-ops, mounted at /var/lib/dlm-ops |
| Model | Qwen3.5-4B-Q4_K_M.gguf, 2.55 GiB, downloaded on the first start and checked against a SHA-256 value |
If port 8090 is taken, publish another host port and tell DeepLM Ops, so the printed link is right:
bash
-p 127.0.0.1:8091:8090 -e DLM_OPS_PUBLIC_URL=http://127.0.0.1:8091The model
The model is not in the image. You can reuse a file you already have by mounting it read-only. It is still checked:
bash
-v /path/to/Qwen3.5-4B-Q4_K_M.gguf:/opt/models/Qwen3.5-4B-Q4_K_M.gguf:roBehind a proxy, add -e HTTPS_PROXY=http://proxy:port. On a host with no internet, add -e DLM_OPS_OFFLINE=1 and mount the file as above.
Everyday commands
bash
# Show the logs and the sign-in link
docker logs dlm-ops
# Get a new sign-in link
docker restart dlm-ops
# Check the setup: the GPUs it sees, the engine, and the cluster connection
docker exec dlm-ops dlm-ops doctor
# Stop and remove the container
docker rm -f dlm-ops
# Also delete the downloaded model and all state
docker volume rm dlm-ops