# Compose stacks for phy-srv-gpu01

Rendered by `ironicbadger.docker_compose_generator` (pinned to 2.0.1) into
`{{ docker_compose_generator_output_path }}/compose.yaml` on the host.
The fragments here are **Jinja2 templates**, not plain compose files — Ansible
variables are substituted at render time.

| Stack | Contains |
| --- | --- |
| `01-vllm` | vLLM, serves the chat model on port 8000 (OpenAI API) |
| `02-embeddings` | second vLLM instance, embedding model on port 8001 |
| `03-openwebui` | Open WebUI (port 80) + pgvector Postgres |
| `99-network` | the shared `llmnet` bridge — must sort last |

Ordering comes from the `NN-` prefix; `99-network` keeps `networks:` at the
bottom of the generated file. The leading comment in `99-network/compose.yml`
is load-bearing: the role indents only the first line of each fragment, and the
comment absorbs that indentation so `networks:` stays top-level.

## Deploy

```sh
just compose phy_srv_gpu01     # renders ~/docker/compose.yaml on the host
```

The role only **writes** the file. Bringing the stack up is deliberate:

```sh
ssh sbxadmin@192.168.66.69 'cd ~/docker && docker compose up -d'
```

## Before the first run

- Secrets must exist in `group_vars/secrets.yml` (`just vault edit`):
  `vault_owui_secret_key`, `vault_owui_db_password`, `vault_ldap_bind_password`
- LDAP parameters in `group_vars/phy_srv_gpu01.yml` need the real bind DN and
  base DN from the customer
- `vllm_image` must point at a build that works on SM120 (Blackwell) — verify
  before deploying, the official image is not guaranteed to work