# 1. Apply env overrides to the profile .env file
# 2. docker compose --env-file .env config > resolved.yml   (dry-run)
# 3. Review resolved.yml
# 4. docker compose -f resolved.yml up -d

# 1. GPU visible
nvidia-smi --query-gpu=index,name --format=csv,noheader

# 2. NVIDIA runtime in Docker
docker info 2>/dev/null | grep -i "runtimes"

# 3. NVIDIA runtime works end-to-end
docker run --rm --gpus all ubuntu:22.04 nvidia-smi 2>&1 | head -5

If this is the host's first deploy, the `docker compose down`
line is a no-op (exit 0 with no containers to stop) — safe to run
unconditionally.

### Step 1 — Gather context

Discover what's available on the host and cross-reference with the
[VSS prerequisites page](https://docs.nvidia.com/vss/3.1.0/prerequisites.html)
to choose a deployment shape that fits.

| Value | How to determine |
|---|---|
| **Profile** | Match user intent to routing table above. Default: `base` |
| **Repo path** | Find `video-search-and-summarization/` on disk |
| **Hardware** | `nvidia-smi --query-gpu=name,memory.total --format=csv,noheader` → look up per-GPU VRAM against the prerequisites page |
| **LLM/VLM placement** | Pick `local_shared`, `local`, or `remote` per LLM/VLM based on available GPUs + `$LLM_REMOTE_URL` / `$VLM_REMOTE_URL` / `$NGC_CLI_API_KEY`. If no combination on this host satisfies the prerequisites, stop and report the blocker instead of silently picking another shape. |
| **API keys** | `NGC_CLI_API_KEY` for local NIMs, `NVIDIA_API_KEY` for remote |
| **Host IP** | `hostname -I \| awk '{print $1}'` |

**Hardware profile mapping:**

| GPU name contains | HARDWARE_PROFILE | Recommended LLM path |
|---|---|---|
| H100 | `H100` | Nano 9B v2 (NIM) |
| L40S | `L40S` | Nano 9B v2 (NIM) |
| RTX 6000 Ada, RTX PRO 6000 | `RTXPRO6000BW` | Nano 9B v2 (NIM) |
| GB10 (DGX Spark) | `DGX-SPARK` | **Edge 4B** (vLLM) — see [`references/edge.md`](references/edge.md) |
| IGX | `IGX-THOR` | **Edge 4B** (vLLM) — see [`references/edge.md`](references/edge.md) |
| AGX | `AGX-THOR` | **Edge 4B** (vLLM) — see [`references/edge.md`](references/edge.md) |
| Other | `OTHER` | — |

**Minimum GPU count per (profile × mode × platform).** Canonical source
is the [VSS prerequisites page](https://docs.nvidia.com/vss/3.1.0/prerequisites.html);
reproduced here so the skill can fail fast when the host is too small:

| Profile | Mode | H100 / RTX PRO 6000 (Blackwell) | L40S | DGX-Spark / IGX-Thor / AGX-Thor |
|---|---|---|---|---|
| `base` | shared (`local_shared` LLM + VLM) | **1** | — (48 GB/GPU too small) | **1** (Edge 4B + VLM, unified memory) |
| `base` | dedicated (`local` LLM + VLM) | **2** | **2** | — |
| `base` | `remote-llm` | **1** (VLM local) | **1** (VLM local) | **1** (remote LLM only) |
| `base` | `remote-vlm` | **1** (LLM local) | **1** (LLM local) | — |
| `base` | `remote-all` | **0** | **0** | **0** |
| `lvs` | shared | **1** | — | - |
| `lvs` | dedicated | **2** | **2** | — |
| `lvs` | `remote-llm/vlm` | 1 | 1 | - |
| `lvs` | `remote-all` | 0 | 0 | - |
| `alerts` (verification / CV) | shared | **2**  | — | — |
| `alerts` (verification / CV) | dedicated | **3** | **3**  | — |
| `alerts` (verification / CV) | `remote-all` | 1 | 1 | 1 |
| `alerts` (verification / CV) | `remote-llm/vlm` | 2 | 2 | 1 |
| `alerts` (real-time / VLM) | shared | **2** | — | — |
| `alerts` (real-time / VLM) | dedicated | **3** | **3**  | — |
| `alerts` (real-time / VLM) | `remote-llm` | 2 | 2 | 1 |
| `search` | shared | **2** | — | - |
| `search` | dedicated | **3** | **3**  | — |
| `search` | `remote-*` | **2**  | **2** | - |

A few hard rules encoded in the table:

- **L40S can't do `shared`.** 48 GB is not enough VRAM for LLM + VLM
  on a single GPU. Fall back to `dedicated` or a `remote-*` mode.
- **L40S needs +1 GPU for alerts / search vs H100** because the
  shared-on-one-GPU trick doesn't work — RT-CV / Embed1 must take
  their own GPU, and LLM+VLM still need a second.
- **DGX-Spark / Thor are early-access for most profiles.** Only
  `base` + `lvs` are expected to fully land locally; `alerts` /
  `search` currently require a remote LLM. See
  [`references/edge.md`](references/edge.md).

If the host's (GPU count × VRAM) combination doesn't appear above,
**stop and report the blocker** — don't silently pick a different
mode.

> **Edge shared mode requires Edge 4B + `HF_TOKEN`.** On DGX Spark and AGX/IGX
> Thor, both LLM and VLM must fit in unified memory, AND the standard
> `nvcr.io/nim/nvidia/nvidia-nemotron-nano-9b-v2:1` image has a broken arm64
> manifest. You must run `NVIDIA-Nemotron-Edge-4B-v2.1-EA-020126_FP8` as a
> standalone vLLM container on port 30081 with the agent pointed at it via
> `--use-remote-llm`. Full recipe and the mandatory `HF_TOKEN` verification
> step are in [`references/edge.md`](references/edge.md).

### Step 1b — Prepare the data directory

The data directory layout (asset paths, ownership, mount points, profile-specific subdirs) is documented in [`references/data-directory.md`](references/data-directory.md). Read that file before deploying for the first time on a host or when changing profiles.
# Profile-specific subdirs:
#   alerts → mkdir -p "$DATA/data_log/vss_video_analytics_api" "$DATA/videos/dev-profile-alerts" "$DATA/models/rtdetr-its" "$DATA/models/gdino"
#   search → mkdir -p "$DATA/models"
chmod -R 777 "$DATA/data_log" "$DATA/agent_eval"
# If you created $DATA/models above, also: chmod -R 777 "$DATA/models"

sudo rm -rf "$DATA/data_log/vst/postgres"  # postgres re-initializes on next start
docker restart centralizedb-dev

REPO=/path/to/video-search-and-summarization
PROFILE=base
ENV_FILE=$REPO/deployments/developer-workflow/dev-profile-$PROFILE/.env

# Read current .env, apply overrides, write back
# (read lines, update matching keys, append new keys, write)

# Resolve compose
cd $REPO/deployments
docker compose --env-file $ENV_FILE config > resolved.yml

cd $REPO/deployments
docker compose -f resolved.yml up -d

# Container status
docker ps --format 'table {{.Names}}\t{{.Status}}\t{{.Ports}}'

# Logs for a specific service
docker compose -f $REPO/deployments/resolved.yml logs --tail 50 <service>

cd $REPO/deployments
docker compose -f resolved.yml down

# 1. All expected containers Up
docker ps --format 'table {{.Names}}\t{{.Status}}'

# 2. Agent API + UI responding
curl -sf http://localhost:8000/docs >/dev/null && echo "agent OK"
curl -sf http://localhost:3000/ >/dev/null && echo "ui OK"

# 3. VLM NIM responding (base/lvs profiles)
curl -sf http://localhost:30082/v1/models | python3 -m json.tool

# 4. LLM NIM responding
curl -sf http://localhost:30081/v1/models | python3 -m json.tool