Two related changes shipped together. The stack rename is independent
but adding `vllm-reward` to the existing `vllm-qwen3` would have made
that name actively misleading.
**Rename:** `stacks/vllm-qwen3/ → stacks/vllm/`. Updated all in-repo
references (README.md root, servers/ana-ml2/, stacks/llama-swap/,
configs/restic/ana-ml2/, docs/runbooks/disaster-recovery.md). Two
intentional history mentions retained (servers/ana-ml2 + stacks/vllm
README).
**Add `vllm-reward` service:** serves Skywork-Reward-V2-Llama-3.1-8B-AWQ
on port 8003. The AWQ output is a locally-quantized model (not from HF),
so bind-mounts `/tank/aimodels/llm:/local-models:ro` rather than the
shared HF cache. Model config.json declares LlamaForSequenceClassification
which vLLM's pooling runner picks up automatically — produces a single
reward score per input via /classify.
**Flag note:** the user's spec listed `--task classify`, but vLLM 0.19.1
deprecated --task in favor of --runner pooling (model architecture in
config.json drives the classification head). Compose uses --runner
pooling with a comment explaining the substitution.
**GPU memory:** no rebalance needed — production had already tuned
EMBED/RERANK down from 0.40 to 0.20 each (canonical .env.example now
matches reality). Adding REWARD at 0.30 totals 0.70, leaving ~14 GB
headroom on the 48 GB Ada.
**Server-side:** brought existing vllm-qwen3 down, mv'd
/opt/docker/compose/vllm-qwen3 → /opt/docker/compose/vllm, appended
REWARD_* lines to existing .env (preserving API_KEY/HF_TOKEN), deployed
new compose via scripts/deploy-stack.sh, brought all 3 services up.
**Smoke tests:**
- /health on 8001/8002/8003 → 200
- /v1/models on 8003 → lists Skywork/Skywork-Reward-V2-Llama-3.1-8B-AWQ
with max_model_len 16384
- /classify with a sample conversation → returns LABEL_0 with prob 0.9999
(single-output regression-style reward score, expected shape for a
reward model)
rest-server-ana README now describes the /mnt/backup/restic/repo/
top-level NFS mount and its three per-site subdirs:
ana/ — live data served by this rest-server (per-host repos +
.htpasswd) — what DATA_DIR points at
esh/ — mirror destination for ESH-site backups (pending)
nh3/ — mirror destination for NH3 Synology's tree (pending)
ana-ml2 README gains a proper "Recreating the repo" section with the
correct /mnt/backup/restic/repo/ana/ana-ml2/ path for wiping the old
repo after a lost passphrase, and two paths for regenerating keys:
- interactive: type a user-generated passphrase at restic's init
prompt, then install it into /etc/restic/password via `cat > file`
+ Ctrl-D (no shell history or transcript exposure)
- scripted: openssl rand -base64 48, passphrase prints once and must
be captured into the password manager immediately
Cross-site replication snippet in rest-server-ana README updated to
use the unified /mnt/backup/restic/repo/{esh,nh3}/ destinations
instead of the earlier restic-mirror-*/ staging paths.
Original used `sudo env \$(cat /etc/restic/restic.env) …` but the `cat`
in command substitution runs as the login user, not root. Since the
env-file is root:600, the substitution silently yielded an empty
RESTIC_REPOSITORY and restic errored with "Please specify repository
location".
Wrap the whole dance in `sudo bash -c "…"` so the env-file read and
the restic call both happen as root.
An ana-ml2 user and repo were created during the original backup
pipeline pass. Reuse keeps snapshot history consolidated and avoids
duplicate infrastructure. Reworks the README to:
- Skip `restic init` (repo exists) and the htpasswd step (user exists)
- Install the two existing secrets (REST URL w/ htpasswd password,
repo passphrase) into /etc/restic/{restic.env,password}
- Verify credentials against the existing repo via `restic snapshots`
Fresh-setup flow retained below as a fallback for zero-state rebuilds.
ana-ml2 is not on any Proxmox hypervisor, so vzdump doesn't touch it.
This closes the biggest single backup gap per the 2026-04-20 pipeline
audit.
Sources: /opt/docker (~110 MB), /etc, /root, /var/lib/docker/volumes.
Excludes /tank/* (model weights — regenerable from Hugging Face and
would blow repo size budget). No pre-backup DB hook — none of the
llama-swap / vllm / comfyui / kokoro / parakeet / vibevoice stacks
use relational databases.
README walks through the one-time setup: rest-server .htpasswd entry,
restic init with fresh passphrase, resticprofile install, systemd timer
generation, verification against the Backrest UI.
vh