Captures the May 18-20 session arc:
- Skaldsong CD shipped on ana-docker:8300 via registry-pull pattern
(differs from althing/asset-engine which build-on-host); three
first-deploy footguns surfaced + canonical-patched.
- Worldtree CD disk-hygiene PR #184 (vh/Worldtree) — watermark gate,
eager post-deploy SHA prune, set -e hardening in SSH blocks.
Approved by worldtree-dev, pending merge.
- mead-hall Bifrost v0.3 wire-layer + data-flow end-to-end smoke
green (task #32 closed) — resolves the "stalled mid-Worldtree"
state in the prior snapshot.
- Volva systemd install — three-stage debug (PATH, ReadWritePaths,
handle pinning); impersonation footgun caught via PID/journal
cross-reference.
- Per-project user keys minted against personal Worldtree (:8081)
for skaldsong, althing, mead-hall + skaldsong-prod. Personal
becomes multi-consumer dev iteration instance; demo stays for
isolation.
File now at 434 lines (over 300 soft cap); next /snapshot run will
surface archival candidates.
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)
vh