Retire qwen35-vl (Qwen3.5-9B); add qwen36-vl serving the official FP8
Qwen3.6-35B-A3B vision MoE on :8007 under its TRUE name only — no alias.
qwen3.5-9b-fp8 is killed at vLLM AND the litellm gateway (404/400); a model is
never served under a prior model's name. Consumer (comfy-dev/arbo) notified +
migrated; arbo vkeys flipped to all-proxy-models; shared all-agents-local key
repointed to qwen3.6-35b-a3b.
GPU-1 rebalance for the heavier FP8 weights (~34 GB): granite 0.35->0.24 /
131K->64K, embed/rerank 0.05->0.03 (reclaimed util-reservation waste). Verified:
vision correct, 20-concurrent/endpoint load test = no OOM (~7.5 GB headroom).
Drop the llama-swap qwen3.5-9b GPU-0 pin (GPU 0 freed for the creative-writing
hot-swap card). NVFP4 was the lighter fit (~21 GB) but its vLLM ModelOpt-MoE
loader is broken (KeyError w2_input_scale / lm_head.input_scale, vllm #44081);
revisit when fixed.
ana-ml2 was upgraded 2026-06 from dual RTX 6000 Ada (46GB, cc 8.9) to
dual RTX PRO 6000 Blackwell Max-Q (96GB, cc 12.0 / sm_120). Update the
stale hardware facts across the workspace:
- CLAUDE.md servers table row
- servers/ana-ml2/README.md hardware spec (+ refreshed system-details.txt)
- stacks/vllm compose + .env.example FP8/KV comments (Ada cc 8.9 -> Blackwell cc 12.0)
- stacks/llama-swap config VRAM-budget comment (48GB -> 96GB, GPU-0 pin)
Also corrects the adjacent stale 'Phi-4-mini' comment in the granite
service block (the service has been Granite 4.1 8B since 34a43a0).
Doc/comment-only; no runtime change.
ana-ml2's Ada->Blackwell swap (2x96GB) frees GPU 0 entirely. Pin llama-swap
to GPU 0 via NVIDIA_VISIBLE_DEVICES so on-demand large-model hot-loads land
there, off GPU 1 where the always-on vLLM services (granite + embed/rerank/
reward) live. Closes the long-standing 'pin llama-swap to GPU 0' item.
IBM Granite 4.1 dense instruct (3B + 8B), unsloth Q4_K_M GGUF via -hf
syntax, 64K ctx with q8 KV. Auto-exposed through the LiteLLM gateway
wildcard (ana-docker:4000) and direct on llama-swap (:9292).
Replaces the retired irv-ml1 Ollama granite4.1 — Ollama is now banned
fleet-wide; serving consolidates onto the sanctioned llama-swap/vLLM
substrate.
Worldtree-dev's --reasoning-format gemma suggestion isn't supported in
the deployed llama.cpp build (accepts only none|deepseek|deepseek-legacy).
Falling back to deepseek, which also populates reasoning_content — the
field Worldtree's GemmaProvider fallback path checks. Verified via test
inference: 608 reasoning_content deltas + 199 content deltas + 0 raw
<|channel> marker leaks.
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)
Playbook handles models, datasets, and spaces (via --var repo_type=...)
since 3025d49 — the "-model" suffix was misleading. Renaming to match
actual scope.
Updates the single in-repo reference (changelog comment in
stacks/llama-swap/conf/config.yaml). config.yaml was scp'd to ana-ml2;
no docker compose restart needed (comment-only).
AtlaAI's Selene-1-Mini judge model for evaluation/scoring tasks.
Llama 3.1 8B base, mradermacher imatrix-quantized Q6_K (~6.5GB,
quality-leaning quant). Apache-2.0. Per Atla cookbook these defaults
hit 84% on RAGTruth hallucination eval.
New 'JUDGE / EVAL MODELS' section between the dense chat models and
the embedding models — separate category from chat/reasoning since
the run-params shape is different (deterministic-leaning: temp 0.01,
top-p 1.0, no repeat penalty).
q8_0 KV cache to fit 32K ctx cleanly on the 3090 with headroom.
Pre-pulled into the shared HF cache via the new
playbooks/pull-hf-model.yaml playbook (canonical replacement for
ad-hoc huggingface_hub.snapshot_download calls; see CHANGELOG).
Smoke-tested 2026-05-13: GET /v1/models lists selene-1-mini-8b,
POST /v1/chat/completions returns expected output cleanly.
Symptom: granite-4-small and qwen3.6-27b were evicting each other
when called in alternation. granite is the news-digest curator (fires
twice daily on cron) — being evicted means a cold reload (~5s) on
every digest tick, plus visible churn whenever the user uses 27b
concurrently.
Added granite-4-small to the `pinned` group as a persistent member.
~5-6 GB at Q4_K_M + 120K KV ≈ comfortable inside the existing pin
budget (qwen3.5-9b ~6 GB → ~12 GB total persistent). Single RTX 6000
Ada is 48 GB, leaves ~36 GB headroom for whichever non-pinned model
the user invokes (qwen3.6-27b at ~30 GB fits cleanly).
Updated the pinned group's docstring to capture the current member set
+ VRAM math + the historical context (qwen3.6-35-a3b was here, was
too heavy, got removed yesterday). Marked the granite ttl: 0 with the
matching "pinned — never unloads" comment as the other group members.
Symptom: qwen3.6-35-a3b refused to deload when other models needed
the VRAM, even with the model itself at ttl: 0. The pinning came from
the `pinned` group's `persistent: true` flag, which exempts members
from eviction by the scheduler regardless of memory pressure. The
model's ttl: 0 only governs idle-timeout, NOT scheduler eviction —
those are separate concerns.
Removed qwen3.6-35-a3b from the group's members. Kept ttl: 0 on the
model itself: still no idle-unload, but the scheduler CAN now evict
it when another non-coexistent model is requested. qwen3.5-9b stays
pinned (~6 GB at Q4 — cheap to hold).
Updated the inline comment + the group-header docstring to reflect
the new semantics so future-Claude doesn't undo this.
The base qwen3.6-35-a3b is already ttl: 0 via the `pinned` group.
The three other Qwen 3.6 variants (abliterated, heretic, 27b) had
ttl: 600 → llama-swap auto-unloaded them after 10 min idle, costing
the next request a full reload (~5-15s). Removed so they stay loaded
once warm. Still get evicted by the normal swap when another
non-pinned model is requested — these aren't joining the pinned group,
just losing their idle-unload timer.
Decision recorded in CLAUDE.md ("Stack tree convention") and memory
(convention_stacks_vs_mirror.md):
stacks/<stack>/ canonical / intent. git-tracked.
deploy-stack.sh reads from here.
stacks-mirror/<host>/<stack>/ snapshot / reality. gitignored.
sync-stacks.sh writes here. Used
for drift inspection only — never
a deploy source.
Bug this fixes: deploy-stack.sh was reading from the mirror, so edits
to stacks/llama-swap/config.yaml never reached ana-ml2. Today's
two new model entries (qwen3.6-35-a3b-heretic + qwen3.6-27b) lived
in the canonical for hours but the deploy reported "in sync" because
the script only diffed mirror vs server.
Changes:
* deploy-stack.sh: source switched from MIRROR_DIR/$HOST/$STACK to
STACKS_DIR/$STACK. Header comment + error message updated.
* sync-stacks.sh: header explicitly identifies its role as drift
detection; documents the diff command for comparing canonical vs
mirror.
* stacks/llama-swap/{config.yaml → conf/config.yaml}: matches the
deploy mapping (conf/ in canonical → /opt/docker/conf/ on host).
* CLAUDE.md: "Stack mirror (pull / push)" section rewritten as
"Stack tree convention (canonical vs mirror)" with the role table
+ workflow rules + diff recipe. Layout diagram updated.
Both models pre-pulled into /tank/aimodels/huggingface (HF_HOME=/hfcache
inside the container) via huggingface_hub.snapshot_download with
hf_transfer for parallel chunked download — heretic's 29 GB landed in
~4 min, unsloth's 26.5 GB in ~3:46 (~118 MB/s each).
heretic: llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-GGUF:Q6_K
27b: unsloth/Qwen3.6-27B-GGUF:UD-Q6_K_XL
Both repos include mmproj-BF16.gguf alongside the main GGUF, and
llama-server's -hf flag auto-loads the mmproj when present in the same
repo (-hf docs: "mmproj is also downloaded automatically if available").
So both entries get vision (image-text-to-text) without needing an
explicit --mmproj path. ttl: 600 (10-min idle unload), matching the
existing abliterated entry's style.
Adds a new `pinned` group with swap: false (models coexist in VRAM),
exclusive: false (group shares with other groups), persistent: true
(never unload). Each member also gets ttl: 0 so the per-model
idle-timeout can't drop them either — belt + suspenders.
Pair is currently qwen3.5-9b (~6 GB Q4) + qwen3.6-35-a3b (~29 GB Q6).
Plus the 128K KV caches, roughly 50-60 GB VRAM resident. Appropriate
for an A6000/H100-class card; verify fit after deploy.
Committed as a canonical change; push + restart still needed on
ana-ml2.
Captures the full workspace state built up to this point:
- CLAUDE.md + README.md describing conventions and the four-host fleet
(ana-ml2, ana-docker, nh3-docker, esh-docker-vm).
- Per-host notes under servers/<host>/ with ssh-target fallback files
and latest system-details snapshots (two in-compose credential leaks
scrubbed; the upstream compose files still need to move those to .env).
- scripts/: server_inspect.sh (read-only remote diagnostic),
refresh-server-info.sh (dir-driven discovery + snapshot capture with
validation warnings), add-host.sh, sync-stacks.sh (pull
compose/conf trees), deploy-stack.sh (push with per-file diff + prompt).
- stacks/: canonical compose for backrest, beszel, dozzle, llama-swap,
rest-server-ana, rest-server-nh3, vllm-qwen3, plus the retired
infinity reference. All use the .env-driven + traefik-net + homepage
label pattern.
- configs/restic/ana-docker/: first resticprofile config + pre-backup
hook (Synapse pg_dump, Seafile mysqldump, Vaultwarden SQLite); templates
for the other three hosts to come.
- docs/pfi/: general infrastructure reference carried over.
- .gitignore excludes .env, stacks-mirror/, and assorted secret/state
filenames to prevent re-leaks on later commits.
vh