The abliterated/NVFP4 Qwopus 122B "gen" model (+ qwen-large / summarizer-large
aliases) had no repetition control in its sampling defaults, causing degenerate
repetition loops. Add presence_penalty: 1.0 (Qwen-documented anti-repetition
lever, range 0-2) to all 7 qwen3.5-122-a10b gateway records. Overrideable
default; bake into the vLLM serving def once the value is validated.
Native-allocator expandable segments to cut Qwen-Image-Edit fragmentation
OOMs on the A6000 (a ~2 GB alloc failing with 1.75 GB free while 45 GB sat
allocated + reserved-but-unallocated). Cache-preserving — packs better
without unloading the checkpoint, so no edit-latency hit. Paired with the
existing --disable-cuda-malloc (incompatible with cudaMallocAsync).
Deployed + recreated on irv-ml1; verified env present, PyTorch reads it,
container healthy. comfy-dev request 2026-06-25.
Langfuse's ClickHouse member spewed ~94 GB of unrotated logs and filled ana-docker's
root disk (took the fleet host to 100%, 28/48 containers unhealthy). Its trace UI was
redundant with LiteLLM's native logging — store_prompts_in_spend_logs:true already
captures full prompts/responses/tokens/cost/latency at :4000/ui — and nothing used its
unique trace-grouping/eval features (it only received flat gateway success_callbacks).
Removed the callbacks (gateway observability stays fully native) and tore down the
6-container langfuse stack + volumes on ana-docker. Re-add the callbacks if it returns.
Capability aliases for the PAID frontier tier, mirroring glm-5.2 / glm-5.2-
reasoning (thinking off / on) → openai/glm-5.2 @ z.ai. Worldtree binds these for
a frontier-grade generation/reasoning capability so the backing frontier model
can be swapped gateway-side (operator jump-started WT's request). PAID: only
all-proxy-models / explicitly-scoped keys reach them; the free all-agents-local
key stays fenced off z.ai spend. Verified both resolve + route to GLM 5.2.
llama-swap (ana-ml2:9292) is decommissioned (:9292 confirmed down), so the
catch-all wildcard routed every unmatched / typo'd / stale model name to a DEAD
backend, surfacing a misleading "Connection error" instead of a clean
"model not found". This is the exact footgun that silently swallowed Worldtree's
defunct model names (mistral-small-4 etc.) instead of erroring. Removed (operator
call) so unknown models now 404 loudly. Verified: gateway healthy post-restart,
a bogus model name now returns a clean not-found error, real aliases (gen) still
serve. Re-add explicit per-model entries if a swappable zoo ever returns.
Stand up the gateway-side capability aliases for the role→capability model
indirection (worldtree-dev's transparent-swap direction; operator: no wt-
prefix, reuse the existing summarizer/classifier/gen alias convention).
- chat-judge -> selene-1-mini-8b (mode chat) — WT selene-judgment role.
- reranker -> qwen3-reranker (mode rerank) — generic name for the cap.
- scalar-judge -> Skywork-Reward-V2 via a pass_through_endpoint to ana-ml2:8003
(LiteLLM has no reward/pooling MODE, so it's a passthrough, gateway-key-gated;
consumers hit /scalar-judge/<route> e.g. /score|/pooling|/classify).
Deliberately NO generic `embedding` alias: embedding vectors are model-specific
(not swap-transparent), so that capability stays `qwen3-embedding` — the model-
specific name is the guardrail against treating it as freely swappable. Verified
all three live (chat-judge 200, reranker present, scalar-judge passthrough 200
returning a Skywork reward). Deployed + gateway health-gated.
Empirical follow-up to the streaming /tts smoke test on the 3090. OmniVoice
is diffusion: a ~fixed per-call overhead (~1.5s at 32 steps, ~0.7s at 16)
dominates regardless of chunk length, so the upstream-claimed 40x RTF does
NOT hold here (measured ~2.8x/32-step, ~5.6x/16-step) and the chatterbox-
tuned scheduler over-chunks and starves.
- Streaming /tts defaults to num_step=16 (TTFA ~1.5s -> ~0.7s); batch
/v1/audio/speech stays num_step=32 for quality. Per-request override intact.
- Scheduler prior raised to rtf_prior=20 (env OMNIVOICE_STREAM_RTF_PRIOR,
wired through compose + .env.example) so it packs whole-text-minus-first-
sentence into a few chunks: validated ~3 chunks, no starvation, total wall
~= one-shot, less per-chunk silence padding.
- Docs corrected: the "sub-second / 40x" claims were wrong; streaming has a
diffusion TTFA floor (~0.7s) and wins mainly on long replies. chatterbox-
fast (autoregressive, ~0.5s TTFA) stays the lowest-latency front-end;
OmniVoice is the multilingual / voice-design complement.
Add a live-consumer streaming path and text sanitation to the OmniVoice
wrapper, so it can front speech-to-speech chat engines (not just the
asset-engine's batch WAV use).
- POST /tts: chunked 24 kHz mono s16le PCM (or open-ended WAV), driven by
the adaptive buffer-ratchet scheduler. Emits the first sentence
immediately, then ratchets chunk size up on OmniVoice's ~40x realtime
headroom -> sub-second time-to-first-audio. Wire-compatible with
chatterbox-fast /tts (both 24 kHz mono PCM). Batch /v1/audio/speech is
unchanged for asset/file callers.
- scheduler.py: VENDORED byte-faithful copy of chatterbox-fast's pure-
Python (torch-free) scheduler, pinned to commit 7631462 (v0.1.0/v0.1.1).
Vendor-copy over a shared package (operator call 2026-06-19): the module
has no GPU deps, so reuse it without dragging chatterbox-fast's torch
tree into this image. Promote to a shared package only on a 3rd consumer
or real drift.
- sanitize.py: language-safe TTS sanitizer run on both endpoints. Strips
markdown, <think> blocks, HTML, and model control tokens; deliberately
SKIPS the fork's English-only number/phone normalization that would
corrupt OmniVoice's 600-language input. Preserves [laughter]-style tags.
- Refactor: shared GenParams base for SpeechRequest + TTSStreamRequest;
single GEN_LOCK serializes generation (single-stream interactive).
- Dockerfile/playbook: copy + upload the two new modules; build-time
`import app` smoke; correct stale "Gradio demo / no FastAPI" comments.
Stands up tools/gateway-chat.html as a permanent URL on ana-docker (http://10.250.50.70:8091)
via a tiny nginx:alpine static container (no GPU, no DB). conf/index.html is a deployed
mirror of tools/gateway-chat.html (re-sync one-liner in README). Homepage tile + tnet per
convention. The enhanced tool (auto-discovers /v1/models, system prompts, streaming +
reasoning, image upload for vision) is now always-on for smoking new gateway models.
Duplicate-entry aliases. classifier -> granite-4.1-8b (:8004, same backend as the
existing summarizer alias). summarizer-large -> gen/qwen3.5-122-a10b (:8013, thinking
off) for heavier summarization on the 122B Qwopus. summarizer -> granite already
existed (no-op). Config-staged + deployed without bouncing the gateway; like any
config-add these activate on the next restart (no live-add performed).
Source + deployed config cleaned without bouncing the gateway. NOTE: these were
config-loaded models, which the /model/delete API can't remove (DB-only -> 'not
found in db'), so the LIVE gateway still serves them until its next restart, at
which point the cleaned config drops them. No bounce performed.
Same Qwopus gen model as gen / gen-reasoning (served-name qwen3.5-122-a10b @
:8013, thinking off/on respectively), but each bakes a dummy 'noop' function tool
+ tool_choice:none into litellm_params so a NON-EMPTY tools array always reaches
vLLM — for consumers where the global strip_empty_tools hook isn't the right fix
(they need a valid tools structure present, not stripped). tool_choice:none means
the noop is never called. api_base = the real LAN endpoint http://10.250.50.54:8013
(the requested http://vllm:8000 template wouldn't resolve from the ana-docker
litellm container). Verified: gen-nt + gen-reasoning-nt both survive a client
tools:[] send; noop never invoked; reasoning split intact.
Replaces the bjk110 text-only qwen3.5-122b as the `gen` model on ana-ml2 GPU 0.
OpenYourMind/Qwopus3.5-122B-A10B-Kimi-K2.6-destilled-abliterated-NVFP4 — Kimi-
distilled, abliterated, NVFP4, and crucially VISION-INTACT (serves as plain
multimodal, no text-only patch). Served as qwen3.5-122-a10b so the litellm
gen / gen-reasoning / qwen-large records route here unchanged.
Tuned for full native context on the 96GB Blackwell:
- stable vLLM image + fp8 KV → 11GB pool = 870,014 tokens = 3.32x concurrency
at the full 262144 (256K) window. Nightly+turboquant-4bit was unnecessary.
- CUDA graphs ON (no --enforce-eager) → 92.7 tok/s warm single-stream.
- util 0.95 + PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True — 0.96 OOM'd by
0.1GB on the 3.09GB FusedMoE transient workspace (the hard floor; defrag
reclaims the 4.2GB fragmentation, 0.95 adds margin).
- max-num-seqs 16 (short reqs fan out ~16x32k; 256K reqs pool-limit to 3.32x).
- text + image + video all enabled; tool-calling via qwen3_coder (XML), verified.
gen/gen-reasoning tool-calling 400'd (operator + brokkr's capability battery both caught it):
the bjk110 serve command shipped --reasoning-parser qwen3 but no tool flags. Qwen3.5 emits XML
tool calls (<tool_call><function=NAME><parameter=K>V</parameter></function></tool_call>), NOT
Hermes JSON — so `hermes` mis-parsed to raw text; `qwen3_xml` is the correct parser. Reasoning +
tools coexist (gen-reasoning keeps its thinking split). Verified live: a get_weather request
returns tool_calls=[get_weather {"city":"Paris"}].
bjk110/Qwen3.5-122B-A10B-abliterated-NVFP4 on ana-ml2 GPU 0 (heretic downed):
- stacks/qwen3.5-122b/ — vLLM serve via the repo's text-only patch (Qwen3.5 MoE is a
multimodal arch but this checkpoint is text-only weights), --reasoning-parser qwen3,
GPU 0 pin, :8013; entrypoint+patch mounted from the model dir.
- serve-qwen3.5-122b.yaml — displace heretic + serve + verify.
- litellm: REMOVED dead mistral-small-4 / -reasoning; added qwen3.5-122-a10b[-reasoning]
+ aliases qwen-large[-reasoning] + repointed gen[-reasoning] -> qwen (thinking split via
chat_template_kwargs.enable_thinking + --reasoning-parser qwen3).
Verified live: qwen healthy on :8013; gen / qwen-large / qwen3.5-122-a10b route, and
gen-reasoning returns reasoning_content; mistral-small-4 removed.
NOTE: Worldtree character backend (was bound to mistral-small-4) is dark until repointed
(operator-acknowledged).
Duplicate-entry aliases (NOT router_settings.model_group_alias — that's hidden from
/v1/models and can be silently ignored in config per litellm #15020/#5524):
- summarizer -> granite-4.1-8b
- gen -> mistral-small-4
- gen-reasoning -> mistral-small-4-reasoning (reasoning_effort:high preserved)
Each alias is a real model_name co-located with its target (keep api_base in sync).
Verified live: all 3 in /v1/models + route end-to-end; gen-reasoning returns
reasoning_content.
Wrapper /v1/audio/speech now accepts OmniVoice's whole surface:
- voice (clone, now OPTIONAL) and/or instruct (voice DESIGN). instruct is a CONTROLLED
vocabulary (gender/age/pitch/accent/whisper tags, comma-separated), not free prose —
discoverable at the new /v1/audio/instruct-items endpoint (23 items).
- language (Auto + 647, new /v1/audio/languages endpoint), speed, duration.
- diffusion controls: num_step, guidance_scale, denoise, preprocess_prompt,
postprocess_output; plus a generation_overrides JSON passthrough for expert
GenerationConfig knobs (t_shift, layer_penalty_factor, position/class temperature,
audio_chunk_*).
- at least one of voice/instruct required (else 400).
Catalog (services.yaml): omnivoice v1 -> v2, 13 schema-valid fields; instruct as a
controlled-vocab text field sourced from the items endpoint.
Verified live on irv-ml1: clone, voice-design (instruct-only), and tuned-param synths
all -> 24 kHz PCM_16 WAV; 647 languages; 23 instruct items.
- app.py: thin FastAPI wrapper exposing OpenAI /v1/audio/speech (+ /v1/audio/voices,
/healthz) around OmniVoice's Python API; precomputes a voice-clone prompt per voice
at startup (loaded Whisper auto-transcribes each reference). Replaces the Gradio demo.
- Dockerfile/compose: run the uvicorn wrapper, /healthz healthcheck, project name pinned
to "omnivoice" so the asset-engine liveness probe matches.
- deploy-omnivoice.yaml: stage chatterbox /refs/*.wav as clone voices (skip _* artifacts)
+ verify the API surface.
- services.yaml: catalog entry (id omnivoice, :8199/v1/audio/speech, voice list sourced
live from /v1/audio/voices) + reproducibility_audit row.
Verified live on irv-ml1: /healthz ok, 33 voices loaded, test synth -> 24kHz PCM_16 WAV.
Zero-shot, massively-multilingual (600+ language) voice-cloning + voice-design
TTS (diffusion-LM, Apache-2.0). No official image, so a thin CUDA container
around the pip package running upstream's own Gradio demo (no FastAPI wrapper).
Pinned to GPU 0 (3090) — the A6000 is ComfyUI-exclusive — port 8199. Built +
verified live on irv-ml1 (Gradio 200, container healthy). Surface is the Gradio
UI + Gradio API, NOT OpenAI-compat /v1/audio/speech (wrap later if asset-engine
should consume it). deploy-omnivoice.yaml builds local + verifies.
ComfyUI 0.24.1 added native attention selection; the node-based
BlehGlobalSageAttention errors "does not support the new ComfyUI attention
changes". Add --use-sage-attention to COMFY_CMDLINE_EXTRA so the in-image
sageattention v2.2.0 sm_86 build (rebuilt vs pinned torch 2.12.1) binds via
the native path. OOM flags preserved. Deployed to irv-ml1 + recreated; log
confirms "Using sage attention", container healthy, serving 200.
(comfy-dev request, thread 01KVE89T2DKC)
Operator consolidation (2026-06-18): give ComfyUI the full 48 GB A6000 and move the
audio/TTS zoo to the 3090.
- comfyui: NVIDIA_VISIBLE_DEVICES all -> 1 (A6000 only), + DISABLE_UPGRADES=true to
pin torch at 2.12.1+cu129 so the mmartial boot script stops auto-upgrading it and
the compiled SageAttention kernels stay matched (comfy-dev torch-pin, approved).
- parakeet: NVIDIA_VISIBLE_DEVICES all -> 0 (3090).
Other GPU reassignments are deployment-side (not repo compose): chatterbox-fast via
its .env CBF_GPU_DEVICES=0; vibevoice device_ids ["1"]->["0"] (deployed from
/worktank/vibevoice/build); yt-voice-clipper worker via its override. dia2-2b,
ace-step, csm-expressiva downed (stale/unused).
Result: A6000 = ComfyUI alone (48.3 GB free); 3090 = chatterbox + parakeet + the
on-demand audio (vibevoice/ytvc/kokoro). SageAttention rebuilt against the pinned
torch; OOM cmdline (COMFY_CMDLINE_EXTRA) preserved; /object_info still lists the 9
acceleration nodes.
Serves the in-house abliterated Mistral Small 4 (heretic NVFP4, vision-intact)
under --served-model-name mistral-small-4 on ana-ml2 GPU0:8010 — a true drop-in
for the official mistral-small-4 backend. Both litellm entries (mistral-small-4 +
mistral-small-4-reasoning) route here with no litellm change.
GPU0 fits one mistral-class model, so this is a backend swap, not a co-tenant:
bring up after downing the official stack; revert by downing this and up-ing the
official. Verified live through the gateway: standard returns clean answers,
reasoning populates reasoning_content (the [THINK] split). Checkpoint built per
tools/mistral-small4-nvfp4/.
GLM 5.2 released ~2026-06; confirmed reachable with our existing
Z_AI_API_KEY (z.ai /models lists glm-5.2; a live completion returned
clean). Added two model_list entries mirroring the glm-5.1 pattern:
glm-5.2 (thinking DISABLED by default, per the 2026-06-11 operator call)
and glm-5.2-reasoning (thinking ENABLED, opt-in). Deployed to
ana-docker /opt/docker/conf/litellm/config.yaml, litellm restarted,
both verified through the gateway (disabled -> reasoning_tokens 0;
reasoning -> 234).
/healthz catalog_version read the BAKED package version (importlib.metadata),
so a catalog/frontend-only webhook deploy (no image rebuild) left it stale —
v0.12.4 data went live but /healthz still reported 0.12.3. comfy-dev's v0.12.5
reads comfy_catalog.__version__ from the repo-root pyproject.toml; mounting it
on the same checkout mount (catalog/graphs/frontend) makes /healthz report the
MOUNTED version after a catalog-restart. Falls back cleanly if absent.
Pushed to irv-ml1's host compose + validated via `docker compose config`
(bind -> /app/pyproject.toml:ro resolves). Recreate deferred to comfy-dev's
imminent v0.12.5 rebuild (the mount is inert for v0.12.5 itself, which bakes
its own version — it only matters for subsequent catalog-only deploys — so no
separate prod blip). Requested by comfy-dev (engine owner), althing thread
01KV95R88A3Y.
vLLM's OpenAI server 400s on an empty tools array ("tools must not be an
empty array"), which broke every gateway call carrying tools:[] (clients
that send it to mean "no tools" -- OpenAI tolerates it, vLLM does not).
drop_params doesn't help: it drops unsupported PARAMS, not empty VALUES.
Add a CustomLogger async_pre_call_hook (conf/strip_empty_tools.py) that
pops an empty/None tools field (+ orphaned tool_choice) before forwarding,
registered globally via litellm_settings.callbacks so it covers every
vLLM-backed model, not just mistral-small-4. Mounted at
/app/strip_empty_tools.py beside config.yaml (LiteLLM resolves callbacks
relative to the config dir). Surgical: only fires when tools is present
and empty; real tools pass through untouched.
Verified on live gateway (1.87.0): mistral-small-4 and granite-4.1-8b
with tools:[] now 200 (were 400); no-tools baseline unchanged; a real
tool still passes through.
Put GPU1's idle ~11 GB buffer to work on the two KV-bound models that gained
live consumers from the worldtree migration (granite + the pooling models
under-use their util, so growing them is wasted):
- selene 0.13→0.17: KV 2.53→6.33 GiB, concurrency 1.27x→3.16x @32K (Domari judge)
- qwen36 0.32→0.34: KV 7.73→9.63 GiB, concurrency 2.92x→3.64x @131K (arbo judge +
worldtree actor/echo + gateway)
GPU1 free now ~5.6 GB (safe floor for single-service recreates).
selene: AtlaAI Selene-1-Mini-Llama-3.1-8B judge restored on vLLM after the
llama-swap teardown took its Q6_K GGUF offline. FP8 (dynamic --quantization
fp8; FP8 >= the validated Q6_K fidelity, and text-only Llama so no vision-
tower-noise risk; NVFP4's W4A4 too aggressive for a precision judge). GPU1
util 0.13 (8.51 GiB weights + 2.53 GiB KV, 32K ctx, 1.27x concurrency),
~11 GB GPU1 buffer left. Gateway selene-1-mini-8b → :8011 (shadows the *
wildcard that used to reach it via llama-swap). Judge smoke: scored an
unfaithful claim 1/5 correctly.
mistral-small-4: max-model-len 131072 → 262144 (full native 256K) for
novel-length consistency-checking. KV pool is util-bound (~862K tokens), so
256K costs no extra VRAM — max concurrency just drops to 3.29x at full length.
max-num-seqs 64 → 32 keeps the warmup transient flat (scales with seqs × len),
so it fits the tight GPU0 (free unchanged at 5.2 GB). Verified loaded + healthy.
Operator needs a verified-working vision tower as the abliteration/tuning
baseline. vLLM 0.23.0 crashes Mistral multimodal at startup (#44911
fetch_images regression, ~0.22.1+). Pinned the Mistral container to
v0.22.0 — the last pre-regression release — which loads the NVFP4
(compressed-tensors) AND serves vision: verified a half-blue/half-red
image read correctly ('left blue, right red'). Dropped --limit-mm
(vision re-enabled). qwen36 stays on 0.23.0 (separate container; needs it
for its ModelOpt NVFP4).
- gateway: add mistral-small-4-reasoning. Operator asked for effort=medium
but Mistral's reasoning_effort is BINARY (none/high only — medium 400s);
set to 'high' (sole reasoning-ON level). NOTE: reasoning fires but
reasoning_content-splitting is unreliable on v0.22.0 (lands in content);
clean split would need 0.23.0, which breaks vision — vision prioritized.
- mistral-small-4 (instant) + mistral-small-4-reasoning both gateway-live.
Mistral-Small-4-119B-2603-NVFP4 (119B/6.5B-active MoE, 65.3 GiB) on the
freed GPU 0 (dedicated 96 GB Blackwell), vLLM 0.23.0, :8010. NVFP4 is the
only variant that fits one card (FP8 ~119 GB / bf16 ~238 GB need 2 GPUs).
- TEXT-ONLY: vLLM 0.23.0's Mistral multimodal processor crashes at startup
(fetch_images bug); loaded with --limit-mm-per-prompt image/video=0.
Remove the flag to restore vision once vLLM patches it.
- MLA attn (TRITON_MLA), mistral tool-call + reasoning parsers, util 0.93,
max-len 131072 (capped from native 256K), image pinned by 0.23.0 digest.
- litellm: mistral-small-4 → :8010, shadows the * wildcard.
- GPU 0 reassigned from the (now-offline) llama-swap zoo per operator.
The qwen3.6-35b-a3b VL checkpoint is a single hybrid model with a per-
request enable_thinking switch (Qwen3-style), defaulting thinking ON.
Make the default non-thinking and add an opt-in reasoning variant,
mirroring the existing glm-5.1 / glm-5.1-reasoning gateway split.
- qwen36-vl compose: add --reasoning-parser qwen3 (model-matched) so the
single :8007 endpoint splits <think> into reasoning_content when on and
routes all output to content when off — serving both modes cleanly.
- litellm gateway: base qwen3.6-35b-a3b pins chat_template_kwargs
enable_thinking=false (non-thinking default); new qwen3.6-35b-a3b-thinking
pins enable_thinking=true (opt-in reasoning). Same upstream checkpoint,
no extra VRAM/container.
Deployed + verified on ana-ml2 (vLLM recreated, healthy) and ana-docker
(litellm reloaded): default returns a direct answer with no reasoning_content;
-thinking returns cleanly-separated reasoning_content, no raw tag leak.
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.
Operator's call: keep the arbo stack in eshpfi and version its deploy machinery
alongside the compose (was host-only on irv-ml1 = recoverability foot-gun).
- arbo-webhook.py: :9009 HMAC listener (secret externalized to host file, not git)
- arbo-deploy.sh: internal-route fetch + catalog-only targeted restart
Document both in the README Q5 section + the internal-gitea-route gotcha.
Operator decision 2026-06-13 (relayed by comfy-dev, confirmed in-session):
turn off the prod arbo engine's bearer auth and rely on the WireGuard
perimeter. Reverses ADR-0001's open-auth-hole-closed posture (comfy-dev owns
the ADR update on the vh/arbo side).
The app's protected-gate no-ops only when ENGINE_TOKEN is ABSENT — an empty
string still gates (verified: ENGINE_TOKEN="" -> /workflows 401). So both
inject paths are removed: the compose environment line is commented out and
the .env line deleted on the host. Result: tokenless GET /workflows 200 (was
401), matching the dev engine. Original token preserved in the host's
.env.pre-auth-off.bak for re-enable.
playbooks/arbo-disable-engine-token.yaml captures the reversible procedure.
The arbo-prompt-enhance vkey was extended to reach qwen3.5-9b-fp8 for the
hero auto-judge step (v0.11.3+), not granite-only. Confirmed via /v1/models
for the key. Docs-only; no version bump.
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.
Benched granite prefix caching at ~6.5x faster TTFT (45ms cached vs 292ms
uncached) on a shared ~4.5k-token summarizer template. granite already had it
on by vLLM-v1 default; pinned explicit so a version flip can't silently disable
it. qwen (nightly) defaulted it OFF -> flipped on (free for the text-chat path,
marginal for vision where each image is a distinct prefix). Soft/evictable KV,
zero memory change (GPU1 still ~3.7GB free), all 5 services healthy.
ComfyUI's ~325G model tree moved off the near-full worktank NVMe (97%->26%,
342G free) to /storetank/arbo (roomy SATA SSD on irv-ml1), overlay-mounted
back at /basedir/models so ComfyUI behaviour is unchanged. rsync byte-verified
(src==dst), one comfyui restart, worktank original removed. Inventory of the
set in docs/arbo-comfyui-model-catalog.md for the retain decision. The older
919G /storetank/image-models/comfy archive is untouched (separate reclaim).
Qwen3.5-9B vision-language served FP8 on ana-ml2 GPU1 (co-located with
granite + the embed/rerank/reward trio; GPU0 kept free for hot-loading
large models), :8007, fronted by LiteLLM as qwen3.5-9b-fp8.
Pinned to vllm/vllm-openai nightly@sha256:49211ab2 — :latest (v0.19.1)
quantizes the VL vision tower under fp8 and garbles vision; the nightly
correctly excludes it (LM stays FP8, vision tower BF16). util 0.40
(~38GB) on the shared card (vLLM needs free>=util*total here). Vision
verified end-to-end through the gateway.
SGLang 0.5.13 confirmed to support our formats on Blackwell sm_120
(compressed-tensors NVFP4 W4A4, fp8, modelopt_fp4, petit_nvfp4, fp4_e2m1 KV),
so the bench can be a real NVFP4 head-to-head. Parameterized compose (model/
quant/GPU via .env) + a common streaming load generator (bench.py: agg tok/s,
TTFT p50/p99, TPOT) so both engines are driven identically on an exclusive GPU.
Bench-oriented; promote to a real stack only if SGLang wins. Launch deferred
until the NVFP4 eval frees a GPU.
Ada->Blackwell swap doubled card VRAM, so the Ada-era fractions (0.07/0.07/
0.18) reserved ~2x the bytes for the same models. Empirically re-floored via
0.01-step climb until each service was stable under 20x parallel inference:
embed/rerank 0.05 (load-floor for the 0.6B models), reward 0.10 (the real
over-provision). Frees ~11 GB on GPU 1. Live .env on ana-ml2 already applied.
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.
Extends the catalog/graphs git-pull-mount pattern to the SPA frontend so
frontend changes reach prod via git pull + restart, no image rebuild.
Delivers the v0.11.2 auth-on catalog-load fix without a rebuild; baked
image frontend stays the fallback.
vh