astro-orbiter's vLLM primary model changed 2026-09-01 from
DeepSeek-R1-Distill-Qwen-32B-AWQ to Gemma-4-26B-A4B-it-AWQ (Google,
Apache 2.0, US-origin). Same endpoint/API key — only the served model
name changed. max_model_len also bumped to 65536 (was 32768).
Gemma 4 does not emit an always-on <think> reasoning trace like
DeepSeek-R1 did, so HINDSIGHT_API_RETAIN_MAX_COMPLETION_TOKENS=4096
should have more effective headroom for real extraction output than
before, not less.
Retired DeepSeek-R1-Distill-Qwen-32B after confirming its auto tool-choice
reliability is a known, documented DeepSeek-R1-distillation limitation
(trained on pure reasoning traces, no function-calling data — upstream
GitHub-confirmed, not a config gap). Model choice moved to Gemma 4 26B A4B
(Google, Apache 2.0, US-origin, matches Ryan's model-origin preference):
- cyankiwi/gemma-4-26B-A4B-it-AWQ-4bit — MoE (25.2B total / 3.8B active),
chosen over the dense 31B variant for smaller on-disk footprint
(~17.2GB vs ~20.9GB), buying more KV-cache headroom on this 24GB card
- max_model_len=65536 (comfortably over Hermes's 64K floor; native
context is 256K, no extension trick needed)
- Native gemma4 tool-call-parser + gemma4 reasoning-parser (both
registered in this host's vLLM 0.28.0) — purpose-built for this
model's actual output format, not a same-family approximation
- kv_cache_dtype: int4_per_token_head from the outset (learned from the
DeepSeek swap's fp16->fp8->int4 trial-and-error escalation)
Bug found and fixed during deployment: the repo's config.json declares
quant_method 'compressed-tensors' (llm-compressor output) despite the
repo name saying 'AWQ-4bit'. Passing --quantization awq explicitly
caused a hard pydantic ValidationError on every startup attempt.
Fix: omit the quantization field entirely and let vLLM auto-detect from
the model's own config.json — confirmed clean single-attempt start,
NRestarts=0, once removed.
Verified live:
- /health 200, /v1/models confirms max_model_len=65536
- Live completion: correct answer, no unwanted reasoning trace by default
- tool_choice=auto with a clear trigger prompt: correct tool_calls
response with valid JSON args — the exact test DeepSeek-R1-Distill
failed (it either answered in plain text or burned tokens reasoning
about how to call the tool instead of calling it)
- tool_choice=auto with an irrelevant tool present: correctly answered
in plain text, did not over-trigger the tool
- Ansible idempotent re-run confirmed: changed=0, NRestarts=0, clean
journalctl (zero error/traceback lines) after a fresh restart
Known follow-up (not done here): Hindsight's HINDSIGHT_API_LLM_MODEL
cluster config still references the retired DeepSeek-R1-Distill-Qwen-32B
(itself a follow-up from the prior Qwen2.5-32B swap) — needs another
GitOps update to point at Gemma-4-26B-A4B-it-AWQ.
astro-orbiter's vLLM primary model changed 2026-09-01 from
Qwen2.5-32B-Instruct-AWQ to DeepSeek-R1-Distill-Qwen-32B-AWQ
(single-model deployment, see ansible/roles/deploy-vllm/README.md).
Same endpoint (http://astro-orbiter:8000/v1) and API key — only the
served model name changed.
Flagged for follow-up, not fixed here: DeepSeek-R1 emits a
<think>...</think> reasoning trace before its final answer; Hindsight's
fact-extraction prompt/parsing has not been specifically verified
against this reasoning-model output shape. Watch extraction quality
after this rolls out; HINDSIGHT_API_RETAIN_MAX_COMPLETION_TOKENS left
at 4096 pending confirmation that's enough headroom for a real retain
to complete past the reasoning trace.
Single-model deployment per Ryan's direction:
- Primary model: casperhansen/deepseek-r1-distill-qwen-32b-awq, max_model_len=32768
- nomic-embed-text-v1.5 and Qwen3-8B-AWQ both disabled (single-model requirement)
- kv_cache_dtype: int4_per_token_head required to fit 32768 ctx on 24GB RTX 3090
(fp16 KV: 26GB needed, doesn't fit at any utilization; fp8 KV: OOM'd during
FlashInfer warmup with ~50-150MB margin; int4 KV: clean single-attempt start)
- Added kv_cache_dtype / kv_cache_memory_bytes as new optional per-model
template fields in vllm.service.j2 (guarded, no effect on other models)
Verified live: /health 200, /v1/models confirms max_model_len=32768,
live /v1/completions smoke test + manual chat completion both passed
(genuine <think> reasoning trace, correct arithmetic). NRestarts=0,
steady-state VRAM 23.2GB/24.576GB. Ansible idempotent re-run confirmed
changed=0.
Known follow-up (not done here): Hindsight's HINDSIGHT_API_LLM_MODEL
cluster config still references the retired Qwen2.5-32B-Instruct-AWQ —
needs separate GitOps update to point at the new model.
Companion commit to 6bfcc76 (vllm permanent residency cutover). Now that
llama-swap is stopped+disabled and vLLM is the permanent serving layer:
- HINDSIGHT_API_LLM_BASE_URL: llama-swap :8001 -> vLLM :8000
- HINDSIGHT_API_LLM_MODEL: Qwen3.8-27B-Q4_K_M -> Qwen2.5-32B-Instruct-AWQ
- Added HINDSIGHT_API_RETAIN_MAX_COMPLETION_TOKENS=4096 (vLLM's
max_model_len=8192 is far below the 64000 Hindsight retain default;
uncapped this 400s on long retains -- found in t_e6facb19's validation)
- ExternalSecret HINDSIGHT_API_LLM_API_KEY source: 1Password 'nous' item
(unused fallback) -> 1Password 'vllm' item's api-key (vLLM validates
its bearer token, unlike llama-swap which accepted anything)
Embeddings intentionally UNCHANGED -- Hindsight stays on its bundled
local 384-dim embedder (HINDSIGHT_API_EMBEDDINGS_PROVIDER still unset).
Switching to vLLM's nomic-embed (768-dim) is a separate, destructive,
explicitly-approved-only migration (~1300 rows across ~20 banks) --
out of scope here, not part of the dashboard's 'stop llama-swap /
start vLLM' instruction.
Dashboard decision (t_5508360a): 'stop and disable llama-swap and start
vLLM and its 3 models' -- explicit approval of a breaking change.
Applied:
- llama-swap stopped + disabled on astro-orbiter (systemd unit removed
from multi-user.target.wants, files left in place -- full teardown is
t_6dff1ecc, separate task)
- vllm_service_enabled/state flipped to true/started -- vLLM is now the
permanent, boot-persistent serving layer (was shadow-only/staged)
- Attempted enabling Qwen3-8B-AWQ (the 3rd model) -- does NOT fit.
Qwen2.5-32B-Instruct-AWQ (~18.6GB) + nomic-embed (~0.8GB) leaves only
~1.25GiB free on the 23.55GiB usable budget, below the 3.53GiB floor
Qwen3-8B-AWQ needs even at gpu_memory_utilization=0.15. Confirmed via
journalctl: identical ValueError on 7/7 consecutive restart attempts,
not a transient crash-loop. Reverted Qwen3-8B-AWQ to enabled: false.
2 of the 3 requested models fit permanently, not 3.
Verified live: /health 200 on both :8000 and :8020, live completion and
live embedding both returned correct real output, NRestarts=0 on both
services after a clean idempotent re-run (changed=0).
Critical finding: flipping vllm_service_enabled/state=true/started and
restarting llama-swap alongside it broke llama-swap's ability to load
ANY of its own generative models -- every /v1/chat/completions request
against Qwen3.8-27B-Q4_K_M or Qwen3-8B aux models failed with
'upstream command exited prematurely' (llama-server OOM at spawn,
~1.8GB free on this 24GB card once vLLM's ~22.8GB was claimed).
Confirmed by direct A/B: same request 500s with vLLM running, 200s
seconds after stopping it.
This breaks 21 Hermes agent profiles' aux-model tasks (skills_hub,
approval, mcp, title_generation, profile_describer, compression) plus
OpenViking's VLM -- a far larger blast radius than Hindsight's single
LLM endpoint. Reverted:
- vllm_service_enabled/state back to role defaults (false/stopped) --
vLLM stays staged, startable for a brief validated shadow window,
NOT safe to leave resident in production.
- Hindsight's HINDSIGHT_API_LLM_BASE_URL back to llama-swap
(astro-orbiter:8001, Qwen3.8-27B-Q4_K_M) and the API key secret
source back to the Nous fallback item (pre-task state) --
the vLLM cutover, while functionally validated in isolation
(health, /v1/chat/completions, and a live hindsight_retain+recall
round-trip all succeeded), requires continuous vLLM availability
which is now known to be unsafe on this card.
Comment posted on t_6dff1ecc: teardown remains correctly blocked --
full cutover is not achievable within this card's VRAM budget as
currently scoped. Needs a human decision on aux-model migration
strategy (see roles/deploy-vllm README's 'Critical architectural
finding' section) before any further progress.
Default 64000 exceeded vLLM Qwen2.5-32B-Instruct-AWQ's --max-model-len
8192, causing every retain call to 500 with
'max_tokens=64000 cannot be greater than max_model_len=8192'.
llama-swap's Qwen3.8-27B ran at ctx=65536 so this never surfaced
before the vLLM cutover. Lowered to 4096.
HINDSIGHT_API_EMBEDDINGS_PROVIDER=openai pointed at vLLM's
nomic-embed-text-v1.5 (768-dim) crash-looped hindsight-api:
'Cannot change embedding dimension from 384 to 768: memory_units
table contains 1289 rows with embeddings.' Hindsight was never
actually using astro-orbiter for embeddings (defaults to a bundled
local 384-dim sentence-transformers model when the env var is unset)
-- this was a genuine architecture assumption error, not a config
typo. Re-embedding all existing memory data across ~20 agent banks is
a destructive, irreversible operation requiring its own explicit,
approved task -- not something to trigger as a side effect of an
infra migration. Keeps the LLM cutover (astro-orbiter:8000, vLLM
Qwen2.5-32B-Instruct-AWQ) which is safe and already validated.
- vllm.service.j2: branch on role==embedding for --runner pooling
--convert embed, --no-enable-prefix-caching, per-model
trust_remote_code toggle (needed for nomic-embed-text-v1.5's custom
NomicBertModel code), and enforce_eager toggle (needed to avoid CUDA
graph capture OOM when co-resident with another vLLM process on this
24GB card).
- tasks/verify.yml: split completions vs embedding smoke tests --
embedding-mode instances don't serve /v1/completions. Assert a
non-empty embedding vector, not just HTTP 200.
- host_vars/astro-orbiter: enable nomic-embed-text-v1.5 (port 8020),
lower primary model's gpu_memory_utilization 0.95->0.90 + add
enforce_eager after finding 0.95 crash-looped 6-7x before stabilizing
with co-resident nomic-embed (real fix, confirmed via NRestarts=0
after clean stop/start, not luck).
- Hindsight (values.yaml + externalsecret.yaml): cut LLM + embeddings
over to vLLM (:8000, :8020), wire the previously-unset
HINDSIGHT_API_EMBEDDINGS_* env vars for the first time, and swap the
API key secret source from the Nous fallback item to vllm/api-key
(vLLM enforces real auth, llama-swap did not).
- README: document the embedding-mode branch, VRAM findings, and a
genuine architecture gap -- vLLM's one-model-per-process design
cannot replace llama-swap's 5-model LRU roster on this 24GB card, so
21 Hermes profiles' aux-model consumers (Qwen3-8B-no_think,
Phi-3.5-mini, Meta-Llama-3.1-8B, Qwen2.5-Coder-14B) and OpenViking's
VLM stay on llama-swap. Full teardown (t_6dff1ecc) needs a human
decision on the aux-model strategy before it can proceed.
Idempotent vLLM OpenAI-compatible serving role, staged-first (does not
start/enable the systemd unit or touch production traffic by default).
Validated end-to-end against astro-orbiter in a brief shadow window
(llama-swap stopped ~5 min, per homelab-llm-inference skill's documented
shadow-validation pattern):
- /health 200, /v1/models returns Qwen2.5-32B-Instruct-AWQ,
/v1/completions live smoke test passes, clean journalctl
- 3 consecutive full-role runs confirmed changed=0 (idempotent)
- production restored: llama-swap active, /v1/embeddings against
nomic-embed-text-v1.5 confirmed still working (Hindsight retain path)
Deviates from the original spec's model choices (Qwen2.5-32B-Instruct /
Qwen3-8B-Instruct bf16) to use the official Qwen AWQ pre-quantized variants
instead -- vLLM does not do safe on-the-fly quantization on this host
(bitsandbytes OOM history) and unquantized bf16 32B does not fit 24GB VRAM.
Two real bugs found+fixed during first-start validation (systemd-only
repro, not visible via interactive SSH testing):
1. ninja not on systemd's minimal PATH -> vLLM torch.compile
FileNotFoundError. Fixed via explicit PATH env in the unit.
2. FlashInfer sampler JIT fails to compile on RTX 3090 (SM86) --
known upstream issue class (vLLM GH #23023, #44305). Fixed via
VLLM_USE_FLASHINFER_SAMPLER=0 (falls back to native sampler).
Also fixed a real idempotency bug: force-upgrading setuptools to latest
fought with vLLM's own setuptools<81.0.0 pin, causing an install/downgrade
flip-flop (changed:true) on every run.
vllm_service_enabled defaults to false -- a host reboot must not
auto-start vLLM and VRAM-collide with the still-live llama-swap production
service. Cutover (enabling + starting + migrating consumers) is an
explicit, separate step outside this role, gated on adding embedding-mode
support (--task embed) for nomic-embed-text-v1.5, which this role does
not yet implement (Hindsight retain still depends on llama-swap's
nomic-embed until that follow-up lands).
Role: roles/deploy-vllm/ (defaults/handlers/meta/tasks/templates/README)
Playbook: playbooks/day1_deploy_vllm.yml
t_d7f8cd65 — astro-orbiter is a single serial llama-swap process; Hindsight's
default LLM_MAX_CONCURRENT=32 burst N parallel retains -> node rejects/times
out extras -> 502. Cap global + retain concurrency to 1 (native semaphore
config, no code change). Raise client timeout 120->600s and ingress
proxy-read/send 300->600s so a single long retain (~90s under load) survives.
stepfun rejects Hindsight tagged fact-extraction (400 'missing tags');
solar-pro4:free is the previously-verified-good Nous model for retain/recall
(t_e3375410 / t_d0dffc3d). base_url + provider unchanged.
Scoped to the hindsight ingress (cosmic-rewind.local.mk-labs.cloud) only.
A slow agentic reflect (~132s on solar-pro4) exceeded the 60s default
proxy-read-timeout, causing nginx to return 504 before the app delivered 200.
Raising read+send to 300s lets the response land. No global configmap change.
Refs t_e3375410.
astro-orbiter llama-server-router retired :8002 on 2026-08-18
(llama-swap now serves the OpenAI-compatible API on :8001).
OpenViking still pointed at the dead :8002, so every search/find
call failed: dense embedding -> Connection error -> 500 INTERNAL
server error on /api/v1/search/search and /api/v1/search/find.
Verified from the pod: :8002 = connection refused,
:8001/v1/embeddings (nomic-embed-text-v1.5) = 200 with vectors.
Repoints embedding.dense.api_base and vlm.api_base to :8001.
t_abc05db5
Resolved conflicts between local bafd76a (t_c5cef2b2, Qwen3-8B single-variant)
and origin/main 5cf4468 (t_664289a0, Qwen3-8B dual-thinking deployment).
Conflict resolution strategy: took origin/main version throughout — it is the
authoritative result from t_664289a0 which deployed the live no_think config to
astro-orbiter and already reflects the correct production state.
Changes incorporated from origin/main:
- defaults/main.yml: Qwen3-8B-Q4_K_M-no_think model entry (port 8107, n_gpu_layers=99,
chat_template_file), row6 matrix entry, dual-thinking comment block.
- templates/llama-server-router-preset.ini.j2: [Qwen3-8B-Q4_K_M] with sleep-idle-seconds=60
plus new [Qwen3-8B-Q4_K_M-no_think] section with chat-template-file directive.
- templates/qwen3-no-think.jinja.j2: new file — Qwen3 template with enable_thinking=false.
- tasks/models.yml: template deploy task for qwen3-no-think.jinja (chat_templates tag).
- tasks/swapmode.yml: GATE 2 assert updated to 7 models.
- templates/llama-swap-config.yaml.j2: chat_template_file flag support.
Also pulled in monitoring defaults (llm_monitoring_enabled, VRAM exporter settings,
Grafana dashboard vars, Prometheus scrape config) from origin/main monitoring branch.
- Remove global --n-gpu-layers from router unit ExecStart in preset mode
(llama.cpp CLI arg outranked per-model INI n-gpu-layers=0; root cause from
War Machine's run 1). Flag now emitted only in --models-dir mode.
- All 5 preset INI sections carry explicit n-gpu-layers:
Qwen3.8=99, Phi=99, nomic=99, Coder=0, Llama=0.
- host_vars/astro-orbiter: llm_router_models_max 2 -> 4 so CPU-offloaded
models count as loaded without LRU-evicting Qwen3.8.
- defaults: llm_router_coder_gpu_layers / llm_router_llama_gpu_layers = 0.
- verify.yml: fix pre-existing .meta attribute crash in router mode.
- New playbook day2_cpu_offload_aux_models.yml.
Deployed + verified on astro-orbiter (gates A-E PASS): concurrent residency
achieved, Qwen3.8 stays GPU-resident. Measured CPU throughput Llama 9.0 /
Coder 4.7 tok/s. VRAM note: llama.cpp 6ea215d allocates ~1.4-1.7GB CUDA-context
per CPU model even at n-gpu-layers=0 -> ~24,004 MiB steady-state, below the
24,576 MiB physical limit. Comments corrected to match the measurement.
Report: friday/inbox/ryan/2026-08-17-llm-cpu-offload-coder-llama-deployed.md
Ryan-directed model swap (kanban t_f5f7e9ad, 2026-08-16).
Changes:
- Replace [Qwen3.6-35B-A3B-UD-Q4_K_S] with [Qwen3.8-27B-Q4_K_M] in
llama-server-router-preset.ini.j2 (production model slot).
- Qwen3.8-27B: dense 27B VLM, Apache-2.0, Alibaba Aug 2026.
Unsloth Dynamic V3.0 GGUF quantization.
Q4_K_M chosen: 17,106,775,008 bytes, 17.1GB. Measured VRAM: 17,068 MiB
at ctx=32768 (q4_0 KV cache).
- ctx-size set to 32768 (32K) via new variable llm_router_qwen38_ctx_size.
Native context is 262K; 32K chosen to maintain eviction headroom on 24GB RTX 3090.
- models-max reduced 4 -> 2 in host_vars. Qwen3.8 (17.6GB) + nomic-embed
(558MB) exhaust the 24GB card; no auxiliary model can co-reside with Qwen3.8.
LRU eviction handles model switching with ~30-60s cold-load latency.
- llm_router_expected_model_id updated to Qwen3.8-27B-Q4_K_M.
- Qwen3.6 GGUF retained at /opt/models/Qwen3.6-35B-A3B-UD-Q4_K_S.gguf
(not deleted — pending stable period and explicit cleanup task).
- day2_swap_qwen38.yml playbook added for Ansible idempotent redeployment.
Architecture note: Qwen3.8 uses Gated DeltaNet; llama.cpp 6ea215d logs
'fused Gated Delta Net (chunked) not supported, set to disabled'. Inference
works correctly on the non-fused fallback. A llama.cpp update may improve
throughput on the GDN layers.
Smoke test passed: model responded via router endpoint (http://10.1.71.130:8002).
VRAM: 17,630 MiB (Qwen3.8) + 5,928 MiB (Llama-8B concurrent) = 23,558 MiB.
Also commits accumulated but unpushed changes:
- nomic-embed batch-size/rope-scaling fix (t_openviking_embed_batch)
- per-model ctx-size day2 playbook (day2_per_model_ctx_size.yml)
- llama-server-router.service.j2 minor update
- vlm.model was 'llama3.1-8b' which doesn't exist on astro-orbiter's
/v1/models, causing every summarization call to 400 and endless
circuit-breaker retries. Correct id: Meta-Llama-3.1-8B-Instruct-Q4_K_M.
- embedding.max_input_tokens=1536 still let chunks through that actually
tokenized to 2000-2860 real tokens (estimator undercounts vs llama.cpp's
tokenizer by 1.35x-1.86x on this corpus). Lowered to 1024 for real margin
under the 2048 n_ctx ceiling.
astro-orbiter's llama.cpp router hard-caps nomic-embed-text-v1.5 effective
context at 2048 tokens regardless of ctx-size (known nomic-bert/RoPE limitation
in llama.cpp, not fixable server-side). OpenViking chunks observed at
2000-3400 tokens were tripping 400 exceed_context_size_error and endless
circuit-breaker re-enqueue for viking://temp/default/08140552_5f1c9e/homelab.tar/*.
Set embedding.max_input_tokens: 1536 (well under 2048) since OpenViking's
chunk-time token estimator uses a different tokenizer than llama.cpp's context
counter, so token counts won't match 1:1 - 1536 leaves ~25% headroom.
Approved by Ryan as lowest-risk mitigation (does not touch astro-orbiter/
llama.cpp serving config, which is War Machine's domain and already fixed
separately for the ubatch-size issue).
- Build/push image: the-seas.local.mk-labs.cloud/library/maelstrom-ui:v0.3.17-1
(upstream volcengine/openviking web-studio/, pinned to commit 3cd1d4e9)
- Deployment + Service serving the static SPA via nginx (reverse-proxies
/api, /health, /ready to openviking backend; /bot deliberately NOT proxied)
- Ingress at maelstrom.local.mk-labs.cloud (TLS via letsencrypt-internal)
- ExternalSecret wiring scoped maelstrom-ui-key from
op://mk-labs/openviking/maelstrom-ui-key into the pod env (MAELSTROM_UI_KEY)
Per approved plan: inbox/ryan/2026-08-14-maelstrom-ui-deployment-plan.md
Key mint + approval: system/inbox/agents/nick-fury/2026-08-14-maelstrom-ui-key-mint-complete.md
Ryan approval: inbox/ryan/2026-08-14-maelstrom-key-approval.md
The letsencrypt-internal ClusterIssuer does not exist on the cluster.
Only letsencrypt-prod and letsencrypt-staging are available.
Both use DNS-01 via Cloudflare for the mk-labs.cloud zone,
so they work for internal-only hosts with no public HTTP reachability.
Fixes: https://github.com/volcengine/openviking/issues/...
Closes: kanban task t_1c2cc2db
Bug 1: ExternalSecret referenced three separate 1Password items
(openviking-root-api-key, openviking-embedding-api-key, openviking-vlm-api-key)
but Ryan created ONE item 'openviking' with three fields inside.
Changed all remoteRef.key values to 'openviking' and corrected field property names.
Bug 2: values.yaml had two invalid embedding config fields:
- encoding_format: 'float' (not in upstream schema, removed)
- max_concurrent under embedding.dense (wrong nesting, moved to embedding level)
Verified against upstream chart schema at github.com/volcengine/openviking
Remove resource-level sync-wave annotations that caused ArgoCD deadlock.
The wave 8 annotation was meant for platform-level ordering (apps-of-apps)
but was incorrectly applied to individual resources within the Application.
This caused ArgoCD to apply Deployment (wave 0) before ExternalSecret (wave 8),
resulting in CreateContainerConfigError since the pod needed the secret first.
Changes:
- namespace.yaml: Remove sync-wave annotation, add clarifying comment
- externalsecret.yaml: Change sync-wave from 8 to -1 (must sync before Deployment at wave 0)
- application.yaml: Remove sync-wave annotation, document as platform-level only in comments
This ensures:
1. ExternalSecret syncs first (wave -1)
2. Deployment uses it immediately (implicit wave 0)
3. No deadlock
Task: t_3906c41a
- Changed source 1 from 'chart: deploy/helm/openviking' to 'path: deploy/helm/openviking'
- ArgoCD multi-source now correctly resolves the Helm chart from the git repo
- targetRevision: main now correctly refers to a git branch, not a chart version
- Fixes: invalid revision 'main': improper constraint error
Platform Knowledge Infrastructure pilot - context database for large
file trees, shared skills, and long-term logs to reduce agent token
consumption. Pilot scope: two corpora (hermes/ skills library,
personal/homelab/), two consumer profiles (Wong, Shuri) for before/after
token comparison.
- namespace.yaml: openviking namespace, sync-wave 8 (after Harbor wave 7)
- externalsecret.yaml: credentials from 1Password via onepassword-connect
ClusterSecretStore (Wong, t_32766900)
- values.yaml: Helm overrides - px-fa-direct-access storage (30Gi),
embedding (nomic-embed-text-v1.5) + VLM (Llama-3.1-8B) via astro-orbiter
router (:8002), internal-only ingress
- application.yaml: multi-source ArgoCD Application, Harbor pattern
(Peter Parker, t_eefdcc17 + reconciled in t_3e54efa8)
Prerequisites verified complete before this commit:
- nomic-embed-text-v1.5-Q4_K_M live on astro-orbiter router (War Machine,
t_34b96e83, commit ad70b34)
- All 3 1Password items provisioned (root/embedding/vlm api keys)
- Storage class corrected to px-fa-direct-access after live PV audit
showed pure-block/pure-file have zero provisioned volumes (t_77b3ff79)
- Dry-run validated against live cluster prior to commit
Constraint: vault (~/friday) remains canonical source of truth; OpenViking
index is a derived cache, rebuilt from vault source files.
Honcho/lincoln explicitly out of scope for this work.
/metrics?model=Qwen3.6 forces the router to attempt loading Qwen3.6 every
90s scrape cycle, triggering a CUDA OOM error since VRAM is already consumed
by the resident Llama3+Phi3.5 models. This produces real GPU power spikes
(~110W load-attempt), not a benign counter read like the Llama3/Phi3.5 jobs.
nvidia_gpu_exporter (:9835) already provides GPU power/VRAM/utilization at
zero wake cost. No Grafana dashboard panel references Qwen3.6 model-specific
llama-server metrics. Ryan approved full removal.
Job is commented out (not deleted) for easy revert if Qwen3.6 is ever
re-added as a resident model. Llama3/Phi3.5 scrape jobs untouched.
Each /metrics?model=<id> request on the llama.cpp router wakes the GPU
sub-server to P2 (~110W). At 15s with two active jobs (llama3 + phi35),
combined scrape frequency (~7-8s effective) keeps the GPU continuously at
P2 despite zero real inference requests.
At 90s: each scrape wakes GPU for ~5-10s then it drops to P8 (~20W) for
~80s. Verified via iptables block test on 2026-08-13 (t_e7d547ea):
Before block: 110-115W P2 continuously
After block: 19-21W P8 consistently
After unblock: returned to 110W P2 within seconds
qwen3 interval also set to 90s (model not loaded so moot, but consistent).
Ref: t_e7d547ea
Port 8000 (gemma-2-27b-it-GGUF) is dead after the day2 router cutover on
2026-08-12. Production inference now runs through llama-server-router on
port 8002.
The router exposes per-model Prometheus metrics via /metrics?model=<id>.
Since a single /metrics request without ?model returns HTTP 400, replaced
the single stale job with three per-model jobs — one per model registered
in the router per /v1/models:
- Qwen3.6-35B-A3B-UD-Q4_K_S (currently unloaded but registered)
- Meta-Llama-3.1-8B-Instruct-Q4_K_M (loaded)
- Phi-3.5-mini-instruct-Q8_0 (loaded)
Static 'model' label carries the canonical llama.cpp model id (not alias).
Added 'endpoint: astro-orbiter-router' to identify the scrape origin.
Removed dead :8000 target entirely.
- host_vars/astro-orbiter/vars.yml: add Qwen2.5-Coder-14B-Instruct-Q4_K_M.gguf
to llm_staged_models (size_bytes=8988111072, bartowski GGUF public repo).
Updated VRAM note to reflect 4-model roster and LRU eviction semantics.
- defaults/main.yml: add llm_router_coder_ctx_size=16384 and
llm_router_coder_flash_attn=true variables for per-model ctx tuning.
- templates/llama-server-router-preset.ini.j2: add [Qwen2.5-Coder-14B-Instruct-Q4_K_M]
section with alias=Qwen2.5-Coder-14B-Instruct-4bit, ctx-size=16384, flash-attn=true.
- playbooks/day2_add_coder_alias.yml: new playbook that downloads the GGUF (if
absent/mismatched), deploys updated preset INI and systemd unit, restarts
llama-server-router, and verifies all 4 models in /v1/models.
VRAM: Coder ~9GB. Full 4-model co-residency impossible on 24GB — LRU eviction
handles this automatically. Qwen3.6-35B <-> Coder switches incur ~30-60s cold load.
- Add Meta-Llama-3.1-8B-Instruct-4bit alias on Meta-Llama-3.1-8B-Instruct-Q4_K_M
entry in the preset INI (per Ryan dashboard note).
- Document known issue: Phi-3.5-mini json_schema grammar sampler incompatibility
in router mode (GH #23460 variant — chat template token format mismatch).
Meta-Llama works with json_schema response_format; confirmed via live test.
- Phi-3.5-mini-instruct-8bit alias remains working for model routing;
structured output (json_schema) fails due to the model's token format.
Changes:
- host_vars/astro-orbiter/vars.yml: add llm_router_models_max: 4 (overrides
conservative default of 1). Detailed VRAM OOM risk note included inline:
worst-case 3-model co-residency ~31GB > 24GB RTX 3090. LRU eviction
mitigates in single-user operation; flagged for Ryan's review.
- playbooks/day2_bump_router_models_max.yml: new targeted playbook; deploys
updated router unit, restarts the live service, verifies /health 200 and
/v1/models lists all 3 GGUFs post-restart.
- group_vars/all/semaphore.yml: add llm_router_update_unit template pointing
at the new playbook.
- roles/llm-inference-multimodel/defaults/main.yml: update comment to reflect
the var is now overridden in host_vars rather than 'hardcoded to 1'.
- roles/llm-inference-multimodel/templates/llama-server-router.service.j2:
correct stale 'HARDCODED TO 1' comment — value is variable-driven.
Constraints honored:
- --parallel 1 left untouched (not in scope, not modified anywhere)
- No ad-hoc SSH/systemctl/curl state mutation; all execution via Semaphore
- No installed/vendored code patched
Adds idempotent, data-driven GGUF staging for the two new router models on
astro-orbiter alongside the production Qwen3.6-35B-A3B-UD-Q4_K_S. Both files
were already staged live (byte-exact); this commit codifies them in Ansible so
future re-runs and any new model adds are version-controlled and audit-friendly.
Changes:
- roles/llm-inference-multimodel/tasks/stage_model.yml (NEW)
Idempotent per-model task: stat -> exact byte-size guard -> conditional
get_url -> ownership/mode ensure -> notify router restart handler only on
actual download. Loops from models.yml; nothing hardcoded.
- roles/llm-inference-multimodel/tasks/models.yml
Appends the stage_model.yml loop (tagged: models) after the existing Qwen3.6
download tasks. Data driven from host_vars/astro-orbiter/vars.yml.
- roles/llm-inference-multimodel/defaults/main.yml
Adds llm_staged_models: [] default (empty = safe no-op for hosts with no
staged model list defined).
- roles/llm-inference-multimodel/handlers/main.yml
Adds 'restart llama-server-router on new GGUF' handler. Only fires when
stage_model.yml performs an actual download or corrects ownership/mode.
Normal idempotent re-runs (files already correct) do NOT fire this handler.
- host_vars/astro-orbiter/vars.yml
Adds llm_staged_models list with the two new models:
* Phi-3.5-mini-instruct-Q8_0.gguf (4,061,222,688 bytes,
bartowski/Phi-3.5-mini-instruct-GGUF)
* Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf (4,920,739,232 bytes,
bartowski/Meta-Llama-3.1-8B-Instruct-GGUF)
- playbooks/day1_deploy_llm_inference_multimodel.yml
Updates header comment: removes stale 'Semaphore broken' note, documents
the correct execution channel (Semaphore template
llm_inference_multimodel_stage_models, --tags models).
- group_vars/all/semaphore.yml
Adds llm_inference_multimodel_stage_models template entry (config-as-code).
Template is scoped to --tags models explicitly. Phase 4 (verify) is
EXCLUDED: verify.yml starts llama-server-qwen on :8002, which would collide
with the production llama-server-router.service already running on :8002.
Semaphore template created via API: project 1 / template id 19.
Execution: triggered immediately after this commit via Semaphore REST API.
