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.
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.
- 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
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
- 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.
Promote llama-server-router to production on port 8002.
- Stops + disables llama-server-qwen (unit preserved as rollback target)
- Redeploys llama-server-router unit with --port 8002 (not 8003)
- PORT DECISION: rebind router to :8002 — zero Hermes profile config changes needed
- Re-runs validation gates 1-3 against production endpoint (hard gates)
- Gate 4: bundled SvelteKit UI check (HTTP 415 from ansible.builtin.uri is
a false negative — llama.cpp requires Accept-Encoding: gzip; real browsers work)
- Rollback via --tags cutover_rollback (uses 'never' Ansible tag to prevent
accidental execution during normal cutover run)
All 4 gates PASSED on cutover run 2026-08-12:
Gate 1: n_ctx=65536 >= 64000 PASS
Gate 2: finish_reason=tool_calls PASS
Gate 2b: finish_reason=stop (no spurious tool_calls) PASS
Gate 3: 20410 MiB / 23000 MiB ceiling PASS
Gate 4: SvelteKit HTML confirmed via curl + gunzip PASS
Production endpoint: llama-server-router on :8002 (enabled + running)
Rollback target: llama-server-qwen unit at /etc/systemd/system/ (stopped, disabled)
- Retire llama-server-aux (Phi-4, 8000) and llama-server-toolcall (Mistral-Small-24B, 8001): stopped, disabled, unit files removed from host and Ansible role
- Promote llama-server-qwen (Qwen2.5-14B-Instruct-1M, port 8002) to sole production model, serving both friday and war-machine Hermes profiles
- Verified live: n_ctx=65536/n_ctx_train=1010000, and tool_calls response via /v1/chat/completions probe (no hallucination)
- Deleted superseded GGUF weights (phi-4, mistral-small, orphaned base-Qwen, gemma-2-27b) from astro-orbiter, ~45GB reclaimed
- Updated friday and war-machine Hermes profile configs (model + compression + skills_hub aux) to point at 10.1.71.130:8002
- Ryan explicitly accepted single-model tradeoffs for both profiles
Mistral (llama-server-toolcall, 8001) stopped temporarily on astro-orbiter to
free ~6.2GB VRAM headroom for this test window per Ryan/JARVIS approval.
Not a permanent decommission of Mistral.
- New llama-server-qwen systemd unit template, gated by llm_qwen_service_enabled (default false)
- Idempotent GGUF download task (bartowski Qwen2.5-14B-Instruct-Q5_K_M, stat-guarded)
- Launch flags per local-llm-64k-context-recommendation.md: ctx-size 65536, flash-attn, q8_0 KV cache, batch 2048/ubatch 512, jinja, parallel 1
- verify.yml only starts/verifies the qwen unit when llm_qwen_service_enabled=true
- README: documents live VRAM gate finding (nvidia-smi 2026-08-06: Phi-4+Mistral already ~16.6/24GB, ~7.5GB free -- insufficient for Qwen weights concurrently) and options
- Does NOT touch llama-server-aux (8000) or llama-server-toolcall (8001) service state
discover.yml sets llm_existing_gemma_unit_found, but main.yml imports each
phase file with import_tasks + a distinct per-phase tag. Tags on
import_tasks apply to the whole file, so --tags verify (a supported,
documented way to re-run just this phase) skips discover.yml, leaving
the fact undefined. The stop task's 'default(false)' silently no-op'd,
so re-running verify alone against a host with Gemma still running would
start both new instances on top of it -- the OOM this task exists to
prevent.
Fix: gather service_facts and set the fact locally in verify.yml too,
only when not already defined, so the guard works regardless of which
tags were selected.
Phase 2 (systemd tag) notified per-service restart handlers and then
called meta: flush_handlers itself, so any run where either unit's
template content changed (including first apply) restarted BOTH
live services immediately in Phase 2 -- before Phase 3 firewall
scoping or Phase 4 smoke tests ran. This contradicted the phase's
documented purpose (units land on disk only, nothing starts/restarts
until Phase 4).
Fix: Phase 2 only reloads the systemd daemon and registers each
template task's changed result. Phase 4 (verify.yml) now decides
start vs restart per-service based on that recorded change, so
restarts remain independent per instance and never fire before
Phase 4.
- Prometheus scrape configs for node/gpu/llama-server exporters on
astro-orbiter now declared in cluster/applications/monitoring/values.yaml
(additionalScrapeConfigs), applied via ArgoCD sync instead of an
imperative kubectl secret patch from the Ansible role.
- Grafana dashboard for astro-orbiter LLM inference added as a ConfigMap
in cluster/applications/monitoring/dashboards.yaml (grafana_dashboard=1
sidecar label), replacing the role's ad-hoc kubectl apply of a rendered
Jinja template.
- ansible/roles/llm-inference/tasks/monitoring.yml: removed the kubectl
get/patch/apply tasks and orphaned grafana-llm-dashboard.json.j2
template; role now only stands up node_exporter + nvidia_gpu_exporter
and verifies they're reachable — cluster-facing config lives in Git.
- host_vars/vars.yml + inventory.yml: finalize astro-orbiter as the
llama.cpp/RTX 3090 host (jarvis user, ssh key), drop stale Ollama/AMD
vars and ollama_server inventory group superseded by the ATX rebuild.
bitsandbytes quantizes on-the-fly: loads full bf16 weights (~54GB RAM peak)
before compressing to int4. Kills the 40GB OptiPlex on torch.compile warmup.
Fix in next commit: switch serve phase to llama.cpp + GGUF Q4_K_M.
Pre-quantized weights load directly — peak RAM ~16GB, no compile overhead.
Migration from Ubuntu 3.x to upstream 6.x is complete. The explicit
removal task was firing changed on every run. state: latest on the
install task handles upgrades going forward.
Rules were documented but never deployed — /dev/jmri/nce was missing
entirely, only /dev/jmri/loconet existed (created manually).
Adds:
- templates/99-jmri-devices.rules.j2: LocoBuffer-NG -> loconet,
NCE Power Pro (Microchip CDC) -> nce
- Task to deploy rules + trigger udev settle
- Trigger udev handler (reload-rules alone is insufficient)
