homelab/ansible/roles/deploy-vllm/README.md

deploy-vllm

Idempotent Ansible role that deploys a vLLM OpenAI-compatible inference server. Written for astro-orbiter (RTX 3090, 24GB VRAM, 64GB RAM, Ubuntu 24.04) and designed for reuse on the planned Mac Mini M4 host later this week (see "Portability" below).

Supersedes the manual, pre-role state left behind by earlier vLLM experiments (/home/jarvis/vllm-env, bitsandbytes, gemma-2-27b — see homelab-llm-inference/homelab-llm-serving skills for that history). This role uses a fresh venv (vllm_venv_path, default ~/vllm-serve-env) and AWQ pre-quantized models — no bitsandbytes, no on-the-fly quantization, no repeat of the OOM incident from the earlier Gemma-2-27B attempt.

Phases

Phase	File	What it does
1	tasks/dependencies.yml	System Python 3.10+, dedicated venv, pip install vllm>=0.5.0, verifies nvidia-smi and torch.cuda.is_available()
2	tasks/models.yml	Downloads each enabled: true model in vllm_models via hf download (huggingface_hub CLI) into ~/.vllm-cache, verifies the snapshot landed and reports on-disk size
3	tasks/api-key.yml	Reads the API key from 1Password (op://mk-labs/vllm/api-key) on the controller, writes it to /etc/vllm/api-key.env (root:root, 0600) on the target
4	tasks/systemd.yml	Renders and installs one systemd unit per enabled model (vllm.service for the role: primary model, vllm-<id>.service for others)
5	tasks/verify.yml	Only runs when vllm_service_state=started. Waits for /health (up to 5 min — torch.compile warmup), checks /v1/models, runs a live completion, scans journalctl for errors

Run all phases: ansible-playbook -i inventory.yml playbooks/day1_deploy_vllm.yml --limit astro-orbiter Run one phase: --tags vllm-dependencies / vllm-models / vllm-api-key / vllm-systemd / vllm-verify

Deliberate staging-first default

vllm_service_state defaults to stopped. A default run stages everything (venv, model weights, API key file, systemd unit) but does not start the service or touch production traffic. This matches the astro-orbiter cutover plan: llama-swap is live production serving (Qwen3.8-27B

• nomic-embed for Hindsight) — vLLM must be deployed and validated on a side port/inactive unit before anything is cut over.

To start and validate:

ansible-playbook -i inventory.yml playbooks/day1_deploy_vllm.yml \
  --limit astro-orbiter --extra-vars "vllm_service_state=started"

This starts the systemd unit(s), enables them, and runs Phase 5 verification (health, /v1/models, live completion, clean journalctl).

Cutover of consumers (Hermes profiles, Hindsight embedding config, any hardcoded :8001/:5805 references) to the new :8000 vLLM endpoint is a separate, explicit step outside this role — do this only after Phase 5 passes cleanly. Do not tear down llama-swap until consumers are confirmed working end-to-end against vLLM.

Model roster (vllm_models in defaults/main.yml)

vLLM 0.5.x-0.28.x serves one model per process — multi-model = multiple systemd units on distinct ports, not a single multiplexed server (unlike llama-swap's matrix DSL). Today's phase enables only the primary model; flip enabled: true on the others as VRAM allows (see "Phased Strategy"):

id	hf_repo	role	port	quant	enabled
Qwen2.5-32B-Instruct-AWQ	Qwen/Qwen2.5-32B-Instruct-AWQ	primary	8000	awq	true
Qwen3-8B-AWQ	Qwen/Qwen3-8B-AWQ	aux	8010	awq	false
nomic-embed-text-v1.5	nomic-ai/nomic-embed-text-v1.5	embedding	8020	none	false

Note on the original spec's model choices: the task body named Qwen/Qwen2.5-32B-Instruct and Qwen/Qwen3-8B-Instruct (bf16, unquantized). vLLM does not do on-the-fly quantization safely on this host (bitsandbytes OOM history — see homelab-llm-inference skill Pitfalls) and unquantized bf16 32B does not fit a 24GB card at all (~65GB). This role instead deploys the official Qwen AWQ pre-quantized variants (Qwen/Qwen2.5-32B-Instruct-AWQ, Qwen/Qwen3-8B-AWQ), which vLLM natively supports (--quantization awq) and which fit the VRAM budget:

• Qwen2.5-32B-Instruct-AWQ: ~19.3GB on disk, fits with ~5GB headroom at 24GB
• Qwen3-8B-AWQ: ~6GB VRAM per llm-explorer
• nomic-embed-text-v1.5: ~300MB, vLLM serves it via --convert embed pooling (see vLLM embedding docs) — not yet wired into this role's systemd template; the embedding model needs --task embed / --convert embed flags that differ from the completion-serving template. Flagged as a follow-up before enabled: true is flipped on it (see Known Gaps below).

Known Gaps / Follow-ups

• Quarterly API key rotation is documented (/etc/vllm/API_KEY_ROTATION.md on the target, rendered by tasks/api-key.yml) but not automated — no cron job exists to force rotation on a schedule. Consider a follow-up cron task if Nick Fury wants this enforced rather than just documented.
• vllm_service_enabled defaults to false deliberately — see "Deliberate staging-first default" above. Flip together with the cutover step, not before.
• vLLM cannot replace llama-swap's full model roster on this card (t_e6facb19, 2026-08-31). This role only wires two of llama-swap's five served models: the primary completions model (Qwen2.5-32B-Instruct-AWQ, replacing Qwen3.8-27B) and the embedding model (nomic-embed-text-v1.5). llama-swap ALSO serves Qwen3-8B-Q4_K_M(-no_think), Phi-3.5-mini-instruct-Q8_0, Meta-Llama-3.1-8B-Instruct-Q4_K_M, and Qwen2.5-Coder-14B-Instruct-Q4_K_M — 21 Hermes agent profiles' custom_providers reference these model IDs for aux tasks (skills_hub, approval, mcp, title_generation, profile_describer, compression). vLLM 0.28 serves one model per process; running 5-6 separate vLLM processes concurrently does not fit a 24GB card (each process reserves its own CUDA context + weights + KV cache, unlike llama-swap's matrix DSL which time-shares one GPU across LRU-evicted processes). Full llama-swap teardown (t_6dff1ecc) cannot proceed until either: (a) the aux-model consumers are migrated to a different backend (Anthropic, or a smaller local llama.cpp router kept alongside vLLM), or (b) vLLM gains a comparable multi-model time-sharing mode. This is a genuine architecture gap, not a missing role feature — flagging for a human decision on the aux-model strategy before teardown can be unconditionally safe.

Embedding-mode support (t_e6facb19, 2026-08-31)

vllm.service.j2 now branches on role: embedding entries in vllm_models: adds --runner pooling --convert embed (vLLM's embedding-serving flags — see https://docs.vllm.ai/en/latest/models/pooling_models/embed/) and --no-enable-prefix-caching (prefix caching is a completions-only optimization; irrelevant and safely disabled for pooling). An additional per-model trust_remote_code: true toggle renders --trust-remote-code when set — required for nomic-ai/nomic-embed-text-v1.5, which ships custom NomicBertModel modeling code on its HF repo.

Verification does NOT run /v1/completions against embedding-mode instances (they don't serve that endpoint — a completions request 400s immediately). tasks/verify.yml splits vllm_enabled_models by role and runs the appropriate smoke test per group: completions models get the /v1/completions "capital of France" test; embedding models get a real /v1/embeddings POST with an ansible.builtin.assert on a non-empty data[0].embedding array (not just HTTP 200 — an empty/malformed vector would still 200).

Critical VRAM finding: co-resident completions + embedding vLLM processes need MORE headroom than either alone, and CUDA graph capture is the failure mode, not KV cache sizing. Enabling nomic-embed-text-v1.5 alongside the primary Qwen2.5-32B model at the role-default gpu_memory_utilization: 0.95 crash-looped repeatedly:

• First failure: torch.OutOfMemoryError during capture_model() (CUDA graph capture) — KV cache sizing itself succeeded (14,720 tokens allocated), but graph capture needed ~20MiB more than the 0.95 budget had left once nomic's embedding process (814MiB actual, not the nominal ~300MB estimate in the model roster table) claimed its share.
• Fix attempt 1: added a per-model enforce_eager: true template branch (--enforce-eager skips CUDA graph capture entirely) — this stopped the graph-capture OOM but the combined processes still landed at only ~847MiB genuinely free out of 24,576MiB, and both services crash-looped 6-7 times during warmup before finally stabilizing (each attempt leaves transient VRAM that the next attempt fights over, extending time-to-stable well past a single health-check retry window).
• Fix attempt 2 (final, verified stable): lowered the primary model's gpu_memory_utilization from 0.95 to 0.90 (host_vars override) in addition to enforce_eager: true. Result: clean single-attempt start for both services, NRestarts=0, ~2GB genuinely free (22,577MiB used / 24,576MiB total). Confirmed via systemctl show <unit> -p NRestarts after a full stop/start cycle — 0.95 was NOT a fluke of Restart=always masking the underlying fragility; 0.90 is a real, reproducible fix.
• Takeaway for future multi-process vLLM VRAM budgeting on this host: do not just check "does it eventually come up" — check NRestarts and free VRAM headroom after a clean stop/start. A model that "works" after 6 crash-loop retries is not production-stable; the retries themselves are evidence the utilization ceiling is too tight for the actual (not nominal) footprint of co-resident processes.

Consumer cutover status (t_e6facb19, 2026-08-31)

Cut over (validated end-to-end):

• Hindsight (cluster/applications/hindsight/values.yaml + externalsecret.yaml): HINDSIGHT_API_LLM_BASE_URL → vLLM :8000 (Qwen2.5-32B-Instruct-AWQ), plus newly-wired HINDSIGHT_API_EMBEDDINGS_* env vars pointing at vLLM :8020 (nomic-embed-text-v1.5). Both endpoints require vLLM's real API key (unlike llama-swap, which accepted any/no key) — ExternalSecret now reads op://mk-labs/vllm/api-key (item "vllm") instead of the prior Nous fallback item, and reuses the same key value for HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY (both vLLM endpoints share one key file per tasks/api-key.yml).

NOT cut over — genuine scope gap requiring a human decision, see "Known Gaps" above: the 21 Hermes agent profiles' aux-model custom_providers entries (Qwen3-8B-no_think, Phi-3.5-mini, Meta-Llama-3.1-8B, Qwen2.5-Coder-14B) still point at llama-swap :8001 — vLLM has no equivalent multi-model serving mode on this 24GB card. llama-swap MUST stay up to serve these until that gap is resolved. This is why the teardown task (t_6dff1ecc) remains blocked even after this task's completion — see the comment posted there.

• OpenViking (cluster/platform/openviking/values.yaml): still points at llama-swap :8001 (Meta-Llama-3.1-8B-Instruct-Q4_K_M VLM + nomic-embed for dense embeddings). Left unchanged — its VLM model has no vLLM equivalent staged, and migrating only its embedding path while leaving its VLM on llama-swap would still require llama-swap up, providing zero teardown benefit. Flagged, not touched, per the same aux-model gap above.

Validation Log (2026-08-31, t_ca1af9fb)

Full Phase 1-5 run executed against astro-orbiter in a brief shadow-validation window (llama-swap stopped ~5 min, per the homelab-llm-inference skill's documented shadow-validation pattern — production traffic could not be tested concurrently with vLLM's VRAM footprint on this 24GB card).

Two real bugs found and fixed during first-start validation (not present in the original spec, discovered only by actually starting the service):

1. ninja not on systemd's PATH. vLLM's torch.compile path shells out to the bare ninja command. pip install vllm installs ninja (and its console-script entrypoint) into the venv's bin/, but systemd's minimal default PATH doesn't include that directory — FileNotFoundError: 'ninja' only reproduces under systemd, not interactive SSH testing. Fixed by setting Environment="PATH=<venv>/bin:...standard dirs..." in the unit template.
2. FlashInfer sampler JIT fails to compile on RTX 3090 (SM86). flashinfer/data/csrc/sampling.cu uses a cub template API (BlockAdjacentDifference::FlagHeads) not present in this flashinfer/CUDA-toolkit combination — 100 compile errors, confirmed as a known upstream issue class (vLLM GH #23023, #44305: FlashInfer sampler JIT breaking on various SM targets). Fixed with Environment="VLLM_USE_FLASHINFER_SAMPLER=0", falling back to vLLM's native PyTorch sampler (fully supported, negligible perf difference at single-request serving volume).

Also corrected vllm_gpu_memory_utilization from 0.90 to 0.95 — at 0.90 the KV cache allocation failed (2.0 GiB KV cache needed, 1.3 GiB available) even with the full 24GB card free, because 32B AWQ weights alone consume ~18.4GB, leaving too little headroom at a 90% cap.

Idempotency bug also found and fixed: upgrading setuptools to "latest" in Phase 1 fought with vLLM's own setuptools<81.0.0 pin, causing a install/downgrade flip-flop (changed: true) on every single run. Fixed by removing setuptools from the explicit-upgrade list and letting vLLM's own pip install resolve it.

Final validated result, once these fixes were applied:

• systemctl status vllm.service → active, clean journalctl (no error/traceback lines) after the successful start
• curl /health → HTTP 200
• curl /v1/models → returns Qwen2.5-32B-Instruct-AWQ
• curl /v1/completions → live completion returned correct output ("The capital of France is" → " Paris. Correct! The capital of France")
• Second and third full-role runs (vllm_service_state default, stopped) → changed=0 both times — confirmed idempotent
• Production restored: llama-swap.service active, /health 200, /v1/embeddings against nomic-embed-text-v1.5 returns a valid vector — Hindsight retain path confirmed still working after the shadow window
• Post-restore VRAM: 486 MiB used / 24,576 MiB total (normal quiescent state)

Testing this role (idempotency)

Second-run test (staging phases only, safe to run repeatedly):

ansible-playbook -i inventory.yml playbooks/day1_deploy_vllm.yml \
  --limit astro-orbiter --tags vllm-dependencies,vllm-models,vllm-api-key,vllm-systemd
# Run it again immediately — expect changed=0 (or only handler-driven
# restarts if vllm_service_state=started and the API key file rotated)

Confirmed 2026-08-31 (t_ca1af9fb): Phase 1 (dependencies) ran once with changed=3 (venv create, pip upgrade, vllm install); a second run reported changed=0 for those three tasks — venv creates: guard and pip module's own idempotency both held.

Portability — Mac Mini M4 (planned, end of week)

This role's host-specific assumptions live in defaults/main.yml (all overridable via host_vars/<host>/vars.yml) plus one hard assumption baked into tasks/dependencies.yml: an NVIDIA GPU (nvidia-smi check, CUDA wheels). Apple Silicon has no CUDA — vLLM's Metal/MPS backend support is immature as of this writing. Before reusing this role for the Mac Mini M4:

1. Fork tasks/dependencies.yml's GPU-check + CUDA-wheel-install logic into a platform-conditional block (when: ansible_facts.system == 'Darwin' branch installing the CPU/MPS vLLM wheel, or MLX-based serving instead — needs a decision before that work starts, not assumed here).
2. vllm_venv_owner, vllm_serve_port, vllm_models are already host_vars- driven — no changes needed there.
3. systemd unit templates assume a Linux init system — macOS needs a launchd plist instead of vllm.service.j2.

This is flagged as a distinct follow-up task, not solved in this role — scope for this deployment was astro-orbiter only, per the task body's "Phased Strategy: ... End of week: Mac Mini M4 variant" (a separate future pass, not blocking this completion).

Files

roles/deploy-vllm/
├── defaults/main.yml               # all tunables — host overrides go in host_vars
├── handlers/main.yml                # reload systemd / restart vllm services
├── meta/main.yml
├── tasks/
│   ├── main.yml                     # phase orchestrator
│   ├── dependencies.yml             # Phase 1
│   ├── models.yml                   # Phase 2
│   ├── api-key.yml                  # Phase 3
│   ├── systemd.yml                  # Phase 4
│   └── verify.yml                   # Phase 5
├── templates/
│   ├── vllm.service.j2              # one instance per enabled model
│   └── vllm-workspace.sh.j2         # debugging helper deployed to the target
└── README.md                        # this file