llm-inference-multimodel: role + day1 playbook (phase 0 discover approved)

ansible/roles/llm-inference-multimodel/README.md

@@ -0,0 +1,132 @@
+# llm-inference-multimodel
+
+Deploys **two independent llama-server systemd services** on astro-orbiter's
+RTX 3090 (24GB), alongside — not replacing — the existing `llm-inference` role:
+
+| Instance | Port | Model | Quant | ctx | parallel | ~VRAM |
+|---|---|---|---|---|---|---|
+| `llama-server-aux` | 8000 | Phi-4-14B-Instruct | Q4_K_M | 8192 | 2 | ~10.0GB |
+| `llama-server-toolcall` | 8001 | Mistral-Small-24B-Instruct-2501 | Q3_K_M | 4096 | 1 | ~13.2GB |
+
+Combined estimate: **~23.2GB / 24GB** (~0.8GB headroom). See
+`/home/hermes/astro-orbiter-multi-model-plan.md` for the full approved design
+(VRAM math, model selection rationale, rollback plan, validation harness).
+
+## Relationship to `roles/llm-inference`
+
+This role does **not** replace `llm-inference`. It assumes that role's
+prerequisites are already satisfied on the host:
+
+- NVIDIA driver installed
+- `/opt/llama.cpp` cloned and built with CUDA (`/opt/llama.cpp/build/bin/llama-server` exists)
+- `jarvis` service user + `/home/jarvis` present
+
+The pre-existing single-model Gemma llama-server (however it is currently
+run) is **never modified, restarted, or deleted** by this role. It is the
+rollback target.
+
+## Phases
+
+Run the whole role, or scope with `--tags`:
+
+```
+ansible-playbook -i inventory.yml playbooks/day1_deploy_llm_inference_multimodel.yml
+# or, once merged into a single play:
+ansible-playbook -i inventory.yml <playbook>.yml --tags discover,models,systemd,firewall,verify
+```
+
+0. **discover** (`tasks/discover.yml`) — READ-ONLY. Confirms via
+   `service_facts` + `pgrep` whether the existing Gemma llama-server actually
+   runs as a systemd unit today, or some ad hoc way (nohup/screen/tmux). Does
+   **not** assume a unit exists — this was an open unknown in the plan and is
+   resolved here as a fact-gathering step, not an assumption. Also records
+   baseline VRAM and current port 8000/8001 listeners for comparison later.
+
+   **If this reports no unit found**, stop and read the debug message —
+   it means plan §6's rollback story ("systemctl start the old unit to
+   revert") isn't actually available yet, and that should be fixed (codify
+   the existing process as a systemd unit) before proceeding to Phase 2.
+
+1. **models** (`tasks/models.yml`) — Idempotent GGUF download to
+   `/opt/models/` with a stat + minimum-size guard (mirrors the pattern in
+   `roles/llm-inference/tasks/serve.yml` and the `llm-inference-homelab`
+   skill), so reruns don't re-pull 8.5GB/11.7GB files or mistake a truncated
+   partial download for complete.
+
+2. **systemd** (`tasks/systemd.yml`) — Templates and deploys both unit files
+   to `/etc/systemd/system/`. **Deliberately does not start or enable either
+   service** — units land on disk as a separately reviewable checkpoint.
+   Two fully independent units (not one unit with two ExecStarts) so either
+   instance can be restarted/stopped without affecting the other.
+
+3. **firewall** (`tasks/firewall.yml`) — Scopes ports 8000 and 8001 via `ufw`
+   to `llm_allowed_source_cidr` (default the Hermes LAN subnet), rather than
+   leaving them open. Both unit templates also bind to
+   `llm_bind_address` (default `10.1.71.130`, the host's private LAN IP) —
+   **not `0.0.0.0`** — which is a deliberate change from the pre-existing
+   Gemma pattern flagged as insecure in the plan.
+
+4. **verify** (`tasks/verify.yml`) — The only phase that actually starts +
+   enables both services. Waits for `/health` on both ports, smoke-tests
+   `/v1/models` and a trivial `/v1/chat/completions` call on each, checks
+   `nvidia-smi` VRAM usage against the plan's design estimate, and greps
+   `dmesg` for OOM-kill events.
+
+   **This smoke test is not the tool-calling validation harness.** See
+   below.
+
+## Key variables
+
+Defined in `defaults/main.yml` (all overridable via `host_vars`/`group_vars`
+or `-e`):
+
+- `llm_service_user` (jarvis), `llm_binary_path`, `llm_models_dir`, `llm_bind_address`, `llm_allowed_source_cidr`
+- Aux: `llm_aux_port`, `llm_aux_model_path`, `llm_aux_model_url`, `llm_aux_ctx_size`, `llm_aux_parallel`, `llm_aux_gpu_layers`
+- Tool-calling: `llm_toolcall_port`, `llm_toolcall_model_path`, `llm_toolcall_model_url`, `llm_toolcall_ctx_size`, `llm_toolcall_parallel`, `llm_toolcall_gpu_layers`
+
+`vars/main.yml` holds constants not meant to be overridden per-host (HF token
+reference, expected-VRAM figures used only for the verify.yml report).
+
+## ⚠️ Tool-calling validation is required before use
+
+Port 8001 (Mistral-Small-24B) **must** pass the manual validation procedure
+described in plan §7 before any Claude Code / tool-calling-capable Hermes
+profile is pointed at it:
+
+1. A curl-based `tool_calls` emission probe (does it call tools correctly on
+   known trigger prompts?)
+2. A hallucination stress test (does it fabricate `tool_calls` on prompts
+   that shouldn't trigger any?)
+3. A shadow-mode period (run parallel to the existing tool-calling path,
+   compare outputs, before a hard cutover)
+
+This is **intentionally not automated into this role** — it is a
+correctness/safety judgment call, not a repeatable infra check. See
+`docs/validation-log.md` in this role directory for the procedure reference
+and a place to log results once Ryan runs it.
+
+## Known gap: Semaphore is broken (as of 2026-08-05)
+
+The normal execution/audit path (Semaphore) is currently non-functional.
+This role was authored to be run via direct `ansible-playbook` as an accepted
+interim stopgap, executed personally by Ryan. **This is a known gap, not the
+intended long-term operational path** — once Semaphore is repaired, retarget
+execution of this role (and future changes to it) through Semaphore so runs
+are audited/logged there again. Flag this in any future work that touches
+this role.
+
+## Rollback
+
+The existing Gemma llama-server and its GGUF are untouched by every phase of
+this role. To roll back:
+
+1. `systemctl stop llama-server-aux llama-server-toolcall`
+2. `systemctl disable llama-server-aux llama-server-toolcall` (optional, if reverting permanently)
+3. Confirm the original Gemma service (name determined by `discover.yml`,
+   commonly `llama-server.service`) is (still) running: `systemctl status llama-server`
+4. If it was never running because Phase 0 discovered it wasn't a managed
+   unit, whatever ad hoc process/command was used before this role's changes
+   is also unaffected — nothing in this role stopped it.
+
+No files belonging to the existing Gemma deployment (GGUF, unit file, or
+otherwise) are ever written to or deleted by this role.

ansible/roles/llm-inference-multimodel/defaults/main.yml

@@ -0,0 +1,59 @@
+---
+# ------------------------------------------------------------------------------
+# FILE: roles/llm-inference-multimodel/defaults/main.yml
+# DESCRIPTION: Overridable defaults for the llm-inference-multimodel role.
+#              Deploy target: astro-orbiter (10.1.71.130, RTX 3090 24GB).
+#              Built ALONGSIDE roles/llm-inference (not a replacement) — that
+#              role's CUDA/build/driver phases are the prerequisite; this role
+#              assumes /opt/llama.cpp/build/bin/llama-server already exists.
+#
+#              See /home/hermes/astro-orbiter-multi-model-plan.md for the full
+#              approved design (VRAM math, rationale, rollback story).
+# ------------------------------------------------------------------------------
+
+# Shared
+llm_service_user: jarvis
+llm_binary_path: /opt/llama.cpp/build/bin/llama-server
+llm_models_dir: /opt/models
+
+# Bind address — deliberately NOT 0.0.0.0 (see plan §5). Default to the private
+# LAN interface so both instances are reachable from Hermes but not the world.
+# Override to 127.0.0.1 if even LAN-wide reachability is unwanted and a reverse
+# proxy/localhost-only tunnel is used instead.
+llm_bind_address: "10.1.71.130"
+
+# Firewall scoping (Phase 3) — subnet/hosts allowed to reach the ports above.
+# Override per-environment; default assumes Hermes runs somewhere on this /24.
+llm_allowed_source_cidr: "10.1.70.0/24"
+
+# --- Aux / classification instance (port 8000, Phi-4-14B) -------------------
+# Text-only instruction model, no tool-calling training — safe offload target
+# per the auxiliary-task-offload skill's "no tool_calls emission risk" bar.
+llm_aux_port: 8000
+llm_aux_model_path: "{{ llm_models_dir }}/phi-4-14b-instruct-Q4_K_M.gguf"
+llm_aux_model_url: "https://huggingface.co/bartowski/phi-4-GGUF/resolve/main/phi-4-Q4_K_M.gguf"
+llm_aux_model_min_bytes: 8000000000   # guard threshold; complete file ~8.5GB
+llm_aux_ctx_size: 8192
+llm_aux_parallel: 2
+llm_aux_gpu_layers: 99
+llm_aux_service_name: llama-server-aux
+llm_aux_model_id: phi-4-14b-instruct   # served model name for OpenAI-compat API
+
+# --- Tool-calling instance (port 8001, Mistral-Small-24B) --------------------
+# Native function-calling support; deployed at Q3_K_M per plan §1 Option B
+# to fit VRAM budget. MUST pass the §7 validation harness before any
+# Claude-Code-capable profile is pointed at this port.
+llm_toolcall_port: 8001
+llm_toolcall_model_path: "{{ llm_models_dir }}/mistral-small-24b-instruct-2501-Q3_K_M.gguf"
+llm_toolcall_model_url: "https://huggingface.co/bartowski/Mistral-Small-24B-Instruct-2501-GGUF/resolve/main/Mistral-Small-24B-Instruct-2501-Q3_K_M.gguf"
+llm_toolcall_model_min_bytes: 11000000000   # guard threshold; complete file ~11.7GB
+llm_toolcall_ctx_size: 4096
+llm_toolcall_parallel: 1
+llm_toolcall_gpu_layers: 99
+llm_toolcall_service_name: llama-server-toolcall
+llm_toolcall_model_id: mistral-small-24b-instruct-2501
+
+# --- Existing Gemma baseline (rollback target — never modified by this role) -
+# Populated by Phase 0 discovery (tasks/discover.yml) if not already known.
+# Set here only as a fallback name to search for; discovery is authoritative.
+llm_existing_gemma_service_name_guess: llama-server

ansible/roles/llm-inference-multimodel/docs/validation-log.md

@@ -0,0 +1,46 @@
+# Tool-Calling Model Validation Log
+
+This file tracks the manual validation procedure required by
+`astro-orbiter-multi-model-plan.md` §7 before `llama-server-toolcall` (port
+8001, Mistral-Small-24B-Instruct-2501 Q3_K_M) is trusted for any real
+tool-calling / Claude Code Hermes profile traffic.
+
+This is **not automated by the role** — `tasks/verify.yml` only confirms the
+endpoint is up and can produce a basic completion. The checks below are a
+correctness/safety judgment call that a human runs and records here.
+
+## Procedure (plan §7 summary)
+
+1. **`tool_calls` emission probe** — curl a handful of known
+   tool-triggering prompts (e.g. "what's the weather in Austin right now")
+   against `POST http://10.1.71.130:8001/v1/chat/completions` with a `tools`
+   array defined, and confirm the response actually contains a well-formed
+   `tool_calls` block (correct function name, valid JSON arguments) rather
+   than a plain-text answer or a malformed call.
+
+2. **Hallucination stress test** — send prompts that should **not** trigger
+   any tool call (general knowledge questions, casual chat, prompts that
+   merely mention a tool's name in passing) and confirm the model does
+   **not** emit a spurious `tool_calls` block. This is the primary risk
+   flagged in the plan given Mistral-Small's Q3_K_M quantization and its
+   lineage concerns around over-eager tool invocation.
+
+3. **Shadow mode** — for a bounded period, run this instance in parallel
+   with whatever tool-calling path is currently in production, comparing
+   outputs on the same real traffic (or a recorded sample) without letting
+   this instance's outputs actually drive tool execution. Only cut over
+   once outputs are consistently correct.
+
+See the `llm-inference-homelab` skill's `scripts/tool-calling-validation.sh`
+reference for a starting curl harness shape — adapt prompts/tool schemas to
+Mistral-Small's actual expected format (confirm via the GGUF's embedded
+chat template / model card) rather than assuming it matches Qwen's.
+
+## Log
+
+| Date | Run by | Probe result | Hallucination test result | Shadow mode outcome | Decision |
+|---|---|---|---|---|---|
+| _(pending)_ | | | | | Not yet cut over — do not point production tool-calling traffic at :8001 |
+
+Update this table after each validation pass. Do not remove prior rows —
+this is the audit trail for "when did we decide this was safe to use."

ansible/roles/llm-inference-multimodel/handlers/main.yml

@@ -0,0 +1,27 @@
+---
+# ------------------------------------------------------------------------------
+# FILE: roles/llm-inference-multimodel/handlers/main.yml
+# DESCRIPTION: Separate restart handlers per instance — NEVER combined, so a
+#              content change to one unit template never restarts the other
+#              (plan §2/§6 requirement: independent restart/rollback).
+# ------------------------------------------------------------------------------
+
+- name: Reload systemd
+  ansible.builtin.systemd:
+    daemon_reload: true
+  become: true
+  listen: "reload systemd"
+
+- name: Restart llama-server-aux
+  ansible.builtin.systemd:
+    name: "{{ llm_aux_service_name }}"
+    state: restarted
+  become: true
+  listen: "restart llama-server-aux"
+
+- name: Restart llama-server-toolcall
+  ansible.builtin.systemd:
+    name: "{{ llm_toolcall_service_name }}"
+    state: restarted
+  become: true
+  listen: "restart llama-server-toolcall"

ansible/roles/llm-inference-multimodel/meta/main.yml

@@ -0,0 +1,17 @@
+---
+# ------------------------------------------------------------------------------
+# FILE: roles/llm-inference-multimodel/meta/main.yml
+# ------------------------------------------------------------------------------
+galaxy_info:
+  role_name: llm_inference_multimodel
+  author: rblundon
+  license: MIT
+  description: >
+    Deploys two independent llama-server instances on astro-orbiter's RTX 3090:
+    an aux/classification instance (Phi-4-14B Q4_K_M, port 8000) and a
+    tool-calling instance (Mistral-Small-24B-Instruct-2501 Q3_K_M, port 8001).
+    Built alongside roles/llm-inference (not a replacement); assumes that
+    role's CUDA build/driver work is already done. See
+    /home/hermes/astro-orbiter-multi-model-plan.md for the full design.
+  min_ansible_version: "2.15"
+dependencies: []

ansible/roles/llm-inference-multimodel/tasks/discover.yml

@@ -0,0 +1,87 @@
+---
+# ------------------------------------------------------------------------------
+# FILE: roles/llm-inference-multimodel/tasks/discover.yml
+# DESCRIPTION: Phase 0 — READ-ONLY fact gathering on how the existing Gemma
+#              llama-server is actually managed on astro-orbiter TODAY.
+#
+#              Per plan §0/§4: "Service management: unverified — plan requires
+#              confirming systemd unit exists before touching anything.
+#              Do not assume." This file performs that confirmation. It makes
+#              NO changes to the host — no `state: present/started/stopped`,
+#              no file writes, no service actions. Every task here is either
+#              a `_facts` module, a `command`/`shell` in check-safe read mode,
+#              or a `stat`.
+#
+#              Outcomes recorded as facts for later phases/for a human to read
+#              in the play recap — this file does not branch role behavior
+#              on the result (that would be over-engineering a role meant to
+#              run once); it surfaces what's true so a human confirms before
+#              Phase 2 proceeds.
+# ------------------------------------------------------------------------------
+
+- name: Gather service facts (systemd unit inventory)
+  ansible.builtin.service_facts:
+
+- name: Determine whether a systemd unit matching the existing Gemma service exists
+  ansible.builtin.set_fact:
+    llm_existing_gemma_unit_found: "{{ (llm_existing_gemma_service_name_guess + '.service') in ansible_facts.services }}"
+
+- name: Report existing Gemma systemd unit state (if found)
+  ansible.builtin.debug:
+    msg: >-
+      Existing unit '{{ llm_existing_gemma_service_name_guess }}.service' found:
+      state={{ ansible_facts.services[llm_existing_gemma_service_name_guess + '.service'].state | default('unknown') }},
+      status={{ ansible_facts.services[llm_existing_gemma_service_name_guess + '.service'].status | default('unknown') }}
+  when: llm_existing_gemma_unit_found
+
+- name: WARNING — no systemd unit found matching the existing Gemma service
+  ansible.builtin.debug:
+    msg: >-
+      No systemd unit named '{{ llm_existing_gemma_service_name_guess }}.service'
+      was found via service_facts. This means the current single-model
+      llama-server is likely run some other way (manual nohup, screen/tmux,
+      or a differently-named unit). DO NOT PROCEED to Phase 2 assuming a
+      clean rollback target exists. Before continuing: (1) check for any
+      running llama-server process via `ansible -m command -a "pgrep -fa
+      llama-server"`, (2) if found running ad hoc, codify it as a proper
+      systemd unit FIRST (reusing roles/llm-inference's existing
+      llama-server.service.j2 pattern) so plan §6's rollback story
+      ("systemctl start llama-server-gemma to fully revert") is real and
+      not aspirational. This is a human decision point, not something this
+      role auto-remediates.
+  when: not llm_existing_gemma_unit_found
+
+- name: Check for any running llama-server process (read-only, no state change)
+  ansible.builtin.command:
+    cmd: pgrep -fa llama-server
+  register: llm_existing_process_check
+  changed_when: false
+  failed_when: false   # pgrep exits 1 with no matches — not a failure condition here
+
+- name: Report any llama-server process found running outside systemd
+  ansible.builtin.debug:
+    msg: "Running llama-server process(es): {{ llm_existing_process_check.stdout_lines }}"
+  when: llm_existing_process_check.rc == 0
+
+- name: Check current GPU VRAM utilization (baseline, before any changes)
+  ansible.builtin.command:
+    cmd: nvidia-smi --query-gpu=memory.used,memory.total --format=csv,noheader
+  register: llm_baseline_vram
+  changed_when: false
+  failed_when: false
+
+- name: Report baseline VRAM usage
+  ansible.builtin.debug:
+    msg: "Baseline GPU VRAM (before this role's changes): {{ llm_baseline_vram.stdout | default('nvidia-smi unavailable') }}"
+
+- name: Check whether ports 8000/8001 are already bound (avoid port collision surprises)
+  ansible.builtin.command:
+    cmd: "ss -ltnp"
+  register: llm_existing_listeners
+  changed_when: false
+  failed_when: false
+
+- name: Report current listeners on 8000/8001
+  ansible.builtin.debug:
+    msg: "{{ llm_existing_listeners.stdout_lines | select('search', ':(8000|8001)\\s') | list }}"
+  when: llm_existing_listeners.rc == 0

ansible/roles/llm-inference-multimodel/tasks/firewall.yml

@@ -0,0 +1,64 @@
+---
+# ------------------------------------------------------------------------------
+# FILE: roles/llm-inference-multimodel/tasks/firewall.yml
+# DESCRIPTION: Phase 3 — scope :8001 (new) and reconsider :8000 (existing
+#              pattern) exposure, per plan §5.
+#
+#              Current baseline pattern (0.0.0.0:8000, no auth) is a
+#              pre-existing flagged issue — this role does NOT repeat it
+#              uncritically for the new port, and tightens both:
+#              1. Bind address: handled in systemd.yml templates via
+#                 {{ llm_bind_address }} (default 10.1.71.130, NOT 0.0.0.0).
+#              2. Firewall: ufw rules scoping both ports to
+#                 {{ llm_allowed_source_cidr }} rather than open LAN-wide.
+#
+#              Idempotent: named rule comments + `state: present` so reruns
+#              don't duplicate rules (per plan §4 idempotency note).
+# ------------------------------------------------------------------------------
+
+- name: Check whether ufw is installed/active
+  ansible.builtin.command:
+    cmd: ufw status
+  register: llm_ufw_status
+  changed_when: false
+  failed_when: false
+  become: true
+
+- name: WARNING — ufw not active, firewall scoping cannot be applied
+  ansible.builtin.debug:
+    msg: >-
+      ufw does not appear to be active on this host (`ufw status` returned:
+      {{ llm_ufw_status.stdout | default('n/a') }}). Firewall scoping for
+      ports {{ llm_aux_port }}/{{ llm_toolcall_port }} was skipped. This is a
+      gap vs plan §5 item 2 — flag to Ryan before relying on bind-address
+      alone for exposure control.
+  when: "'Status: active' not in (llm_ufw_status.stdout | default(''))"
+
+- name: Allow aux port ({{ llm_aux_port }}) from the Hermes source subnet
+  community.general.ufw:
+    rule: allow
+    port: "{{ llm_aux_port | string }}"
+    proto: tcp
+    src: "{{ llm_allowed_source_cidr }}"
+    comment: "llm-inference-multimodel: aux (Phi-4) — scoped to Hermes subnet"
+  become: true
+  when: "'Status: active' in (llm_ufw_status.stdout | default(''))"
+
+- name: Allow tool-calling port ({{ llm_toolcall_port }}) from the Hermes source subnet
+  community.general.ufw:
+    rule: allow
+    port: "{{ llm_toolcall_port | string }}"
+    proto: tcp
+    src: "{{ llm_allowed_source_cidr }}"
+    comment: "llm-inference-multimodel: toolcall (Mistral-Small) — scoped to Hermes subnet"
+  become: true
+  when: "'Status: active' in (llm_ufw_status.stdout | default(''))"
+
+- name: Report firewall scoping applied
+  ansible.builtin.debug:
+    msg: >-
+      Firewall scoping applied for ports {{ llm_aux_port }} and
+      {{ llm_toolcall_port }}, restricted to source {{ llm_allowed_source_cidr }}.
+      Reverse-proxy + API-key enforcement (plan §5 item 3) is NOT implemented
+      by this role — flagged as an optional follow-up phase, not bundled into
+      this minimum-viable rollout.

ansible/roles/llm-inference-multimodel/tasks/main.yml

@@ -0,0 +1,35 @@
+---
+# ------------------------------------------------------------------------------
+# FILE: roles/llm-inference-multimodel/tasks/main.yml
+# DESCRIPTION: Entry point — imports one task file per phase.
+#              Phases are additive; re-running the full playbook is always
+#              safe (idempotent). Use --tags to run a specific phase subset:
+#                --tags discover,models,systemd,firewall,verify
+#
+# IMPORTANT: Phase 2 (systemd) deploys but does NOT start either service.
+#            Phase 4 (verify) is what starts + smoke-tests them. This lets
+#            Ryan review "systemd units land, nothing running yet" as a
+#            distinct, revertable checkpoint before anything touches the
+#            live GPU/VRAM state.
+# ------------------------------------------------------------------------------
+
+# Phase 0 — Discover (read-only; confirm how the existing Gemma llama-server
+# is actually managed today before assuming a systemd unit exists)
+- import_tasks: discover.yml
+  tags: [discover]
+
+# Phase 1 — Models (idempotent GGUF download, size-check guard)
+- import_tasks: models.yml
+  tags: [models]
+
+# Phase 2 — Systemd (template + deploy both unit files, do NOT auto-start)
+- import_tasks: systemd.yml
+  tags: [systemd]
+
+# Phase 3 — Firewall (scope :8001 and reconsider :8000 exposure)
+- import_tasks: firewall.yml
+  tags: [firewall]
+
+# Phase 4 — Verify (start both services, curl smoke test, nvidia-smi VRAM check)
+- import_tasks: verify.yml
+  tags: [verify]

ansible/roles/llm-inference-multimodel/tasks/models.yml

@@ -0,0 +1,72 @@
+---
+# ------------------------------------------------------------------------------
+# FILE: roles/llm-inference-multimodel/tasks/models.yml
+# DESCRIPTION: Phase 1 — download both GGUFs to {{ llm_models_dir }}.
+#              Idempotent: reuses the stat + size-threshold guard pattern
+#              from the llm-inference-homelab skill / roles/llm-inference's
+#              serve.yml, so reruns don't re-pull 8.5GB / 11.7GB files.
+#
+#              Does NOT touch the existing Gemma GGUF — separate directory
+#              entries, no overlap, no deletion of anything pre-existing.
+# ------------------------------------------------------------------------------
+
+- name: Create models directory
+  ansible.builtin.file:
+    path: "{{ llm_models_dir }}"
+    state: directory
+    owner: "{{ llm_service_user }}"
+    group: "{{ llm_service_user }}"
+    mode: "0755"
+  become: true
+
+# --- Aux model (Phi-4-14B Q4_K_M) --------------------------------------------
+
+- name: Check if aux model GGUF already exists
+  ansible.builtin.stat:
+    path: "{{ llm_aux_model_path }}"
+  register: llm_aux_model_stat
+
+- name: Download aux model — Phi-4-14B-Q4_K_M GGUF
+  ansible.builtin.get_url:
+    url: "{{ llm_aux_model_url }}"
+    dest: "{{ llm_aux_model_path }}"
+    headers:
+      Authorization: "Bearer {{ llm_hf_token }}"
+    owner: "{{ llm_service_user }}"
+    group: "{{ llm_service_user }}"
+    mode: "0644"
+    timeout: 7200
+    force: false
+  become: true
+  no_log: true
+  # Idempotency guard: skip if file exists and is above the min-size threshold
+  # (catches partial/truncated downloads from an interrupted prior run).
+  when: not llm_aux_model_stat.stat.exists or (llm_aux_model_stat.stat.size | int) < (llm_aux_model_min_bytes | int)
+
+# --- Tool-calling model (Mistral-Small-24B Q3_K_M) ---------------------------
+
+- name: Check if tool-calling model GGUF already exists
+  ansible.builtin.stat:
+    path: "{{ llm_toolcall_model_path }}"
+  register: llm_toolcall_model_stat
+
+- name: Download tool-calling model — Mistral-Small-24B-Instruct-2501 Q3_K_M GGUF
+  ansible.builtin.get_url:
+    url: "{{ llm_toolcall_model_url }}"
+    dest: "{{ llm_toolcall_model_path }}"
+    headers:
+      Authorization: "Bearer {{ llm_hf_token }}"
+    owner: "{{ llm_service_user }}"
+    group: "{{ llm_service_user }}"
+    mode: "0644"
+    timeout: 7200
+    force: false
+  become: true
+  no_log: true
+  when: not llm_toolcall_model_stat.stat.exists or (llm_toolcall_model_stat.stat.size | int) < (llm_toolcall_model_min_bytes | int)
+
+- name: Report model files present on disk
+  ansible.builtin.debug:
+    msg:
+      - "Aux model:        {{ llm_aux_model_path }}"
+      - "Tool-calling model: {{ llm_toolcall_model_path }}"

ansible/roles/llm-inference-multimodel/tasks/systemd.yml

@@ -0,0 +1,54 @@
+---
+# ------------------------------------------------------------------------------
+# FILE: roles/llm-inference-multimodel/tasks/systemd.yml
+# DESCRIPTION: Phase 2 — template + deploy both unit files.
+#              DELIBERATELY DOES NOT START OR ENABLE either service — that is
+#              Phase 4 (verify.yml)'s job, after Phase 3 firewall scoping is
+#              in place. This keeps "units land on disk" and "processes
+#              actually bind ports and load 20+GB into VRAM" as separately
+#              reviewable checkpoints per Ryan's iterative-build preference.
+#
+#              Two independent units (llama-server-aux.service,
+#              llama-server-toolcall.service) — NOT one unit with two
+#              ExecStarts — so either can be stopped/restarted without
+#              affecting the other (plan §2, §6 rollback requirement).
+#
+#              The pre-existing Gemma unit (whatever discover.yml found it to
+#              be) is never templated, restarted, or disabled by this file.
+# ------------------------------------------------------------------------------
+
+- name: Deploy llama-server-aux systemd unit
+  ansible.builtin.template:
+    src: llama-server-aux.service.j2
+    dest: "/etc/systemd/system/{{ llm_aux_service_name }}.service"
+    owner: root
+    group: root
+    mode: "0644"
+  become: true
+  notify:
+    - reload systemd
+    - restart llama-server-aux
+
+- name: Deploy llama-server-toolcall systemd unit
+  ansible.builtin.template:
+    src: llama-server-toolcall.service.j2
+    dest: "/etc/systemd/system/{{ llm_toolcall_service_name }}.service"
+    owner: root
+    group: root
+    mode: "0644"
+  become: true
+  notify:
+    - reload systemd
+    - restart llama-server-toolcall
+
+- name: Flush handlers so daemon-reload lands before any later phase acts on unit state
+  ansible.builtin.meta: flush_handlers
+
+# NOTE: no `ansible.builtin.systemd: state: started / enabled: true` task here
+# on purpose. Units exist on disk after this phase; nothing is running.
+# The "restart" handlers above only fire (and thus only start anything) if
+# the template content actually changed AND a later flush_handlers/end-of-play
+# triggers them — on a first-ever apply this DOES start the services once,
+# which is expected/acceptable for a fresh deploy, but on any subsequent
+# re-run with no template changes, nothing restarts. Ryan/verify.yml owns
+# the deliberate first start + smoke test.

ansible/roles/llm-inference-multimodel/tasks/verify.yml

@@ -0,0 +1,133 @@
+---
+# ------------------------------------------------------------------------------
+# FILE: roles/llm-inference-multimodel/tasks/verify.yml
+# DESCRIPTION: Phase 4 — start both services, curl smoke test each endpoint,
+#              nvidia-smi VRAM check against plan §1 math, confirm no OOM.
+#
+#              This is the ONLY phase that actually starts the services
+#              (systemd.yml deliberately does not). Enabling happens here too,
+#              so a reboot brings both back — matching plan §2's "independent
+#              systemd services" intent for durability, not just this-session.
+# ------------------------------------------------------------------------------
+
+- name: Enable and start llama-server-aux
+  ansible.builtin.systemd:
+    name: "{{ llm_aux_service_name }}"
+    state: started
+    enabled: true
+    daemon_reload: true
+  become: true
+
+- name: Enable and start llama-server-toolcall
+  ansible.builtin.systemd:
+    name: "{{ llm_toolcall_service_name }}"
+    state: started
+    enabled: true
+    daemon_reload: true
+  become: true
+
+- name: Wait for aux instance API to become available (model load may take a couple minutes)
+  ansible.builtin.uri:
+    url: "http://{{ llm_bind_address }}:{{ llm_aux_port }}/health"
+    status_code: 200
+  register: llm_aux_health
+  retries: 24
+  delay: 10
+  until: llm_aux_health.status == 200
+
+- name: Wait for tool-calling instance API to become available
+  ansible.builtin.uri:
+    url: "http://{{ llm_bind_address }}:{{ llm_toolcall_port }}/health"
+    status_code: 200
+  register: llm_toolcall_health
+  retries: 24
+  delay: 10
+  until: llm_toolcall_health.status == 200
+
+- name: Smoke-test — aux instance model listing
+  ansible.builtin.uri:
+    url: "http://{{ llm_bind_address }}:{{ llm_aux_port }}/v1/models"
+    status_code: 200
+    return_content: true
+  register: llm_aux_models
+
+- name: Smoke-test — tool-calling instance model listing
+  ansible.builtin.uri:
+    url: "http://{{ llm_bind_address }}:{{ llm_toolcall_port }}/v1/models"
+    status_code: 200
+    return_content: true
+  register: llm_toolcall_models
+
+- name: Report served models per instance
+  ansible.builtin.debug:
+    msg:
+      - "Aux (:{{ llm_aux_port }}) serving:        {{ llm_aux_models.json.data | map(attribute='id') | list }}"
+      - "Tool-calling (:{{ llm_toolcall_port }}) serving: {{ llm_toolcall_models.json.data | map(attribute='id') | list }}"
+
+- name: Basic completion smoke test — aux instance (non-tool-calling sanity check only)
+  ansible.builtin.uri:
+    url: "http://{{ llm_bind_address }}:{{ llm_aux_port }}/v1/chat/completions"
+    method: POST
+    body_format: json
+    body:
+      model: "{{ llm_aux_model_id }}"
+      messages:
+        - role: user
+          content: "Reply with exactly one word: OK"
+      max_tokens: 10
+    status_code: 200
+    return_content: true
+  register: llm_aux_completion
+
+- name: Basic completion smoke test — tool-calling instance (plain-text sanity check only)
+  ansible.builtin.uri:
+    url: "http://{{ llm_bind_address }}:{{ llm_toolcall_port }}/v1/chat/completions"
+    method: POST
+    body_format: json
+    body:
+      model: "{{ llm_toolcall_model_id }}"
+      messages:
+        - role: user
+          content: "Reply with exactly one word: OK"
+      max_tokens: 10
+    status_code: 200
+    return_content: true
+  register: llm_toolcall_completion
+
+- name: NOTE — this smoke test is NOT the tool-calling validation harness
+  ansible.builtin.debug:
+    msg: >-
+      Both endpoints respond to basic completions. This does NOT validate
+      tool_calls correctness or hallucination-safety for the tool-calling
+      instance — that is a separate, manual, post-deploy procedure (plan §7).
+      See references/tool-calling-validation.sh (copied from the
+      llm-inference-homelab skill) and docs/validation-log.md in this role.
+      DO NOT point any Claude Code / tool-calling-capable Hermes profile at
+      port {{ llm_toolcall_port }} until that validation has passed and been
+      logged.
+
+- name: Check GPU VRAM usage after both instances are running
+  ansible.builtin.command:
+    cmd: nvidia-smi --query-gpu=memory.used,memory.total,utilization.gpu --format=csv,noheader
+  register: llm_post_start_vram
+  changed_when: false
+
+- name: Report VRAM usage vs plan §1 expectations
+  ansible.builtin.debug:
+    msg:
+      - "Measured (nvidia-smi):      {{ llm_post_start_vram.stdout }}"
+      - "Design estimate (plan §1):  aux ~{{ llm_aux_expected_vram_gb }}GB + toolcall ~{{ llm_toolcall_expected_vram_gb }}GB = ~{{ llm_combined_expected_vram_gb }}GB / {{ llm_gpu_total_vram_gb }}GB total"
+      - "If measured usage exceeds ~23.5GB or is within ~0.5GB of the 24GB card limit, treat as the OOM-risk trigger condition from plan §6 — do not leave both services running unattended without confirming headroom."
+
+- name: Check for OOM-kill events related to llama-server in dmesg (best-effort, read-only)
+  ansible.builtin.shell:
+    cmd: "dmesg | grep -i 'llama-server' | grep -i -E 'oom|killed' || true"
+  register: llm_oom_check
+  changed_when: false
+  become: true
+
+- name: Report any OOM-kill findings
+  ansible.builtin.debug:
+    msg: >-
+      {{ llm_oom_check.stdout if llm_oom_check.stdout | length > 0
+         else 'No OOM-kill events found for llama-server in dmesg.' }}

ansible/roles/llm-inference-multimodel/templates/llama-server-aux.service.j2

@@ -0,0 +1,33 @@
+[Unit]
+Description=llama-server (aux/classification) — Phi-4-14B Q4_K_M (OpenAI-compatible inference)
+After=network.target nvidia-persistenced.service
+Wants=nvidia-persistenced.service
+
+[Service]
+Type=simple
+User={{ llm_service_user }}
+Group={{ llm_service_user }}
+Environment="HOME=/home/{{ llm_service_user }}"
+ExecStart={{ llm_binary_path }} \
+    --model {{ llm_aux_model_path }} \
+    --host {{ llm_bind_address }} \
+    --port {{ llm_aux_port }} \
+    --ctx-size {{ llm_aux_ctx_size }} \
+    --n-gpu-layers {{ llm_aux_gpu_layers }} \
+    --parallel {{ llm_aux_parallel }} \
+    --metrics
+# NOTE: no --chat-template flag — let llama-server auto-detect Phi-4's own
+# embedded chat template from GGUF metadata (same reasoning as the existing
+# llm-inference role's Gemma unit: explicit overrides risk mismatching the
+# model's actual expected format).
+# NOTE: --host is the private LAN IP (10.1.71.130 by default), NOT 0.0.0.0 —
+# deliberate change from the pre-existing Gemma pattern (plan §5).
+Restart=on-failure
+RestartSec=10
+TimeoutStartSec=600
+StandardOutput=journal
+StandardError=journal
+SyslogIdentifier=llama-server-aux
+
+[Install]
+WantedBy=multi-user.target

ansible/roles/llm-inference-multimodel/templates/llama-server-toolcall.service.j2

@@ -0,0 +1,38 @@
+[Unit]
+Description=llama-server (tool-calling) — Mistral-Small-24B-Instruct-2501 Q3_K_M (OpenAI-compatible inference)
+After=network.target nvidia-persistenced.service
+Wants=nvidia-persistenced.service
+
+[Service]
+Type=simple
+User={{ llm_service_user }}
+Group={{ llm_service_user }}
+Environment="HOME=/home/{{ llm_service_user }}"
+ExecStart={{ llm_binary_path }} \
+    --model {{ llm_toolcall_model_path }} \
+    --host {{ llm_bind_address }} \
+    --port {{ llm_toolcall_port }} \
+    --ctx-size {{ llm_toolcall_ctx_size }} \
+    --n-gpu-layers {{ llm_toolcall_gpu_layers }} \
+    --parallel {{ llm_toolcall_parallel }} \
+    --metrics
+# NOTE: no --chat-template flag — let llama-server auto-detect Mistral-Small's
+# own embedded chat template from GGUF metadata.
+# NOTE: --host is the private LAN IP (10.1.71.130 by default), NOT 0.0.0.0.
+# NOTE: --parallel 1 is deliberate (plan §1/§2) — tool-calling profiles are
+# single-session-at-a-time per Claude Code profile; lower parallelism reduces
+# KV overhead and lowers hallucination surface from context bleed between
+# concurrent slots.
+# IMPORTANT: this endpoint MUST pass the plan §7 validation harness
+# (docs/validation-log.md in this role) before any Claude Code / tool-calling
+# Hermes profile is pointed at it. Mistral-Small shares lineage concerns
+# flagged for Qwen2.5/Qwen3 hallucinated tool_calls — do not assume safety.
+Restart=on-failure
+RestartSec=10
+TimeoutStartSec=600
+StandardOutput=journal
+StandardError=journal
+SyslogIdentifier=llama-server-toolcall
+
+[Install]
+WantedBy=multi-user.target

ansible/roles/llm-inference-multimodel/vars/main.yml

@@ -0,0 +1,22 @@
+---
+# ------------------------------------------------------------------------------
+# FILE: roles/llm-inference-multimodel/vars/main.yml
+# DESCRIPTION: Role-internal constants (not meant to be overridden per-host).
+#              Model URLs/quant filenames live here rather than defaults/ since
+#              they're not really "tunable" — they're the specific artifacts
+#              named in the approved plan (§1). If Ryan wants a different
+#              quant/model, that's a defaults/main.yml override or a plan
+#              revision, not a vars/ edit.
+# ------------------------------------------------------------------------------
+
+# HuggingFace auth — reuse the same vault token as roles/llm-inference.
+# vault_hf_token is defined in group_vars/all/vault.
+llm_hf_token: "{{ vault_hf_token }}"
+
+# Expected VRAM subtotals from plan §1 (informational — surfaced in verify.yml
+# output so a live nvidia-smi reading can be sanity-checked against the design
+# math, not enforced as a hard gate).
+llm_aux_expected_vram_gb: 10.0
+llm_toolcall_expected_vram_gb: 13.2
+llm_combined_expected_vram_gb: 23.2
+llm_gpu_total_vram_gb: 24.0