pve-cluster: when-Listen in Anführungszeichen (YAML-Flow-Sequenz)

platform/proxmox/ansible/roles/infra-stack/tasks/targets.yml

@@ -1,15 +1,19 @@
-# file_sd-Ziele der GPU-VMs. Auch einzeln aufrufbar: playbooks/monitoring-targets.yml
+# file_sd-Ziele der GPU-VMs (node, gpu, llamacpp). Auch einzeln aufrufbar: playbooks/monitoring-targets.yml
 # (nach refresh-vm-access.sh, wenn sich DHCP-Adressen geändert haben). Prometheus liest die Dateien selbst neu.
 - name: file_sd-Ziele aus tofu output erzeugen (auf ops)
   ansible.builtin.shell: |
     set -euo pipefail
     tofu output -json gpu_vms | python3 {{ infra_repo_src }}/platform/proxmox/scripts/monitoring-targets.py \
-      --registry {{ infra_repo_src }}/sites/{{ knaster_site }}/registry.yaml --out {{ infra_targets_dir }}
+      --registry {{ infra_repo_src }}/sites/{{ knaster_site }}/registry.yaml --out {{ infra_targets_dir }} \
+      --llamacpp '{{ dict(_llamacpp_hosts | zip(_llamacpp_hosts | map('extract', hostvars, 'llamacpp_profiles'))) | to_json }}'
   args:
     chdir: "{{ infra_tofu_dir }}"
     executable: /bin/bash
   delegate_to: localhost
   run_once: true
+  vars:
+    # Profile je VM aus dem Inventar (Gruppe gpu_llamacpp, Task 6.1)
+    _llamacpp_hosts: "{{ groups['gpu_llamacpp'] | default([]) }}"
   changed_when: false
 
 - name: file_sd-Verzeichnis
@@ -23,4 +27,4 @@
     src: "{{ infra_targets_dir }}/{{ item }}.json"
     dest: "{{ infra_state_dir }}/file_sd/{{ item }}.json"
     mode: "0644"
-  loop: [node, gpu]
+  loop: [node, gpu, llamacpp]

platform/proxmox/ansible/roles/pve-cluster/tasks/main.yml

@@ -38,7 +38,7 @@
 
 - name: Admin-Gruppe darf alles (Administrator auf /)
   ansible.builtin.command: pveum acl modify / --groups {{ knaster_pve.access.admin_group }} --roles Administrator
-  when: ['/', '@' ~ knaster_pve.access.admin_group, 'Administrator'] not in pve_access.acl
+  when: "['/', knaster_pve.access.admin_group, 'Administrator'] not in pve_access.acl"
   changed_when: true
 
 # Ohne Passwort: das setzt scripts/pve-admin-user.sh vom Laptop (zusammen mit TOTP und Recovery-Keys)
@@ -57,7 +57,7 @@
 
 - name: Healer darf GPU-VMs im Pool gpu schalten
   ansible.builtin.command: pveum acl modify /pool/gpu --users {{ pve_healer_user }} --roles {{ pve_healer_role }}
-  when: ['/pool/gpu', pve_healer_user, pve_healer_role] not in pve_access.acl
+  when: "['/pool/gpu', pve_healer_user, pve_healer_role] not in pve_access.acl"
   changed_when: true
 
 - name: Healer-Token lesen

platform/proxmox/scripts/monitoring-targets.py

@@ -6,11 +6,18 @@ Eingabe ist `tofu output -json gpu_vms` (stdin), dazu die Registry für die Slot
 
     tofu output -json gpu_vms | python3 monitoring-targets.py --registry sites/knaster/registry.yaml --out DIR
 
-Schreibt DIR/node.json (node_exporter, Port 9100) und DIR/gpu.json (nvidia_gpu_exporter, Port 9835).
+Schreibt DIR/node.json (node_exporter, Port 9100), DIR/gpu.json (nvidia_gpu_exporter, Port 9835) und
+DIR/llamacpp.json (llama-server /metrics je Dienst, Task 6.1).
 
 Labels je Ziel: instance (VM-Name, bleibt bei IP-Wechsel gleich), vm, pve_node, vmid, slot, slots.
 `slot` ist der erste Slot der Registry mit dieser vmid (Reihenfolge der Registry), `slots` alle,
 kommagetrennt. VMs ohne Slot bekommen slot="none" (der Healer markiert dann keinen echten Slot).
+
+llama.cpp: Welche VM welches Profil fährt, steht im Ansible-Inventar (Gruppe gpu_llamacpp), nicht in der
+Registry (die kennt je Slot nur eine VM, die zweite Embedding-Instanz fehlt dort). Übergabe als JSON:
+`--llamacpp '{"gpu-pve44": ["embed"], ...}'`. Die Slots eines Dienstes kommen aus der Registry
+(engine llamacpp, gleicher Container). Labels zusätzlich docker_host (= VM) und container, damit der
+Healer die tatsächliche VM trifft. Diese Slots zählen auch für node/gpu der VM.
 """
 
 from __future__ import annotations
@@ -23,6 +30,11 @@ from pathlib import Path
 import yaml
 
 PORTS = {"node": 9100, "gpu": 9835}
+# Dienste je Profil wie in deploy/compose/gpu-llamacpp (feste Ports, Host-Netz)
+LLAMACPP_PROFILES = {
+    "embed": [("llamacpp-embed", 8081), ("llamacpp-rerank", 8082)],
+    "chat": [("llamacpp-chat", 8080)],
+}
 
 
 def slots_by_vmid(registry: dict) -> dict[str, list[str]]:
@@ -34,16 +46,31 @@ def slots_by_vmid(registry: dict) -> dict[str, list[str]]:
     return out
 
 
-def build_targets(gpu_vms: dict, registry: dict) -> dict[str, list[dict]]:
+def llamacpp_slots_by_container(registry: dict) -> dict[str, list[str]]:
+    out: dict[str, list[str]] = {}
+    for name, slot in (registry.get("slots") or {}).items():
+        container = (slot.get("ops") or {}).get("container")
+        if slot.get("engine") == "llamacpp" and container:
+            out.setdefault(container, []).append(name)
+    return out
+
+
+def build_targets(gpu_vms: dict, registry: dict,
+                  llamacpp: dict[str, list[str]] | None = None) -> dict[str, list[dict]]:
     by_vmid = slots_by_vmid(registry)
-    result: dict[str, list[dict]] = {job: [] for job in PORTS}
+    by_container = llamacpp_slots_by_container(registry)
+    llamacpp = llamacpp or {}
+    result: dict[str, list[dict]] = {job: [] for job in [*PORTS, "llamacpp"]}
     for name in sorted(gpu_vms):
         vm = gpu_vms[name]
         if not vm.get("ipv4"):
             print(f"Warnung: {name} hat keine IPv4 (Gast-Agent?), übersprungen", file=sys.stderr)
             continue
         vmid = str(vm["vmid"])
-        slots = by_vmid.get(vmid, [])
+        services = [svc for profile in llamacpp.get(name, []) for svc in LLAMACPP_PROFILES[profile]]
+        slots = list(by_vmid.get(vmid, []))
+        for container, _ in services:
+            slots += [s for s in by_container.get(container, []) if s not in slots]
         labels = {
             "instance": name,
             "vm": name,
@@ -54,6 +81,11 @@ def build_targets(gpu_vms: dict, registry: dict) -> dict[str, list[dict]]:
         }
         for job, port in PORTS.items():
             result[job].append({"targets": [f"{vm['ipv4']}:{port}"], "labels": labels})
+        for container, port in services:
+            cslots = by_container.get(container, [])
+            result["llamacpp"].append({"targets": [f"{vm['ipv4']}:{port}"], "labels": {
+                **labels, "docker_host": name, "container": container,
+                "slot": cslots[0] if cslots else "none", "slots": ",".join(cslots)}})
     return result
 
 
@@ -61,11 +93,16 @@ def main() -> int:
     ap = argparse.ArgumentParser(description=__doc__.splitlines()[0])
     ap.add_argument("--registry", required=True, type=Path)
     ap.add_argument("--out", required=True, type=Path)
+    ap.add_argument("--llamacpp", default="{}", help='JSON {"<vm>": ["embed"|"chat", ...]} aus dem Inventar')
     args = ap.parse_args()
 
     gpu_vms = json.load(sys.stdin)
     registry = yaml.safe_load(args.registry.read_text(encoding="utf-8"))
-    targets = build_targets(gpu_vms, registry)
+    llamacpp = json.loads(args.llamacpp)
+    unknown = {p for profiles in llamacpp.values() for p in profiles} - set(LLAMACPP_PROFILES)
+    if unknown:
+        ap.error(f"unbekannte llama.cpp-Profile: {sorted(unknown)}")
+    targets = build_targets(gpu_vms, registry, llamacpp)
     args.out.mkdir(parents=True, exist_ok=True)
     for job, entries in targets.items():
         path = args.out / f"{job}.json"