monitoring: llama-swap GPU/LLM stack (v250) — PrometheusRule, Grafana dashboard, scrape config, VRAM exporter

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

@@ -0,0 +1,53 @@
+{#
+  FILE: roles/llm-inference-multimodel/templates/llama-swap.service.j2
+  DESCRIPTION: llama-swap systemd unit template.
+               Single Go binary, no subprocess management — just a /usr/local/bin/llama-swap
+               process reading /etc/llama-swap/config.yaml.
+  
+  Design:
+  - Type=simple (no forking)
+  - User={{ llm_swapmode_service_user }} (jarvis)
+  - Restart=on-failure, RestartSec=10
+  - Logs to journald (StandardOutput/StandardError=journal)
+  - After nvidia-persistenced.service (NVIDIA driver dependency)
+  
+  Config location: /etc/llama-swap/config.yaml (rendered by swapmode_config phase)
+  Listen address: 127.0.0.1 inside the container (exposed by --listen flag)
+#}
+[Unit]
+Description=llama-swap — hot-swap model proxy (port {{ llm_swapmode_port }})
+Documentation=https://github.com/mostlygeek/llama-swap
+After=network.target nvidia-persistenced.service
+Wants=nvidia-persistenced.service
+
+[Service]
+Type=simple
+User={{ llm_swapmode_service_user }}
+Group={{ llm_swapmode_service_user }}
+Environment="HOME=/home/{{ llm_swapmode_service_user }}"
+
+ExecStart=/usr/local/bin/llama-swap \
+    --config {{ llm_swapmode_config_file }} \
+    --listen {{ llm_swapmode_bind_address }}:{{ llm_swapmode_port }}
+
+# LLAMA-SWAP NOTES (2026-08-18, t_c1e44190):
+# - Single Go binary, zero runtime dependencies (llama.cpp statically linked).
+# - Upstream servers (llama-server instances) are spawned on-demand per config.yaml model definitions.
+# - --listen can override config.yaml's listen key; this flag takes precedence.
+#   Double-check consistency between ExecStart and config.yaml.
+# - CUDA_VISIBLE_DEVICES can be set via Environment= if GPU isolation is needed.
+#   Default: inherit from parent (systemd likely has it unset, picks all GPUs).
+# - No jinja flag needed: llama.cpp model templates are embedded in each model's GGUF.
+
+Restart=on-failure
+RestartSec=10
+TimeoutStartSec=600
+StandardOutput=journal
+StandardError=journal
+SyslogIdentifier=llama-swap
+
+# Resource limits (optional; adjust per VRAM budget)
+# MemoryMax=24G  # Enforce hard limit; uncomment if runaway is a concern
+
+[Install]
+WantedBy=multi-user.target