[Unit]
Description=llama-server — Gemma 2 27B-it Q4_K_M (OpenAI-compatible inference)
After=network.target nvidia-persistenced.service
Wants=nvidia-persistenced.service

[Service]
Type=simple
User={{ llm_venv_owner }}
Group={{ llm_venv_owner }}
Environment="HOME=/home/{{ llm_venv_owner }}"
ExecStart=/opt/llama.cpp/build/bin/llama-server \
    --model {{ llm_gguf_path }} \
    --host {{ llm_serve_host }} \
    --port {{ llm_serve_port }} \
    --ctx-size {{ llm_max_model_len }} \
    --n-gpu-layers {{ llm_gpu_layers }} \
    --parallel {{ llm_parallel_slots }} \
    --metrics
# NOTE: no --chat-template flag — llama-server auto-detects and uses the
# GGUF's own embedded Jinja chat template (verified correct Gemma-2
# start_of_turn/end_of_turn format for bartowski's gemma-2-27b-it-Q4_K_M).
# The built-in "--chat-template gemma" name does NOT match this model's
# expected format on this llama.cpp build and produced garbled completions.
Restart=on-failure
RestartSec=10
TimeoutStartSec=120
StandardOutput=journal
StandardError=journal
SyslogIdentifier=llama-server

[Install]
WantedBy=multi-user.target