Sanitized systemd units for a multi-model llama.cpp fleet on Strix Halo. These are the live configs from the reference box with deployment-specific details removed: generic unit names, placeholder paths, no API keys (units bind 127.0.0.1; put auth in front with a proxy). Placeholders to replace: /opt/llama/bin/llama-server-rocm - your ROCm build (docs/BUILD.md) /opt/llama/bin/llama-server - CPU-only build (no HIP needed) /opt/models/... - your model paths /etc/llama-fleet.env - optional shared env file (created by you) /var/lib/llama/hotcache// - slot save path (create, chown llama) User/Group - a dedicated service user is good hygiene; needs render + video groups for /dev/kfd The MemoryMax=/MemoryLow= budgets are the actual fleet values. They are load-bearing on a unified-memory APU - see docs/PITFALLS.md before changing them.