[Unit]
Description=FastFlowLM NPU runtime (OpenAI-compatible server)
Documentation=https://github.com/ROCm/FastFlowLM
After=network.target
Wants=network.target

[Service]
Type=simple
User=fastflowlm
Group=fastflowlm
# Shared model store (FLM_MODEL_PATH), group writable for other flm users.
StateDirectory=fastflowlm
StateDirectoryMode=2770
Environment="FLM_MODEL_PATH=/var/lib/fastflowlm"
EnvironmentFile=-/etc/default/fastflowlm
Environment="FLM_MODEL=qwen3:0.6b"
Environment="FLM_HOST=127.0.0.1"
Environment="FLM_PORT="
Environment="FLM_PMODE=performance"
Environment="FLM_ASR=1"
Environment="FLM_EMBED=1"
Environment="FLM_EXTRA_OPTS="
# FLM mlocks NPU buffers: memlock must be unlimited.
LimitMEMLOCK=infinity
# The conditional --port flag and positional model argument need shell
# parameter expansion; $$ escapes the dollar so systemd passes the
# literal $ through to /bin/sh, which then expands.
ExecStart=/bin/sh -c 'exec /usr/bin/flm serve --host "$$FLM_HOST" $${FLM_PORT:+--port "$$FLM_PORT"} --pmode "$$FLM_PMODE" -a "$$FLM_ASR" -e "$$FLM_EMBED" $$FLM_EXTRA_OPTS "$$FLM_MODEL"'
Restart=always
RestartSec=5s

NoNewPrivileges=true
PrivateTmp=true
LockPersonality=true
RestrictSUIDSGID=true
RestrictRealtime=true
ProtectSystem=full
ProtectHome=true

[Install]
WantedBy=multi-user.target