tinyagentos-rknn-sd.service

[Unit]
Description=TinyAgentOS RKNN Stable Diffusion server (RK3588 NPU)
After=network.target
Wants=network.target
# Want to start after tinyagentos.service so the main app registers the
# backend in BackendCatalog as soon as this comes up, not before.
After=tinyagentos.service

[Service]
Type=simple
User=jay
WorkingDirectory=/home/jay/tinyagentos

# Default backend: ez_rknn_async (tp_mode=all on unet/vae_decoder).
# To revert to rknn-toolkit-lite2 / darkbit1001's rknnlcm.py, add:
#   Environment=RKNN_SD_LEGACY_WRAPPER=1
Environment=RKNN_SD_MODEL_DIR=/home/jay/.local/share/tinyagentos/rknn-sd/model
Environment=RKNN_SD_WRAPPER=/home/jay/.local/share/tinyagentos/rknn-sd/rknnlcm.py
Environment=RKNN_SD_HOST=127.0.0.1
Environment=RKNN_SD_PORT=7863
Environment=PYTHONPATH=/home/jay/tinyagentos
Environment=PYTHONUNBUFFERED=1

# The wrapper loads 2.3 GB of RKNN weights onto the NPU at startup,
# which takes ~20s on the RK3588. Give systemd time before it assumes
# the unit is dead.
TimeoutStartSec=90

ExecStart=/usr/local/bin/python3.12 -m tinyagentos.services.rknn_sd_server

Restart=always
RestartSec=5
StandardOutput=append:/home/jay/tinyagentos/data/rknn-sd-server.log
StandardError=inherit

# The UNet alone is 1.8 GB and the full pipeline peaks around 5.6 GB
# during inference. Leave some headroom.
MemoryMax=7G

[Install]
WantedBy=multi-user.target