service/run.sh

#!/bin/bash
# The demo service. One card holds VGGT-Omega + VAE + the finetuned transformer + the DMD LoRA (~77 GiB).
#   CARD=0 service/run.sh --port 8412
export CUDA_VISIBLE_DEVICES=${CARD:-0}
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
export PYTORCH_ALLOC_CONF=expandable_segments:True
cd "$(dirname "$0")/.."
# Exit code 3 = the service found its CUDA context poisoned and asked to be reloaded; anything else stops.
while :; do
  python service/app.py "$@"
  [ $? = 3 ] || break
  sleep 2
done