#!/usr/bin/env bash
set -euo pipefail

MODEL_ID="${MODEL_ID:-meta-llama/Llama-3.1-8B-Instruct}"
mkdir -p /tmp/vllm
# Docker Spaces expect the configured port to answer while the batch workers warm up.
python3 -m http.server 7860 --bind 0.0.0.0 --directory /app &
for GPU in 0 1 2 3; do
  PORT=$((8000 + GPU))
  CUDA_VISIBLE_DEVICES="$GPU" vllm serve "$MODEL_ID" --host 127.0.0.1 --port "$PORT" \
    --dtype auto --max-model-len 6144 --gpu-memory-utilization 0.92 \
    --enable-prefix-caching --served-model-name compact-retracer >"/tmp/vllm/$GPU.log" 2>&1 &
done
python3 /app/wait_for_vllm.py
python3 /app/generate.py
# Keep the Space alive after a bounded run, so its logs and status stay available.
wait