#!/bin/bash # One-shot warmup (supervisord program, autorestart=false): # waits for all three services to be healthy, then sends a tiny request to each # vLLM instance so the models are hot in VRAM before the first user request. set -u PORT_OCR="${PORT_OCR:-8001}" PORT_VLM="${PORT_VLM:-8002}" LAYOUT_PORT="${LAYOUT_PORT:-8003}" SERVED_OCR="${SERVED_OCR:-glm-ocr}" SERVED_VLM="${SERVED_VLM:-qwen3-vl}" wait_http() { local url="$1" max="${2:-240}" i=0 until curl -sf "$url" >/dev/null 2>&1; do i=$((i + 1)) if [ "$i" -ge "$max" ]; then echo "[warmup] TIMEOUT waiting for $url" return 1 fi sleep 5 done echo "[warmup] $url is ready" return 0 } echo "[warmup] Waiting for vllm-ocr on :${PORT_OCR} (model load can take several minutes)..." wait_http "http://127.0.0.1:${PORT_OCR}/health" 240 || exit 1 echo "[warmup] Waiting for vllm-vlm on :${PORT_VLM}..." wait_http "http://127.0.0.1:${PORT_VLM}/health" 240 || exit 1 echo "[warmup] Waiting for layout service on :${LAYOUT_PORT}..." wait_http "http://127.0.0.1:${LAYOUT_PORT}/health" 120 || exit 1 echo "[warmup] Sending warmup request to ${SERVED_OCR}..." curl -s -m 600 -X POST "http://127.0.0.1:${PORT_OCR}/v1/chat/completions" \ -H 'Content-Type: application/json' \ -d "{\"model\":\"${SERVED_OCR}\",\"messages\":[{\"role\":\"user\",\"content\":\"Hi\"}],\"max_tokens\":8}" \ >/dev/null || echo "[warmup] WARNING: OCR warmup request failed (first request will be slower)" echo "[warmup] Sending warmup request to ${SERVED_VLM}..." curl -s -m 600 -X POST "http://127.0.0.1:${PORT_VLM}/v1/chat/completions" \ -H 'Content-Type: application/json' \ -d "{\"model\":\"${SERVED_VLM}\",\"messages\":[{\"role\":\"user\",\"content\":\"Hi\"}],\"max_tokens\":8}" \ >/dev/null || echo "[warmup] WARNING: VLM warmup request failed (first request will be slower)" echo "[warmup] Done." exit 0