#!/bin/bash set -eo pipefail echo "==========================================================" echo " vLLM dual-model + PP-DocLayoutV3 (CPU) on HF Space (T4) " echo "==========================================================" # 1. GPU & CUDA environment validation echo "[1/4] Validating GPU and CUDA environment..." if ! python3 /scripts/gpu_detect.py; then echo "[CRITICAL ERROR] GPU environment validation failed." echo "Deployment requires >= ${MIN_VRAM_MB:-15000} MB VRAM and CUDA >= ${MIN_CUDA_MAJOR:-11}.${MIN_CUDA_MINOR:-8}." exit 1 fi echo "[OK] GPU environment verified." # 2. Authentication configuration echo "[2/4] Configuring Nginx Basic Authentication..." if [ -z "${NGINX_PASSWORD:-}" ]; then echo "[CRITICAL ERROR] NGINX_PASSWORD is not defined." echo "Set NGINX_PASSWORD in Hugging Face Space Secrets." exit 1 fi python3 /scripts/generate_htpasswd.py echo "[OK] Nginx authentication configured." # 3. Model cache on LOCAL disk (fast); /data bucket stays mounted for other persistent data echo "[3/5] Preparing local model cache at ${HF_HOME:-/tmp/hf}..." mkdir -p "${HF_HOME:-/tmp/hf}" \ /tmp/nginx/body /tmp/nginx/proxy /tmp/nginx/fastcgi /tmp/nginx/uwsgi /tmp/nginx/scgi \ /tmp/torch_extensions /tmp/.cache /var/log/nginx echo "[OK] Cache directory ready." # 4. Preload models from the HF Hub (hf_xet, fast, parallel) into local disk — avoids # slow reads through the FUSE-mounted bucket (was ~12 Mo/s, >10 min per model) echo "[4/5] Preloading models from HF Hub into ${HF_HOME:-/tmp/hf} (hf_xet)..." if ! /scripts/preload_models.sh; then echo "[WARNING] preload failed; vLLM will fall back to downloading on the fly" fi # 5. Hand over to supervisord: vllm-ocr, layout (CPU), vllm-vlm, warmup, nginx # supervisord comes from pip (/usr/local/bin, Python 3.12 compatible), NOT from apt # (the apt package 4.2.1 crashes on Python 3.12: asynchat was removed from the stdlib) echo "[5/5] Starting supervisord..." exec supervisord -n -c /etc/supervisord.conf