start_server.sh

#!/bin/bash
set -ex
echo "--- Starting setup and server ---"

# Use the exact snapshot path resolved at build time so Python can import the model code.
# The HF cache's model root doesn't contain the files; they're under snapshots/<commit>.
HF_MODEL_PATH_FILE="/home/user/app/model_path.txt"
if [ -f "$HF_MODEL_PATH_FILE" ]; then
    export HF_MODEL_PATH="$(cat "$HF_MODEL_PATH_FILE")"
fi
# Ensure Python sees the model's modules (e.g., modeling_dots_ocr_vllm.py).
export PYTHONPATH="${HF_MODEL_PATH}:${PYTHONPATH}"

if [ -n "$HF_MODEL_PATH" ] && [ ! -e "DotsOCR" ]; then
    ln -s "$HF_MODEL_PATH" DotsOCR
fi

python3 - <<'PY'
import sys, os
print("HF_MODEL_PATH:", os.environ.get("HF_MODEL_PATH"))
print("PYTHONPATH:", sys.path)
from DotsOCR import modeling_dots_ocr_vllm
print("DotsOCR import OK")
PY

echo "Starting server..."
exec vllm serve ${HF_MODEL_PATH} \
    --host 0.0.0.0 \
    --port ${PORT} \
    --tensor-parallel-size 1 \
    --gpu-memory-utilization 0.8 \
    --chat-template-content-format string \
    --served-model-name dotsocr-model \
    --trust-remote-code