#!/usr/bin/env bash
# =============================================================================
# Qwen3-Omni vLLM-Omni 服务启动脚本
# 
# 使用方法 (在远程服务器上):
#   1. 安装 vllm-omni:
#      pip install vllm-omni
#   
#   2. 下载模型:
#      modelscope download --model Qwen/Qwen3-Omni-30B-A3B-Instruct \
#          --local_dir ./models/Qwen/Qwen3-Omni-30B-A3B-Instruct
#
#   3. 启动服务:
#      bash start_qwen3_omni.sh
#
# 端口: 8904
# API:
#   - WebSocket 实时: ws://127.0.0.1:8904/v1/realtime
#   - Chat Completions: http://127.0.0.1:8904/v1/chat/completions
#   - Health: http://127.0.0.1:8904/health
# =============================================================================
set -euo pipefail

SCRIPT_DIR="$(cd -- "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
MODEL_PATH="${SCRIPT_DIR}/models/Qwen/Qwen3-Omni-30B-A3B-Instruct"
PORT="${PORT:-8904}"

echo "============================================"
echo "  Qwen3-Omni vLLM-Omni 服务启动"
echo "============================================"
echo "  模型路径: ${MODEL_PATH}"
echo "  端口: ${PORT}"
echo "============================================"

# 检查模型是否存在
if [ ! -d "${MODEL_PATH}" ]; then
    echo ""
    echo "❌ 模型未找到: ${MODEL_PATH}"
    echo ""
    echo "请先下载模型:"
    echo "  pip install modelscope -U"
    echo "  modelscope download --model Qwen/Qwen3-Omni-30B-A3B-Instruct \\"
    echo "      --local_dir ${MODEL_PATH}"
    exit 1
fi

# 检查 vllm-omni 是否安装
if ! python3 -c "import vllm_omni" 2>/dev/null; then
    echo ""
    echo "❌ vllm-omni 未安装"
    echo ""
    echo "请安装:"
    echo "  pip install vllm-omni"
    echo ""
    echo "或从源码安装:"
    echo "  git clone https://github.com/vllm-project/vllm-omni.git"
    echo "  cd vllm-omni"
    echo "  pip install -e ."
    exit 1
fi

echo ""
echo "🚀 启动 vLLM-Omni 服务..."

# 停止旧进程 (如果存在)
OLD_PID=$(lsof -t -i:${PORT} 2>/dev/null || true)
if [ -n "${OLD_PID}" ]; then
    echo "  停止旧进程 (PID: ${OLD_PID})..."
    kill ${OLD_PID} 2>/dev/null || true
    sleep 2
fi

# 启动服务
# --omni: 启用多阶段编排 (Thinker → Talker → Code2Wav)
# 默认配置: vllm_omni/deploy/qwen3_omni_moe.yaml
# 异步 chunk 流式传输默认开启
vllm serve "${MODEL_PATH}" \
    --omni \
    --port "${PORT}" \
    --host 0.0.0.0 \
    --gpu-memory-utilization 0.85 \
    --max-model-len 32768 \
    2>&1 | tee "${SCRIPT_DIR}/qwen3_omni_server.log"

echo ""
echo "服务已停止."