#!/usr/bin/bash
set -e
LLAMA_DIR="/home/brassaikao/llama.cpp"
PROCESS_NAME="llama-server"
echo "------------------------------------------------"
echo "🤖 [1/4] 檢查服務狀態..."
echo "------------------------------------------------"
if pgrep -x "$PROCESS_NAME" > /dev/null; then
    echo "⚠️ 偵測到 $PROCESS_NAME 正在執行中，正在自動優雅關閉..."
    # 發送 SIGINT (Ctrl+C 訊號) 讓 llama-server 釋放 VRAM 並正常關閉
    pkill -2 -x "$PROCESS_NAME" || true
    
    # 最多等待 15 秒，確保程序完全釋放
    WAIT_COUNT=0
    while pgrep -x "$PROCESS_NAME" > /dev/null; do
        sleep 1
        WAIT_COUNT=$((WAIT_COUNT+1))
        if [ $WAIT_COUNT -ge 15 ]; then
            echo "⚡ 等待超時，強制終止程序 (SIGKILL)..."
            pkill -9 -x "$PROCESS_NAME" || true
            break
        fi
    done
    echo "✅ 舊服務已順利關閉，VRAM 已釋放！"
else
    echo "ℹ️ $PROCESS_NAME 目前未執行。"
fi
cd "$LLAMA_DIR"
echo ""
echo "------------------------------------------------"
echo "🔍 [2/4] 比對 GitHub 遠端版本..."
echo "------------------------------------------------"
git fetch origin
LOCAL_HASH=$(git rev-parse HEAD)
REMOTE_HASH=$(git rev-parse @{u})
NEED_BUILD=0
if [ "$LOCAL_HASH" != "$REMOTE_HASH" ]; then
    echo "🚀 發現新版本！"
    echo "  本地 Hash: ${LOCAL_HASH:0:7}"
    echo "  遠端 Hash: ${REMOTE_HASH:0:7}"
    echo "正在拉取最新原始碼..."
    git pull
    NEED_BUILD=1
else
    echo "✅ 當前已是最新版本 (${LOCAL_HASH:0:7})，無需重新編譯！"
fi
echo ""
echo "------------------------------------------------"
echo "⚙️ [3/4] 編譯階段..."
echo "------------------------------------------------"
if [ $NEED_BUILD -eq 1 ]; then
    echo "正在進行本機 CUDA (AMD 5600G, RTX 3050) 極致原生編譯..."
    export PATH=/opt/cuda/bin:$PATH
    export CUDA_PATH=/opt/cuda
    export CC=gcc
    export CXX=g++
    cmake -B build \
        -DGGML_CUDA=ON \
        -DGGML_NATIVE=ON \
        -DGGML_CUDA_FA_ALL_QUANTS=ON \
        -DCMAKE_CUDA_ARCHITECTURES=86 \
        -DCUDAToolkit_ROOT=/opt/cuda \
        -DCMAKE_BUILD_TYPE=Release
    cmake --build build --config Release -j $(nproc)
    echo "🎉 編譯成功！"
else
    echo "⏩ 跳過編譯，直接使用現有極致優化二進制檔。"
fi
echo ""
echo "------------------------------------------------"
echo "🚀 [4/4] 啟動最新版 llama-server..."
echo "------------------------------------------------"
./build/bin/llama-server \
  -m /home/brassaikao/ai-models/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-IQ4_NL.gguf \
  --mmproj /home/brassaikao/ai-models/mmproj-Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-f16.gguf \
  --alias Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-IQ4_NL \
  -np 1 -t 6 -tb 6 \
  --flash-attn on \
  --image-min-tokens 1024 \
  --load-mode none \
  --port 8080 \
  --host 172.17.0.1 \
  --kv-unified \
  -c 131072 \
  --jinja \
  --cont-batching \
  --reasoning-preserve