#!/bin/bash
# filepath: zhch/launch_model_vllm_macos.sh

# Apple Silicon 优化的 vLLM 启动脚本
export model_path="./weights/DotsOCR_CPU_bfloat16"  # 使用 bfloat16 优化的模型路径
export model_name="DotsOCR"
export PYTHONPATH=$(dirname "$model_path"):$PYTHONPATH

echo "🍎 Starting vLLM on Apple Silicon with optimizations..."
echo "Model path: $model_path"
echo "PyTorch version: $(python -c 'import torch; print(torch.__version__)')"
echo "MPS available: $(python -c 'import torch; print(torch.backends.mps.is_available())')"

# Apple Silicon 优化参数
vllm serve ${model_path} \
    --host 0.0.0.0 \
    --port 8101 \
    --max-model-len 32768 \
    --max-num-batched-tokens 4096 \
    --max-num-seqs 8 \
    --gpu-memory-utilization 0.8 \
    --trust-remote-code \
    --chat-template-content-format string \
    --served-model-name ${model_name} \
    --uvicorn-log-level info \
    --limit-mm-per-prompt image=5 \
    --tensor-parallel-size 1 \
    --disable-log-stats \
    --disable-log-requests