MELON 实验完全本地化

1. Insights

1. 将论文中Embedding model 替换为 BAAI/bge-small-en-v1.5 ，无需调用openai API
2. 将 LLM 替换为Qwen3.5-9B，更加轻量化。单张5090即可跑起来

2. 目录

服务器上默认使用：

/code/melon/agentdojo
/root/autodl-tmp/models/Qwen3.5-9B
/root/autodl-tmp/models/bge-small-en-v1.5
/root/autodl-tmp/hf-cache

3. 配置环境

AgentDojo

conda create -n agentdojo python=3.11 -y
conda activate agentdojo
cd /code/melon/agentdojo
python -m pip install -U pip
pip install -e ".[transformers]"

vLLM

conda create -n vllm-qwen35 python=3.12 -y
conda activate vllm-qwen35
python -m pip install -U pip
pip install -U vllm --extra-index-url https://wheels.vllm.ai/nightly

如果是 5090 / Blackwell，不稳时用：

pip install -U vllm --extra-index-url https://wheels.vllm.ai/nightly/cu130

4. 模型检查

启动前确认：

ls -lah /root/autodl-tmp/models/Qwen3.5-9B/config.json
ls -lah /root/autodl-tmp/models/bge-small-en-v1.5/config.json

5. 启动 vLLM

conda activate vllm-qwen35
export HF_HOME=/root/autodl-tmp/hf-cache
unset OMP_NUM_THREADS
export OMP_NUM_THREADS=1

vllm serve /root/autodl-tmp/models/Qwen3.5-9B \
  --host 0.0.0.0 \
  --port 8000 \
  --tensor-parallel-size 1 \
  --gpu-memory-utilization 0.88 \
  --max-model-len 65536 \
  --max-num-seqs 1 \
  --max-cudagraph-capture-size 128 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --reasoning-parser qwen3 \
  --default-chat-template-kwargs '{"enable_thinking": false}' \
  --language-model-only 2>&1 | tee vllm.log

检查：

curl http://127.0.0.1:8000/health
curl http://127.0.0.1:8000/v1/models

6. 最小实验

conda activate /root/autodl-tmp/conda/envs/agentdojo
cd /code/melon/agentdojo
./util_scripts/run_melon_minimal.sh

默认：

• VLLM_PARSED
• tool_knowledge
• melon
• slack

7. 全量实验

conda activate /root/autodl-tmp/conda/envs/agentdojo
cd /code/melon/agentdojo
./util_scripts/run_melon_full.sh

默认：

• workspace slack travel banking
• tool_knowledge
• melon
• VLLM_PARSED

8. 可选参数

最小实验

SUITE=workspace ./util_scripts/run_melon_minimal.sh
LOGDIR=./runs/test_min ./util_scripts/run_melon_minimal.sh
MELON_EMBED_DEVICE=cuda ./util_scripts/run_melon_minimal.sh

全量实验

SUITES="slack workspace" ./util_scripts/run_melon_full.sh
FORCE_RERUN=1 ./util_scripts/run_melon_full.sh
LOGDIR=./runs/test_full ./util_scripts/run_melon_full.sh

9. 日志

vLLM

tail -f vllm.log

最小实验

默认：

runs/qwen35-9b-melon-local-embed/benchmark.log

全量实验

默认：

runs/qwen35-9b-melon-full/benchmark_workspace.log
runs/qwen35-9b-melon-full/benchmark_slack.log
runs/qwen35-9b-melon-full/benchmark_travel.log
runs/qwen35-9b-melon-full/benchmark_banking.log

Acknowledgements

This project is built on MELON (Zhu et al., ICML 2025).

@inproceedings{zhu2025melon,
  title={MELON: Provable Defense Against Indirect Prompt Injection Attacks in AI Agents},
  author={Zhu, Kaijie and Yang, Xianjun and Wang, Jindong and Guo, Wenbo and Wang, William Yang},
  booktitle={International Conference on Machine Learning},
  year={2025}
}