README.md

# Track B Phase 3 Submission

Team: Works on my agent

This archive contains the runnable submission for Track B Phase 3.

## Environment

Python 3.11 is recommended for the inference runner:

```bash
python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.txt
```

Our local validation environment used Huawei Ascend 910B hardware.

The runner does not require internet access at runtime. It connects only to the local vLLM OpenAI-compatible API and the local Track B sandbox APIs.

Default endpoints:

```text
vLLM API:          http://localhost:8001/v1
Command API:       https://localhost:8080/ip/api/agent/execute
Device discovery:  http://localhost:8080/ip/api/agent/get_devices_list
```

## Directory Structure

```text
.
├── README.md
├── run.sh
├── run.py
├── requirements.txt
├── src/
│   └── track_b_agent/
│       ├── submission_run.py
│       ├── submission_io.py
│       ├── run_agent.py
│       ├── agent/
│       ├── config/
│       ├── generalized/
│       ├── knowledge/
│       ├── prompts/
│       └── tools/
└── models/
    ├── deploy.sh
    ├── config.json
    ├── tokenizer.json
    ├── tokenizer_config.json
    ├── model.safetensors.index.json
    ├── model-00001-of-00016.safetensors
    ├── ...
    └── model-00016-of-00016.safetensors
```

## Deploy the Model

The model files are stored in `models/`. Start the vLLM OpenAI-compatible server with:

```bash
bash models/deploy.sh
```

The deployment script serves the model as:

```text
Qwen3.5-35B-A3B
```

It uses `models/` as the model path and starts vLLM on port `8001` with `nohup`. Logs are written to:

```text
models/Qwen3.5-35B-A3B_vllm_output.log
```

If there are any vLLM-environment-related issues, please contact:

```text
250010135@slai.edu.cn
```

## Run Inference

Run the submission on the provided Track B test file:

```bash
bash run.sh --input /path/to/test.json --output result
```

`run.sh` uses `.venv/bin/python` from the local virtual environment created above.

The runner uses 5-way scenario concurrency by default. The implementation uses `asyncio` with a 5-worker semaphore rather than the ThreadPoolExecutor shown in the example guide; the effective concurrent dispatch limit is still 5.

## Expected Output

The run command writes the required files into the output directory:

```text
result/
├── traces.json
├── results.csv
└── runtime.json
```

`results.csv` contains one final prediction per scenario:

```csv
scenario_id,prediction
```

`traces.json` contains all generated completions recorded during inference.

`runtime.json` contains per-problem runtime in seconds.