To build the test image with dataset, please follow the instruction below.

1. test docker installation:

docker run --rm --gpus all nvcr.io/nvidia/cuda:12.8.1-cudnn-devel-ubuntu22.04 nvidia-smi

2. Run the following command to build docker container:

docker build -f scripts/docker/dockerfile -t ajet:latest .

3. By the way, for users in China, using the alternative dockerfile script can optimize the download speed using alibaba public cloud.

docker build -f scripts/docker/dockerfile_zh -t ajet:latest .

4. To run build-in tests, please follow instructions to mount test models and datasets.

• Download model manually, or use the helper script python ./scripts/download_model.py.
• For example, if your model is in ./modelscope_cache/Qwen/Qwen2___5-14B-Instruct.
• Run the instruction below to run the first training program

  clear && docker run -it --gpus all --shm-size="64g" --rm \
      -v "$(pwd)/modelscope_cache/Qwen/Qwen2___5-7B-Instruct:/mnt/data_cpfs/model_cache/modelscope/hub/Qwen/Qwen/Qwen2___5-7B-Instruct" \
      -e SWANLAB_API_KEY="xxxxxxxxxxxxxxxxxx" \
      -e DASHSCOPE_API_KEY="sk-xxxxxxxxxxxxxxxxxxxx" \
      -e CUDA_VISIBLE_DEVICES="4,5,6,7" \
      -e VERL_PYTHON="/opt/venv/bin/python" \
      -e NCCL_NVLS_ENABLE=0 \
      ajet:latest \
      python -m pytest -s tests/bench/benchmark_math/execute_benchmark_math.py::TestBenchmarkMath::test_01_begin_verl