Installation

Requirements

• Linux or WSL2
• Python 3.11
• NVIDIA GPU with CUDA 12.x
• FFmpeg available on PATH

The paper experiments used three NVIDIA L40 GPUs with 48 GB memory each. One-GPU evaluation is supported but takes longer; Qwen3-VL-8B and long ActivityNet videos may require a 48 GB GPU or a lower input-token limit.

Create the environment

git clone https://github.com/JiaqiLi404/SemVID.git
cd SemVID

conda create -n openvps python=3.11 -y
conda activate openvps

sudo apt-get update
sudo apt-get install -y ffmpeg

python -m pip install --upgrade pip setuptools wheel
python -m pip install torch==2.8.0 torchvision==0.23.0 \
  --index-url https://download.pytorch.org/whl/cu128
python -m pip install -r requirements.txt
python -m pip install flash-attn==2.8.3 --no-build-isolation

Install PyTorch for a different CUDA runtime only when required by your system. Keep the package versions in requirements.txt, especially Transformers, unchanged unless you also validate the vendored model implementations.

Verify the environment

python - <<'PY'
import torch
import transformers

print("torch:", torch.__version__)
print("CUDA runtime:", torch.version.cuda)
print("visible GPUs:", torch.cuda.device_count())
print("transformers:", transformers.__version__)
assert transformers.__version__.startswith("4.57.")
PY

Model access

The public configs use these Hugging Face IDs:

• Qwen/Qwen3-VL-4B-Thinking
• Qwen/Qwen3-VL-8B-Thinking
• Qwen/Qwen2.5-VL-7B-Instruct
• llava-hf/llava-onevision-qwen2-7b-ov-hf

They download on first use. To use a local checkpoint, pass --model-path /path/to/checkpoint when evaluating.

Troubleshooting

• flash-attn build failure: install PyTorch first, confirm the CUDA compiler matches the PyTorch build, then reinstall with --no-build-isolation.
• Transformers import/model errors: run python -c "import transformers; print(transformers.__version__)"; versions 5.0 and newer are unsupported.
• CUDA out of memory: use a smaller backbone, lower eval_args.max_model_len, or reduce script_args.total_pixels_token_length through --cfg-options.
• No GPU visible: verify nvidia-smi, the WSL NVIDIA driver, and CUDA_VISIBLE_DEVICES before launching evaluation.

Next: prepare datasets, then run evaluation.