Installation and model download
Requirements
- Python 3.10.16
- PyTorch 2.9.0
- CUDA 12.6
Install PyTorch and the runtime dependencies:
pip install torch==2.9.0 torchvision==0.24.0 torchaudio==2.9.0 \
--index-url https://download.pytorch.org/whl/cu126
pip install websockets einops diffusers==0.36.0 transformers==4.55.2 \
accelerate msgpack opencv-python matplotlib ftfy easydict
pip install flash-attn --no-build-isolation
Model checkpoints
| Model | Hugging Face | ModelScope | Use |
|---|---|---|---|
lingbot-va-base | View model | View model | Base checkpoint with a shared backbone |
lingbot-va-posttrain-robotwin | View model | View model | Checkpoint post-trained on RoboTwin |
lingbot-va-posttrain-libero-long | View model | View model | Checkpoint post-trained on LIBERO-Long |
Post-training datasets
| Dataset | Hugging Face | ModelScope | Use |
|---|---|---|---|
robotwin-clean-and-aug-lerobot | View dataset | View dataset | Cleaned and augmented RoboTwin data in LeRobot format |
libero-long-lerobot | View dataset | View dataset | LIBERO-Long data in LeRobot format |
Configure the attention mode
The model reads attn_mode from <model-path>/transformer/config.json. Set it for the task before starting:
| Task | attn_mode |
|---|---|
| Training | "flex" |
| Inference or evaluation | "torch" or "flashattn" |
"flex" is only supported during training.
After installation and checkpoint setup, continue to Inference and evaluation or Data and post-training.