Installation

Requirements

The code is built and tested with Python 3.11, CUDA 12.8, and PyTorch 2.10.0.

Preparation

1. Clone the repository

git clone https://github.com/InternRobotics/InternVLA-A-series.git
cd InternVLA-A1.5

2. Create Conda environment

conda create -y -n internvla_a1_5 python=3.11
conda activate internvla_a1_5
pip install --upgrade pip

3. Install system dependencies

We use FFmpeg for video encoding/decoding and SVT-AV1 for efficient storage.

conda install -c conda-forge ffmpeg svt-av1 -y

4. Install PyTorch for CUDA 12.8

pip install torch==2.10.0 torchvision==0.25.0 \
  --index-url https://download.pytorch.org/whl/cu128

5. Install Python dependencies

pip install transformers==5.2.0
pip install -e .

6. Install attention kernels

pip install flash-attn==2.8.3 flash-linear-attention==0.5.0 causal-conv1d==1.6.1 \
  --no-build-isolation

If the installation fails on your CUDA/PyTorch platform, follow the upstream build instructions and install the matching versions from flash-attn, flash-linear-attention, and causal-conv1d.

7. Patch HuggingFace Transformers

InternVLA-A1.5 uses custom model code for Qwen3.5 and robot-learning policies. Copy the replacement modules into the installed Transformers package:

TRANSFORMERS_DIR=${CONDA_PREFIX}/lib/python3.11/site-packages/transformers/

cp -r src/lerobot/policies/pi0/transformers_replace/models ${TRANSFORMERS_DIR}
cp -r src/lerobot/policies/pi05/transformers_replace/models ${TRANSFORMERS_DIR}
cp -r src/lerobot/policies/internvla_a1_5/transformers_replace/models ${TRANSFORMERS_DIR}

Make sure ${TRANSFORMERS_DIR} exists before copying.

8. Configure environment variables

export HF_TOKEN=your_token
export HF_HOME=path_to_huggingface
export HF_LEROBOT_HOME=${HF_HOME}/lerobot

9. Optional: download VGM weights

If you need to train with the VGM (video generation model) branch enabled, download Wan2.2-TI2V-5B to ${HF_HOME}/hub:

huggingface-cli download Wan-AI/Wan2.2-TI2V-5B \
  --local-dir ${HF_HOME}/hub/Wan2.2-TI2V-5B

This step can be skipped for action-only inference/evaluation or runs that disable video foresight supervision.

10. Link local LeRobot cache

If your datasets are stored under ${HF_HOME}/lerobot, link them into this repository:

ln -s ${HF_HOME}/lerobot data

This allows the training scripts to access datasets through ./data/.