Installation

Requirements

• Python 3.10
• CUDA 12.8 (or compatible version)
• Conda (recommended)

Step 1: Create Conda Environment

conda create -n shaper python=3.10
conda activate shaper

Step 2: Set Up CUDA Environment

Ensure CUDA is properly configured. Adjust paths based on your system:

export CUDA_HOME=/path/to/cuda
export CUDA_INCLUDE=$CUDA_HOME/include
export CUDA_LIB=$CUDA_HOME/lib
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CUDA_LIB
export LIBRARY_PATH=$LIBRARY_PATH:$CUDA_LIB
export CFLAGS="-I$CUDA_HOME/include"
export CXXFLAGS="-I$CUDA_HOME/include"
export CPATH="$CUDA_HOME/include:$CPATH"

Step 3: Install Python Packages

conda install -c conda-forge gcc_linux-64=11 gxx_linux-64=11
conda install -c conda-forge sparsehash
pip install wheel setuptools ninja
pip install numpy tqdm hydra-core matplotlib opencv-python imageio easydict munch plyfile
pip install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu128
pip install transformers trimesh scikit-image diffusers gradio peft einops
pip install flash-attn --no-build-isolation --no-cache-dir
pip install "imageio[ffmpeg]" "imageio[pyav]"
pip install pymeshlab sophuspy fast_simplification scikit-learn timm plotly torchdiffeq sentencepiece protobuf pyrender jupyter

Step 4: Install Torch-Cluster and Torchsparse

pip install torch-cluster -f https://data.pyg.org/whl/torch-2.7.1+cu128.html
pip install --verbose git+https://github.com/nihalsid/torchsparse@legacy --no-build-isolation

Note: The legacy version of torchsparse is required. Newer versions have incompatible data structures. This step takes a while, be patient.

Step 5: Verify Installation

python -c "import torch; from torchsparse import SparseTensor;
x = SparseTensor(coords=torch.tensor([[1,2,3,0], [4,5,6,1]], dtype=torch.int32), feats=torch.randn(2, 4));
x = x.cuda();
print('Installation successful!')"

Troubleshooting

Torchsparse build fails

• Ensure sparsehash headers are in your include path (CPATH)
• Ensure CUDA environment variables are set correctly
• Try building with verbose output to see specific errors

Flash attention build fails

• Ensure you have a compatible GPU (Ampere or newer recommended)
• Check that CUDA toolkit matches your PyTorch CUDA version

Import errors

• Verify all packages installed correctly: pip list | grep torch
• Check CUDA availability: python -c "import torch; print(torch.cuda.is_available())"