Skip to content

Installation.md

Latest commit

 

History

History
79 lines (57 loc) · 2.05 KB

Installation.md

File metadata and controls

79 lines (57 loc) · 2.05 KB

Installation Guide

Building from Source

Prerequisites for building from source:

  • Operating System: Linux
  • Python Version: >= 3.7
  • CUDA Version: >= 12.1
  • LLVM: < 20 if you are using the bundled TVM submodule

We currently provide three methods to install TileScale:

(optional)Prepare the container:

docker pull nvcr.io/nvidia/pytorch:25.03-py3
docker run --name tilescale --ipc=host --network=host --privileged --cap-add=SYS_ADMIN --shm-size=10g --gpus=all -it nvcr.io/nvidia/pytorch:25.03-py3 /bin/bash
echo -n > /etc/pip/constraint.txt
bash Miniconda3-latest-Linux-x86_64.sh # install conda
conda install -c conda-forge libstdcxx-ng
  1. Clone the Repository:
git clone --recursive https://github.com/tile-ai/tilescale
cd tilescale
  1. Install Project:
pip install cuda-python==12.9 # should align with your nvcc version
pip install scikit-build-core CMake torch ninja Cython
pip install -e . --no-build-isolation
  1. Verify Installation:

Verify that TileScale is working correctly:

python -c "import tilelang; print(tilelang.__version__)"

You can now run TileScale examples and develop your applications.

Example Usage:

You can run TileScale examples:

cd /home/tilelang
TILELANG_USE_DISTRIBUTED=1 python examples/distributed/example_allgather_gemm_overlapped.py

To use NVSHMEM APIs

Before running the examples using NVSHMEM APIs (e.g., example_allgather.py), you need to build NVSHMEM library for device-side code generation.

pip install mpich  # building NVSHMEM needs MPI
export NVSHMEM_SRC="your_custom_nvshmem_dir" # default to 3rdparty/nvshmem_src
cd tilelang/distributed
source build_nvshmem.sh

You also need to install the pynvshmem package, which provides wrapped host-side Python API for NVSHMEM.

cd ./pynvshmem
python setup.py install
export LD_LIBRARY_PATH="$NVSHMEM_SRC/build/src/lib:$LD_LIBRARY_PATH"

Then you can test python import:

python -c "import pynvshmem"