Skip to content

Installation

This page covers the base TeleFuser package. Model checkpoints, LiveKit Server, and the optional tf-kernel distribution are installed separately.

Requirements

Component Requirement
Operating system Linux recommended
Python 3.10 through 3.13
PyTorch 2.6 or newer
CUDA toolkit 12.8 or newer for the maintained CUDA development path
ROCm 7.x with a PyTorch +rocm build for AMD GPUs; see the ROCm note under verification
GPU Depends on the selected model; check its Cookbook guide

An example may impose stricter versions or GPU architecture requirements. In particular, locally built tf-kernel artifacts are tied to their recorded PyTorch, CUDA, ABI, and GPU-family configuration.

Install the Package

python -m pip install --upgrade pip
python -m pip install telefuser

git clone https://github.com/Tele-AI/TeleFuser.git
cd TeleFuser
python -m pip install -e .

git clone https://github.com/Tele-AI/TeleFuser.git
cd TeleFuser
python -m pip install -e ".[dev]"
pre-commit install

Optional dependency groups are ui, distributed, docs, and dev. Install only the groups required by the workflow, for example python -m pip install -e ".[distributed]" for Ray-backed execution.

Verify the Installation

python -c "import torch, telefuser; print(torch.__version__); print(torch.cuda.is_available())"
telefuser --help

Model execution expects torch.cuda.is_available() to print True. If it does not, verify the installed PyTorch build and visible NVIDIA driver before diagnosing TeleFuser.

On AMD ROCm hosts, install a PyTorch +rocm build instead of the CUDA toolkit path. A HIP build also prints True for torch.cuda.is_available() (check torch.version.hip to distinguish it), and TeleFuser's platform layer detects ROCm before CUDA. Examples ending in _rocm.py (for example examples/wan_video/wan21_1_3b_text_to_video_rocm.py) are the validated entry points; see Hardware Platforms for per-platform capabilities and backend availability.

Model Checkpoints

TeleFuser does not bundle model weights. The Supported Models page links to each Cookbook guide, where the validated Hugging Face and ModelScope identifiers, directory layout, and additional artifacts are documented. A Hugging Face model ID can be passed directly only when the selected example explicitly supports hub loading.

Optional Components

  • TF-Kernel documents the separate Makefile-based build and compatibility checks.
  • Attention lists optional attention backends and their hardware requirements.
  • Stream Server covers the separately operated LiveKit service.

Continue with the LingBot-World v2 WebRTC Core Experience, or use the lighter Basic Inference Quickstart to verify a single-GPU pipeline and batch API.