WorldKernels

GPU-first world model simulation engine — serve learned world models (DiT, VAE) as interactive sessions.

⚠️ Early Development — This package is in pre-alpha. APIs may change.

Installation

pip install worldkernels

For full functionality:

pip install worldkernels[all]       # Everything
pip install worldkernels[serve]     # HTTP/WebSocket server
pip install worldkernels[diffusers] # HuggingFace Diffusers support

Quick Start

from worldkernels import WorldKernel, Action, WorldConfig

# Initialize engine
wk = WorldKernel(device="cuda")

# Load a world model from HuggingFace Hub
wk.load_world("Etched/oasis-500m")

# Create an interactive session
session = wk.create_session(
    world="oasis-500m",
    config=WorldConfig(height=360, width=640, fps=20),
)

# Step through the simulation
for _ in range(100):
    action = Action("keyboard", {"keys": ["W", "SPACE"]})
    obs = session.step(action)
    # obs.frames contains generated video frames

session.close()
wk.shutdown()

Features (Planned)

• 🎮 Session-based API — Stateful simulation with checkpoint/branch
• 🚀 GPU-optimized — Pre-allocated buffers, CUDA graphs, torch.compile
• 🔌 HuggingFace native — Load models directly from the Hub
• 🌐 HTTP/WebSocket server — REST API and real-time streaming
• 🧩 Extensible backends — PyTorch eager, torch.compile, TensorRT

Documentation

Coming soon at worldkernels.dev

License

LGPL-2.1 — see LICENSE