salesforce/warp-drive

Implements environment logic in CUDA or Numba kernels running directly on GPU, eliminating CPU-GPU data transfers while supporting both on-policy (A2C, PPO) and off-policy (DDPG) training via PyTorch. Scales from single-agent classic control to complex multi-agent scenarios with thousands of concurrent agents across distributed multi-GPU setups, achieving millions of environment steps per second on a single A100 GPU.

500 stars. No commits in the last 6 months.