Chronos – Fair GPU Time-Sharing (Side Project) (github.com)

Chronos is a new open-source tool that enforces time-based, fair GPU partitions with automatic expiration — think guaranteed, lease-like GPU access for users sharing a single card. It provides a simple Python API and CLI (pip package chronos-gpu or Docker), so you can request, monitor and auto-release partitions (e.g., device=0, memory=0.5, duration=3600s). Chronos targets research labs, small teams and classrooms to eliminate resource conflicts, manual coordination and wasted compute while keeping sub-second expiration accuracy and per-user isolation. Technically, Chronos uses a C++ core with Python bindings, an OpenCL runtime layer for multi-vendor support (NVIDIA/AMD/Intel/Apple Silicon), lock files and a monitor thread for auto-expiration and cleanup. It enforces memory quotas, works with PyTorch/TensorFlow/JAX, and reports low overhead: ~3.2 ms partition create, ~1.8 ms release, ~0.8% GPU compute overhead in benchmarks and a 24-hour stress test with 1.2M ops and zero failures. Compared with NVIDIA MIG/MPS and traditional time-slicing, Chronos uniquely offers time-based allocation and auto-expiration across vendors and OSes. It’s Apache 2.0-licensed, hosted on GitHub, and encourages contributions — practical for anyone needing predictable, low-overhead, shareable GPU access.