PhoenixOS (PhOS) checkpoints and restores a running GPU process — its live CUDA context, loaded kernels, and GPU memory — without killing the process. This lets you park an idle GPU application off the GPU and later bring it back in place, so a single GPU can be time-shared across far more workloads than physically fit at once.
PhOS v1 is a preview refactor of SJTU-IPADS/phoenixos that focuses on ease of use and adds checkpoint-and-restore support for vLLM. It offers two modes:
- GPU-only checkpoint/restore (
detach/attach): only the GPU worker is torn down while the host (CPU-side) process stays alive holding its state. A shared context service (standby) keeps kernels warm across processes so restores skip reloading them. For LLM serving this means many models can be parked using zero GPU memory and packed onto one GPU — where plain sleep would pin each model's CUDA context resident. - Whole-process checkpoint/restore (with CRIU): PhOS snapshots the GPU state while CRIU checkpoints the host process, so the entire process — CPU side included — can be killed and later restored from disk.
See the vLLM guide and examples for a hands-on walkthrough of both.
- Remoting component and
config.tomldocs: docs/remoting.md - Environment docs: environment/README.md
- vLLM examples: tests/apps/vllm/README.md