Optimize gVisor GPU checkpoint restore

[luke-lombardi]

Summary

• use gVisor's native NVIDIA checkpoint path for GPU workloads
• make automatic gVisor checkpoints terminal and overlap background CPU page loading with CUDA restore
• pin the tested Beam gVisor amd64 release with SHA-512 verification
• retain the official gVisor release for arm64 workers

The gVisor patch is isolated in beam-cloud/gvisor at 4272182b and is also open upstream as google/gvisor#13753. It prefaults only MemoryFile ranges mapped by restored CUDA processes before NVIDIA process resume. There are no vLLM-specific flags, KV-cache changes, eager mode, or workload-state reduction.

Staging results

RTX 5090, untouched Qwen/Qwen2.5-1.5B-Instruct BF16 vLLM deployment, existing 7.3 GiB automatic checkpoint:

• baseline gVisor runtime restore: roughly 25.7-29 seconds
• optimized runtime restore: roughly 6.2-6.6 seconds
• warm image load: 37 ms
• warm checkpoint path: 7.16 seconds
• request to exact valid chat completion: 7.62 seconds
• first request after worker recycle: 8.94 seconds

The exact release artifact and original checkpoint produced HTTP 200 responses with exact requested completion text. The checkpoint remained available and no replacement snapshot was created.

Verification

• go test ./pkg/runtime ./pkg/worker
• gVisor //pkg/sentry/mm:mm_test
• gVisor //pkg/sentry/pgalloc:pgalloc_test
• gVisor //runsc:runsc
• amd64 gvisor Docker stage build with checksum and version verification
• repeated staging checkpoint restores through /v1/chat/completions

---

Summary by cubic

Optimize gVisor GPU checkpoint/restore by using its native NVIDIA path and making auto-checkpoints terminal, overlapping CPU page loading with CUDA resume. Cuts restore time on RTX 5090 vLLM from ~26–29s to ~6.2–6.6s without changing checkpoints.

• New Features

  • Use gVisor nvproxy via a GPU bundle annotation; bind cuda-checkpoint; pass --cuda-checkpoint-path on checkpoint and --nvproxy on restore; remove per-PID CUDA freezing.
  • Run restore --background, wait for completion, and add safe force-delete cleanup.
  • Treat gVisor auto-checkpoints as terminal (same as runc); add VLLM_HOST_IP and VLLM_LOOPBACK_IP loopback env vars for service checkpoints.

• Dependencies

  • Pin amd64 runsc to beam-cloud/gvisor release-20260714.0-beam.1 with SHA-512 digest verification; keep official gVisor release for arm64.

^{Written for commit f59099b. Summary will update on new commits.}