vLLM 0.31.0 Cuts Engine Restart Time

This is a meaningful operator-focused release: faster restarts directly improve iteration speed and service recovery for self-hosted inference. It is still a substantial version bump, so production teams should stage it rather than upgrade solely for preload.

• –`vllm preload` can reduce weight-loading time from minutes to seconds by reusing resident GPU memory.
• –Zero-copy sharing avoids adding another full model copy during engine restarts.
• –CUDA and ROCm are supported, along with tensor, expert, and data parallelism, including multi-node deployments.
• –Pipeline parallelism is not supported when launching the preload daemon.
• –The release’s broad scope means model-runner and hardware-specific changes deserve regression testing.