Zero-Downtime Model Updates: How Atomic Weight Swap Works

Blue-green deployments for LLMs double your GPU memory footprint. Ghost Weights atomic swap replaces weights inside the serving process between requests — no second instance, no downtime, no traffic splitting.