SGLang's Weight Cache Daemon Cuts 1T Model Restart Time from 8.8 min to 32 sec
xiaosun86 · x · 2026-08-22
SGLang, in collaboration with Ant Group and Alibaba, introduced the Weight Cache Daemon to drastically reduce model reload times. By keeping post-quantized weights persistent in GPU memory and using CUDA IPC zero-copy mapping, the new system achieves:
- Weight Loading: 495s → 0.63s (approx. 785x speedup) on a Ling-2.6-1T FP8 model.
- Total Startup: 8.8min → 32s (93.9% reduction).
Key Features:
- Zero-Copy Loading: New engine instances map directly to existing GPU memory.
- Multi-Instance Sharing: Eliminates redundant disk I/O and transforms for multiple instances on the same GPU.
- Active-Standby Failover: Enables sub-second failover without dedicating full GPU memory to standby engines.
This is the first phase of their Fast Engine Recovery Framework, targeting <10s cold restarts and <1s warm standby switches for production LLM serving.
More from Infra
- Meta releases Muse Spark Open: 816B-param open-weight model with 1M context, natively multimodal — IanAndrewsDC · 2026-10-16
- llama.cpp adds day-0 support for Meta's 816B Muse Spark Open within 30 minutes — ngxson · 2026-10-16
- Copper to optical: optical backplanes seen as necessary as racks move to optics — BenBajarin · 2026-10-16
- Polymarket prices 28% odds a US state enacts data center moratorium by year-end — Polymarket · 2026-10-16
- Pennsylvania residents reject $10,000-per-household offer for massive Amazon data center — Polymarket · 2026-10-16
- PyTorch now runs natively on AWS Trainium with zero code changes — PyTorch · 2026-10-16