SGLang adds experimental unified memory mode

SGLang has introduced an experimental unified memory system for hybrid AI models that replaces rigid static sub-pools with a single GPU memory arena and dynamic capacity allocation. The design eliminates pre-configured memory ratios, supports zero-fragmentation page compaction, and packs multi-layer token state into single entries. Results on models like GPT-OSS and Qwen3.5 show that unified memory increases concurrent request capacity, eliminates costly tuning errors, and improves throughput. Open article →