This article introduces Tair-KVCache-HiSim, a high-fidelity CPU-based simulator for optimizing multi-tier KV Cache configurations in LLM inference.
This article introduces hierarchical sparse attention: the full KV Cache is stored on the CPU, while the GPU keeps only a Top-k LRU Buffer.
This article introduces a dual memory-pool inference framework enabling efficient hybrid Transformer-Mamba model execution by resolving conflicting caching mechanisms.
This article shows how SGLang RBG + Mooncake enable production-grade, cloud-native LLM inference with PD-disaggregation.
This article introduces HiCache, a hierarchical KVCache infrastructure developed by Alibaba Cloud Tair and SGLang to optimize performance and memory capacity for long-context "agentic" LLM inference.
