Full-Pipeline Inference Optimization for MiMo-V2.5 Series

MiMo-V2.5 series uses Hybrid Sliding Window Attention to reduce KVCache storage to roughly 1/7 that of traditional full attention, with the 70-layer MiMo-V2.5-Pro split between 10 global attention layers and 60 local window layers. The optimization enables long-context inference with significantly lower memory overhead through system-level improvements to cache management and scheduling.