January 2026 (2026-01) monthly summary for kvcache-ai/sglang. Key accomplishments focused on documentation improvements for CUDA attention backend selection: - Delivered comprehensive CUDA attention backend documentation detailing the automatic selection logic for different model architectures (MHA/MLA) and the GPU-architecture defaults. - Clarified defaults and behavior to reduce user confusion and misconfigurations when selecting backends. - Noted FP8 KV cache support within the CUDA attention backend docs, aligning with performance optimization roadmap (commit eb38d6441375322878a428761e1d298cbe98a73b). Major bugs fixed: - None reported or pushed to this month; stability maintained through documentation and process improvements. Overall impact and accomplishments: - Improved developer onboarding and user experience by clarifying CUDA attention backend choices and defaults across GPU architectures. - Reduced support overhead and risk of misconfiguration, enabling faster, correct deployments. - Established documentation groundwork for FP8 KV cache support and related performance enhancements. Technologies/skills demonstrated: - CUDA backend reasoning and model architecture awareness (MHA/MLA). - Documentation best practices, clear impact communication, and cross-repo consistency. - Attention to performance-oriented features (FP8 KV cache) and roadmap alignment.