single mention · inference & serving
Pre-scheduled tile chunking
Pre-scheduled tile chunking is a KV-cache kernel scheduling method that splits each KV tile along its query dimension into small chunks.1 It fans hot tiles, selected by many queries, across many CTAs sharing the load, and preassigns each (query, chunk) pair an output slot so partials are written without atomics.2
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
splits each KV tile along its query dimension into chunks
— MiniMax Sparse Attentionfanning hot tiles across many CTAs that share the same load
— MiniMax Sparse Attention
Filed under inference & serving › KV cache management