llama.cpp Release b11411 Fixes KV Cache Rotation Metadata Handling
llama.cpp has released build b11411, updating its KV cache implementation to save exact attention rotation metadata. The update explicitly rejects attempts to restore state files with mismatched rotation parameters. Saving and validating exact rotation metadata prevents subtle state corruption and inference errors when saving and resuming model sessions. This improves the reliability of state management across different KV cache configurations during local LLM deployment. The patch incorporates KV state rotation checks directly into the `test-save-load-state` matrix across tested models. Models without attention rotation pass the test vacuously, while unsupported KV cache types are automatically skipped.
## BACKGROUND
Key-Value (KV) caching is a critical LLM inference optimization that stores computed attention keys and values to avoid recomputing past context for each generated token. Modern LLMs often utilize Rotary Position Embedding (RoPE) or attention rotation techniques, requiring context state saving mechanisms to accurately track rotation offsets.