github-actions released this 24 Feb 01:13

v0.3.27-Metal-macos-20260224

01ced97

LLama: Optimize KV cache management for multi-round conversations

- Implements prefix-matching logic to truncate stale "ghost" tokens in C++ KV cache
- Prevents attention misalignment and context poisoning during multi-turn interactions
- Reduces memory overhead by reusing matched prefixes efficiently

Signed-off-by: JamePeng <jame_peng@sina.com>

github-actions released this 21 Jan 09:04

main-metal

f1c876d

Update Submodule vendor/llama.cpp 10c98cb..d1e3556