Deprecation of n-gram logic in kv-cache for ggml-org/llama.cpp

The latest release of ggml-org/llama.cpp, titled b10750, introduces significant changes to the kv-cache mechanism. The update removes n-gram logic from the kv-cache and optimizes the lookup process for previous tokens. The seq_pos index now stores (pos, cell) pairs in a std::set, ensuring distinct entries for repeated positions and enabling efficient removal of specific cells. The seq_pos_tok_le() function replaces the old window lookup and fallback mechanism, providing logarithmic-time token retrieval. The get_prev_tokens() function has been streamlined to perform direct lookups, and the for_each_token_in() function has been removed entirely. Performance benchmarks on Qwen3.8-Flash-Next UD-Q4_K_XL show a throughput improvement of 4.9% for tg, while pp remains unchanged at ~2720 t/s.

Operators should verify whether any custom integrations or workflows rely on the removed n-gram logic or the now-deprecated for_each_token_in() function. These changes may require updates to dependent codebases or workflows that interact with the kv-cache. Additionally, testing the new kv-cache behavior in staging environments is recommended to ensure compatibility and performance improvements before deploying this release in production.