b9286

B9286 introduces support for HybridDNATokenizer as its own vocabulary type (LLAMA_VOCAB_TYPE_HYBRIDDNA), enabling proper tokenization of DNA sequences by routing raw text through a DNA-aware splitter while handling non-DNA text through standard BPE. The release also fixes a VRAM leak in the server's multi-token prediction models by ensuring proper cleanup of speculative decoder resources during context destruction.

Source: llama.cpp Releases | 2026-05-22