b9270

Release b9270 adds support for the HybridDNATokenizer used by the Carbon-3B model family, implementing a new BPE pre-type for tokenizing DNA sequences. The tokenizer handles DNA k-mers with fixed 6-mer chunking while processing text through standard BPE, with bases outside ACGT mapped to a special token. The release also fixes a VRAM leak in the server context where speculative decoder and draft model resources were not properly freed during sleep/resume cycles, which could cause out-of-memory errors.

Source: llama.cpp Releases | 2026-05-21