Your Laptop's Free AI Just Got Better at Finding the Right Answer

llama.cpp put out a pre-release update (b11223, released 27 Sep) that fixes reranking for causal LLM rerankers like Qwen3 and Qwen3-VL. Rerank models — the ones that re-sort results so the best ones come first — come in two flavors: bidirectional cross-encoders (BERT and similar) that need all tokens in a single physical batch, and causal LLMs repurposed as rerankers that can use chunked prefill like any other decoder. The server previously rejected all RANK-pooling inputs larger than n_ubatch, and the graph builder hardcoded QWEN3/QWEN3VL architecture checks to decide last-token pooling — which broke long-document and multimodal reranking for causal models. The fix exposes llama_get_causal_attn so the server can check the effective runtime attention type, exposes llama_model_is_causal for the static architectural property from GGUF metadata, and lets can_split() permit chunked prefill for RANK pooling when the context is causal. The inline architecture check in the graph builder was replaced with the same cparams.causal_attn predicate, removing duplication.