What changed
Signed prerelease b11223 allows RANK-pooling batches to split for causal LLM rerankers such as Qwen3 and Qwen3-VL. The server now checks the effective runtime attention mode instead of hardcoded architectures, so long-document and multimodal reranking can use chunked prefill rather than failing above the microbatch limit.
Why it matters
The change removes a correctness and availability boundary for long or multimodal reranking while respecting runtime attention overrides. It matters to retrieval pipelines whose inputs exceed `n_ubatch`.
What to do
Canary b11223 with the exact Qwen reranker, attention override and longest expected document or image batch. Compare rankings against a trusted baseline, test inputs around `n_ubatch`, capture memory and latency, and retain b11222 for rollback.
Caveat
This is a fast-moving prerelease, not a stable checkpoint. The signed artifact and project explanation establish the code change, but cross-model ranking quality, performance and regression behavior remain independently unverified.