llama.cpp server gains image-text embeddings for Qwen3-VL

llama.cpp's HTTP server can now take combined image and text inputs on its embedding and rerank endpoints, opening multimodal retrieval workflows for Qwen3-VL embedding and reranker models.

Timothy Wang added the support for RAG pipelines that need to embed or score content mixing pictures and prose. Embeddings follow OpenAI-compatible conventions on the usual embeddings route; reranking follows Jina and Text Embeddings Inference conventions on the rerank routes and their aliases.

The change also relaxes batching for causal decoder rerankers in the Qwen3 family. Earlier rerankers were bidirectional and required every candidate in one batch; Qwen3-style models are repurposed language models and do not need that constraint, so the server can split those jobs across batches.

Embedding and rerank calls are now fully stateless. The server no longer reuses KV-cache state between discrete requests, which closed a CUDA memory fault that appeared once vision chunks were in play: skipped re-encoding left stale pointers that crashed the backend.