BERT-class encoder pipelines (MiniLM, E5, cross-encoder reranker) on DirectML with CPU fallback, embeddings, and reranking with per-layer coalesced GPU submissions.