Neutral wdmlpack runtime-package contracts for win-directml: the *.wdmlpack reader/writer/manifest, runtime + source tensor catalogs, and the model artifact lifecycle (inspect/plan/convert) shared by the runtime, encoder and generation modules.

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

Neutral, swing-free Java 21 local generation runtime for Windows DirectML/WARP. Public API in com.aresstack.windirectml.inference.api (GenerationRuntime, GenerationModelHandle, catalog-driven family adapters for Qwen, SmolLM2, Gemma 3, Phi-3 and ...