MindIE - Ascend inference engine product page

PRIMARY VENDOR SOURCE for the software side. MindIE 3.1.0 community release 2026-08-21. Architecture: MindIE Motor (Ascend inference serving), MindIE SD (multimodal generation, open-sourced 2025-12-17), MindIE LLM, MindIE Turbo model acceleration library, on top of CANN, with MindSpore and PyTorch as AI frameworks. MindIE Motor fronts third-party engines vLLM and SGLang for text generation - so vLLM/SGLang run ON Ascend via CANN, which is the practically important fact for anyone porting. Also documents load-aware PD-disaggregation scheduling, K8s templated deployment, and recovery targets (60 s self-healing faults, 7 min system-level). Source repos on gitcode.com/Ascend.