Cambricon Completes Day 0 Adaptation for SenseNova Loop-9B Model
Cambricon has officially announced the Day 0 adaptation of SenseTime's newly released SenseNova Loop-9B model on its MLU platform using vllm_mlu. This integration successfully establishes the inference pipeline for this novel recurrent model architecture. This development highlights the growing ecosystem support for non-standard, memory-efficient recurrent architectures on domestic AI hardware. By enabling Day 0 adaptation, hardware platforms can rapidly support state-of-the-art models for complex reasoning, coding, and multi-modal tasks. The adaptation involves static loop expansion to convert 32 physical decoding layers into 56 sequential calculations, MLU operator mapping for attention and multi-modal modules, and mHC routing optimization using Triton kernels to minimize launch overhead.
## BACKGROUND
Cambricon MLU is a series of artificial intelligence accelerator chips developed by Cambricon, which utilize dedicated software plugins like vllm_mlu to run large language models. SenseNova Loop-9B is an open-source multimodal language model developed by SenseTime that reuses model layer parameters via a Loop structure to extend computational depth with minimal memory requirements.