IT之家 · 智能时代T2·Infra
58 pts
Cambricon announced Day 0 adaptation of SenseTime's newly released SenseNova Loop-9B, bringing the model's inference pipeline onto its MLU platform via vllm_mlu. The work covers static unrolling of the Loop layers, reuse of existing Qwen3-Next MLU operators, and Triton-fused optimization of mHC routing, enabling low-memory deployment of this loop-computing architecture on domestic accelerators.