The China Telecom Research Institute, in conjunction with the Computing Institute of the Chinese Academy of Sciences, Huawei Shengteng, Mu Xi, etc., successfully completed the industry's first cross-architecture large-scale model inference technology verification for domestic computing power, relying on China Telecom's cloud network integration technology pilot verification platform. Through self-developed Triton's cross-architecture compiler and unified large model operator library, the same set of operator source codes can run seamlessly on Nvidia, Shengteng, and Mu Xi chips. End-to-end inference tests of more than 10 mainstream open source models on three types of chips have been completed, and the adaptation cycle of large model operators has been reduced from “weekly” to “advanced”, and the performance has reached 90% of the native operator library.

The China Telecom Research Institute, in conjunction with the Computing Institute of the Chinese Academy of Sciences, Huawei Shengteng, Mu Xi, etc., successfully completed the industry's first cross-architecture large-scale model inference technology verification for domestic computing power, relying on China Telecom's cloud network integration technology pilot verification platform. Through self-developed Triton's cross-architecture compiler and unified large model operator library, the same set of operator source codes can run seamlessly on Nvidia, Shengteng, and Mu Xi chips. End-to-end inference tests of more than 10 mainstream open source models on three types of chips have been completed, and the adaptation cycle of large model operators has been reduced from “weekly” to “advanced”, and the performance has reached 90% of the native operator library.