Ascend AI Works with FlagOpen to Promote Foundation Model Ecosystem Innovation

Recently, Ascend AI has carried out in-depth cooperation with FlagOpen of Beijing Academy of Artificial Intelligence (BAAI), an open source platform for AI foundation models, and has made significant progress. Both parties have completed the test and evaluation of Ascend hardware on the ResNet50 model by using the integrated AI hardware evaluation engine, FlagPerf. The results proved that Ascend hardware boasts excellent performance. Ascend AI supports single-server and multi-server training of Aquila, an open source large language model (LLM) independently developed by BAAI, and the training performance is industry-leading. The two parties plan to evaluate more popular models such as GLM and LLaMA in the second half of the year to jointly flourish the open source ecosystem for foundation models.

As a renowned AI cutting-edge research institute, BAAI takes the lead in foundation model research in China and also in building an open source foundation model ecosystem. Its AI foundation model Wu Dao is the first language model in China and also the largest one (with 1.75 trillion parameters) in the world. Wu Dao 3.0 is now fully open sourced. FlagOpen is a one-stop, top-notch open source software system collaboratively developed by BAAI along with several enterprises, universities, and scientific research institutes. It comprises foundation model algorithms, models, data, tools, and evaluation, and is designed to facilitate collaborative innovation and open competition. FlagOpen is committed to building a vibrant open source foundation model ecosystem.

In the FlagOpen system, the integrated evaluation engine FlagPerf plays a vital role. Designed by BAAI and multiple AI software and hardware vendors, FlagPerf is an open source, flexible, fair, and objective tool that can efficiently adapt to and evaluate AI software and hardware. Additionally, it can address challenges such as poor compatibility, high heterogeneity of technology stacks, and complex application scenarios. FlagPerf has adapted to nearly 20 classic models covering CV, NLP, voice, and foundation models, and supports training and inference for AI hardware evaluation. In the future, FlagPerf will continue to evolve to support more models and fields, including but not limited to AI servers and graph computing scenarios. It is dedicated to help comprehensively evaluate vendors' software and hardware strengths and keep pace with the development of the AI industry through continuous iteration and upgrade.

FlagPerf has cooperated with four technical vendors including Ascend AI to jointly promote AI hardware evaluation.

Ascend AI is built on the Ascend AI software and hardware platform. The platform includes Atlas series hardware and partner-branded hardware, heterogeneous computing architecture CANN, all-scenario AI framework MindSpore, MindX for Ascend application enablement, one-stop development platform ModelArts, and unified toolchain MindStudio. Ascend AI not only unleashes computing power, but also provides various efficient development tools to enable full-process development of foundation models and accelerate native innovation of big models.

Ascend AI offers a one-stop migration tool that allows foundation models with tens of billions of parameters to be migrated within just two days and model accuracy and performance to be optimized within two weeks. Based on the Ascend inference engine, inference APIs and service systems of various frameworks can be connected to support online distributed inference and quick application rollout.

Currently, Ascend AI supports more than 30 L0 foundation models, including GPT-3, LLaMA-2, GLM, and iFLYTEK Spark. What's more, Huawei has worked with industries such as finance, carrier, Internet, and energy to incubate multiple L1 industry models and L2 scenario-specific models. Ascend AI will adhere to openness and enable customers to develop their own industry and scenario-specific models.

It is a practical and promising direction to combine big data, computing power, and algorithms to build foundation models. To this end, Ascend AI and BAAI will deepen their cooperation in base software and hardware, as well as ecosystem co-development, to give full play to their advantages, achieve capability complementation, promote foundation model innovation, and jointly build a new open source ecosystem for foundation models.