According to reports, the training and inference integration deployment of Spark all-in-one machine can be used for question answering system, dialogue generation, knowledge graph construction, intelligent recommendation and other fields of application, with large model pre-training, multi-mode understanding and generation, multi-task learning and migration capabilities.

Spark all-in-one machine also tailor-made the hardware for the training algorithm and reasoning application of Spark cognitive intelligence large model, which can greatly reduce the use cost of enterprises. It can directly provide 5 customized optimization modes such as dialogue development, task choreography, plug-in execution, knowledge access, and prompt engineering, as well as more than 10 out-of-the-box rich scenario packages such as office, code, customer service, operation and maintenance, marketing, and procurement, and supports 3 model sizes for users to choose.

It is worth noting that last month, iFlytek announced that iFlySpark will join forces with Centeng AI to create a new base of general intelligence based on China’s independent innovation. On the one hand, iFlyspark cognitive large model is based on the

1771-IVN

integrated design of training reasoning, achieving technical breakthroughs in the sparse and low-precision quantification of large models, which can effectively adapt to Senteng AI and accelerate the application and iteration of large models in the industry. On the other hand, with Centeng AI as the core, software and hardware collaborative optimization, to build a large model training cluster with concentrated computing power, superior performance, stable supply and data security.

In the “IFlyspark Cognition Large Model V2.0 upgrade Conference” speech on the same day, Liu Qingfeng explained in more detail, “Huawei and IFlytek jointly combined the high computing power AI chip, high-performance operator library, multi-card high-speed interconnection and distributed storage on the software and hardware platform and software support tools of Centeng AI. In particular, we jointly identify and polish the most important operator libraries needed for artificial intelligence. Then on this basis, the architecture of iFLYtek’s training and data closed-loop full process design, as well as training and reasoning integrated design of self-developed large model training platform, the middle is to support large-scale heterogeneous computing power compatibility, but also support hybrid cloud architecture easy to expand, so that we see today iFLYtek Spark V2.0 demonstration and all products, architecture on a safe and controllable platform.”

It is reported that in addition to continuously improving the general capability base, “IFlyspark Cognition large model V2.0” focuses on breaking through code capabilities and multi-modal interaction capabilities. According to Liu Qingfeng, on HumanEval, a public test set of code capabilities built by OpenAI, IFlystar Fire model V2.0’s ability to write code based on Python and C is close to the level of ChatGPT, with a gap of only 1% and 2%, and is expected to fully surpass ChatGPT on October 24 this year. GPT-4 will be officially launched in the first half of next year.