Author丨Poplar
Editor: Lin Xi
On September 7, the Tencent Hybrid Model was officially unveiled. According to reports, this is a general-purpose large language model self-developed by Tencent Full-link, with a scale of more than 100 billion parameters, a pre-trained corpus of more than 2 trillion tokens, and a strong ability to create Chinese, logical reasoning in complex contexts, and reliable task execution capabilities.
In the field of general large models, Tencent is almost the last technology giant. Low-key has always been Tencent's style, but not competing for popularity does not mean that there is nothing to do, before the advent of Mixed Yuan, there have been many iterations within Tencent.
The Chinese market has shown a fanatical side to large models, and as of July, more than 130 large models have been released in China. However, the scale and quality of these large models are uneven, especially after more than half a year of industry exploration, the industry has become more and more aware that China's large models need to look up at the stars and be down-to-earth.
Among them, "looking up at the starry sky" is to benchmark against the world's top models, catch up or even surpass, there are not many domestic companies with this kind of strength, so technology giants including Tencent are placed on high hopes. However, because the GPU available in China is not the latest generation, the shortage of computing power is inevitable, which means that enterprises need longer time to train models.
In the face of this force majeure, Chinese companies are also looking for some innovative ways to make up for the lack of computing power, such as Tencent Mix, which uses a more stable network and higher quality data to train to improve training efficiency.
And for Tencent Mixed Yuan to carry out these behind-the-scenes support is Tencent Cloud. Qiu Yuepeng, Vice President of Tencent Group, COO of Cloud and Smart Industry Business Group, and President of Tencent Cloud, said, "Since we started supporting large model training, we have comprehensively upgraded our cloud infrastructure, from storage, network to computing. ”
Qiu Yuepeng
Build the foundation for large models
For enterprises that want to build large models by themselves, the first thing to solve is the problem of computing power. On the cloud, the threshold for using computing power can be greatly reduced.
As large models enter the era of trillions of parameters, the demand for computing power is also growing, and the computing power provided by a single server is very limited, so it is necessary to connect a large number of servers to create a large-scale and distributed high-performance computing cluster. At this time, how to collaboratively optimize the computing power, network architecture and storage performance of a single machine will also determine the final computing power level.
Qiu Yuepeng said that based on the joint optimization of computing, network and storage, Tencent Cloud has created a new generation of HCC high-performance computing power cluster, which can improve the computing power performance of the computing power cluster by three times compared with the previous generation by using the latest generation of Xingxinghai's self-developed server and equipped with the latest generation of domestic GPU cards.
In October last year, Tencent completed the first trillion-parameter AI model training, shortening the training time from 50 days to 11 days under the same dataset. If based on the new generation cluster, the training time will be further reduced to 4 days.
In the past year, in addition to Tencent Mix, Tencent Cloud has also helped startups such as Baichuan Intelligence, Zhipu Technology, and MiniMax quickly build their own large models. Practical data shows that whether it is access to mature end-to-end solutions or direct access to high-performance computing power, the cloud is the best carrier for enterprises to build large models.
In addition to computing power, building large models requires massive amounts of high-quality data. To this end, Tencent Cloud has built a cloud-native data lake warehouse and vector database on the cloud. They act like "filters" that can clean and classify large amounts of raw data.
According to Qiu Yuepeng, in terms of performance, Tencent Cloud's cloud-native data lake warehouse has taken the lead in supporting millions of data updates per second and terabyte-level massive throughput capabilities. Combined with the newly released vector database, it can achieve a scale of 1 billion vector retrieval and 10 billion levels of offline data cleaning, and control the delay in milliseconds.
The measured results show that compared with traditional methods, Tencent Cloud's high-performance data processing engine can improve the raw data cleaning performance by more than 40% and reduce the overall operation cost by 50%.
In addition, in response to the needs of many enterprises to build exclusive industry large models based on general large models, Tencent Cloud has also further improved the training inference framework Angel and TI platform toolchains in the past few months.
Qiu Yuepeng said that at the level of training inference, Tencent's self-developed machine learning framework Angel helped Tencent efficiently complete the training of mixed elements. Now enterprises and developers can also use the framework through Tencent Cloud to train their own large models.
The tools provided by the TI platform cover the entire link of environment preparation, code debugging, performance evaluation, and deployment, which can be finely tuned and accelerated for large models in every link. Therefore, with the help of the TI platform, enterprise users can quickly try a variety of large models, and fine-tune their own model solutions in a short time according to the needs of their own business scenarios.
It can be seen that whether it is for general large models or industry large models, Tencent Cloud is constantly improving the relevant foundation capabilities. On this basis, Qiu Yuepeng also mentioned a topic of great concern to the outside world, that is, the security of large models.
In response to the privacy security of large models, Tencent Xuanwu Lab has developed the industry's first lightweight privacy protection solution, which can be deployed by enterprises on the terminal side. It is reported that this scheme adopts generative two-way desensitization technology, which can desensitize part of the prompt content when the user interacts with a large model, thereby protecting the security of the user's private data.
In terms of content security, Tencent's Tianyu AIGC full-link content security solution covers the entire process of content security construction from model training to content generation and post-optimization, ensuring that large models are trusted, reliable, and available.
Qiu Yuepeng said that in the face of the trend of large models, each enterprise will make its own choices, and what Tencent Cloud needs to do is to hope that each layer of the large model, such as the computing power layer, data layer, model layer, application layer, etc., can provide corresponding products.
Redefine tools on the cloud
For the relationship between the cloud and the large model, Qiu Yuepeng proposed that on the one hand, high-performance cloud computing power has become the best booster for the large model, and the application of the large model capability has also been implemented through cloud services. On the other hand, the big model will also redefine the cloud tools and significantly improve their efficiency, so that enterprises can obtain more cost-effective, convenient and easy-to-use cloud products.
For example, in business security scenarios, risk control modeling is a very important link. Through the in-depth application of the concept of large models, the "universal" risk control capabilities can be transformed into customized risk control systems and capabilities that are optimized for the enterprise's own scenarios, so as to achieve dynamic risk governance.
Tencent Cloud's big risk control model can automatically generate custom models based on the "sample prompt" mode. Taking the financial risk control big model as an example, financial institutions only need a small number of prompt samples to build a risk control model that adapts to their own business characteristics, from sample collection, model training to deployment and online without manual participation, and the modeling that originally took 2 weeks to complete can now be completed in only 2 days.
In the code creation scenario, Tencent Cloud's AI code assistant can help developers create all code in one IDE, including answering technical questions, generating business code, unit testing, and diagnosing code defects.
According to Tencent Cloud, in order to achieve this goal, Tencent Cloud uses a large number of internal and external selected corpus for training and fine-tuning at the model layer, and makes a lot of targeted optimization for typical coding scenarios at the application layer.
Measured data shows that in the most practical code completion scenario in production, the code generation rate increases to more than 30%, that is, more than 300 lines of code submitted by engineers are generated by AI, and the code adoption rate exceeds 30%.
In fact, Tencent Cloud's products are very agile, both in terms of technical support for large models and in the application of large model technology. This is not unrelated to the product-focused strategy proposed by Tencent Cloud and the Smart Industry Business Group.
Qiu Yuepeng pointed out that Tencent Cloud's transformation goal is to do a good job in products and build capabilities around product needs. In this process, the first thing to do well is the selection work. "Tencent Cloud does not do all products, and we are gradually making choices to see which products Tencent really has accumulated and really has extensive market demand, which are the focus of our efforts."
For example, some time ago, the chairman of a company came to him to develop an app based on Tencent Cloud's capabilities. However, Tencent Cloud did not answer this demand, Qiu Yuepeng said, Tencent Cloud can not help enterprises make apps, but can provide the ability to develop apps, such as cloud development, instant messaging, database, etc.
"For Tencent Cloud at this stage, this is a very critical point," Qiu Yuepeng said, "Many customers put forward specific solution capabilities, rather than general product capabilities." Tencent Cloud now hopes to continue to iterate on the product's capabilities, but if the requirements are not clearly identified, it is likely to complete a project, which will drag down the long-term development of the product. ”
This ability to continuously iterate on products has also been demonstrated in the big model. For example, when the wave of AI came, Tencent Cloud found that customers had shortcomings in their upstream and downstream retrieval and understanding capabilities, so it quickly productized the vector database used by Tencent internally and quickly introduced it to the market.
Qiu Yuepeng said frankly that to do transformation, the team must be unified and form a common goal. "In this process, we also spent half a year 'loosening the soil', so that everyone can realize the meaning of change and adjustment, and do the difficult but right things. This is indeed more torturous, a bit like weight loss and sugar abstinence, the body needs to adapt to the process. ”
But after more than a year, Tencent Cloud has become more determined and more adaptable to the transformation. Especially when technological innovation comes, Tencent Cloud can clearly determine which products should be provided by cloud vendors. For example, in the face of large models, Qiu Yuepeng said that the research and investment in the relevant capabilities of large models must be right, and all cloud vendors must do a good job in supporting general artificial intelligence, which is a mandatory question, and no one can not answer this question.
SFC
Editor of this issue Liu Xueying