The "new king bomb" of the large-scale model era is here
Just yesterday
Baidu has released an epoch-making product
↓
An intelligent computing operating system in the era of large models
Baidu Intelligent Cloud Wanyuan
Why is this an "epoch-making" product?
It's just that times have changed a lot
In the traditional IT era, we have Linux and Windows
In the era of cloud computing, we have a cloud platform and a cloud operating system
Nowadays, large models and AI are unstoppable
We also need one that was born for this era
"Intelligent Computing Operating System"
And "Wanyuan" came into being just in time
Become the industry's first "next-generation intelligent computing operating system"
↓
The overall structure of "Wanyuan".
It is mainly divided into three layers
(1) Kernel layer, (2) Shell layer, and (3) Utility layer
↓
Traditional OS runs on a PC or server
Wanyuan also has to run on the infrastructure base
This base is the cloud infrastructure of the last era
However, in addition to the traditional cloud service storage and computing network capabilities
Additional intelligent computing infrastructure will be added
The first layer is the kernel layer
Further up is the first layer of the most critical operating system
That is, the kernel, the kernel layer
Friends who have played Linux know it
The kernel is the basic disk of the operating system
Whether the kernel is stable or not, and whether the architecture is good or not, determines the lower limit of the OS
And the kernel layer of "Wanyuan".
It consists of two main parts
↓
❶Baidu Baige · AI heterogeneous computing platform
Baige's core competency
It's about hiding the underlying complex cloud-native infrastructure
Cloud infrastructure is already complex
In order to support intelligent computing, it is necessary to introduce massive computing power chips
For example, train a good large model
Usually, you need to schedule computing power clusters with the size of 10,000 cards
This is for traditional cloud platforms that are only good at manipulating general-purpose computing power
Cluster design, scheduling, and fault tolerance are all major challenges
And Baige can easily control tens of thousands of computing power acceleration cards
Like a card, work together and scale linearly
The proportion of effective training time reached 98.8%
Linear acceleration ratio and bandwidth effectiveness of more than 95%
At the same time, due to the instability of the supply of intelligent computing chips
In practice, multi-chip heterogeneity must be considered
Get rid of dependence on a single specific chip
So as to avoid supply chain risks
Now, on the Baige platform, it has been realized
Heterogeneous chip hybrid training on the scale of 100 cards and a single task
In other words, Baige has achieved the difference between shielding chips
It allows users to freely choose different chip combinations
❷ Various preset large models
Another important component of the kernel is the large model
Pre-trained large models are the soul of the AIGC era
Wanyuan kernel layer has a variety of built-in large models
(Powerful ERNIE4.0, 3.5, lightweight ERNIE Speed/Lite/Tiny, Wenxin visual large model, third-party large model)
As long as the developer has an idea for the big model
The kernel layer can be used to the bottom
The second layer is the shell layer
Getting the kernel right is like getting through the underlying logic
So, how do you call the kernel for the user?
Traditional OS has a shell, and Wanyuan is no exception
Qianfan ModelBuilder is equivalent to this shell layer
Help users realize kernel invocation and model secondary development
ModelBuilder provides a fine-tuning tool for models
The user can easily make a move on the base model
There is no need to train large models from scratch
It even needs to be based on the Prompt project
You can quickly and easily come up with a model that suits your business
At the same time, ModelBuilder can be based on the MoE pattern
Depending on the difficulty of the specific task
Flexibility to select and combine the most appropriate model in the core
Achieve the optimal combination of effect and cost
The average inference cost is reduced by more than 30%.
The third layer is the development tools layer
Next, we move on to the top layer: the tool layer
It can be understood as application software/utility software in traditional OS
In the era of large models, Wanyuan OS provides a golden combination
AppBuilder和AgentBuilder
They are both powerful & easy-to-use application development platforms
Integrate workflow orchestration and management plug-ins
Rapid development of AI-native applications and agents
Developers don't need to face traditional OSs like they used to
You need to be proficient in Python, C, and Java
Only need to be based on two "Builders"
Adopt a natural language that is "homely".
It can achieve rapid development and easy launch
In this way, Wanyuan goes from the kernel to the shell to the development tools
The organic combination between layers has been achieved
Comparing the structure of the traditional OS, we can see:
A new operating system in the era of large models
It's got an inner taste
However, Wanyuan continues to evolve
For developers (including ordinary people).
They can ignore all the complexity of computing infrastructure
This also includes the complexity of various languages and architectures
It only needs to face Wanyuan and carry out minimalist human-computer interaction
所想即所得,所Prompt即所得
And for the majority of partners and customers of Baidu Intelligent Cloud
It is also possible to span two generations of OS in one step
On top of "Wanyuan".
Reshape the digital world with AI-native applications and agents
As Baidu Intelligent Cloud Shen Dou said
"Simplicity is the ultimate sophistication"
Baidu is using its deep understanding of the cloud and AI
Step by step, step flat, dedication
Let intelligence generate infinite possibilities