A cloud computing operating system automatic generation method and system based on a self-evolution code large model

Abstract

This invention discloses an automatic generation method and system for a cloud computing operating system based on a self-developed, self-evolving large-scale code model. By constructing a large-scale code model with a hybrid expert structure, grouped query attention, and self-evolution capabilities, and combining technologies such as local low-bit loading, real-time memory release, intelligent code cleaning, and modular scheduling generation, this invention automatically generates the core module code for a cloud computing operating system, including kernel, scheduling, network, storage, and security modules, and automatically completes the compilation and construction. This invention solves the problems of non-standardized code generation, easy memory overflow, inability to generate a complete system, and lack of model evolution in existing large-scale model generation methods. It has advantages such as local high efficiency, clean code, full automation, and self-iterative upgradeability of the model, enabling rapid generation of industrial-grade deployable cloud computing operating systems and significantly reducing the development cost and cycle time of underlying systems.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence large models, automatic code generation, distributed systems, and cloud computing operating systems. Specifically, it relates to a method and system for automatically generating an industrial-grade cloud computing operating system through local inference based on a self-developed self-evolving large code model. Background Technology

[0002] With the deep integration of cloud computing and artificial intelligence, cloud operating systems, as the core foundation for the unified scheduling of computing power, storage, and network resources, face high development challenges, long development cycles, and high costs. Current related technologies suffer from the following significant drawbacks: The code generated by the general large model is messy and contains a lot of markup language, natural language interpretation and redundant content. It cannot be compiled and run directly and needs to be manually cleaned up. Large models consume a lot of video memory for local inference, and are prone to video memory overflow when generating long code or system-level code, making it impossible to generate a complete operating system. Existing code generation models can only generate code snippets and lack systematic architecture capabilities, making it impossible to automatically generate a complete cloud operating system based on modules such as microkernel, scheduling, network, and storage. Most models have a fixed structure and cannot self-optimize or self-evolve after training, resulting in poor adaptability to the generation of underlying system code. The system lacks a fully automated mechanism covering the entire process from code generation and format cleaning to automatic compilation and deployment, resulting in insufficient engineering.

[0003] To address the aforementioned common technical problems in the industry, this invention proposes a method and system for automatically generating a cloud computing operating system based on a self-developed and self-evolving large code model. Summary of the Invention

[0004] (a) Purpose of the invention This invention aims to overcome the shortcomings of existing technologies and provide a fully automated cloud operating system generation solution that is locally deployable, automatically generated, has a clean format, is stable and does not overflow, and has a self-evolving model, thereby achieving unmanned execution of the entire cloud operating system from code generation to compilation and deployment. The system of this invention mainly includes two core components: a self-evolving code large model module and a cloud computing operating system automatic generation engine module. It employs a hybrid expert (MoE) architecture and a grouped query attention mechanism to construct a large model structure specifically for code generation; it uses long context position encoding to support the stable generation of ultra-long code sequences; it improves the model's learning ability and generation quality through distributed training; it has a built-in gradient sparsity analysis mechanism that can automatically identify and prune inefficient expert modules; it supports the model to autonomously generate optimized network structures and dynamically replace original modules, achieving model self-evolution; and it has a built-in code agent that can automatically generate code, write it to files, and perform verification. It includes a local low-bit quantization loading unit, a real-time memory management and release unit, a code intelligent cleaning and format purification unit, a cloud operating system modular scheduling and generation unit, and an automatic saving and compilation building unit.

[0005] The overall process is as follows: Build and train a large self-evolving code model, complete structural optimization and expert pruning, and achieve model self-evolution; Load the model in a low-bit quantization manner in the local environment to reduce hardware resource consumption; Based on the functional modules of the cloud computing operating system, a specialized prompting project is constructed to drive the model to generate corresponding module code; The model output is intelligently cleaned, removing non-code content and formatting tags; Release video memory promptly after each code generation to avoid memory overflow; The code for each module is saved according to language and a compilation script is automatically generated, enabling one-click compilation and deployment. Generates clean and standardized code, free of redundant text and formatting contamination, and can be compiled directly; Local inference is lightweight, with low memory usage, and long code generation does not overflow or interrupt; The model has self-evolution capabilities, automatically pruning and optimizing its structure to continuously improve the generation effect; The entire process is automated, and a complete cloud computing operating system can be generated without human intervention. It supports modular system-level code generation, which can fully cover core modules such as kernel, scheduling, network, storage, and security; The generated cloud operating system natively supports distributed environments and heterogeneous computing power scheduling, and has industrial-grade availability. IV. Detailed Implementation The model is built on a deep learning framework, employing a hybrid expert feedforward structure and a grouped query attention structure, and using long context positional encoding to extend the generation length. During the training phase, distributed parallel training and gradient accumulation strategies are used to improve training stability and efficiency. After training, the model enters a self-evolution phase: it analyzes parameter activity through gradient information, prunes inefficient modules, and autonomously generates optimized network structure code to dynamically replace existing modules, achieving structural upgrades without human intervention. The model has a built-in code execution agent that can write the generated code to a file and run it automatically, forming a generation-verification closed loop. Configure the local model path and code output directory, and initialize the cleaning rules and generation process; Load the model using low-bit quantization to achieve efficient local inference; Based on the cloud computing operating system architecture, the code for modules such as kernel, scheduler, network, storage, security, and main control is generated sequentially. Clean the output by removing markers and invalid text, keeping only the pure code; After each generation, the intermediate tensors are released and the video memory is cleared. Save the code by language and automatically generate compilation scripts to build executable programs with one click.

[0007] The system loads a self-developed, self-evolving large-scale code model onto the local hardware environment, initiating an automatic cloud operating system generation process. The system sequentially generates code for each core module, cleans it, and saves it as a project file. An automatically generated compilation script then completes the compilation, resulting in a runnable cloud operating system program that enables resource scheduling, distributed communication, high-availability storage, inherent security, and intelligent self-healing. The entire process requires no manual code modification, avoids memory overflow, and the generated result can be directly deployed and used.

[0008] Instruction manual with accompanying drawings

[0009] Appendix Figure 1 A cloud computing operating system based on a self-developed, self-evolving large-scale code model automatically generates the overall system architecture diagram. Brief description: This diagram is a vertical arrow text architecture diagram, showing the hierarchy and flow relationship of the overall system modules and units.

[0010] Detailed explanation: The diagram uses vertical arrows to connect the self-evolving code large model module, the cloud computing operating system automatic generation engine module, and all functional units under these two modules. The arrows represent the direction of data and instruction transmission, clearly demonstrating the overall architectural logic of the collaborative cooperation of each unit to complete the fully automatic generation of the cloud computing operating system.

[0011] Appendix Figure 2 Schematic diagram of the internal structure of the self-evolving code large model Brief description: This diagram is a vertical arrow text structure diagram, showing the connection relationship between the internal hierarchy and functional units of the model.

[0012] Detailed explanation: The diagram uses vertical arrows to label the network layers and functional units within the model in sequence. Bidirectional arrows represent the interactive optimization relationship between the self-evolving dynamic replacement unit and the hybrid expert feedforward layer. Unidirectional arrows represent the data processing flow, fully presenting the internal working logic of the entire process of model encoding, inference, gradient analysis, self-evolving replacement, and code execution.

[0013] Appendix Figure 3 Flowchart of an automatic generation method for cloud computing operating systems based on a self-developed and self-evolving large code model Brief description: This diagram is a vertical arrow text flowchart, showing the complete execution steps and sequence of the method.

[0014] The diagram uses vertical unidirectional arrows to connect all execution steps. Starting from the process, it sequentially goes through model training, self-evolution optimization, local loading, code generation, cleaning, GPU memory release, code saving, compilation and building, and finally outputs the cloud computing operating system and ends the process. The arrows clearly indicate the order of execution of the steps, fully presenting the entire technical path of this invention from model preparation to system generation.

Claims

1. A cloud computing operating system automatic generation method based on self-research and self-evolution code large model, characterized in that, Includes the following steps: (1) Construct and train a self-evolving code large model, adopt a hybrid expert architecture and group query attention mechanism, use gradient information to prune inefficient modules, and have the model autonomously generate an optimized structure to complete dynamic replacement, thereby realizing model self-evolution; (2) Load the trained model locally with low-bit quantization to reduce resource consumption; (3) Based on the functional modules of the cloud computing operating system, construct the prompts and drive the model to generate the corresponding industrial-grade code for the modules; (4) Perform intelligent cleaning on the model output content, removing non-code text and formatting tags; (5) Release intermediate cache and video memory after each code generation to avoid resource overflow; (6) Save code by module and programming language, and automatically generate compilation scripts to realize one-click compilation of cloud computing operating system.

2. The method of claim 1, wherein, The self-evolving code large model adopts a long context position encoding method, which supports the continuous generation of ultra-long sequence codes.

3. The method of claim 1, wherein, The model self-evolution includes: identifying and pruning inefficient experts based on gradient sparsity, automatically generating an optimized network structure, and dynamically replacing the original model modules.

4. The method of claim 1, wherein, The intelligent code cleaning includes removing marker symbols, filtering natural language interpretations, eliminating redundant language, and retaining pure code that can be directly compiled.

5. The method according to claim 1, characterized in that, The memory release includes deleting intermediate tensors and forcibly clearing the device cache to ensure that continuous generation of long code does not overflow.

6. The method according to claim 1, characterized in that, The cloud computing operating system module includes a kernel, global scheduling, distributed network, distributed storage, intelligent scheduling engine, security engine, and system main control entry point.

7. A cloud computing operating system automatic generation system based on a self-developed, self-evolving large-scale code model, characterized in that, include: The self-evolving code large model module is used for model building, training, pruning, and structural self-evolution. The local quantization loading module is used to load the model locally in a low-bit manner; A modular scheduling generation module is used to generate code in modules according to the cloud operating system architecture; The code cleaning module is used to clean up the output content and retain standard compilable code; The video memory management module is used to release resources in real time and prevent overflow during the inference process; Automatic build modules are used to categorize and save code and generate build scripts.

8. The system according to claim 7, characterized in that, The self-evolving code large model module includes an encoding layer, a grouped query attention layer, a hybrid expert layer, a gradient analysis unit, and a dynamic replacement unit.

9. The system according to claim 7, characterized in that, The code cleaning module removes formatting tags and invalid content through rule matching, outputting clean, engineering-grade code.

10. The system according to claim 7, characterized in that, The system can automatically generate, organize, and compile a complete cloud computing operating system without human intervention.

Images