Method, medium and equipment for constructing container cloud agent based on MiniMax-Text-01 large model

Abstract

The invention provides a method for constructing a container cloud agent based on a MiniMax-Text-01 large model, a medium and equipment, belongs to the technical field of automatic operation and maintenance, and can perform fault detection, flow prediction and resource elastic expansion and contraction according to real monitoring data. The intelligent agent constructed by the invention can automatically pre-judge and give an optimal scheme for capacity expansion or migration when finding that the load of the server is abnormal, so that the frequency and the duration of manual intervention are reduced. In a multi-cloud and hybrid cloud architecture, an intelligent agent can cooperatively manage calculation, storage and network resources on a heterogeneous platform, and cross-cloud load balancing and cost optimization are realized. In addition, the intelligent agent can help a data analysis and machine learning team to better utilize a cloud GPU / TPU cluster and automatically select a proper computing power and storage scheme, so that the training and prediction efficiency of mass data is remarkably improved. In the scene of Internet of Things and edge computing, the intelligent agent can link the cloud and the edge nodes, intelligently dispatch the computing power and bandwidth of each node, and find and correct abnormal data streams in time.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of automated operation and maintenance, and particularly relates to a method, medium and device for constructing a container cloud agent based on a MiniMax-Text-01 large model. BACKGROUND

[0002] Container clouds are in a stage of rapid development and popularization, and more and more enterprises and organizations choose to run their services, store data and deploy artificial intelligence applications in the cloud. In the face of a large pool of resources and diversified computing needs, the concept of "agent" has gradually attracted attention: it can not only accurately understand user instructions and respond in natural language, but also perform cloud resource management, scheduling and collaborative optimization tasks. To achieve this vision, it is necessary to have strong language understanding and reasoning capabilities at the bottom level; and MiniMax-Text-01 is exactly a large model with good language generation and understanding capabilities, which can provide excellent multi-round dialogue, inference and instruction execution capabilities.

[0003] Through the capabilities of MiniMax-Text-01, an agent suitable for the container cloud scenario can be constructed from data preparation, model pre-training, instruction fine-tuning, cloud deployment and reasoning optimization, etc. However, how to make the agent in the cloud, in the face of multi-source and diverse user needs, not only generate reliable natural language replies, but also efficiently execute actual operations including resource automatic expansion and reduction, task scheduling, log analysis, operation and maintenance alarms, and meet the requirements of enterprise-level applications in terms of security compliance and scalability, there is still no effective solution in the prior art. SUMMARY

[0004] The present application provides a method, medium and device for constructing a container cloud agent based on a MiniMax-Text-01 large model, which overcomes the deficiencies of the prior art. The agent constructed by the MiniMax-Text-01 model has stronger performance and consumes less computing power than traditional LLM agents. The optimized linear attention mechanism replaces the traditional Transformer mechanism to realize lightweight agent construction and improve system efficiency.

[0005] To achieve the above purpose, the application adopts the following technical solutions:

[0006] In a first aspect, the application provides a method for constructing a container cloud agent based on a MiniMax-Text-01 large model, comprising the following steps:

[0007] S101: Collect training required corpus and process it;

[0008] S102: Select the MiniMax-Text-01 large model to construct the model, and pre-train it with autoregressive language modeling as the pre-training target;

[0009] S103: Fine-tune the pre-trained model with instructions to obtain a trained model;

[0010] S104: Deploy the trained model to a container cloud environment;

[0011] S105: Perform inference optimization on the model deployed to the container cloud environment.

[0012] Optionally, in step S101, the corpus includes: public cloud and hybrid cloud technology white paper; DevOps, microservice architecture, container orchestration Docker, Kubernetes, virtualization technology VMWare related documents; product description and user manual of major cloud service providers; architecture design and developer documentation of open source project OpenStack;

[0013] The corpus also includes operation and maintenance dialogue logs, troubleshooting cases and cloud platform API examples, and is arranged into a dialogue form dataset.

[0014] Optionally, in step S101, the collected corpus is processed as follows: data cleaning and labeling; adding labels to key dialogues or operation logs; shuffling and dividing the data into shards.

[0015] Optionally, in step S102, multiple GPUs A100 or TPU clusters are selected for the training environment, and a combination strategy of data parallelism and tensor parallelism is used; the learning rate scheduling adopts the Warmup and Cosine Decay strategies; the training is set with indicators including language modeling perplexity and container cloud question and answer accuracy.

[0016] Optionally, in step S102, the model uses a hybrid attention mechanism of Lightning Attention and Softmax Attention; a dynamic routing module based on task priority is designed in the model, which is used to dynamically select attention calculation methods from Lightning Attention and Softmax Attention according to the complexity of the task, the context length and real-time requirements;

[0017] Lightning Attention adopts a distributed file system Ceph or a dedicated communication protocol RDMA; in distributed training, gradient compression techniques Top-k gradient compression and quantized gradient transmission are used to realize the process: at each parameter update, the current node calculates the full gradient of the model, and according to the absolute value of the gradient, the top-k gradients with the largest absolute value are selected; the indexes of the selected Top-k gradients are recorded; the selected Top-k gradients are quantized; the quantized Top-k gradients and the corresponding indexes are sent to other nodes through the communication network, and the receiving node restores the sparse gradient matrix according to the received index and quantized gradient value, and combines the local unselected gradient to update the parameters;

[0018] Lightning Attention separates data loading, preprocessing and model calculation; performs attention calculation on new data segments, and caches and reuses the attention weights of historical data; distributes preprocessing tasks to edge devices for execution;

[0019] Lightning Attention dynamically adjusts the allocation of attention calculation tasks by monitoring the real-time load and resource usage of the nodes;

[0020] Lightning Attention designs dedicated embedding layers and attention weight sharing mechanisms for different modal data;

[0021] Lightning Attention is designed as a cloud-native model service and encapsulated as a container image.

[0022] Optionally, in step S103, the collected dialogue form data set is sorted into "instruction-answer" pairs, and a certain proportion is randomly selected for review, and the "instruction-answer" pairs after review are used to fine-tune the model.

[0023] Optionally, in step S104, the trained model includes Tokenizer, weight file and inference script, which are packaged into a Docker image, and a microservice is constructed to provide API externally; when deployed, a unified API gateway is built for identity authentication, traffic control and security detection.

[0024] Optionally, in step S105, the model is quantized to INT8, and the attention caching mechanism is used for multi-round dialogue scenarios.

[0025] In a second aspect, the present application provides a computer readable storage medium storing a computer program, which causes a computer to execute the method of constructing a container cloud agent based on the MiniMax-Text-01 large model according to the first aspect.

[0026] In a third aspect, the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method for constructing a container cloud agent based on a MiniMax-Text-01 large model according to the first aspect when executing the computer program.

[0027] The present application has the following beneficial effects:

[0028] (1) The present application can perform fault detection, traffic prediction, and resource elasticity according to real monitoring data. The agent can automatically predict and provide the best solution for expansion or migration when discovering abnormal server load, thereby reducing the frequency and duration of manual intervention. In a multi-cloud and hybrid cloud architecture, the container cloud agent can collaboratively manage computing, storage, and network resources on heterogeneous platforms, achieve cross-cloud load balancing and cost optimization, and make enterprises more flexible in migration and deployment. In addition, the container cloud agent can also help data analysis and machine learning teams better utilize cloud GPU / TPU clusters, automatically select appropriate computing power and storage solutions, and significantly improve the training and prediction efficiency of massive data. In the Internet of Things and edge computing scenarios, the agent can link the cloud and edge nodes, intelligently schedule the computing power and bandwidth of each node, and timely discover and correct abnormal data flow, providing strong background support for smart cities, smart factories, and other applications.

[0029] (2) The present application uses a container cloud agent based on a MiniMax-Text-01 large model to solve the problem. Relying on conventional large-scale language models (LLM) and their standard multi-head attention (Multi-Head Attention) mechanism, it not only occupies a large amount of computing power during training and inference, but also cannot well balance lightweight and efficiency. The present application replaces the traditional Transformer mechanism with an optimized linear attention mechanism and incorporates the technical concept of MiniMax-Text-01 into the model architecture, achieving lightweight agent construction. While ensuring stronger model performance, it significantly reduces computing power consumption, thereby providing more efficient and flexible service capabilities in the cloud environment. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 is a method flowchart for constructing a container cloud agent based on a MiniMax-Text-01 large model. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the present application will be described in detail below with reference to the accompanying drawings in the embodiments of the present application.

[0032] In an embodiment, the present application proposes a method for constructing a container cloud intelligent agent based on a MiniMax-Text-01 large model, as shown in Figure 1 The method mainly includes the following steps:

[0033] S101: Data preparation and corpus construction.

[0034] First, the multi-source and multi-level corpus required for training needs to be collected and organized. Specifically, it includes general language data such as Wikipedia, Internet articles, e-books, etc., and professional documents in the container cloud field such as public cloud, hybrid cloud technology white paper, DevOps, container orchestration, virtualization related materials, product manuals of major cloud service providers, and development documents of open source project OpenStack. At the same time, operation dialogue logs, solutions to common alarm scenarios and automated script annotation information must be included to enhance the model's understanding and operation ability in container cloud actual scenarios. After the data collection is completed, strict cleaning, deduplication and filtering are required: remove invalid lines, duplicate lines, sensitive information and obviously low-quality text; delete non-compliant content; and make necessary context annotations to key dialogues or operation logs. Finally, the processed data set is scattered and fragmented according to certain rules to prepare for large-scale distributed training.

[0035] S102: Model pre-training and basic architecture.

[0036] This step aims to build the basic language understanding and generation ability of the intelligent agent. The MiniMax-Text-01 large model is selected, which adopts the Transformer Decoder-only structure and takes the autoregressive language modeling (CausalLM) as the main pre-training target, so that the model can learn rich context reasoning and multi-round dialogue ability in a large amount of text corpus.

[0037] The training environment can use multiple GPUs (such as A100) or TPU clusters, and use a combination strategy of data parallelism + tensor parallelism to share the calculation and communication load of a large number of parameters. In mixed precision (such as bfloat16) mode, not only can the training stability be guaranteed, but also the memory usage can be greatly reduced. In terms of learning rate scheduling, the Warmup+Cosine Decay strategy is adopted; during training, set indicators such as language modeling perplexity and container cloud question and answer accuracy to regularly check the training effect and save model snapshots to ensure that the model continuously improves on general understanding and container cloud professional knowledge.

[0038] S103: Instruction fine-tuning and container cloud scenario alignment.

[0039] After the basic pre-training is completed, the model needs to be fine-tuned for instructions to adapt to the actual needs of the container cloud scenario. The operation and maintenance dialogue logs, troubleshooting cases and cloud platform API examples collected earlier are organized into "instruction-answer" pairs for supervised fine-tuning (SFT). Through manual review of part of the data, the accuracy and compliance of the instructions and answers are ensured, and the processing of security, fault and permission related content is strengthened in the data set. In this process, the model learns to give specific solutions to container cloud operations for natural language instructions, such as listing the surviving Pods in the Kubernetes cluster, creating a GCP virtual machine and opening the port, analyzing the alarm log and providing repair ideas, etc. At the same time, as soon as it receives inappropriate or dangerous instructions, the model can give a refusal or warning in time, so as to balance functionality and compliance at the application level.

[0040] S104: Deployment in the container cloud environment.

[0041] This step deploys the trained model (including tokenizer, weight file, inference script, etc.) to the container cloud environment to provide services to the outside. First, the above contents need to be packaged into a Docker image, and then combined with container orchestration platforms such as Kubernetes and DockerSwarm to realize the elastic expansion of model services. Once the user request volume rises, the system can automatically create more replicas to share the load; after the peak is over, it will automatically shrink. If the model parameter volume is large, layered parallelism (different layers are deployed on different GPUs) or more tensor parallelism techniques can be used to reduce the single machine memory pressure; at the same time, load balancing is used for multiple replica routing to ensure that the response delay is minimized in high concurrency scenarios. When deploying, a unified API gateway (such as RESTful or gRPC interface) can be built and identity authentication, traffic control and security detection can be performed to intercept potential malicious or sensitive requests, improving service security and controllability.

[0042] S105: Inference optimization and security monitoring.

[0043] To improve inference performance and service reliability, several optimizations and monitoring measures are required. First, the model can be quantized with INT8 or lower precision, and the corresponding option can be enabled during the inference phase to reduce memory usage and computational load. For multi-turn dialogue scenarios, an attention caching (KV caching) mechanism can be used to maximize the reuse of previous decoding results and accelerate the continuous generation process. At the security and monitoring level, a real-time logging and auditing mechanism must be established to dynamically monitor critical operations or external inputs called by the model. Once high-risk instructions or abnormal traffic are detected, the system can automatically trigger rejection or rate limiting policies to prevent the abuse or misuse of large models. Simultaneously, combined with automated alerts and a visualization dashboard, administrators can intervene when necessary to ensure visibility and traceability of the entire inference process and operation calls. This container cloud agent based on MiniMax-Text-01 completes the entire process from data preparation and model training to deployment and inference optimization.

[0044] This is an integrated solution encompassing basic data preparation and model training, as well as deployment and application examples on a container cloud platform. The specific steps are as follows:

[0045] 1. Data preparation and corpus construction

[0046] When building intelligent agents for container cloud scenarios, the core task is to obtain high-quality, multi-layered training data. To ensure the model possesses broad natural language understanding and expression capabilities, a large amount of general-purpose corpora from Wikipedia, internet articles, ebooks, etc., is required. After collection through web scraping or public datasets, rigorous filtering and deduplication must be performed to exclude duplicate, low-quality, or non-compliant text. Simultaneously, for multilingual scenarios or cross-cultural user needs, while retaining the primary language (such as Chinese and English), a small amount of text in other languages ​​can be supplemented to ensure the model can handle international scenarios.

[0047] A dataset built with efficient and useful product documentation is more professional. To give the model a deep understanding of container clouds, a significant amount of container cloud and related technical documentation needs to be added, specifically including: white papers on public cloud and hybrid cloud technologies; documentation on DevOps, microservice architecture, container orchestration (Docker), Kubernetes, virtualization technologies (VMWare), etc.; product descriptions and user manuals from major cloud service providers such as AWS, Azure, Alibaba Cloud, and Huawei Cloud; and the architecture design and developer documentation of the open-source project OpenStack. This data allows the large model to understand professional knowledge such as cloud resource orchestration, high availability, elastic scaling, and network configuration.

[0048] The establishment of the agent requires the combination of dialogue ability and executable operation, so it is very important to prepare the operation and maintenance dialogue log, the solution of common alarm scene and the annotation information of automatic script. The interactive records of historical operation and maintenance personnel and monitoring system, cloud platform API document and example script can be sorted into dialogue form dataset, and be used in subsequent instruction fine-tuning, so that the model has the ability of "container cloud context inference" and "automated operation".

[0049] After completing the above data collection, data cleaning and labeling are needed to remove invalid rows, duplicate rows, potential sensitive information and obviously low-quality text; strictly filter non-compliant content. For key dialogues or operation logs, labels can be added to help the model identify context intent during subsequent training. Finally, the data is shuffled and divided into appropriate size shards for efficient reading by distributed training in the future.

[0050] 2. Model pre-training and infrastructure

[0051] In order to enable the agent to have a solid foundation of language understanding and generation, the MiniMax-Text-01 model is first pre-trained on a large scale. This model uses the Transformer Decoder-only structure, which can learn rich contextual associations and multi-round dialogue capabilities in autoregressive language modeling tasks. Based on the previous corpus collection and cleaning results, autoregressive language modeling Causal LM can be used as the main pre-training target.

[0052] It is recommended to use cloud computing power including multiple GPU A100 or TPU clusters for distributed training. If the model size reaches hundreds of billions or even thousands of billions of parameters, it is necessary to have high-bandwidth network interconnection and large-capacity video memory to support massive tensor calculations and communications. Data parallelism and tensor parallelism are combined to perform parallel calculations on moderately segmented model parameters, and a large batch size is used to improve training efficiency.

[0053] Using mixed-precision bfloat16 can greatly save video memory usage while ensuring training stability.

[0054] Warmup+Cosine Decay learning rate scheduling is used, which involves warming up for a few steps with a lower learning rate, and then slowly decreasing in a large number of steps to ensure that the model gradually converges to a stable state. During training, several benchmark tests are needed to optimize language modeling perplexity and container cloud question and answer accuracy, and model parameters are saved at fixed intervals to monitor the trend of the validation set indicators. Through curve changes, it is determined whether to adjust the learning rate, gradient clipping or training duration, to ensure that the model gradually improves in language understanding and container cloud expertise.

[0055] To this end, the embodiment adopts an optimized and improved linear attention mechanism Lightning Attention, which has slightly better performance than Transformer.

[0056] Lightning Attention is a linear attention mechanism designed to improve the efficiency of Transformer models. Its core lies in reducing the computational complexity of traditional self-attention mechanisms from quadratic to linear, effectively handling ultra-long sequence inputs. The computational complexity of traditional self-attention mechanisms is O(n 2 ), where n is the sequence length. Lightning Attention reduces this complexity to O(n) by applying a linear kernel trick, enabling efficient processing of long sequence inputs.

[0057] To further optimize computational efficiency, Lightning Attention divides the sequence into multiple blocks and processes intra-block and inter-block attention calculations separately. The intra-block part uses regular attention calculation, while the inter-block part uses the kernel trick of linear attention for processing. Lightning Attention adopts an I / O-aware optimization strategy, fully utilizing GPU hardware characteristics, reducing memory bandwidth bottlenecks, and improving computational efficiency.

[0058] Lightning Attention is combined with the Mixture of Experts (MoE) architecture, forming a model with 32 experts and a total parameter size of 456 billion, where each token activates approximately 459 billion parameters. To balance efficiency and performance, the model uses a hybrid attention mechanism (Hybrid-lightning), replacing Lightning Attention with softmax attention every 8 layers to improve the model's expansion capability and retrieval performance.

[0059] During training and inference, Lightning Attention employs batch kernel fusion, separate pre-padding and decoding execution, multi-level padding, and cross-step batch matrix multiplication expansion optimization strategies to improve computational efficiency and reduce resource consumption. Thanks to the above design, the model using Lightning Attention can have a context length of up to 4 million tokens during inference, significantly exceeding the limitations of traditional models and meeting the demand for long context processing in practical applications.

[0060] Through the above design, Lightning Attention significantly improves the processing capacity for ultra-long sequences while maintaining model performance, providing capabilities beyond the Transformer architecture for large-scale language model applications.

[0061] Based on the agent construction in the container cloud scenario, Lightning Attention needs to make the following adjustments: In order to adapt to the needs of agents based on the container cloud scenario, Lightning Attention architecture can be optimized specifically to address the characteristics of multi-task concurrency, dynamic resource scheduling, efficient data transmission, and distributed computing in the container cloud environment. The following are the specific contents and expansion of the new improvement scheme based on the container cloud scenario:

[0062] Agents in the container cloud scenario often need to handle multiple tasks simultaneously, such as natural language processing, recommendation systems, data analysis, etc. To this end, a multi-task dynamic scheduling mechanism can be introduced in Lightning Attention. Design a dynamic routing module based on task priority, dynamically select the appropriate attention calculation method: Lightning Attention or Softmax Attention, according to the different complexity of the task, the length of the context and the real-time demand. For longer task context, Lightning Attention is preferred, while for short sequence high-precision calculation, Softmax Attention is used. By introducing a task-aware module, the task feature input mode, target type is used to dynamically adjust the depth and block size of attention calculation, reducing resource waste and improving execution efficiency.

[0063] The container cloud environment usually relies on distributed computing architecture to support high-performance model training and inference, so it is necessary to optimize the performance of Lightning Attention in the distributed environment. Global attention calculation is decomposed into local subtasks and distributed to different nodes for parallel computation. For example, based on the sequence slicing method, the sequence is divided into several parts, and each node is responsible for attention calculation of a slice, and finally the results are aggregated through communication mechanisms. Use shared memory technology across nodes to store attention weights and intermediate results in shared memory to avoid redundant data transmission. Specific implementation can use distributed file system Ceph or dedicated communication protocol RDMA. In distributed training, gradient compression techniques such as Top-k gradient compression or quantized gradient transmission are used to reduce communication bandwidth occupancy and improve the efficiency of large-scale model distributed training.

[0064] The implementation process of Top-k gradient compression and quantized gradient transmission: At each parameter update, the full gradient of the model is calculated, and the top-k gradient elements with the largest absolute values are selected according to the absolute value of the gradient. These Top-k gradients are considered to contribute most to model optimization. The indices of the selected Top-k gradient elements are recorded to accurately restore the sparse distribution of the gradient matrix at the receiving end. The selected Top-k gradients are quantized from high-precision floating-point numbers (such as 32 bits) to low-precision formats (such as 16 bits or 8 bits), further reducing the size of the transmitted data. The quantized Top-k gradients and corresponding indices are sent to other nodes through the communication network, significantly reducing the amount of communication data. The receiving end restores the sparse gradient matrix according to the received indices and quantized gradient values, and combines the local unselected gradients for parameter update, ensuring the accuracy and consistency of the optimization process.

[0065] Agents in container cloud environments usually rely on real-time transmission and processing of large-scale data. LightningAttention needs to optimize its data management and preprocessing modules for this requirement. Through the design of a hierarchical data pipeline, data loading, preprocessing, and model calculation are separated. Asynchronous data flow technology is used to ensure that data loading and preprocessing do not become a bottleneck for computation. For dynamic data streams in cloud scenarios, an incremental attention mechanism is designed to perform attention calculation only on new data segments, while the attention weights of historical data are cached and reused, thereby reducing redundant calculations. Combined with cloud-edge collaboration technology, part of the preprocessing tasks (such as data cleaning and feature extraction) are assigned to edge devices for execution, reducing data transmission and reducing the computational pressure on the cloud.

[0066] Agents in container cloud environments usually need to adapt to dynamic changes in resource allocation and load demand, so an elastic expansion mechanism needs to be introduced into Lightning Attention. By monitoring the real-time load and resource usage of the nodes, the allocation of attention calculation tasks is dynamically adjusted. For example, when GPU resources are tight, some tasks can be temporarily assigned to CPUs for execution to ensure the stability of the overall system. A computation graph that supports elastic expansion is designed to enable Lightning Attention to dynamically adjust the computation flow in a multi-node environment, such as adaptively adjusting the block size and parallelism according to the number of nodes. In the container cloud scenario, instances of agents often go through a cold start process. Through optimization of the model initialization process (such as delayed loading of parameter weights and hierarchical initialization), the startup delay can be significantly reduced.

[0067] Agents in container cloud often need to process cross-modal data (such as text, images, audio) and collaborate with other agents to complete complex tasks. To this end, Lightning Attention can be extended to support cross-modal modeling and multi-agent collaboration. The cross-modal version of Lightning Attention is designed to handle inputs from different modalities simultaneously. The specific method is to design dedicated embedding layers and attention weight sharing mechanisms for different modalities to ensure efficient fusion of information between modalities. In the multi-agent collaboration scenario, an attention-based communication mechanism is introduced to enable agents to share key context information. For example: the attention weights generated by one agent can be used as input for other agents, enabling efficient information sharing and task division.

[0068] To better adapt to the container cloud environment, Lightning Attention can be designed as a cloud-native model service, supporting standardized deployment and service management. Lightning Attention is encapsulated as a container image (such as Docker) to facilitate rapid deployment and expansion in the cloud environment. Combined with K8S, the model's automatic scaling and high availability are achieved. The attention calculation module is split into multiple microservices, such as intra-block attention calculation, inter-block attention calculation, dynamic scheduling module, etc., each of which is independently deployed and interacts through standard interfaces. Online updating and upgrading of the model are achieved without interrupting service, ensuring that the agent can continuously optimize performance and adapt to new application scenarios.

[0069] 3. Instruction fine-tuning aligns with container cloud scenarios

[0070] After completing the basic pre-training, MiniMax-Text-01 has good general dialogue and text generation capabilities, but to make it a "container cloud agent", further instruction fine-tuning is required to enable the model to accurately interface with user tasks and complete operations in the cloud scenario. With the help of human or professional operations engineers, these materials are organized into "instruction-answer" form, combining the previously prepared operation dialogue logs, container cloud platform API call examples, troubleshooting scenarios, etc.

[0071] Instruction examples:

[0072] "Please help me check all the Pods that are alive in the current Kubernetes cluster."

[0073] "I want to create a new virtual machine instance on GCP and open port 8080."

[0074] "Please analyze the error cause in this alarm log and provide a solution."

[0075] Answering examples:

[0076] List the results of the operation and maintenance commands or API calls, supplemented by necessary text descriptions or subsequent action prompts.

[0077] Optimal instruction fine-tuning strategy: SFT + partial dialogue human review

[0078] To avoid multiple complex solutions, this embodiment directly selects supervised fine-tuning (SFT) combined with small-scale human review as the best practice. A certain proportion of data is randomly selected from the aforementioned instruction-answer dialogue for manual review to ensure the accuracy and standardization of the answers. Supervised fine-tuning is performed using high-quality labeled data to enable the model to provide high-quality and contextually appropriate answers to input instructions. If the instructions involve prohibited operations or sensitive content, appropriate rejection and safety warning responses are added to the fine-tuning data. To address error states and potential security risks in the container cloud scenario, the data set includes content related to operation troubleshooting and security compliance scenarios, and these examples are emphasized during fine-tuning to enable the model to learn to alert and refuse execution when encountering illegal or dangerous instructions.

[0079] Method overview: After completing the basic pre-training, the model needs to be fine-tuned for instructions to adapt to the actual needs of the container cloud scenario. The operation dialogue logs, troubleshooting cases, and cloud platform API examples collected earlier are organized into "instruction-answer" pairs for supervised fine-tuning (SFT). Through manual review of part of the data, the accuracy and compliance of the instructions and answers are ensured, and examples of handling security, fault, and permission-related content are emphasized in the data set. During this process, the model learns to provide specific solutions for container cloud operations in response to natural language instructions, such as listing live Pods in a Kubernetes cluster, creating a GCP virtual machine and opening a port, analyzing alarm logs and providing repair ideas, etc. At the same time, the model can also give a refusal or warning as soon as it receives inappropriate or dangerous instructions, thus balancing functionality and compliance at the application level.

[0080] Through this instruction fine-tuning process, MiniMax-Text-01 can assume the role of "executable dialogue" in the container cloud scenario. It not only understands the required cloud operations based on user natural language descriptions, but also provides specific execution recommendations, possible script examples, and reasonable scheduling solutions for various resources in the cloud environment.

[0081] 4. Deployment and inference optimization in the container cloud environment

[0082] In order to enable the agent to actually serve the majority of container cloud users, the model needs to be deployed and optimized for inference in the cloud environment to balance performance, cost, and high concurrency. The following work needs to be done.

[0083] The trained model includes Tokenizer, weight files, and inference scripts, which are packaged into a Docker image and a microservice is built to provide APIs externally. This allows elastic expansion and automated operation in clusters such as Kubernetes and Docker Swarm. Once the user request volume increases, the container replicas can be increased on demand; when the peak is over, the resources can be automatically recycled.

[0084] If the model parameter size is large, hierarchical parallelism (loading model layers to different GPUs) or tensor parallelism can be used to address the problem of insufficient memory on a single machine. At the same time, multiple replica services are established at the load balancing level, and each replica is dedicated to processing user requests under a certain route to minimize response delay.

[0085] Using INT8 or lower precision quantization for the model helps reduce memory usage and computation during inference. At this time, based on the mixed precision or quantization strategy adapted during the training phase, the corresponding options can be turned on during inference to speed up response time. In addition, attention caching (KV cache) is introduced for multi-turn dialogue scenarios, which can reuse the dialogue state of the previous round during the next generation, further shortening the inference time.

[0086] A unified API gateway should be built when deploying to provide RESTful and gRPC interfaces for external calls. The gateway layer can perform identity authentication and traffic control, and perform security checks before user input reaches the model to prevent sensitive or malicious requests from directly invoking the underlying large model. If high-risk behavior is detected, it can return a denial or limit flow to prevent abuse of the model.

[0087] To improve inference performance and service reliability, multiple optimizations and monitoring are needed. First, the model can be quantized to INT8 or lower precision, and the corresponding options can be turned on during the inference phase to reduce memory usage and computation. For multi-turn dialogue scenarios, the attention caching (KV cache) mechanism can be used to maximize the reuse of the previous decoding results and speed up the continuous generation process. In terms of security and monitoring, real-time logging and auditing mechanisms must be established to dynamically monitor key operations or external inputs of model calls. Once high-risk instructions or abnormal traffic are detected, the system can automatically trigger denial or limit flow strategies to prevent abuse or misuse of the large model. At the same time, automated alerts and visual panels are combined to allow administrators to intervene when needed, ensuring that the entire inference process and operation calls are visualized and traceable. The MiniMax-Text-01 container cloud agent completes the entire process from data preparation, model training to deployment and inference optimization.

[0088] The embodiment, based on the MiniMax-Text-01 large model, relies on the previous high-quality data preparation and customized pre-training and instruction fine-tuning process to build a powerful intelligent agent for container cloud scenarios. During cloud deployment, containerization and layered parallelism are used to balance scalability and throughput efficiency, and high concurrency and security filtering are achieved in the production environment through microservices architecture and API gateway.

[0089] This integrated solution enables the intelligent agent not only to handle routine natural language understanding and dialogue requirements, but also to perform various practical operations in the container cloud platform and tool system, thereby significantly reducing manual operations by operation and maintenance personnel, accelerating problem troubleshooting, and promoting automation of enterprise DevOps processes. With continuous incremental data updates and continuous fine-tuning, the intelligent agent will become increasingly mature in practice and truly become a "smart brain" in the container cloud era, providing convenient and innovative solutions for a wide range of cloud applications.

[0090] The embodiment of how to build an intelligent agent in a container cloud scenario: starting from diversified and high-quality data preparation, running through large model pre-training and deep instruction fine-tuning, and combining containerized deployment, layered parallel reasoning, quantization optimization, API gateway protection, etc., a complete intelligent agent technology system for cloud applications is finally formed. MiniMax-Text-01 can not only become a general language model, but also adapt to the needs of the cloud scenario and become a reliable assistant, fully improving the automation and intelligence level of container cloud services.

[0091] After the completion of the above technical solution, the cloud intelligent agent built can be widely used in various practical application scenarios. Administrators can communicate with the intelligent agent through natural language, such as asking "why does a certain server have high load", and the intelligent agent can retrieve relevant log information and provide possible reasons. If new nodes need to be added, configurations need to be updated, or services need to be restarted, the intelligent agent can also generate Shell scripts or Kubernetes command lists for administrators to execute with one click.

[0092] When an enterprise deploys services on multiple cloud platforms, the intelligent agent can predict load changes based on business indicators and recommend whether to expand Kubernetes Pod scale or temporarily open more virtual machine instances. If the system detects that the unit price of a cloud platform is raised or that resources are tight, it can also perform migration and alert in advance.

[0093] With the large number of operation and maintenance fault and security case library integrated during model training, the intelligent agent can quickly analyze monitoring alarms and propose possible solutions. Clear and actionable instruction prompts are provided before fault propagation. At the same time, if sensitive operation requests are encountered, the intelligent agent will refuse to execute or notify the superior operation and maintenance personnel for secondary confirmation according to the security policy.

[0094] With cloud platforms crammed with technical documents, user manuals, internal wikis, and other materials, intelligent agents can serve as comprehensive knowledge base robots, helping engineers quickly find relevant entries. Through professional fine-tuning, these agents can accurately identify cloud environment terminology and configuration relationships, improving team collaboration efficiency.

[0095] To ensure the long-term availability of cloud-based intelligent agents, newly generated operational dialogues, fault logs, and cloud platform operation and upgrade data are periodically added to the data pool, and the model is fine-tuned again to guarantee the real-time updates of the knowledge base and operational strategies. If the container cloud platform launches new features or interfaces, the corresponding documentation and examples should also be added to the model fine-tuning set in a timely manner to maintain support for new technologies.

[0096] After deployment, user feedback is continuously collected; inaccurate or inappropriate answers are labeled and included in the next round of fine-tuning dataset. Through this feedback loop, the quality of the agent's answers and alignment can be continuously improved.

[0097] In an enterprise environment, it is crucial to ensure that intelligent agents do not leak sensitive information or perform unauthorized operations. Teams can sample and review conversation logs, and integrate this with a permission control system to implement role-based access control. If violations are detected, commands should be immediately modified to fine-tune the strategy, and filtering and monitoring should be further strengthened at the gateway level.

[0098] In the practical implementation of container cloud agents, several scenarios have emerged where they intersect with other new technologies, further promoting the depth and breadth of their applications. Firstly, in the fields of smart cities and the Internet of Things (IoT), a vast number of sensors and edge nodes rely on container cloud agents for distributed data management, real-time alert processing, and automated command issuance. Due to heterogeneous data sources and complex transmission paths, simple scripts and rules are insufficient for efficient control; only the adaptive learning of agents can address this. Secondly, in hybrid cloud and multi-cloud architectures, agents can uniformly manage public clouds, private clouds, and local data centers across platforms, solving the problems of resource fragmentation and soaring multi-cloud operating costs. Thirdly, in DevOps and MLOps processes, agents can be responsible for checking each stage of continuous integration and continuous deployment (CI / CD), monitoring the impact of code changes on system performance in real time, and triggering rollbacks or adjustment strategies upon detecting regressions or anomalies. For machine learning applications, agents can also help automate the allocation of GPU clusters or schedule large batches of training tasks, effectively reducing manual configuration and queuing time.

[0099] In building the intelligent agent, in order to have a solid language understanding and generation ability, first of all, the MiniMax-Text-01 model needs to be pre-trained on a large scale. The model adopts the Transformer Decoder-only structure, which can learn rich contextual associations and multi-turn dialogue capabilities in the causal language modeling (CausalLM) task. With the previous careful collection and cleaning of large corpus, the model can fully absorb multi-domain and multi-type text knowledge in the pre-training stage, laying a solid foundation for subsequent instruction fine-tuning and scene adaptation.

[0100] Compared with the prior art, the main advantage of MiniMax-Text-01 is large-scale self-recurrent pre-training and deep modeling of multi-turn dialogue context. It gradually masters the dynamic mechanism of contextual association and topic extension in the process of predicting the next word through the Decoder-only structure, greatly enhancing the flexibility and accuracy in the dialogue generation scene. At the same time, the model absorbs rich corpus and effectively removes and cleanses, reducing semantic redundancy and repetition errors in the reasoning stage, avoiding generating invalid or off-topic replies in the interaction process.

[0101] In the aspect of creativity, MiniMax-Text-01 not only conducts deep pre-training on general text corpus, but also integrates multi-source dialogue corpus and domain-specific text features, providing new ideas for the scalability and adaptability of the system in multi-industry scenarios. It not only pursues high-quality language generation, but also pays attention to semantic consistency and multi-turn context processing, laying a solid technical foundation for building the next generation of intelligent agents with "thinking" and "collaboration" capabilities. Through a series of innovations and optimizations, the model can exhibit more in-depth and extensive dialogue capabilities in automatic question answering, knowledge retrieval, creative writing and other applications, providing practical and efficient AI solutions for the industry.

[0102] The method proposed in this embodiment is mainly aimed at enterprises and organizations that need to perform automated operation and maintenance, resource scheduling, and multi-tenant management in the cloud. Due to the complex challenges of high concurrency, large traffic, and heterogeneous applications in modern cloud environments, operation and maintenance personnel need to monitor service health and perform elastic scaling at all times, while developers often need to coordinate task execution across multiple service or container platforms. The container cloud intelligent agent based on the MiniMax-Text-01 large model can efficiently analyze operation and maintenance requirements and application deployment scenarios through natural language or standardized instructions, real-time control of multi-level underlying resources, and timely decision-making suggestions, executable scripts or rollback strategies for exceptions or failures, thereby solving the core problems of long analysis time, high misjudgment rate, and difficulty in managing multi-cloud hybrid scenarios.

[0103] The embodiment further integrates container cloud with artificial intelligence. The agent technology will be widely used in many fields. First, it can provide multi-round dialogue CI / CD planning and code review for DevOps teams, quickly locate performance bottlenecks or security risks, and automatically test and deploy. Second, in high availability and disaster recovery solutions, the agent can assist administrators to realize failover and node expansion according to real-time load and historical data analysis, to ensure service stability. Third, for Internet of Things and smart city applications, the container cloud agent can make dynamic coordination of edge nodes and cloud resources through comprehensive analysis of multi-source data, and provide accurate processing for monitoring and alarm. Finally, in the cross-platform management scene of hybrid cloud and multi-cloud, the agent can integrate the APIs of multiple cloud service providers, uniformly schedule core business processes and optimize costs, while maintaining high flexibility and scalability.

[0104] From a business perspective, this container cloud agent technology is expected to significantly reduce operation and maintenance costs and time overhead, while improving service quality and user satisfaction. By introducing natural language human-computer interaction, enterprises can enable more non-professionals to understand and configure cloud environments, narrowing the technical threshold and talent gap. In addition, since the agent has multi-round conversation and large-scale knowledge base support, it can quickly respond to common faults, data analysis and performance optimization, and provide solutions, improving decision-making efficiency. In terms of multi-cloud management and elastic scheduling, it can reduce excessive overhead through fine-grained resource allocation strategies, and provide flexible deployment modes for large enterprises, better responding to sudden business growth and cross-regional collaboration needs. In summary, the agent improves the degree of automation and enterprise competitiveness, while bringing considerable opportunities and potential benefits to the value-added services and ecological cooperation of the container cloud market.

[0105] In another embodiment, the present application provides a computer readable storage medium storing a computer program, which causes a computer to execute the method of constructing a container cloud agent based on the MiniMax-Text-01 large model of the foregoing embodiment.

[0106] In another embodiment, the present application provides an electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the method of constructing a container cloud agent based on the MiniMax-Text-01 large model of the foregoing embodiment when executing the computer program.

[0107] In the embodiments disclosed in the present application, the computer storage medium can be a tangible medium which can contain or store programs for use by or in connection with an instruction execution system, apparatus or device. The computer storage medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or any suitable combination of the above. More specific examples of computer storage medium can include one or more wires, portable computer disks, hard drives, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), fiber optics, compact disk read-only memory (CDROM), optical storage devices, magnetic storage devices, or any suitable combination of the above.

[0108] Those skilled in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed in the present application can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software manner depends on the specific application and design constraints of the technical solutions. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0109] The above is only the preferred embodiment of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solution falling within the concept of the present application shall fall within the protection scope of the present application. It should be noted that, for ordinary skilled persons in the art, some improvements and refinements without departing from the principles of the present application shall be considered within the protection scope of the present application.

Claims

1. A method for constructing a container cloud agent based on a MiniMax-Text-01 large model, characterized in that, The method comprises the following steps: S101: Collecting training required corpus and processing; S102: Selecting MiniMax-Text-01 large model to construct a model, and pre-training with autoregressive language modeling as the pre-training target; S103: Fine-tuning the pre-trained model for instructions to obtain a trained model; S104: Deploying the trained model to a container cloud environment; S105: Inference optimization is performed on the model deployed to the container cloud environment.

2. The method of claim 1, wherein the method is based on a MiniMax-Text-01 large model to construct a container cloud agent. In step S101, the corpus includes: public cloud and hybrid cloud technology white paper; DevOps, microservice architecture, container orchestration Docker, Kubernetes, virtualization technology VMWare related documents; product instructions and user manuals of major cloud service providers; architecture design and developer documents of open source project OpenStack; The operation and maintenance dialogue log, troubleshooting case and cloud platform API example are also included in the corpus, and the data set is arranged in the form of dialogue.

3. The method of claim 1, wherein the method is based on a MiniMax-Text-01 large model to construct a container cloud agent. In step S101, the collected corpus is processed as follows: data cleaning and labeling; adding labels to key dialogues or operation logs; shuffling and dividing the data into segments.

4. The method for constructing a container cloud agent based on a MiniMax-Text-01 large model according to claim 1, characterized in that: In step S102, multiple GPUs A100 or TPU clusters are selected for training environment, and the combination strategy of data parallelism and tensor parallelism is adopted; the learning rate scheduling adopts Warmup and Cosine Decay strategy; The set indicators in training include language modeling perplexity and container cloud question and answer accuracy.

5. The method for constructing a container cloud agent based on a MiniMax-Text-01 large model according to claim 1, characterized in that: In step S102, the model adopts a hybrid attention mechanism of Lightning Attention and Softmax Attention; a dynamic routing module based on task priority is designed in the model, which is used to dynamically select the attention calculation method from Lightning Attention and Softmax Attention according to the complexity of the task, the context length and the real-time demand; Lightning Attention adopts distributed file system Ceph or special communication protocol RDMA; in distributed training, gradient compression technology Top-k gradient compression and quantized gradient transmission are used, the process is as follows: in each parameter update, the current node calculates the full gradient of the model, and according to the absolute value of the gradient, the top-k gradients with the largest absolute value are selected; record the index of the selected Top-k gradient; quantize the selected Top-k gradient; The quantized Top-k gradient and the corresponding index are sent to other nodes through the communication network, and the receiving node restores the sparse gradient matrix according to the received index and quantized gradient value, and combines the local unselected gradient for parameter update; Lightning Attention separates data loading, preprocessing and model calculation; attention calculation is performed on new data segments, and the attention weight of historical data is cached and reused; The preprocessing task is allocated to the edge device for execution; Lightning Attention dynamically adjusts the allocation of attention computation tasks by monitoring the real-time load and resource usage of nodes; Lightning Attention designs dedicated embedding layers and attention weight sharing mechanisms for different modal data; Lightning Attention is designed as a cloud-native model service and encapsulated as a container image.

6. The method for constructing a container cloud agent based on a MiniMax-Text-01 large model according to claim 2, characterized in that: In step S103, the collected data set in the form of dialogue is sorted into "instruction-answer" pairs, and a certain proportion is randomly selected for review. The model is fine-tuned using the reviewed "instruction-answer" pairs.

7. The method for constructing a container cloud agent based on a MiniMax-Text-01 large model according to claim 1, characterized in that: In step S104, the trained model includes Tokenizer, weight file and inference script, which are packaged into Docker image and built into microservice to provide API externally. When deployed, a unified API gateway is built for identity authentication, traffic control and security detection.

8. The method for constructing a container cloud agent based on a MiniMax-Text-01 large model according to claim 1, characterized in that: In step S105, the model is quantized to INT8, and the attention caching mechanism is used for multi-round dialogue scenarios.

9. A computer readable storage medium storing a computer program, characterized in that, The computer program enables the computer to execute the method for constructing a container cloud agent based on the MiniMax-Text-01 large model according to any one of claims 1-8.

10. An electronic device, comprising: Comprise: A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the method for constructing a container cloud agent based on the MiniMax-Text-01 large model according to any one of claims 1-8.

Images