A method and system for intelligent computing power operation and maintenance scheduling based on a large language model agent

Abstract

The application provides a kind of intelligent computing computing power operation and maintenance scheduling method and system based on large language model agent.The application builds object graph semantic runtime to integrate heterogeneous operation and maintenance data, and outputs structured observation context with cluster topology;The large model agent embedded with intelligent computing field knowledge prompts completes chain reasoning, and performs cache-aware expansion screening, five-level event priority arbitration, net profit quantification scoring and conflict resolution in parallel, to generate the optimal scheduling scheme;An independent decision verification layer is added to check the safety and compliance of scheduling actions, and standardized skill library is used to land operation and maintenance operations;A memory storage module is constructed, based on multi-factor retrieval of historical cases, positive and negative examples are combined to assist reasoning, and scheduling strategy can be continuously optimized without retraining the model.The application solves the problems of traditional scheduling data fragmentation, multi-action conflict, lack of profit orientation and poor self-adaptation, and realizes stable, economic and autonomous iteration of computing power operation and maintenance closed-loop scheduling in multiple scenarios such as intelligent computing cluster fault, traffic and cache.

Description

Technical Field

[0001] This invention relates to the cross-technical fields of artificial intelligence intelligent operation and maintenance and cloud computing container resource scheduling, specifically to an intelligent computing power operation and maintenance scheduling method and system based on a large language model intelligent agent. Background Technology

[0002] The existing intelligent computing cluster computing power operation and maintenance scheduling system mainly relies on three types of tools: elastic scaling components such as KubernetesHPA and KEDA, traditional AIOps operation and maintenance systems, and general-purpose large language model intelligent agents. Among them, elastic scaling components can only rely on manually preset single indicator thresholds to complete container scaling, which cannot cope with complex scenarios with multiple coupled factors such as hardware failures, traffic surges, and simultaneous model cache failures. Traditional AIOps achieves automated operation and maintenance based on fixed time-series prediction, anomaly detection, and reinforcement learning scheduling algorithms. The algorithms have single functions and weak generalization ability. When facing new types of failures, the model needs to be redeveloped, and it does not have the comprehensive chain reasoning ability of operation and maintenance experts. As for general-purpose large language model intelligent agents such as GPT-4, Llama, and Qwen, they are only deployed in general question answering and code generation scenarios and have not been deeply adapted to the intelligent computing power scheduling industry. The system lacks dedicated tools for areas such as memory analysis, cache lookup, computing cost and token revenue calculation, making it unable to output optimal scheduling decisions based on commercial benefits. The current mainstream approach in the industry is to deploy the three types of tools separately and add an independent arbitration module to coordinate conflicts between multiple systems. However, under this architecture, the operational data such as monitoring, logs, caching, and container orchestration are scattered and heterogeneous, with inconsistent data formats. It is difficult to obtain complete topological association information of cluster components. When multiple events occur concurrently, actions such as expansion and fault repair are prone to conflict, causing cluster capacity fluctuations and wasted computing resources. At the same time, the entire architecture lacks a memory mechanism to store historical scheduling cases and assist in iterative optimization. The scheduling decisions cannot be continuously and autonomously optimized as the cluster runs. Overall, it has multiple inherent defects such as frequent decision conflicts, ineffective expansion, lack of adaptive reasoning ability, fragmented multi-source data, and lack of economic quantitative scheduling basis. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method and system for intelligent computing power operation and maintenance scheduling based on a large language model intelligent agent.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a method for intelligent computing power operation and maintenance scheduling based on a large language model intelligent agent, the method being as follows: S1. Unified acquisition and semantic structuring of multi-source observation information: An object graph semantic runtime is built as a unified observation context base. Through the object graph semantic runtime, multi-dimensional observation information is collected and encapsulated in a unified and standardized manner. The observation information includes computing device operation indicators, inference service business indicators, anomaly logs, model cache status of each node, cluster entity topology relationships, and historical decision memory data. The object graph semantic runtime predefines the entity objects and topology relationships in the intelligent computing domain, and uniformly connects to multiple heterogeneous data sources such as monitoring, logs, cache, and container orchestration. It outputs a structured observation context with a unified format and carrying topology relationship information. S2, Chain-based Reasoning Analysis of Large Language Model Intelligent Agents: The standardized observation context output by S1 is input into the large language model agent; the large language model agent loads prompt word templates embedded with intelligent computing-specific domain knowledge to carry out chain thinking; the domain knowledge embedded in the prompt words includes cache level and model loading time quantization mapping, multi-instance resource partitioning rules, fault-recovery strategy mapping, SLA service threshold, and cache-aware expansion execution process; without retraining the large model, only the prompt words need to be updated to iterate the domain scheduling rules; S3. The agent generates multiple candidate decision plans and executes four types of core scheduling operations: The large language model agent outputs candidate decision plans containing multiple sets of skill invocation instructions based on chained reasoning. It simultaneously performs four core operations: cache-aware expansion and filtering, event priority arbitration, economic benefit scoring, and conflict resolution. The specific process is as follows: S31, Cache-Aware Expansion Node Filtering Operation: When it is determined that expansion is needed, the cache query skill is invoked to obtain the cache level of the target model of each candidate node, and the loading time of each node model is matched according to the preset mapping relationship. Use load prediction skills to obtain the peak load arrival time; filter nodes whose loading time exceeds the peak arrival time and whose effective service window does not meet the standard, and select the node with the shortest loading time from the remaining feasible nodes; if there are no feasible nodes, generate an alternative scheduling scheme according to resource partitioning, accuracy degradation, and cold loading pre-expansion priority. S32. Unified priority arbitration operation for multiple events: All events in the current period are sorted into five priority levels, from highest to lowest: emergency failure, service breach, capacity recovery, elastic scaling, and cost optimization. Candidate actions are generated in order of priority, and conflict resolution rules are executed: if the estimated capacity for capacity recovery meets the peak load, expansion is suppressed; if there is an expansion action, shrinkage is directly suppressed. S33. Calculation of quantitative scoring for economic benefits: For each candidate scheduling scheme, the changes in computing power cost, word revenue, and default risk cost are calculated separately; only when there are queued requests in the cluster is the added computing power included in the word revenue increment. Net profit per solution = Change in word revenue - Change in computing power cost - Cost of default risk; Select the option with the highest net profit; if all options have a negative net profit, compare the returns of the option with the option of not taking any action and select the best one. S4. Decision Verification Layer Security Check and Plan Correction Filtering: The candidate decision plans output by S3 are sent to the independent decision verification layer, and the compliance of each action is verified based on the preset security rule set. The security rules include the upper limit of replicas for a single expansion, the upper limit of total cluster utilization, the minimum number of replicas of the model after shrinking, target node resource verification, and mutual exclusion of expansion and shrinking of the same model in the same period. Correctable violations such as parameter exceeding limits are automatically adjusted to within the safety threshold, while uncorrectable violations such as insufficient resources are directly filtered. All violation records are cached and pushed to the large language model agent for iterative learning in the next control cycle. S5. Skill execution layer: Landing and scheduling actions and sending back results. Based on the verified safety decision plan output by S4, the corresponding modules in the skill library are called sequentially to perform operation and maintenance operations. The skill set includes log analysis, metric query, cache query, load prediction, scaling up / down execution, precision switching, resource partitioning, economic calculation, container repair, and historical query. After the action is executed, the results, cluster load changes, and cost-benefit data are fully fed back to the large language model agent. S6. Decision data is stored in the database and retrieved from memory to assist in the next round of reasoning: S61. Store the complete context of this scheduling, scene tags, decision plan, execution success or failure indicators, net profit changes, and violation records in a structured form into the memory bank; S62, Memory Retrieval: Extract current scene labels and coarsely screen historical decision records based on the condition that the label intersection is ≥1; calculate a multi-factor comprehensive score for each candidate record; the score includes a weighted sum of label similarity factor, economic effect factor and time decay factor, wherein label similarity is calculated using set intersection-union ratio and time decay is calculated using an exponential decay function; The top K historical cases are selected in descending order of scores, and at least one failed case is forcibly included. The case length is pruned by combining LLM lexical context budget, and the processed historical cases are injected into the agent's reasoning context to achieve continuous optimization without model retraining.

[0005] Preferably, in step S1, the object graph semantic runtime is divided into five layers from top to bottom: agent gateway layer, query service layer, graph storage layer, entity storage layer, and model package layer. The model package layer is used to define entity vocabulary and topological relationships in the intelligent computing domain and supports hot updates. The entity storage layer receives heterogeneous data sources and standardizes entity data. The graph storage layer persists entities and their relationships. The query service layer uses a three-stage pipeline of parser, planner, and executor to process semantic queries. The agent gateway layer provides query and tool call interfaces to large language model agents through standard protocols.

[0006] Preferably, the object graph semantic runtime also includes an agent gateway service layer; the agent gateway layer distinguishes between read-only query tools and write tools with permission control. Metadata and cluster basic topology are exposed through read-only interfaces, while operations such as entity writing and resource modification require separate write permissions to achieve isolation between metadata and runtime data access permissions.

[0007] A computing power operation and maintenance scheduling system based on a large language model intelligent agent, the system comprising an object graph semantic runtime layer, an observation information layer, a large language model intelligent agent layer, a decision verification layer, a skill execution layer, and a memory storage layer; The object graph semantic runtime layer serves as the unified data foundation of the system. It standardizes and integrates various heterogeneous operation and maintenance data sources through five built-in sub-modules, constructs an entity semantic graph with topological associations in the intelligent computing domain, and provides a unified read, write, and query interface based on the intelligent agent gateway. It provides a globally consistent observation context with topological relationships for each upper layer and isolates the differentiated docking logic of multiple data sources. The observation information layer is used to collect runtime metrics of computing devices, business metrics of inference services, log information, model cache status, and cluster events; The large language model intelligent agent layer receives the observation information, performs chain-like thinking analysis based on prompt word templates embedded with intelligent computing domain knowledge, executes the cache-aware scaling node selection algorithm and the unified decision priority arbitration algorithm, and outputs a candidate decision plan containing skill invocation; the decision verification layer performs compliance verification, parameter correction and violation filtering on the candidate decision plan based on a predefined security rule set, and outputs a verified security decision plan; the skill execution layer contains multiple independently callable skill modules; the skill modules include a log analysis module, an indicator query module, a cache query module, a load prediction module, a scaling execution module, and an economic impact calculation module; The memory storage layer is used to store the context, execution results, economic effects and scene labels of historical decisions in a structured format, and provides a multi-factor weighted retrieval interface based on label similarity.

[0008] Preferably, the object graph semantic runtime layer relies on five sub-modules connected from bottom to top internally: model package definition module, entity storage module, graph storage module, query service module, and intelligent agent gateway module. These sub-modules provide a unified and standardized interaction interface to the external world, bidirectionally connecting to the observation information layer. On the one hand, the layer outputs the semantic specifications of intelligent computing domain entities and topological relationships to the observation information layer; on the other hand, the layer receives the raw operation and maintenance data uploaded by the observation information layer and completes the persistence of standardized entities and topological relationships. After integrating the fully structured observation context, the observation information layer pushes it unidirectionally to the large language model agent layer. The observation information layer internally uses a parallel linkage of a cache perception expansion and filtering unit, a five-level event arbitration conflict resolution unit, and a net profit quantification scoring unit to complete comprehensive reasoning and output a candidate decision plan to be sent down to the decision verification layer. After the decision verification layer completes compliance verification, parameter correction and violation filtering for the candidate plan, it sends the security decision plan to the skill execution unit of the skill execution layer to execute various operation and maintenance scheduling actions. After the skill execution unit completes the execution, on the one hand, it writes back the cluster state change data to the object graph semantic runtime layer to update the global topology entity, and on the other hand, it sends the action execution result back to the large language model intelligent agent layer. At the same time, the violation records summarized by the decision verification layer will be pushed to the large language model intelligent agent layer for auxiliary iteration in the next control cycle. The skill execution unit will also write the complete decision context, scene tags, economic benefits, and execution results of this round into the memory storage layer for persistent storage; After receiving the current scene label issued by the large language model agent layer, the memory storage layer performs multi-factor weighted retrieval, filters and prunes historical reference cases, and then injects them back into the large language model agent layer to participate in this round of reasoning. The whole process forms a complete closed-loop link of unified collection, intelligent decision-making, security verification, action execution, and memory iteration. All cross-level interactions are standardized and relayed through the agent gateway of the object graph semantic runtime layer. Each level is decoupled from each other and there is no cross-level direct connection to heterogeneous data source channels.

[0009] Preferably, the planner of the query service module can verify and optimize the topology query statement in advance based on the capability declaration of the graph storage backend, so as to avoid exceeding the performance threshold of the storage backend.

[0010] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, a method for intelligent computing power operation and maintenance scheduling based on a large language model intelligent agent is implemented.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention uses a large language model intelligent agent as a global unified decision-making center to coordinate various cluster events, and has built-in hierarchical event arbitration and conflict resolution logic to avoid the cluster capacity fluctuation problem caused by the mutual interference of multiple operation and maintenance actions from the root. To address the shortcomings of traditional AIOps, such as reliance on fixed algorithms, lack of expert-level comprehensive reasoning capabilities, and inability to adapt to novel faults, this invention leverages the chain-like reasoning capabilities of a large language model to analyze multi-dimensional information such as logs, metrics, and caches to autonomously locate the root cause of faults. It can handle various unknown and abnormal scenarios without the need to redevelop dedicated algorithms. To address the deficiencies of general large-model intelligent agents not being adapted to intelligent computing services, lacking domain-specific scheduling tools, and being unable to balance commercial benefits, this invention incorporates dedicated skills such as cache status query and computing power economic accounting, and selects the optimal scheduling scheme with net profit as the core guide, achieving synergistic optimization of computing power operation and maintenance and business revenue. To address the shortcomings of heterogeneous and fragmented observation data across multi-system architectures, lack of cluster topology association information, and high costs of integrating new data sources, this invention sets up an object graph semantic runtime as a unified data foundation. It standardizes and integrates multiple types of heterogeneous operation and maintenance data and provides entity topology query capabilities, significantly reducing the workload of integrating and modifying new data sources. At the same time, to address the shortcomings of existing technologies that lack continuous optimization mechanisms and cannot iteratively improve scheduling capabilities, this invention adds a memory storage and multi-factor case retrieval mechanism. It reuses historical successful and failed scheduling experience to assist in real-time decision-making, and can continuously improve the quality of scheduling decisions as the cluster runs without retraining large language models. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a diagram of the overall system architecture of the present invention; Figure 3 Flowchart for cache-aware scaling decision-making; Figure 4 To standardize the decision-making and arbitration process flowchart. Detailed Implementation

[0013] To provide a further understanding of the purpose, structure, features, and functions of the present invention, detailed descriptions are provided below with reference to specific embodiments.

[0014] Please refer to the reference. Figures 1 to 4 This invention provides a method and system for intelligent computing power operation and maintenance scheduling based on large language model intelligent agents.

[0015] Example 1: 72B model cache-aware expansion scenario: This embodiment is deployed on a 3-node K8s intelligent computing cluster with 8 A10080G computing cards per node. The control cycle adopts a fixed 30-second cycle combined with an emergency trigger mode for critical events. The underlying layer relies on a self-developed object graph semantic runtime to uniformly connect to the time series indicator library, distributed cache, and log collection system heterogeneous data sources. Internally, the runtime predefines the topology association relationships between computing nodes, inference services, model instances, cache-level entities, and supports hot updates through the model package layer. The entity storage layer receives raw operation and maintenance data and completes standardized entity encapsulation. The graph storage layer persists the node cache association topology. The query service layer optimizes query statements through a three-stage pipeline verification of parser, planner, and executor. The intelligent agent gateway layer distinguishes between read-only query interfaces and authorized write interfaces to achieve metadata and runtime data permission isolation.

[0016] In the S1 phase, the following full observation information was collected: there are only 2 running replicas of the current 72B inference service, queue depth is 18, inference latency is 3.5 seconds and it is close to the SLA default threshold of 4 seconds, node 1 is a 10-second ready memory cache, node 2 is a 300-second cold load, node 3 is a 90-second distributed cache, load prediction reaches peak traffic in 180 seconds, single card computing power cost is 40 yuan / hour, expansion of word units increases revenue by 85 yuan / hour, and default compensation is 30 yuan / hour. This information was then packaged into a unified structured context and pushed to the large language model agent. In the S2 stage, the agent loads prompt word templates containing knowledge of intelligent computing fields such as embedded cache loading duration mapping, five-level event priority, and net profit calculation formula to carry out chain reasoning. Without retraining the large model, it can identify that the current event belongs to P4 elastic scaling event and trigger cache-aware expansion node filtering operation. In the S3 phase, cache query and load prediction skills are first used to compare the loading time and peak arrival time of each node. Node 2 with a loading time exceeding 180 seconds is filtered out. Node 1 with the shortest loading time is selected from Node 1 and Node 3 as the expansion target. Simultaneously, the economic benefit quantification rules generate three candidate schemes and calculate the benefits according to "net profit = change in word revenue - change in computing power cost - default risk cost". Scheme A expands Node 1 by 1 replica with a net profit of 45 yuan / hour, Scheme B switches to 8-bit precision with a net profit of 25 yuan / hour, and Scheme C does not operate and incurs a default loss of 30 yuan / hour. Finally, Scheme A with the highest net profit is selected to generate a candidate decision plan to enter the S4 decision verification layer. In the S4 phase, the verification layer completes the verification based on security rules such as the single expansion limit, the minimum number of replicas, and the mutual exclusion of expansion and contraction of the same model. This expansion of 1 replica has no violations and directly outputs the security execution plan. In the S5 phase, the scaling up and down execution skill is invoked to complete the scaling up of node 1 replica. After the execution is completed, the cluster load change data is written back to the object graph semantic runtime to update the global topology entity, and the execution result is pushed back to the large language model intelligent agent. In the S6 stage, structured data such as scene labels, decision schemes, execution success indicators, and net profit of 45 yuan / hour, including "72B model, queue overflow, memory cache expansion", are stored in distributed memory key-value storage. In the next round of inference, the current scene label is extracted and the historical records are coarsely screened with a label intersection of ≥1. Then, a multi-factor fine ranking is completed by a weighted formula of 60% label similarity, 25% economic effect, and 15% time decay. At least one case of expansion failure is forcibly included and the case length is pruned based on LLM word context budget before being injected into the agent's inference context. Continuous scheduling optimization without model retraining is achieved by relying on memory retrieval.

[0017] Example 2: Hardware failure coupled with a sudden surge in traffic and concurrent scheduling scenario: This embodiment uses the same hardware and software deployment environment as the 3-node A100 intelligent computing cluster. In the S1 phase, the object graph semantic runtime aggregates two types of concurrent cluster events and encapsulates them into a unified observation context. One type is the highest priority P1 emergency failure event, i.e., the hardware failure of node 1 inference container going offline, the container reconstruction takes only 15 seconds and the reconstructed capacity can cover the predicted traffic peak. The other type is the P4 elastic scaling event, i.e., the cluster traffic continues to rise and the request queue depth increases. After the standardized observation information is pushed to the large language model agent, it enters the S2 chain inference process. The agent initiates a unified decision priority arbitration operation, sorting all events according to five priority levels: emergency failure, service breach, capacity recovery, elastic scaling, and cost optimization. First, it generates container reconstruction candidate actions based on the highest priority failure event, then executes conflict resolution rules to determine whether the cluster capacity after capacity recovery can meet the peak load, automatically suppresses redundant expansion actions, and triggers "existence expansion" The "scaling down" constraint is masked by the capacity action, and only a single container reconstruction candidate decision plan is output and sent to the decision verification layer. The verification layer verifies that the number of model replicas after reconstruction is higher than the minimum replica threshold, the target node has sufficient idle memory, and there is no mutual exclusion conflict between the expansion and contraction of the same model. If the plan has no violations, it is directly released. Then, the skill execution layer calls the container repair module to complete the reconstruction of the faulty container. After the execution is completed, on the one hand, the cluster topology entity state in the runtime of object graph semantics is updated, and on the other hand, the execution result without capacity oscillation is fed back to the large language model agent. At the same time, the scenario label of "hardware failure + traffic concurrency, only reconstruction without expansion" and economic benefit data are completely written into the memory storage layer. During subsequent periodic inference, similar historical cases are retrieved to help the agent avoid the problems of expansion and repair action conflicts and severe cluster capacity oscillation caused by traditional multi-system independent scheduling.

[0018] Example 3: A system for intelligent computing power operation and maintenance scheduling based on large language model intelligent agents: The underlying object graph semantic runtime layer relies on a bottom-up interconnected model package definition module, entity storage module, graph storage module, query service module, and intelligent agent gateway module to form a unified data foundation. The model package module completes the definition of entities and topological relationships in the intelligent computing domain and supports hot updates. The entity storage module receives monitoring, logs, cache, and container raw data uploaded from the observation information layer and performs standardized transformation. The graph storage module persists the cluster entity association topology. The query service module's planner optimizes topology query statements based on the storage backend capabilities to avoid performance overflow. The intelligent agent gateway distinguishes read and write permissions, isolates metadata and runtime data access, and bidirectionally connects with the observation information layer to output semantic specifications downwards and receive standardized raw data upwards. The observation information layer collects device operation indicators, business inference indicators, abnormal logs, model cache, and cluster events, integrates the global structured observation context, and pushes it unidirectionally to the large language model intelligent agent layer. The intelligent agent layer relies on built-in domain prompts to complete chained reasoning, and performs parallel execution of cache-aware expansion and filtering, five-level event arbitration conflict resolution, net profit quantification scoring calculation to generate candidate decision plans and distribute them to the decision verification layer. The decision verification layer completes compliance verification, parameter correction, and violation filtering based on the security rule set, and then outputs a security scheduling plan to push to the skill execution layer. The skill execution layer calls independent skill modules such as log analysis, cache query, load prediction, scaling, and economic calculation to perform operation and maintenance actions. After execution, it transmits data bidirectionally. On the one hand, it updates the global topology entity state of the object graph semantic runtime, and on the other hand, it pushes the action execution results and violation records back to the large language model agent for iteration. At the same time, it writes the complete decision context, scene label, net profit change, and execution success or failure identifier of this round into the memory storage layer in a structured way. The memory storage layer receives the scene labels issued by the agent, completes the label intersection coarse screening, multi-factor weighted fine ranking, positive and negative example balanced selection, and word budget case pruning, and then injects the processed historical cases back into the large language model agent to participate in the current reasoning. All cross-level interactions in the entire system are standardized and relayed through the object graph semantic runtime agent gateway. Each level is completely decoupled and there is no cross-level direct connection to heterogeneous data source channels, forming a complete scheduling closed loop of unified collection, intelligent decision-making, security verification, action execution, and memory iteration.

[0019] The present invention has been described in the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.

Claims

1. A method for intelligent computing power operation and maintenance scheduling based on large language model intelligent agents, characterized in that: The method is as follows: S1. Unified acquisition and semantic structuring of multi-source observation information: An object graph semantic runtime is built as a unified observation context base. Through the object graph semantic runtime, multi-dimensional observation information is collected and encapsulated in a unified and standardized manner. The observation information includes computing device operation indicators, inference service business indicators, anomaly logs, model cache status of each node, cluster entity topology relationships, and historical decision memory data. The object graph semantic runtime predefines the entity objects and topology relationships in the intelligent computing domain, and uniformly connects to multiple heterogeneous data sources such as monitoring, logs, cache, and container orchestration. It outputs a structured observation context with a unified format and carrying topology relationship information. S2, Chain-based Reasoning Analysis of Large Language Model Intelligent Agents: The standardized observation context output by S1 is input into the large language model agent; the large language model agent loads prompt word templates embedded with intelligent computing-specific domain knowledge to carry out chain thinking; the domain knowledge embedded in the prompt words includes cache level and model loading time quantization mapping, multi-instance resource partitioning rules, fault-recovery strategy mapping, SLA service threshold, and cache-aware expansion execution process; without retraining the large model, only the prompt words need to be updated to iterate the domain scheduling rules; S3. The agent generates multiple candidate decision plans and executes four types of core scheduling operations: The large language model agent outputs candidate decision plans containing multiple sets of skill invocation instructions based on chained reasoning. It simultaneously performs four core operations: cache-aware expansion and filtering, event priority arbitration, economic benefit scoring, and conflict resolution. The specific process is as follows: S31, Cache-Aware Expansion Node Filtering Operation: When it is determined that expansion is needed, the cache query skill is invoked to obtain the cache level of the target model of each candidate node, and the loading time of each node model is matched according to the preset mapping relationship. Use load prediction skills to obtain the peak load arrival time; filter nodes whose loading time exceeds the peak arrival time and whose effective service window does not meet the standard, and select the node with the shortest loading time from the remaining feasible nodes; if there are no feasible nodes, generate an alternative scheduling scheme according to resource partitioning, accuracy degradation, and cold loading pre-expansion priority. S32. Unified priority arbitration operation for multiple events: All events in the current period are sorted into five priority levels, from highest to lowest: emergency failure, service breach, capacity recovery, elastic scaling, and cost optimization. Candidate actions are generated in order of priority, and conflict resolution rules are executed: if the estimated capacity for capacity recovery meets the peak load, expansion is suppressed; if there is an expansion action, shrinkage is directly suppressed. S33. Calculation of quantitative scoring for economic benefits: For each candidate scheduling scheme, the changes in computing power cost, word revenue, and default risk cost are calculated separately; only when there are queued requests in the cluster is the added computing power included in the word revenue increment. Net profit per solution = Change in word revenue - Change in computing power cost - Cost of default risk; Select the option with the highest net profit; If all options result in a negative net profit, then compare the returns of the option without taking any action and select the best one. S4. Decision Verification Layer Security Check and Plan Correction Filtering: The candidate decision plans output by S3 are sent to the independent decision verification layer, and the compliance of each action is verified based on the preset security rule set. The security rules include the upper limit of replicas for a single expansion, the upper limit of total cluster utilization, the minimum number of replicas of the model after shrinking, target node resource verification, and mutual exclusion of expansion and shrinking of the same model in the same period. Correctable violations such as parameter exceeding limits are automatically adjusted to within the safety threshold, while uncorrectable violations such as insufficient resources are directly filtered. All violation records are cached and pushed to the large language model agent for iterative learning in the next control cycle. S5. Skill execution layer: Landing and scheduling actions and sending back results. Based on the verified safety decision plan output by S4, the corresponding modules in the skill library are called sequentially to perform operation and maintenance operations. The skill set includes log analysis, metric query, cache query, load prediction, scaling up / down execution, precision switching, resource partitioning, economic calculation, container repair, and historical query. After the action is executed, the results, cluster load changes, and cost-benefit data are fully fed back to the large language model agent. S6. Decision data is stored in the database and retrieved from memory to assist in the next round of reasoning: S61. Store the complete context of this scheduling, scene tags, decision plan, execution success or failure indicators, net profit changes, and violation records in a structured form into the memory bank; S62, Memory Retrieval: Extract current scene labels and coarsely screen historical decision records based on the condition that the label intersection is ≥1; calculate a multi-factor comprehensive score for each candidate record; the score includes a weighted sum of label similarity factor, economic effect factor and time decay factor, wherein label similarity is calculated using set intersection-union ratio and time decay is calculated using an exponential decay function; The top K historical cases are selected in descending order of scores, and at least one failed case is forcibly included. The case length is pruned by combining LLM lexical context budget, and the processed historical cases are injected into the agent's reasoning context to achieve continuous optimization without model retraining.

2. The intelligent computing power operation and maintenance scheduling method based on a large language model intelligent agent as described in claim 1, characterized in that: In step S1, the object graph semantic runtime is divided into five layers from top to bottom: intelligent agent gateway layer, query service layer, graph storage layer, entity storage layer, and model package layer. The model package layer is used to define entity vocabulary and topological relationships in the intelligent computing domain and supports hot updates. The entity storage layer receives heterogeneous data sources and standardizes entity data; the graph storage layer persists entities and their relationships; the query service layer uses a three-stage pipeline of parser, planner, and executor to process semantic queries; and the agent gateway layer provides query and tool call interfaces to large language model agents through standard protocols.

3. The intelligent computing power operation and maintenance scheduling method based on a large language model intelligent agent as described in claim 2, characterized in that: The object graph semantic runtime also includes an agent gateway service layer; the agent gateway layer distinguishes between read-only query tools and write tools with permission control. Metadata and cluster basic topology are exposed through read-only interfaces, while operations such as entity writing and resource modification require separate write permissions to achieve isolation between metadata and runtime data access permissions.

4. A smart computing power operation and maintenance scheduling system based on a large language model intelligent agent, characterized in that: The system includes an object graph semantic runtime layer, an observation information layer, a large language model agent layer, a decision verification layer, a skill execution layer, and a memory storage layer; The object graph semantic runtime layer serves as the unified data foundation of the system. It standardizes and integrates various heterogeneous operation and maintenance data sources through five built-in sub-modules, constructs an entity semantic graph with topological associations in the intelligent computing domain, and provides a unified read, write, and query interface based on the intelligent agent gateway. It provides a globally consistent observation context with topological relationships for each upper layer and isolates the differentiated docking logic of multiple data sources. The observation information layer is used to collect runtime metrics of computing devices, business metrics of inference services, log information, model cache status, and cluster events; The large language model intelligent agent layer receives the observation information, performs chain-like thinking analysis based on prompt word templates embedded with intelligent computing domain knowledge, executes the cache-aware scaling node selection algorithm and the unified decision priority arbitration algorithm, and outputs a candidate decision plan containing skill invocation; the decision verification layer performs compliance verification, parameter correction and violation filtering on the candidate decision plan based on a predefined security rule set, and outputs a verified security decision plan; the skill execution layer contains multiple independently callable skill modules; the skill modules include a log analysis module, an indicator query module, a cache query module, a load prediction module, a scaling execution module, and an economic impact calculation module; The memory storage layer is used to store the context, execution results, economic effects and scene tags of historical decisions in a structured format, and provides a multi-factor weighted retrieval interface based on tag similarity.

5. The intelligent computing power operation and maintenance scheduling system based on a large language model intelligent agent as described in claim 4, characterized in that: The object graph semantic runtime layer relies on five sub-modules connected from bottom to top: model package definition module, entity storage module, graph storage module, query service module, and intelligent agent gateway module. These sub-modules provide a unified and standardized interaction interface to the outside world and bidirectionally connect to the observation information layer. On the one hand, the observation information layer outputs the semantic specifications of entities and topological relationships in the intelligent computing domain to the observation information layer. On the other hand, the observation information layer receives the raw operation and maintenance data uploaded by the observation information layer and completes the persistence of standardized entities and topological relationships. After integrating the fully structured observation context, the observation information layer pushes it unidirectionally to the large language model agent layer. The observation information layer internally uses a parallel linkage of a cache perception expansion and filtering unit, a five-level event arbitration conflict resolution unit, and a net profit quantification scoring unit to complete comprehensive reasoning and output a candidate decision plan to be sent down to the decision verification layer. After the decision verification layer completes compliance verification, parameter correction and violation filtering for the candidate plan, it sends the security decision plan to the skill execution unit of the skill execution layer to execute various operation and maintenance scheduling actions. After the skill execution unit completes the execution, on the one hand, it writes back the cluster state change data to the object graph semantic runtime layer to update the global topology entity, and on the other hand, it sends the action execution result back to the large language model intelligent agent layer. At the same time, the violation records summarized by the decision verification layer will be pushed to the large language model intelligent agent layer for auxiliary iteration in the next control cycle. The skill execution unit will also write the complete decision context, scene tags, economic benefits, and execution results of this round into the memory storage layer for persistent storage; After receiving the current scene label issued by the large language model agent layer, the memory storage layer performs multi-factor weighted retrieval, filters and prunes historical reference cases, and then injects them back into the large language model agent layer to participate in this round of reasoning. The whole process forms a complete closed-loop link of unified collection, intelligent decision-making, security verification, action execution, and memory iteration. All cross-level interactions are standardized and relayed through the agent gateway of the object graph semantic runtime layer. Each level is decoupled from each other and there is no cross-level direct connection to heterogeneous data source channels.

6. The intelligent computing power operation and maintenance scheduling system based on a large language model intelligent agent as described in claim 5, characterized in that: The planner of the query service module can verify and optimize the topology query statement in advance based on the capability declaration of the graph storage backend, so as to avoid exceeding the performance threshold of the storage backend.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 3.

Images