Real-time intelligent agent dynamic scheduling method and system composed of multiple large models

Abstract

The invention provides a real-time intelligent agent dynamic scheduling method and system composed of multiple large models, and is applied to the technical field of data processing. The method comprises the following steps: performing dynamic scheduling expansion around a plurality of large-model real-time intelligent agents, firstly receiving task request data of a real-time service scene, performing grouping according to task types, priorities and resource requirements, pre-configuring a scheduling strategy, and generating a pre-scheduling queue and resource pre-planning information; and in combination with the target configuration information, scheduling task information after secondary optimization is obtained through multi-dimensional optimization such as task splitting and priority rearrangement. Dynamically adjusting and generating a target scheduling instruction based on real-time system resources and a model operation state, and cooperatively processing the target scheduling instruction with the large model capability adaptation parameters to form scheduling execution information; and finally, issuing to model nodes according to preset service quality requirements and resource constraints, generating information such as adaptive parameters and priority rules through feature extraction, association processing and the like, and supporting multi-model cooperation to efficiently complete real-time service tasks.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a real-time intelligent agent dynamic scheduling method and system composed of multiple large models. BACKGROUND

[0002] In the field of data processing, with the rapid development of artificial intelligence technology, large models are increasingly widely used in various real-time business scenarios such as natural language processing, data analysis, and image recognition. To meet complex business needs, a single large model has been difficult to cover the processing needs of multiple types of tasks, and multi-large model collaborative scheduling has become the core solution to improve task processing efficiency and accuracy.

[0003] However, the existing multi-large model scheduling technology still has many defects: first, the task scheduling lacks fine preprocessing mechanism, only based on simple task type or priority for preliminary allocation, without combining task resource demand and large model capacity for accurate matching, leading to part of the task being allocated to the model node with low adaptation degree, resulting in waste of computing power or processing precision not meeting the standard; second, the scheduling strategy is obviously static, without fully considering the dynamic changes of real-time system resource state and model running state, when the system resource fluctuates or the model load is too high, it is easy to appear task blocking, long response time and other situations, it is difficult to meet the timeliness requirements of real-time business scenarios; third, there is a lack of perfect conflict mediation, timeout processing and fault tolerance redundancy mechanism in the scheduling process, in the face of resource preemption, task execution exception, model node failure and other sudden situations, there is a lack of effective coping strategies, resulting in insufficient continuity and stability of task execution; fourth, the integration and standardization degree of multi-dimensional scheduling information is low, the association logic between tasks, resources and models is not clear, the permission allocation boundary is fuzzy, affecting the efficiency and orderliness of multi-model collaborative execution.

[0004] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] Other characteristics and advantages of the present application will become apparent from the following detailed description, or will be learned by practice of the application.

[0006] According to an aspect of the present application, a real-time intelligent agent dynamic scheduling method composed of multiple large models is provided, comprising: receiving task request data sent by at least one real-time service scenario, grouping and managing the received task request data based on task type, priority, and resource demand, and pre-configuring scheduling strategy processing to generate pre-scheduling task queue information and resource allocation pre-planning information, wherein different task types include at least one task request data sent by at least one real-time service scenario; receiving target configuration information and optimizing the pre-scheduling task queue information corresponding to each group task, including task splitting, priority rearrangement, model capability matching, resource reservation, conflict mediation, timeout plan, load balancing, fault tolerance redundancy, to generate secondary optimized scheduling task information; dynamically adjusting the secondary optimized scheduling task information based on real-time system resource state data and model running state data to generate target scheduling instruction information; cooperatively processing the target scheduling instruction information and the capability adaptation parameters of each large model to generate real-time intelligent agent scheduling execution information; and based on the preset business service quality requirements and resource scheduling constraints, the real-time intelligent agent scheduling execution information is sent to each large model node, the preset business service quality requirements, resource scheduling constraints, real-time intelligent agent scheduling execution information and large model node attributes are extracted and classified to generate node adaptation parameter information, scheduling priority rule information, task distribution authority information, and task and model association mapping information, so that each large model node cooperatively completes the real-time business task.

[0007] In another aspect of the present application, a real-time intelligent agent dynamic scheduling system composed of multiple large models includes: a data acquisition and preprocessing unit configured to receive task request data sent by at least one real-time service scenario, perform grouping management and scheduling strategy pre-configuration processing on the received task request data based on task type, priority, and resource requirement, generate pre-scheduling task queue information and resource allocation pre-planning information, receive target configuration information, and perform optimization processing on the pre-scheduling task queue information corresponding to each group task, including task splitting, priority rearrangement, model capability matching, resource reservation, conflict mediation, timeout plan, load balancing, and fault tolerance redundancy, to generate secondary-optimized scheduling task information; a data processing and optimization unit configured to perform dynamic adjustment on the secondary-optimized scheduling task information based on real-time system resource state data and model running state data, generate target scheduling instruction information, perform collaborative processing on the target scheduling instruction information and capability adaptation parameters of each large model, and generate real-time intelligent agent scheduling execution information; and a large model node configured to receive the real-time intelligent agent scheduling execution information based on preset service quality of service requirements and resource scheduling constraint conditions, extract and classify the preset service quality of service requirements, resource scheduling constraint conditions, real-time intelligent agent scheduling execution information, and large model node attributes, and generate node adaptation parameter information, scheduling priority rule information, task distribution authority information, and task and model association mapping information, so that each large model node collaboratively completes real-time service tasks.

[0008] According to still another aspect of the present application, an electronic device includes a first processor and a memory configured to store executable instructions of the first processor, wherein the first processor is configured to execute the executable instructions to implement the above-described real-time intelligent agent dynamic scheduling method composed of multiple large models.

[0009] According to still another aspect of the present application, a computer-readable storage medium is provided, which stores a computer program that, when executed by a second processor, implements the above-described real-time intelligent agent dynamic scheduling method composed of multiple large models.

[0010] This application provides a real-time intelligent agent dynamic scheduling method and system composed of multiple large models. The method receives task request data from real-time business scenarios, groups it according to task type, priority, and resource requirements, and pre-configures scheduling strategies to generate a pre-scheduling queue and resource pre-planning information. Combining the target configuration information, it optimizes the scheduling tasks through multi-dimensional processes such as task splitting, priority rearrangement, and model matching to obtain secondary optimized scheduling task information. Based on real-time system resources and model operating status, it dynamically adjusts and generates target scheduling instructions, which are then processed in conjunction with large model capability adaptation parameters to form scheduling execution information. Finally, the instructions are sent to model nodes according to preset requirements, and node adaptation parameters and priority rules are generated through feature extraction and correlation processing to support multi-model collaborative task completion. Through multi-dimensional matrix construction, dynamic algorithm introduction, full-process risk control, and information standardization, it achieves accurate task matching and efficient scheduling.

[0011] By using a model capability adaptation matrix and multi-dimensional feature association, precise binding between tasks and large models is achieved, avoiding wasted computing power and substandard accuracy, and adapting to the needs of complex business scenarios. Based on real-time resources and model status, the scheduling strategy is dynamically adjusted, combined with algorithms such as task scheduling acceleration and load balancing, to reduce response latency and meet the timeliness requirements of real-time business.

[0012] By integrating conflict resolution, timeout contingency plans, and fault-tolerant redundancy mechanisms, it effectively addresses unforeseen circumstances such as resource contention and node failures, ensuring the continuity and reliability of task execution. Through standardized integration of scheduling information and clear definition of authority boundaries, it clarifies the logical connections between tasks, resources, and models, improving the orderly execution and overall efficiency of multi-model collaborative execution, and facilitating the large-scale application of large models.

[0013] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0014] Figure 1 This document illustrates a flowchart of a real-time intelligent agent dynamic scheduling method composed of multiple large models, provided in an embodiment of this application. Figure 2 This illustration shows a schematic diagram of the structure of a real-time intelligent agent dynamic scheduling system composed of multiple large models, provided in an embodiment of this application. Detailed Implementation

[0015] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0016] The following is combined with Figure 1This application describes a method for dynamic scheduling of real-time intelligent agents composed of multiple large models, according to exemplary embodiments thereof. It should be noted that the application scenarios described below are merely illustrative for understanding the spirit and principles of this application, and the embodiments of this application are not limited in any way. Rather, the embodiments of this application are applicable to any suitable scenario.

[0017] In one implementation, Figure 1 A schematic flowchart of a real-time intelligent agent dynamic scheduling method consisting of multiple large models, according to an embodiment of this application, is shown.

[0018] S101, receive task request data sent by at least one real-time business scenario, perform group management and scheduling strategy pre-configuration processing on the received task request data based on task type, priority and resource requirements, and generate pre-scheduled task queue information and resource allocation pre-planning information.

[0019] In one implementation, task request data is received from at least one real-time business scenario. The real-time business scenario refers to various practical application scenarios that require multi-model collaborative processing. The task request data consists of business requirement-related data to be processed in each scenario. The data format must meet the parsing requirements of subsequent group management and scheduling strategy configuration to ensure data integrity and identifiability.

[0020] All received task request data is categorized based on task type, priority, and resource requirements as core grouping dimensions. By task type: Task request data is grouped into corresponding categories according to the functional requirements of different business scenarios. Each task type includes task request data from at least one real-time business scenario, ensuring scenario coverage for all task types. By priority: The priority level of each task request data is clearly defined based on indicators such as task urgency and business importance, providing a basis for subsequent scheduling sequence planning. By resource requirements: The hardware resources and software support conditions, such as computing power, memory, and network bandwidth, required for each task request data are identified, completing the classification based on resource requirements.

[0021] For the grouped task data, scheduling strategies are pre-configured for each group. Based on the task type characteristics, priority level, and resource requirements of each group, a preliminary scheduling plan is developed, clarifying the approximate execution order of tasks, the proposed resource allocation, and the potential large model types to be adapted, ensuring the strategy's effectiveness. Through the above group management and scheduling strategy pre-configuration processing, two types of core information are ultimately generated, as follows: Pre-scheduled task queue information: Based on the group and the priority order of tasks, a task execution sequence is formed, clarifying the order in which each task is processed. Resource allocation pre-planning information: Based on the resource requirements of the grouped tasks and the pre-configured scheduling strategy, a preliminary resource allocation plan is developed, clarifying the allocation ratio and usage plan of various resources among different grouped tasks.

[0022] S102, receive the target configuration information and optimize the pre-scheduled task queue information corresponding to each group task, including task splitting, priority rearrangement, model capability matching, resource reservation, conflict resolution, timeout contingency plan, load balancing, fault tolerance redundancy, and generate secondary optimized scheduling task information.

[0023] In one implementation, a task splitting algorithm is used to granularly decompose and logically reconstruct grouped tasks. Combined with a model capability adaptation matrix, this achieves precise matching between tasks and the larger model, thus building the foundational layer for optimized processing. The structured task splitting algorithm's core logic involves first analyzing the business logic relationships of the grouped tasks, identifying the boundaries and dependencies of functional modules within each task, and then determining the splitting granularity standard based on the task's complexity level (e.g., simple, medium, complex). Complex tasks are split into multiple smallest independently executable sub-task units based on "core functions + auxiliary functions," medium tasks are split based on "step-by-step processes," and simple tasks remain unchanged. Logical reconstruction is performed simultaneously during the splitting process. By establishing a sub-task association graph, the preconditions, execution order, and data interaction interfaces of each sub-task are clarified. This ensures that after independent execution, sub-tasks can be reassembled into a complete task through interfaces, forming a logical closed loop of "split-execution-integration." This fundamentally guarantees the achievability of the overall task objective and provides a foundation for subsequent multi-model parallel processing.

[0024] A comprehensive model capability adaptation matrix is ​​constructed. The horizontal dimension of the matrix covers all large model identifiers participating in scheduling, while the vertical dimension encompasses the core capability indicators of these large models. These indicators include task processing type adaptability (such as compatibility with different task types like natural language processing, data analysis, and image recognition), computing power capacity limits (the amount of data that can be processed in a single batch, and the threshold for concurrent tasks), response speed benchmarks (average task processing latency, and the peak latency limit), accuracy performance thresholds (processing accuracy and allowable error range under different task types), and resource consumption coefficients (the computing power and memory usage ratio required per unit task). The restructured subtasks are then processed according to a unified standard to extract core requirement indicators (such as task type, data volume, accuracy requirements, and latency requirements). These indicators are compared one by one with the capability indicators of each large model in the matrix. A weighted scoring method is used to calculate the adaptation score between the subtask and each model (weights are set according to the importance of each capability indicator in the business scenario). The target large model with the highest adaptation score is selected, achieving precise binding between the task and the large model. This process establishes the foundational layer for optimization, providing a clear execution vehicle for subsequent priority adjustments and resource allocation.

[0025] A dynamic priority reordering model and a resource reservation scheduling protocol are introduced to construct a three-dimensional optimization matrix of task priority, model load, and resource reserve, generating differentiated resource allocation schemes. A multi-dimensional decision-based dynamic priority reordering model is introduced, comprising four core layers: input layer, feature processing layer, decision layer, and output layer. The input layer receives three core data categories: task urgency (e.g., real-time, near-real-time, non-real-time), business value weight (task importance coefficients preset by the business scenario, e.g., core business tasks have higher weights than auxiliary business tasks), and dependency relationships (whether it is a prerequisite for other tasks, and the degree of coupling with other tasks). The feature processing layer standardizes and normalizes the input data, converting indicators of different dimensions into a unified quantitative range (e.g., 0-10 points), and generates a comprehensive feature vector through feature fusion. The decision layer uses the analytic hierarchy process (AHP) to construct a judgment matrix, weighting and prioritizing the comprehensive feature vector. The output layer outputs the dynamically adjusted task priority ranking results. This model supports real-time updates of input data. When the urgency of a task changes, the weight of business value is adjusted, or the dependencies change, the priority is automatically recalculated, breaking the limitations of the initial static priority ranking and ensuring that high-value, high-urgency tasks always receive priority scheduling rights.

[0026] A standardized resource reservation and scheduling protocol is adopted, which clearly defines the triggering conditions, reservation process, release mechanism, and conflict handling rules for resource reservation. First, based on the task's resource requirements (computing power, memory, network bandwidth, storage resources, etc.) and scale (e.g., computing power requirement in FLOPS, memory requirement in GB), combined with the dynamically rearranged task priorities, the scope of tasks requiring resource reservation is determined (high-priority tasks, critical sub-tasks, and pre-tasks with strong dependencies). Then, a resource reservation request is submitted to the system resource management module according to the protocol process, specifying the type, quantity, usage period, and release conditions of the reserved resources. The system resource management module approves the reservation based on the current resource usage status. After approval, the corresponding resources are locked, and usage access is only granted to the reserved task. After the task is completed or the reserved usage period is exceeded, the resources are automatically released according to the protocol and returned to the system resource pool for allocation by other tasks. This process avoids resource preemption for high-priority or critical tasks during execution, ensuring the continuity and timeliness of task execution.

[0027] A three-dimensional optimization matrix is ​​constructed, consisting of task priority, model load, and resource reserve. All three dimensions are represented by quantitative indicators. The "task priority" dimension represents the dynamically reordered priority ranking, quantized using integers from 1 to N (1 being the highest priority and N the lowest). The "model load" dimension represents the real-time load percentage of each model, calculated by weighting indicators such as the number of tasks currently being executed, the ratio of computing power / memory usage, and CPU utilization, with a quantization range of 0-1 (0 representing no load and 1 representing full load). The "resource reserve" dimension represents the statistical data of the system's remaining allocable resources, quantified by resource type (e.g., remaining computing power percentage, remaining memory capacity percentage), also with a quantization range of 0-1. Based on this matrix, multi-dimensional cross-analysis is performed, and resource allocation threshold rules are set (e.g., when a high-priority task matches a low-load model and has a high resource reserve, the optimal resources are allocated; when a medium-priority task matches a medium-low load model and has a medium resource reserve, the appropriate resources are allocated). Through matrix traversal and rule matching, a differentiated resource allocation scheme is generated for each task, clarifying the execution model corresponding to the task, the type and specific quantity of allocated resources, so as to achieve accurate and efficient resource allocation among different tasks and models, and avoid resource waste or insufficient resource problems.

[0028] This system integrates a conflict resolution rule base with a timeout contingency triggering mechanism, dynamically adjusting task execution nodes through a load balancing algorithm to establish a multi-dimensional fault-tolerant redundancy strategy. The system incorporates a multi-scenario conflict resolution rule base, divided into three sub-bases based on conflict type: resource conflict, priority conflict, and model adaptation conflict. Each sub-base includes conflict judgment criteria, conflict level classifications (minor, moderate, and severe), corresponding solutions, and execution processes. The resource conflict sub-base covers judgment conditions for scenarios such as multiple tasks simultaneously requesting the same resource, and conflicts between reserved and immediately allocated resources. Solutions include resource priority arbitration, time-sharing reuse, and resource expansion. The priority conflict sub-base covers judgment conditions for scenarios such as contradictory priority judgment results for different tasks and high-priority tasks being blocked by low-priority tasks. Solutions include priority re-verification, task execution queue adjustment, and conflict task diversion. The model adaptation conflict sub-base covers judgment conditions for scenarios such as mismatch between task and allocated model capabilities, and model failure leading to adaptation failure. Solutions include model re-matching, task migration, and activating backup models. The system monitors data interaction and status changes in the task scheduling process in real time. When an event that meets the conflict judgment criteria is detected, it automatically identifies the conflict type and level, calls the corresponding solution in the rule base, and completes the conflict mediation according to the preset execution process, ensuring that the task scheduling process is not affected by the conflict and proceeds smoothly.

[0029] A full-process timeout contingency trigger mechanism is established, comprising four core components: threshold setting, real-time monitoring, tiered early warning, and contingency plan execution. First, based on task type, complexity, and business requirements, differentiated execution time thresholds are set for each task and subtask, including early warning thresholds (triggered when task execution time reaches this threshold) and timeout thresholds (e.g., a timeout is determined when task execution time exceeds this threshold). Then, the task progress monitoring module collects real-time task execution status data, such as execution time, percentage of progress completed, and current execution stage, and compares this data with the set time thresholds. When the task execution time reaches the early warning threshold, a level-one early warning is triggered, and the system automatically investigates the reasons for slow task execution, such as insufficient resources, excessive model load, and data transmission delays, and generates preliminary optimization suggestions.

[0030] When the task execution time exceeds the timeout threshold, a level 2 warning is triggered, and a preset timeout contingency plan is automatically executed. This plan includes task splitting (splitting unfinished subtasks into other idle models for execution), resource addition (temporarily allocating additional resources to the currently executing model), model switching (migrating the task to a suitable and low-load backup model), and task degradation (reducing the accuracy or complexity of non-critical links without affecting the core objectives). This prevents a single task timeout from blocking the entire scheduling process and ensures overall scheduling efficiency.

[0031] An adaptive load balancing algorithm is employed. This algorithm collects real-time task load data (such as the number of currently executing tasks, CPU utilization, memory usage, and task processing latency) from various model nodes, calculates the load index (quantization range 0-1) for each node, and sets a load balancing threshold (e.g., 0.7, where a load index below this value indicates low load, and above it indicates high load). When the load index of a model node exceeds the balancing threshold, the algorithm automatically analyzes the priority, execution progress, and resource requirements of each task on that node, filters out non-core or low-priority tasks that can be migrated, and calculates the optimal migration path based on the load status of other nodes. The selected tasks are then migrated to nodes with load indices below the balancing threshold. Simultaneously, the algorithm monitors the data transmission status and load changes of the target node during the task migration process in real time to ensure that the migration process does not affect the continuity of task execution. Ultimately, this maintains the load index of all model nodes within the balancing threshold range, achieving load balancing.

[0032] A multi-dimensional fault-tolerant redundancy strategy is established synchronously, including a task-level replica backup mechanism (generating 1-2 replicas for critical tasks, stored on different model nodes, and automatically activating replicas when the main task fails), a model-level standby node deployment mechanism (deploying standby model nodes according to a certain proportion of the number of core models, with standby nodes maintaining consistent capabilities with the main node and synchronizing configuration information in real time), and a system-level rapid failover mechanism. A fault detection module is established to monitor the running status of the main model node in real time. When a node failure or anomaly is detected, a rapid failover process is triggered, migrating tasks on that node to the standby node within milliseconds and updating the task-model association mapping. Through this multi-level fault-tolerant redundancy design, the impact of unexpected events such as model node failures and task execution anomalies on the scheduling process is minimized, ensuring the continuity and stability of task execution.

[0033] The process integrates the splitting results, matching schemes, resource allocation, conflict resolution methods, timeout handling mechanisms, and load balancing strategies to generate standardized scheduling task information after secondary optimization. First, multi-dimensional information is collected, integrating the core processing results from each previous stage. Specifically, this includes: a list of subtasks after task splitting and logical restructuring, containing subtask identifiers, core objectives, preconditions, execution order, and data interaction interfaces; precise matching results between tasks and the large model, including the association between subtasks and the target model, fit scores, and matching criteria; differentiated resource allocation schemes, including the type, quantity, reserved time period, and release conditions of resources allocated to each subtask; specific solutions for various conflicts, including conflict type, occurrence time, handling measures, and handling results; preset timeout handling mechanisms, including time thresholds, warning levels, and corresponding contingency plans for each task; and load balancing adjustment records (including task migration history, node load changes, and migration effect evaluation).

[0034] Then, information standardization processing is carried out, and unified data format specifications and expression standards are formulated: a unified unit and precision are set for all quantitative data (such as the unit of computing power is unified as FLOPS, and the precision is retained to two decimal places); standardized coding and description languages ​​are used for non-quantitative information (such as task type, conflict type, contingency plan name, etc.); and a unified identification system is used for related information (such as sub-task identifier, model node identifier, resource identifier, etc. follow the same coding rules) to ensure that the information in different links has consistency and identifiability.

[0035] Finally, standardized scheduling task information after secondary optimization is generated. This information is presented in the form of a structured document, with each subtask corresponding to an independent scheduling record. It clearly includes the subtask's basic information (identifier, core objective, execution order), execution model information (target model identifier, adaptability, backup model identifier), resource configuration information (resource type, allocation quantity, reserved time period), priority information (priority level, weight coefficient), execution sequence information (planned start time, estimated completion time, time threshold), and risk control information (conflict solutions, timeout contingency plans, fault tolerance mechanisms). This provides a clear, standardized, and directly parsable input basis for subsequent dynamic adjustments based on real-time system resources and model status.

[0036] S103, dynamically adjusts the secondary optimized scheduling task information based on real-time system resource status data and model running status data, and generates target scheduling instruction information.

[0037] In one implementation, a state resolution process is used to perform parameter mapping and rule normalization on real-time system resource data and model runtime data. A resource load assessment model and a dynamic scheduling arbitration algorithm are introduced to achieve structured processing of adjustment strategies. The standardized state resolution process first classifies and collects real-time system resource data and model runtime data—system resource data includes indicators such as real-time occupancy, remaining capacity, and fluctuation range of core hardware resources such as computing power, memory, network bandwidth, and storage capacity; model runtime data includes key runtime indicators such as the number of tasks currently being executed by the model, CPU utilization, memory usage ratio, task processing latency, and accuracy output stability. For the collected multi-source heterogeneous data, parameter mapping processing is performed: raw data of different dimensions and units are mapped to a unified quantization dimension (such as the 0-1 range), establishing a correspondence between data and scheduling adjustment requirements; simultaneously, rule normalization is performed, formulating unified data filtering, cleaning, and standardization rules to remove outliers, duplicates, and invalid data, ensuring consistent data format and semantic uniformity, providing high-quality data input for subsequent model and algorithm applications.

[0038] A resource load assessment model is constructed, comprising four layers: a data input layer, a feature extraction layer, an assessment calculation layer, and a result output layer. The input layer receives system resource data after state parsing and model operating status data. The feature extraction layer extracts time-series features (such as the 5-minute average, 10-minute peak, and variance of resource load) using a sliding window method and integrates static features (such as the model's computing power limit and total resources) to generate a comprehensive load feature vector. The assessment calculation layer uses a weighted summation algorithm to calculate the resource load assessment value based on the weights of each feature's influence on the load assessment (e.g., CPU utilization weight 0.3, memory usage weight 0.25, task processing latency weight 0.2, and remaining resource percentage weight 0.25), with a quantization range of 0-1 (0 representing no load, 1 representing full load). The output layer outputs the load assessment results for each model node and the entire system, clearly defining the load level (low load, medium load, high load, overload). This model supports real-time updates of input data, recalculating the load assessment value every 30 seconds to ensure the timeliness of the assessment results.

[0039] A dynamic scheduling arbitration algorithm is introduced, primarily used to resolve scheduling conflicts between multiple tasks and model nodes, enabling structured decision-making for adjustment strategies. The algorithm's input includes resource load assessment results, task priority information, task resource requirements, and model capability suitability. Internally, the algorithm uses a decision tree model for conflict arbitration: with "load level" as the root node and "task priority," "resource requirement gap," and "model suitability" as branch nodes, decision rules are set (e.g., high-load nodes prioritize migrating low-priority tasks, resource expansion requests are triggered when the resource gap exceeds 30% and there is no suitable model, and model re-matching occurs when model suitability is below a threshold). By traversing the decision tree, structured scheduling adjustment decisions are output (e.g., task migration instructions, resource allocation adjustment instructions, model switching instructions, etc.), ensuring that the adjustment strategy logic is clear and executable, avoiding scheduling chaos.

[0040] By aligning with the intelligent agent scheduling protocol standard, a three-dimensional adjustment matrix of resource type, model state, and task requirements is constructed. A standardized scheduling parameter table is generated through model conversion via scheduling benchmarking, and a dynamic adjustment and update mechanism is established. The system aligns with industry-standard intelligent agent scheduling protocols, clarifying core elements such as supported resource type definitions, model state identifiers, task requirement description specifications, data interaction formats, and command transmission protocols. Resource data, model state data, and task requirement data within the system are formatted and semantically aligned according to the protocol standard to ensure data interoperability between different modules and nodes. Simultaneously, the generation format, transmission path, and execution feedback mechanism of scheduling commands are designed according to protocol requirements to guarantee standardized transmission and accurate execution of scheduling commands, laying the protocol foundation for subsequent three-dimensional adjustment matrix construction and standardized parameter table generation.

[0041] A three-dimensional adjustment matrix is ​​constructed, comprising resource type, model status, and task requirements. All three dimensions are quantified and defined according to protocol standards. The "Resource Type" dimension covers computing power, memory, network bandwidth, and storage resources, with each resource type corresponding to a unique identifier and quantification unit (e.g., computing power in FLOPS, memory in GB). The "Model Status" dimension combines the load level (low, medium, high, overload) output by the resource load assessment model with the model's operating status (normal, warning, abnormal). The "Task Requirements" dimension includes quantified indicators such as the type and quantity of resources required for the task, task priority, task processing latency requirements, and accuracy requirements. Matrix cells represent optimal scheduling adjustment strategy suggestions under the corresponding three-dimensional combination conditions (e.g., "Computing Power Resources - Low-Load Model - High-Priority Task" corresponds to the strategy of "Prioritizing the allocation of sufficient computing power to maintain low-load model operation"), forming a structured mapping relationship of adjustment strategies.

[0042] A scheduling benchmarking and transformation model is introduced. This model takes the strategy recommendations from the three-dimensional adjustment matrix as input and, combined with protocol standards and system configuration parameters, converts the strategy recommendations into standardized scheduling parameters. These parameters include resource allocation parameters (such as allocated computing power and memory quotas), task scheduling parameters (such as task execution node identifiers and task start time windows), and model configuration parameters (such as the number of running threads and cache size). All parameters are encoded and formatted according to protocol standards. Based on these standardized parameters, a standardized scheduling parameter table is generated. This table includes fields such as parameter name, parameter type, parameter value, applicable scenario, effective time, and expiration conditions, achieving standardized management of scheduling parameters.

[0043] A dynamic adjustment and update mechanism is established, clearly defining the update trigger conditions, update cycle, and update process for the parameter table and the three-dimensional adjustment matrix. Update trigger conditions include: changes in total system resources (e.g., adding server nodes), abnormal model status (e.g., a model node failure), addition of new task requirements, resource load assessment results showing a sustained high load state for more than 5 minutes, and historical scheduling adjustment records showing parameter adaptability below 80%. The update cycle is set to automatically update every 10 minutes in normal scenarios, and immediately update when emergency update conditions are triggered. The update process is "data acquisition → load assessment → strategy adaptation → parameter conversion → parameter table replacement → instruction synchronization," ensuring that the standardized scheduling parameter table and the three-dimensional adjustment matrix always match the actual system state and task requirements, guaranteeing the dynamic adaptability of scheduling adjustments.

[0044] Using the task scheduling lifecycle as a time window, a multi-dimensional scheduling feature matrix is ​​constructed by integrating resource type, model status, task requirements, and historical scheduling adjustment records to achieve full-dimensional correlation of adjustment strategies. The time window covers the entire process of a task from entering the scheduling queue, pre-execution preparation, in-execution monitoring, to completion / termination. The window duration is set according to task type and complexity: the time window for short-duration tasks (processing time ≤ 10 minutes) is the entire task lifecycle; the time window for medium-duration tasks (10 minutes < processing time ≤ 1 hour) is divided into 30-minute segments; and the time window for long-duration tasks (processing time > 1 hour) is divided into 1-hour segments, ensuring that scheduling adjustments within each time segment can be decided based on complete status data and historical records.

[0045] Within a defined time window, four core data categories—resource type, model status, task requirements, and historical scheduling adjustment records—are integrated to construct a multi-dimensional scheduling feature matrix. The matrix's rows represent time segments within the time window (e.g., short-duration tasks are divided into 1-minute segments), and the columns represent feature items, specifically including: resource type features (real-time occupancy, remaining capacity, and rate of change for each resource), model status features (load level, operating status, and number of failures), task requirement features (priority, resource requirements, latency requirements, and accuracy requirements), and historical scheduling adjustment features (historical adjustment strategy type, adjustment time, adjustment effect evaluation value, number of conflicts, and solutions). Each matrix cell represents the quantified or encoded value of a specific feature item within the corresponding time segment, achieving full-dimensional, time-series integration of scheduling-related data.

[0046] Based on a multi-dimensional scheduling feature matrix, feature correlation analysis is used to achieve full-dimensional correlation of adjustment strategies. The cosine similarity algorithm is employed to calculate the correlation between different feature items, such as the correlation between remaining resources and task requirements, and the correlation between model load level and historical adjustment strategies, filtering out strongly correlated feature combinations (correlation ≥ 0.7). Correlation rules are set in conjunction with business logic; for example, when the feature combination of "high-priority task + high model load + insufficient remaining resources" occurs, the adjustment strategy of "task diversion + resource addition" is associated. Through feature matrix traversal and correlation rule matching, it is ensured that each scheduling adjustment strategy can be associated with multi-dimensional data such as resources, models, tasks, and historical records within the time window, avoiding adjustment biases caused by single-dimensional decision-making and improving the rationality and comprehensiveness of the adjustment strategies.

[0047] This paper describes a method to supplement missing adjustment rules by using a resource prediction model, extracting key strategies through a dynamic state feedback mechanism, and fusing these strategies with the features of a three-dimensional adjustment matrix to generate a target scheduling instruction set. A resource prediction model is introduced to predict future system resource change trends based on historical data, supplementing missing adjustment rules in the multi-dimensional scheduling feature matrix. The model employs a time-series prediction algorithm, using historical resource data, model state data, and task scheduling data from the multi-dimensional scheduling feature matrix as training inputs. The training period is set to 7 days, the sliding window size to 24 hours, and the model parameters are optimized using gradient descent (learning rate 0.01, iterations 1000). After training, the model can predict the remaining amount of various system resources, model load change trends, and peak task demand periods within the next hour. Based on the prediction results, scenarios not covered by existing adjustment rules are analyzed (e.g., predicting a resource shortage exceeding 50% for a certain period without corresponding adjustment rules). Supplementary adjustment rules are formulated based on business logic and protocol standards to complete the missing adjustment strategies and ensure the integrity of the adjustment rules.

[0048] A dynamic status feedback mechanism is established to collect real-time feedback data on system resource status, model running status, and task execution progress after scheduling adjustments, and compare and analyze this data with the expected effects of the adjustment strategy. Feedback evaluation indicators are set (such as resource utilization improvement rate, model load reduction rate, task latency compliance rate, and conflict resolution success rate). When feedback data shows that the evaluation indicators have not reached the expected thresholds (e.g., resource utilization improvement rate < 10%, task latency compliance rate < 90%), the current adjustment strategy is determined to be a non-critical strategy; when the evaluation indicators reach or exceed the expected thresholds, they are determined to be critical strategies. Through continuous feedback evaluation and screening, the most effective critical adjustment strategies in different scenarios are extracted, forming a critical strategy library, providing a core basis for generating target scheduling instructions.

[0049] The strategies in the key strategy library are fused with the features of the three-dimensional adjustment matrix. A weighted voting method is used to determine the final scheduling adjustment strategy—the strategy corresponding to the three-dimensional adjustment matrix has a weight of 0.6, and the strategy corresponding to the key strategy library has a weight of 0.4. The strategy adaptation score is calculated according to the weights, and the strategy with the highest score is selected as the final execution strategy. Based on the final execution strategy and the parameter configuration in the standardized scheduling parameter table, structured target scheduling instructions are generated. The instructions include core fields such as instruction identifier, execution object (model node / resource module), instruction type (task migration, resource allocation, model switching, etc.), execution parameters (such as migration task identifier, allocated resource quantity, target model identifier), execution time limit, and feedback requirements. All target scheduling instructions are classified and organized according to task type and execution object to form a target scheduling instruction set, providing a clear, standardized, and directly parsable instruction basis for subsequent scheduling execution.

[0050] S104 performs collaborative processing of target scheduling instruction information and capability adaptation parameters of various models to generate real-time agent scheduling execution information.

[0051] In one implementation, target scheduling instruction information and large model capability adaptation parameters are captured based on collaborative processing rules, and features such as instruction execution requirements, model computing power thresholds, task adaptation accuracy, and response latency constraints are extracted. Based on preset collaborative processing rules, two types of core data are captured simultaneously: first, target scheduling instruction information, covering core fields such as instruction identifier, execution object, task type, resource configuration requirements, and execution time limit; second, large model capability adaptation parameters, including key parameters such as the computing power limit of each model, task processing type adaptation range, accuracy output benchmark, response latency interval, and resource consumption coefficient. The collaborative processing rules clearly define the triggering conditions for data capture, such as immediate triggering after the target scheduling instruction is generated, and data verification standards, such as parameter integrity, format standardization, and data association logic (e.g., associating the corresponding model's adaptation parameters according to the instruction task type), ensuring that the captured data is accurately matched and can be directly used for subsequent feature extraction.

[0052] From the captured target scheduling instruction information and the large model capability adaptation parameters, four types of core features are extracted, as follows: Regarding the instruction execution requirement features, it focuses on the core constraints of the instruction on task execution, including the task completion time node, the specific type and quantity of required resources, and the format and accuracy standards of the task output results.

[0053] Based on the model's computing power threshold characteristics, it is used to quantify the computing power carrying capacity of large models, including the maximum amount of data processed in a single batch of tasks, the upper limit of concurrent task processing, and the number of computational operations that can be completed per unit time (FLOPS).

[0054] Task-adaptive accuracy features are used to characterize the accuracy adaptation level of a model to a specific task type, such as the semantic understanding accuracy of natural language processing tasks and the target recognition accuracy of image processing tasks, which are inherent accuracy indicators of the model. In response to the characteristics of response latency constraints, it specifies the task response speed standards required by the instructions, including the maximum delay time from task reception to the start of execution, and the maximum allowable delay for the entire task process.

[0055] A matching analysis is performed between the instruction execution requirements and the model computing power threshold to generate capability matching parameters. Based on the computing power requirement in the instruction execution requirements, a bidirectional matching analysis is conducted with the model computing power threshold in the large model capability adaptation parameters. First, the instruction execution requirements are analyzed to clarify the minimum, recommended, and peak computing power requirements for completing the task. Then, the computing power threshold parameters of the corresponding candidate large models are retrieved, and a quantitative comparison is performed from three dimensions: "minimum computing power satisfaction," "recommended computing power adaptation," and "peak computing power redundancy."

[0056] An algorithm combining interval mapping and weighted calculation is used to generate capability matching parameters. The model computing power threshold is divided into multiple intervals and mapped to the instruction computing power requirement intervals. A basic matching score is calculated (e.g., 1 point for complete coverage of the requirement interval, 0.5-0.9 points for partial coverage, and 0 points for no coverage). Weights are then assigned to "minimum computing power satisfaction", "recommended computing power fit", and "peak computing power redundancy" (e.g., 0.4, 0.4, 0.2), and the weighted sum is used to obtain the final capability matching parameter, with a quantization range of 0-1 (1 represents complete fit, 0 represents no fit), which intuitively reflects the matching level between the model computing power and the instruction computing power requirement.

[0057] The model's accuracy features are calibrated and analyzed to adjust the model adaptation criteria based on the task type, generating an accuracy adaptation factor. The model's accuracy criteria are dynamically adjusted according to differences in task types. Pre-defined accuracy priority rules are established for different task types (e.g., medical data processing tasks have the highest accuracy priority, while ordinary text classification tasks have a medium accuracy priority). Differentiated accuracy adaptation benchmarks are set for tasks of different priorities: high-priority tasks require a model adaptation accuracy of no less than 99%, medium-priority tasks no less than 95%, and low-priority tasks no less than 90%. During the adjustment process, industry standards and business scenario requirements are considered to ensure the rationality and practicality of the adaptation criteria.

[0058] The adjusted model adaptation accuracy standard is compared and calibrated with the task adaptation accuracy characteristics of the large model to generate an accuracy adaptation factor. If the model's inherent accuracy index is higher than the adaptation standard, a positive factor is calculated as "(model accuracy - adaptation standard) / adaptation standard" (e.g., if the model accuracy is 99.5% and the adaptation standard is 99%, the factor is 0.005); if the model's inherent accuracy index is lower than the adaptation standard, a negative factor is calculated as "(model accuracy - adaptation standard) / adaptation standard" (e.g., if the model accuracy is 94% and the adaptation standard is 95%, the factor is -0.0105); if there is a perfect match, the factor is 0. The accuracy adaptation factor has a quantification range of -0.1 to 0.1 and is used to characterize the degree of deviation between the model accuracy and the task requirements.

[0059] The response latency constraint features are validated in a timely manner to detect anomalies where the model's response speed does not meet the instruction requirements, generating a latency deviation coefficient. Using the response latency constraint features in the instruction as the validation benchmark, and combining historical response latency data and real-time running status of the large model, the model's response speed is dynamically validated to ensure it meets the instruction requirements. Validation dimensions include "task start latency" (time from receiving the instruction to starting execution) and "total task processing latency" (time from starting execution to output completion), ensuring that the overall latency conforms to the instruction constraints.

[0060] The latency deviation calculation rules are set as follows: First, obtain the historical average response latency of the model under the same task type and resource configuration. Combine this with the real-time system resource load status (e.g., latency increases by 10%-20% under high load) to predict the model response latency in the current scenario. Then, compare it with the instruction response latency constraint and calculate the latency deviation coefficient according to "(predicted latency - constraint latency) / constraint latency". If the predicted latency ≤ constraint latency, the coefficient is non-positive (indicating no deviation or positive redundancy); if the predicted latency > constraint latency, the coefficient is positive (indicating a latency deviation exists). The quantization range is -0.2 to 0.5, which intuitively reflects the degree of deviation between the model response speed and the instruction requirements.

[0061] Based on preset collaborative processing rules, the capability matching degree parameter, accuracy adaptation factor, and time delay deviation coefficient are processed. A dynamic computing power allocation algorithm is used to supplement the computing power gap, adaptation standard optimization technology is used to improve matching accuracy, and a task scheduling acceleration algorithm is used to correct time delay deviations, generating collaborative optimization results. For the computing power gap reflected by the capability matching degree parameter (a gap is determined to exist when the matching degree parameter < 0.8), a dynamic computing power allocation algorithm is used to supplement it. The algorithm first calculates the missing computing power value (recommended computing power of the instruction - currently available computing power of the model), then schedules supplementation from the system's idle computing power pool and redundant computing power occupied by low-priority tasks, or splits tasks exceeding the computing power capacity of a single model into multiple models for parallel processing through task splitting, ensuring that the computing power gap is accurately filled and the capability matching degree parameter is improved to above 0.8.

[0062] To address the accuracy deviation reflected by the accuracy fitting factor (a factor < 0 indicates a deviation), an adaptation standard optimization technique is employed to improve matching accuracy. For negative accuracy fitting factors, methods such as fine-tuning model parameters (e.g., adjusting the confidence threshold during model inference), enhancing task preprocessing (e.g., denoising and normalizing input data), and post-processing output results (e.g., fusing results from multiple models) are used to improve the model's accuracy performance during actual task execution, bringing the accuracy fitting factor closer to 0 or reaching a positive value.

[0063] To address the latency deviation reflected by the latency deviation coefficient (a coefficient > 0 indicates a deviation), a task scheduling acceleration algorithm is employed for correction. The algorithm reduces task startup latency and total processing latency by optimizing task execution processes (e.g., merging repetitive calculation steps), adjusting model runtime configurations (e.g., increasing the number of model inference threads), and optimizing data transmission paths (e.g., using data transmission from the nearest node). For scenarios with significant deviations, the algorithm can trigger task migration to a faster backup model node, ensuring the latency deviation coefficient drops below 0.

[0064] By integrating various parameters after supplementing computing power, improving accuracy, and correcting latency, a collaborative optimization result is generated, including core components such as optimized model computing power configuration parameters, accuracy assurance scheme, latency control target, and task execution path planning. This ensures that deviations in all dimensions are effectively corrected, meeting the core requirements of multi-model collaborative scheduling.

[0065] The results of the collaborative optimization are integrated to generate real-time intelligent agent scheduling and execution information that meets the requirements of multi-model collaborative scheduling. Key data from the collaborative optimization results are collected, including: optimized capability matching parameters (target ≥ 0.8), accuracy adaptation factor (target ≥ 0), latency deviation coefficient (target ≤ 0), computing power allocation details, accuracy assurance measures, latency optimization schemes, task execution node identifiers, and backup node configurations, ensuring that the integrated information comprehensively covers all key dimensions of task execution.

[0066] According to the preset real-time intelligent agent scheduling execution information format specifications, the integrated information is standardized: the data quantification unit is unified (such as the computing power unit is FLOPS, and the latency unit is millisecond), the parameter encoding rules are standardized (such as the task type is encoded with two digits and the node identifier is encoded with a combination of letters and numbers), and the information association logic is clarified (such as associating the corresponding optimization parameters and execution nodes by task ID), so as to ensure that the information has consistency, identifiability and parsability.

[0067] The final generated real-time intelligent agent scheduling and execution information is a structured document containing core modules such as basic task information (ID, type, priority), execution model information (main execution node, backup node, model configuration parameters), resource configuration information (computing power, memory, bandwidth allocation details), performance guarantee information (accuracy target, latency limit), and execution process information (startup conditions, step sequence, feedback node). This provides a clear and standardized operational basis for subsequent task distribution to large model nodes and collaborative execution.

[0068] S105: Based on preset business service quality requirements and resource scheduling constraints, real-time intelligent agent scheduling execution information is distributed to major model nodes. The preset business service quality requirements, resource scheduling constraints, real-time intelligent agent scheduling execution information and major model node attributes are extracted and classified to generate node adaptation parameter information, scheduling priority rule information, task distribution permission information, and task-model association mapping information, so that major model nodes can work together to complete real-time business tasks.

[0069] In one implementation, features are extracted and categorized from preset service quality requirements, resource scheduling constraints, real-time agent scheduling execution information, and large model node attributes to generate service quality threshold feature information, resource constraint boundary feature information, scheduling execution instruction feature information, node computing power adaptation feature information, node response capability feature information, and task compatibility attribute feature information. The system comprehensively captures and structures four types of core input data: preset service quality requirements define the quality standards for task execution (e.g., response speed, output accuracy, stability requirements); resource scheduling constraints define the allocation boundaries of system resources (e.g., computing power limits, memory quotas, network bandwidth limitations); real-time agent scheduling execution information includes specific instructions for task execution, resource configuration details, and performance guarantee targets; and large model node attributes cover basic information such as node hardware configuration, computing power level, supported task types, response capability benchmarks, and compatibility characteristics. The parsing process ensures data integrity and semantic consistency, laying the foundation for feature extraction.

[0070] Based on data attributes and scheduling requirements, six types of feature information are extracted and classified, as follows: Service quality threshold feature information: Critical standards for quantifying business quality requirements, including quantifiable indicators such as upper limit of task response latency, threshold of output result accuracy, minimum requirement of task execution success rate, and service availability guarantee threshold; Resource constraint boundary feature information: clearly defines the constraints on resource allocation, including the upper limit of computing power allocation per node, memory usage threshold, peak network bandwidth usage, and resource usage duration constraints, etc. Dispatch execution instruction characteristic information: Extract the core execution elements of the dispatch instruction, including task identifier, execution priority, target node range, resource allocation requirements, execution timing requirements, feedback mechanism description, etc. Node computing power adaptation feature information: Characterizes the computing power adaptation capability of large model nodes, including node peak computing power, continuous computing power output level, concurrent task processing limit, computing power consumption coefficient of different task types, etc. Node response capability characteristics: Reflects the node's task response efficiency, including the average latency from task reception to startup, the average time taken to process the task, peak response speed, and the range of response stability fluctuations; Task compatibility attribute characteristics: Reflects the degree of compatibility of a node with different tasks, including a list of supported task types, the processing adaptability of specific tasks, the collaborative compatibility characteristics with other nodes, and software environment adaptation requirements.

[0071] The algorithm performs correlation processing on service quality threshold features, resource constraint boundary features, scheduling execution instruction features, node computing power adaptation features, node response capability features, and task compatibility attribute features to generate node adaptation matching information, scheduling priority ranking information, task distribution permission level information, and model-task adaptation correlation information. A feature association algorithm is used to perform multi-dimensional cross-matching on these six types of features. The core logic is to establish a mapping relationship between features based on scheduling business logic, and to achieve collaborative feature association by calculating feature similarity, dependency, and adaptability. The algorithm first standardizes and quantifies various features (converting non-numerical features into encoded values ​​and normalizing numerical features to the 0-1 range), and then identifies strong correlations between features through cosine similarity calculation and association rule mining.

[0072] By associating node computing power adaptation features, node response capability features, task compatibility attribute features with service quality threshold features, resource constraint boundary features, and scheduling execution instruction features, an adaptation score for each node and task is calculated (weight allocation: computing power adaptation 0.3, response capability 0.25, task compatibility 0.25, quality threshold adaptation 0.1, resource constraint adaptation 0.1), generating a node adaptation matching degree parameter (quantization range 0-1) to characterize the overall adaptation level of the node to the task.

[0073] By combining the initial priority in the scheduling execution instruction characteristics, the urgency requirement in the service quality threshold characteristics, and the resource scarcity in the resource constraint boundary characteristics, the task priorities are reordered using the analytic hierarchy process (AHP) to generate a scheduling priority ranking result and clarify the order of task execution.

[0074] Based on the permission requirements in the scheduling execution instruction characteristics, the resource ownership restrictions in the resource constraint boundary characteristics, and the permission levels in the large model node attributes, the task distribution permissions are classified into levels (such as core permissions, ordinary permissions, and temporary permissions), generating task distribution permission classification information and defining the execution permissions of different nodes for tasks.

[0075] By accurately matching the task type and processing requirements in the node computing power adaptation features, task compatibility attribute features and scheduling execution instruction features, a mapping relationship between tasks and adaptation model nodes is established, and a list of candidate execution nodes for each task is clarified.

[0076] Based on node adaptation matching information, scheduling priority ranking information, task distribution permission level information, and model-task adaptation association information, invalid data in various feature information is filtered and calibrated to generate feature data screening and calibration results. Multi-dimensional invalid data identification rules are set: node adaptation matching parameters below 0.6 are considered invalid adaptation data; scheduling priority ranking with logical conflicts (e.g., low-urgency tasks have higher priority than high-urgency tasks) is considered invalid ranking data; task distribution permission level not matching node permission level is considered invalid permission data; and nodes not supporting task types in model-task adaptation associations are considered invalid association data. Based on these rules, the four types of information after association processing are comprehensively screened to filter invalid data.

[0077] The filtered valid data is calibrated and optimized. For node adaptation matching information, the adaptation score is recalculated by adjusting feature weight coefficients (e.g., the adaptation weight for core task quality improvement thresholds). For scheduling priority ranking information, priority conflicts are resolved (re-ranking is performed using the principle of "urgency priority + resource adaptation assistance"). For task distribution permission hierarchy information, node permission verification results are supplemented, and permission hierarchy is adjusted. For model-task adaptation association information, duplicate association records are removed, and backup node associations are added. Finally, feature data filtering and calibration results are generated, specifically including: Filter nodes with a matching degree parameter ≥ 0.6, and identify the primary execution node and backup node for each task; sort the results according to the calibrated priority to form a hierarchical sequence of "highest level - high level - medium level - low level", and mark the priority level and execution priority number of each task; clarify the distribution permission level of each task, define the scope of nodes that can execute the task and the duration of permission use; establish a one-to-one or one-to-many precise association between tasks and nodes, and clarify the execution adaptability and priority of each task on different nodes.

[0078] The feature data screening and calibration results are integrated and standardized to generate node adaptation parameter information, scheduling priority rule information, task distribution permission information, and task-model association mapping information. First, the feature data screening and calibration results are aggregated from multiple dimensions to ensure the completeness and logical consistency of information regarding the effective adaptation node list, hierarchical priority sequence, permission allocation range, and precise association mapping relationships. Then, the data is standardized according to a preset scheduling information format standard, unifying data encoding rules (e.g., node identifiers use a combination of letters and numbers, and task identifiers use a unified UUID format), quantization units (e.g., latency unit is milliseconds, computing power unit is FLOPS), and expression specifications to ensure the information is resolvable and universal.

[0079] By integrating a list of effective compatible nodes with node compatibility information, the system clarifies the operational configuration standards for each node (such as computing power allocation parameters, memory usage configuration, and response speed thresholds), compatible task types, and compatibility priority ranges, providing a configuration basis for node operation and task allocation. Based on a hierarchical priority sequence, scheduling priority execution rules are formulated, clarifying the resource allocation priority, execution timing rules, and conflict handling mechanisms for tasks of different priorities (such as high-priority tasks being able to preempt resources from low-priority tasks), standardizing the task execution order logic. Combined with the scope of permission allocation, the system clarifies the permission approval process for task distribution, permission activation conditions, and permission revocation mechanisms, defining the boundaries of task execution permissions for different nodes to avoid permission abuse or unauthorized execution. Based on precise association mapping relationships, a structured association mapping table is generated, clarifying the main execution node, backup node, node compatibility, execution priority, and other core association information for each task, achieving precise binding between tasks and model nodes.

[0080] The above four types of information work together to clarify the node operation configuration, task execution priority, permission allocation boundaries and model-task correspondence, providing a comprehensive and standardized execution basis for the collaborative scheduling of major model nodes and supporting the efficient completion of real-time business tasks.

[0081] In one implementation, such as Figure 2 As shown, this application also provides a real-time intelligent agent dynamic scheduling system composed of multiple large models, including: The data acquisition and preprocessing unit 201 is used to receive task request data sent by at least one real-time business scenario, and to perform group management and scheduling strategy pre-configuration processing on the received task request data based on task type, priority, and resource requirements, generating pre-scheduled task queue information and resource allocation pre-planning information; it receives target configuration information and optimizes the pre-scheduled task queue information corresponding to each group of tasks, including task splitting, priority rearrangement, model capability matching, resource reservation, conflict resolution, timeout contingency plan, load balancing, and fault tolerance redundancy, generating secondary optimized scheduling task information; The data processing and optimization unit 202 is used to dynamically adjust the secondary optimized scheduling task information based on real-time system resource status data and model running status data, and generate target scheduling instruction information; to collaboratively process the target scheduling instruction information and the capability adaptation parameters of various models, and generate real-time intelligent agent scheduling execution information; and to distribute the real-time intelligent agent scheduling execution information to various model nodes based on preset business service quality requirements and resource scheduling constraints. It also extracts and classifies the preset business service quality requirements, resource scheduling constraints, real-time intelligent agent scheduling execution information, and large model node attributes to generate node adaptation parameter information, scheduling priority rule information, task distribution permission information, and task-model association mapping information, so that various model nodes can collaboratively complete real-time business tasks.

[0082] The computer-readable storage medium provided in the above embodiments of this application and the real-time intelligent agent dynamic scheduling method composed of multiple large models provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0083] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for evaluating the dynamic scheduling method, system, electronic device, and readable storage medium for real-time intelligent agents composed of multiple large models are basically similar to the embodiments of the dynamic scheduling method for real-time intelligent agents composed of multiple large models described above, and therefore the description is relatively simple. Relevant parts can be referred to in the descriptions of the embodiments of the dynamic scheduling method for real-time intelligent agents composed of multiple large models described above.

Claims

1. A method for dynamic scheduling of real-time intelligent agents composed of multiple large models, characterized in that, include: Receive task request data sent by at least one real-time business scenario, and perform group management and scheduling strategy pre-configuration processing on the received task request data based on task type, priority, and resource requirements to generate pre-scheduled task queue information and resource allocation pre-planning information. The different task types include task request data sent by at least one real-time business scenario. Receive target configuration information and optimize the pre-scheduled task queue information corresponding to each group of tasks, including task splitting, priority rearrangement, model capability matching, resource reservation, conflict resolution, timeout contingency plan, load balancing, fault tolerance redundancy, and generate secondary optimized scheduling task information. Based on real-time system resource status data and model running status data, the scheduling task information after secondary optimization is dynamically adjusted to generate target scheduling instruction information; The target scheduling instruction information and the capability adaptation parameters of various models are processed collaboratively to generate real-time intelligent agent scheduling and execution information. Based on preset business service quality requirements and resource scheduling constraints, real-time intelligent agent scheduling and execution information is distributed to major model nodes. The preset business service quality requirements, resource scheduling constraints, real-time intelligent agent scheduling and execution information, and major model node attributes are extracted and classified to generate node adaptation parameter information, scheduling priority rule information, task distribution permission information, and task-model association mapping information, so that major model nodes can work together to complete real-time business tasks.

2. The method as described in claim 1, characterized in that, Receive target configuration information and optimize the pre-scheduled task queue information corresponding to each group of tasks, including task splitting, priority rearrangement, model capability matching, resource reservation, conflict resolution, timeout contingency plan, load balancing, and fault tolerance redundancy, to generate secondary optimized scheduling task information, including: The task splitting algorithm is used to decompose and reconstruct grouped tasks in a granular manner. Combined with the model capability adaptation matrix, the task and the large model are accurately matched to achieve the construction of the basic layer for optimization. By introducing a priority dynamic reordering model and a resource reservation scheduling protocol, a three-dimensional optimization matrix of task priority, model load, and resource margin is constructed to generate differentiated resource allocation schemes. By integrating a conflict resolution rule base with a timeout contingency triggering mechanism, and dynamically adjusting task execution nodes through a load balancing algorithm, a multi-dimensional fault-tolerant redundancy strategy is established. The results of the splitting, matching schemes, resource allocation, conflict resolution, timeout handling mechanisms, and load balancing strategies are integrated to generate standardized scheduling task information after secondary optimization.

3. The method as described in claim 2, characterized in that, Based on real-time system resource status data and model runtime status data, the secondary optimized scheduling task information is dynamically adjusted to generate target scheduling instruction information, including: The system performs parameter mapping and rule normalization on real-time system resource data and model running status data through a state parsing process, and introduces a resource load assessment model and dynamic scheduling arbitration algorithm to realize structured processing of adjustment strategies. By aligning with the intelligent agent scheduling protocol standard, a three-dimensional adjustment matrix of resource type, model state, and task requirements is constructed. A standardized scheduling parameter table is generated by converting the scheduling model through benchmarking, and a dynamic adjustment and update mechanism is established. Using the task scheduling lifecycle as a time window, a multi-dimensional scheduling feature matrix is ​​constructed by integrating resource type, model status, task requirements, and historical scheduling adjustment records to achieve full-dimensional correlation of adjustment strategies. By supplementing missing adjustment rules through resource prediction models, extracting key strategies through dynamic state feedback mechanisms, and fusing them with the features of the three-dimensional adjustment matrix, a target scheduling instruction set is generated.

4. The method as described in claim 1, characterized in that, The target scheduling instruction information and the capability adaptation parameters of various models are collaboratively processed to generate real-time agent scheduling execution information, including: Based on collaborative processing rules, target scheduling instruction information and large model capability adaptation parameters are captured, and instruction execution requirements, model computing power thresholds, task adaptation accuracy and response latency constraints are extracted. Matching analysis is performed between instruction execution requirements and model computing power thresholds to generate capability matching parameters; The model adaptation accuracy characteristics of the task are calibrated and analyzed, and the model adaptation standard is adjusted according to the task type to generate an accuracy adaptation factor. Timeliness verification is performed on the response delay constraint characteristics to detect abnormal situations where the model response speed does not meet the instruction requirements, and a delay deviation coefficient is generated; Based on the preset collaborative processing rules, the capability matching degree parameter, accuracy adaptation factor and time delay deviation coefficient are processed. The computing power dynamic allocation algorithm is used to supplement the computing power gap, the adaptation standard optimization technology is used to improve the matching accuracy, and the task scheduling acceleration algorithm is used to correct the time delay deviation, so as to generate collaborative optimization results. By integrating the results of collaborative optimization, real-time intelligent agent scheduling and execution information that meets the requirements of multi-model collaborative scheduling is generated.

5. The method as described in claim 4, characterized in that, Based on preset service quality requirements and resource scheduling constraints, real-time agent scheduling and execution information is distributed to all model nodes. The preset service quality requirements, resource scheduling constraints, real-time agent scheduling and execution information, and large model node attributes are extracted and classified to generate node adaptation parameter information, scheduling priority rule information, task distribution permission information, and task-model association mapping information. This enables all model nodes to collaboratively complete real-time service tasks, including: Feature extraction and classification are performed on preset business service quality requirements, resource scheduling constraints, real-time intelligent agent scheduling execution information and large model node attributes to generate service quality threshold feature information, resource constraint boundary feature information, scheduling execution instruction feature information, node computing power adaptation feature information, node response capability feature information and task compatibility attribute feature information; The service quality threshold feature information, resource constraint boundary feature information, scheduling execution instruction feature information, node computing power adaptation feature information, node response capability feature information, and task compatibility attribute feature information are correlated to generate node adaptation matching degree information, scheduling priority ranking information, task distribution permission hierarchical information, and model-task adaptation correlation information. Based on node adaptation matching information, scheduling priority ranking information, task distribution permission hierarchical information, and model-task adaptation association information, invalid data in various feature information is filtered and calibrated to generate feature data screening and calibration results, including a list of effective adapted nodes, hierarchical priority sequence, permission allocation range, and accurate association mapping relationship. The feature data screening and calibration results are integrated and standardized to generate node adaptation parameter information, scheduling priority rule information, task distribution permission information, and task-model association mapping information. Among them, the four types of information are used to clarify the node operation configuration standards, task execution priority logic, permission allocation boundaries, and model-task correspondence, thereby supporting the collaborative and efficient completion of real-time business tasks by various model nodes.

6. A real-time intelligent agent dynamic scheduling system composed of multiple large models, characterized in that, The system includes: The data acquisition and preprocessing unit is used to receive task request data sent by at least one real-time business scenario, and to perform group management and scheduling strategy pre-configuration processing on the received task request data based on task type, priority, and resource requirements, generating pre-scheduled task queue information and resource allocation pre-planning information; it receives target configuration information and optimizes the pre-scheduled task queue information corresponding to each group of tasks, including task splitting, priority rearrangement, model capability matching, resource reservation, conflict resolution, timeout contingency plan, load balancing, and fault tolerance redundancy, generating secondary optimized scheduling task information; The data processing and optimization unit dynamically adjusts the secondary optimized scheduling task information based on real-time system resource status data and model running status data to generate target scheduling instruction information; it collaboratively processes the target scheduling instruction information and the capability adaptation parameters of various models to generate real-time intelligent agent scheduling execution information; based on preset business service quality requirements and resource scheduling constraints, it distributes the real-time intelligent agent scheduling execution information to various model nodes; it extracts and classifies the preset business service quality requirements, resource scheduling constraints, real-time intelligent agent scheduling execution information, and large model node attributes to generate node adaptation parameter information, scheduling priority rule information, task distribution permission information, and task-model association mapping information, so that various model nodes can collaboratively complete real-time business tasks.

7. An electronic device, characterized in that, include: First processor; and memory for storing executable instructions of the first processor; The first processor is configured to execute the real-time agent dynamic scheduling method consisting of multiple large models as described in any one of claims 1 to 5 by executing the executable instructions.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the second processor, it implements the real-time intelligent agent dynamic scheduling method composed of multiple large models as described in any one of claims 1 to 5.

Images