Large model call path intelligent scheduling system and method based on task load and scene

Abstract

The invention belongs to the technical field of large model intelligent scheduling, and particularly relates to a large model calling path intelligent scheduling system and method based on task load and scene, and the method comprises the steps: firstly collecting task sensing data and obtaining resource load data of nodes; then, based on an improved ant lion optimization algorithm, task perception data and resource load data serve as input, an optimal calling path is output through coding mapping, fitness calculation and iterative optimization, and a calling instruction is generated according to the optimal calling path; and finally, issuing the calling instruction to the target node, and executing large model calling. And meanwhile, feedback data are collected and called, and a feedback vector is generated to perform reverse iteration to adjust algorithm parameters in the intelligent scheduling module. According to the method, the problem that effective utilization of resources in an enterprise mixed scene cannot be met due to single dimension and rule solidification of a large model calling scheduling scheme in the prior art can be solved.

Description

Technical Field

[0001] This invention belongs to the field of large model intelligent scheduling technology, and in particular relates to a large model call path intelligent scheduling system and method based on task load and scenario. Background Technology

[0002] With the rapid iteration of generative artificial intelligence technology, the application scenarios of large models in enterprise-level applications are becoming increasingly diversified, ranging from public domain customer service consultations and general content generation to private domain financial data processing and customer privacy analysis. Different scenarios place different demands on the security, response efficiency, and resource adaptability of large model calls. At the same time, the deployment architecture of enterprise large models is gradually evolving towards a hybrid model of "private cloud + authorized public cloud". The performance differences of nodes within the cluster are significant, such as private cloud nodes with high GPU computing power and public cloud nodes with lightweight computing power. The complexity of tasks also shows a hierarchical characteristic. For example, simple question-answering tasks have low computing requirements, while complex data analysis tasks have high requirements for computing power and memory. This poses a greater challenge to the scheduling capabilities of large model calls.

[0003] Current mainstream large-scale model invocation and scheduling solutions still suffer from three major technical bottlenecks, making it difficult to adapt to the needs of enterprise hybrid scenarios and dynamic tasks, including: The scheduling dimension is too singular and the adaptation to scenarios and performance is lacking: Existing solutions mostly use cluster resource load as the sole scheduling basis, and allocate nodes only through indicators such as CPU / GPU utilization and memory usage, without taking into account the attributes of the task itself and the security requirements of the scenario. Fixed path rules and insufficient dynamic adaptability: Most scheduling systems use static routing configuration, which allocates call paths through preset fixed rules, such as binding a certain type of task to a certain group of nodes, and cannot respond to load fluctuations and task changes in real time.

[0004] High computing power costs and the difficulty in balancing compliance and efficiency: To cope with peak task pressure, enterprises need to purchase additional high-performance nodes for expansion, resulting in an average annual increase of 35% in hardware investment costs. The above; meanwhile, to mitigate security risks in confidential scenarios, some enterprises choose to deploy all tasks uniformly on a private cloud, further exacerbating the load pressure on private cloud nodes and creating a contradiction of "rising costs but low resource utilization." Furthermore, in static scheduling mode, routing rules need to be manually adjusted periodically, requiring maintenance personnel to invest 30... The above-mentioned work time is used to handle scheduling anomalies, which increases both management and human resource costs. Summary of the Invention

[0005] The technical problem solved by this invention is to provide an intelligent scheduling system and method for large model call paths based on task load and scenario, so as to solve the problem that existing large model call scheduling schemes cannot meet the effective utilization of resources in mixed enterprise scenarios due to their single dimension and fixed rules.

[0006] The basic solution provided by this invention is: an intelligent scheduling system for large-scale model call paths based on task load and scenario, comprising a task awareness module, a resource monitoring module, an intelligent scheduling module, and an execution and feedback module, wherein: The task awareness module is used to collect task complexity and task scenario attributes in real time and generate task awareness data. The resource monitoring module is used to monitor the resource load of each node in the cluster in real time and generate resource load data; The intelligent scheduling module is used to optimize the antlion algorithm based on the improved algorithm. It takes task perception data and resource load data as input, outputs the optimal calling path through encoding mapping, fitness calculation and iterative optimization, and generates calling instructions from the optimal calling path. The execution and feedback module is used to send the call command to the target node based on the protocol to execute the large model call; at the same time, it collects the call feedback data, generates feedback vectors, and performs reverse iterative adjustment of the algorithm parameters in the intelligent scheduling module.

[0007] Furthermore, the task complexity and task scenario attributes collected in real time in the task awareness module are specifically as follows: Deploy a text parsing service to receive large model call requests and task description texts submitted by business systems; Calculate the number of tokens in the task description text. Calculate the basic complexity The expression is:

[0008] Identify the reasoning requirements of the task and assign additional complexity. Calculate the final task complexity ; Establish a keyword database that includes both classified and unclassified keywords. Use the TF-IDF algorithm to calculate the similarity between the task text and the keyword database. When the similarity exceeds a preset threshold, it is marked as the corresponding scenario, and the task scenario attributes are output. ; Based on the final task complexity and task scenario attributes Generate task feature vectors And it is sent to the intelligent scheduling module via the protocol.

[0009] Furthermore, the resource monitoring module monitors the resource load of each node in the cluster in real time and generates resource load data as follows: Deploy distributed monitoring agents on each private cloud node and each authorized public cloud node in the cluster; Real-time CPU utilization is collected via an agent. GPU utilization Memory usage Broadband network ; Based on the weighted calculation of the overall load rate of the nodes, the expression is:

[0010] Generate node load vector The data is synchronized in real time to the intelligent scheduling module and the cluster load distribution is visualized in the form of a heatmap.

[0011] Furthermore, the intelligent scheduling module, based on the improved Antlion optimization algorithm, takes task-aware data and resource load data as input, and outputs the optimal calling path through encoding mapping, fitness calculation, and iterative optimization as follows: Obtain task feature vector and node load vector The initial population is generated using a mixture of three-dimensional binary and integer encoding. ;in Indicates compliance with scene requirements. Indicates the node ID. This indicates the compatibility between tasks and nodes; A fitness function is constructed based on three constraints: scenario compliance, performance adaptation, and load balancing. The expression is:

[0012] in, Indicates the compliance weight of the scenario. Indicates performance adaptation weights. Indicates the load balancing weight. Indicates compliance with scene requirements. This indicates the compatibility between tasks and nodes. Indicates the overall load rate of the nodes. Indicates the average load rate of the cluster. Indicates the maximum load rate of the cluster; ; Substitute each node into the fitness function Calculate scores and sort them before filtering. Individuals are considered elite antlions; Constructing a dynamic step size factor The optimization accuracy is adjusted based on the number of iterations and load fluctuations, expressed as:

[0013] in, Indicates the initial step size. This represents the current iteration number. Indicates the maximum number of iterations. Indicates the standard deviation of the load factor. Indicates the overall load rate of the nodes; Adjust the ants' search direction based on the location of the elite antlion, and update the ants' positions; Calculate the fitness of the updated ant; if it is higher than the current elite antlion, replace it; continue this process until the maximum number of iterations, then select the individual with the highest fitness and decode. For the target node ID, generate the optimal call path from the business system to the target node to the corresponding large model.

[0014] Furthermore, the acquisition of task feature vectors and node load vector The initial population is generated using a mixture of three-dimensional binary and integer encoding. Specifically: Obtain task feature vector and node load vector Determine the attributes of the task scenario The value, if If the characterization involves classified information, then only private cloud nodes will be included in the coding scope. If the information is not classified, then include it in all nodes; Based on the included nodes, several coded individuals are randomly generated. The expression is:

[0015] in, Indicates compliance with scene requirements; Indicates the node ID; The degree of fit between tasks and nodes is expressed as:

[0016] in, This indicates the GPU utilization of the current node. This represents the maximum GPU utilization among all available compliant nodes. This represents the maximum preset task complexity of the system. This represents a quantified value indicating the complexity of the current task. This represents the scene weight coefficient.

[0017] Furthermore, the step of adjusting the ant's search direction based on the location of the elite antlion and updating the ant's position specifically involves: During the iteration process, the node ID is updated, and the compatibility between the task and the node is updated, while the scenario compliance identifier remains unchanged. The expression for updating the node ID is:

[0018] The expression for updating the fit between tasks and nodes is:

[0019] in, This represents the node ID of the (t+1)th generation of ordinary ants. This represents the node ID of the t-th generation of ordinary ants. This represents the node ID of the t-th generation elite ant. This is the dynamic step size factor. These are random numbers distributed according to a standard normal distribution. This represents the fitness of ordinary ants in generation t+1. This represents the fitness of the t-th generation of common ants; This represents the rounding function.

[0020] Furthermore, the execution and feedback module includes a call execution unit, a data acquisition unit, and a feedback optimization unit, wherein: The execution unit is used to schedule the server to send the large model invocation instruction to the target node according to the gRPC protocol, and to control the target node to load the corresponding large model and perform inference calculation after receiving the invocation instruction; The data acquisition unit is used to collect call response time, node resource utilization, and task completion rate; The feedback optimization unit is used to determine whether the call response time is greater than the preset call response time threshold or the node resource utilization rate is less than the preset node resource utilization threshold. When this condition is met, the initial step size of the dynamic step size factor in the improved Antlion optimization algorithm is adjusted to the next iteration. The fitness function weights are also optimized based on the fitness function value and historical data.

[0021] A method for intelligent scheduling of large model call paths based on task load and scenario, applied to the aforementioned intelligent scheduling system for large model call paths based on task load and scenario, includes: S1: Real-time acquisition of task complexity and task scenario attributes to generate task-aware data; S2: Real-time monitoring of resource load on each node in the cluster and generation of resource load data; S3: Based on the improved Antlion optimization algorithm, it takes task-aware data and resource load data as input, outputs the optimal calling path through encoding mapping, fitness calculation and iterative optimization, and generates calling instructions from the optimal calling path; S4: Based on the protocol, the call command is sent to the target node to execute the large model call; at the same time, the call feedback data is collected, and a feedback vector is generated to perform reverse iterative adjustment of the algorithm parameters in the intelligent scheduling module.

[0022] The principle and advantages of this invention are as follows: In the technical solution of this application, relying on the closed-loop architecture of "perception-computation-scheduling-feedback", the intelligent optimization of the large model call path is achieved through the collaborative operation of four core modules; among them, the task perception module extracts the complexity features and scene attributes of the task and transforms them into structured task perception data; the resource monitoring module collects the CPU / GPU utilization, memory usage and network bandwidth of each node in the cluster in real time through distributed monitoring components, and obtains the comprehensive load status of the nodes through weighted calculation to generate resource load data. The intelligent scheduling module, as the core of the system, introduces an improved Antlion optimization algorithm. It first transforms task characteristics and node attributes into a three-dimensional hybrid code containing "scenario compliance identifier, node ID, and task-node adaptability". Then, by integrating a fitness function that combines scenario compliance constraints, performance adaptation requirements, and load balancing goals, it quantitatively evaluates the merits of different calling paths. Finally, through an iterative optimization process with dynamic step size adjustment, it determines the optimal calling path and generates calling instructions. The execution and feedback module, based on a standardized communication protocol, sends calling instructions to target nodes to execute large model calls. At the same time, it collects response performance, resource utilization efficiency, and task completion status during the calling process to form a feedback vector. This vector is used to iteratively adjust the algorithm parameters of the intelligent scheduling module and continuously optimize scheduling accuracy.

[0023] The advantages lie in achieving technological breakthroughs in multiple aspects through multi-dimensional scheduling logic and dynamic optimization mechanisms, specifically: 1. In terms of scenario adaptation and security assurance, the node call scope of classified and non-classified tasks is strictly distinguished through coding mapping to ensure that classified tasks are only matched with private cloud nodes, thereby avoiding data security risks from the root and accurately meeting scenario compliance requirements. 2. In terms of response efficiency and resource utilization, the improved Antlion optimization algorithm can accurately match high-performance nodes for high-complexity tasks and reasonably fill idle nodes for lightweight tasks, effectively avoiding response delays caused by resource mismatch, while significantly improving the overall resource utilization efficiency of the cluster and reducing computing power waste. 3. In terms of cost control and ease of operation and maintenance, the dynamic load balancing mechanism can avoid the cost pressure caused by additional capacity expansion during peak business periods, and significantly reduce the cost of computing power procurement; at the same time, the feedback iteration mechanism realizes the automatic optimization of algorithm parameters, greatly reduces the frequency of manual intervention, and reduces the manpower and time costs of operation and maintenance management. Attached Figure Description

[0024] Figure 1 This is a functional block diagram of an embodiment of the present invention; Figure 2 This is a flowchart of an embodiment of the present invention. Detailed Implementation

[0025] The following detailed description illustrates the specific implementation method: The basic implementation examples are as follows: Figure 1 As shown: A large-scale model call path intelligent scheduling system based on task load and scenario includes a task awareness module, a resource monitoring module, an intelligent scheduling module, and an execution and feedback module. The task awareness module is used to collect task complexity and task scenario attributes in real time, generating task awareness data. In this embodiment, the process of collecting task complexity is as follows: Deploy NLP text parsing and rule matching services to receive large model call requests and task description texts submitted by business systems; Calculate the number of tokens in the task description text. Calculate the basic complexity The expression is:

[0026] Identify the reasoning requirements of the task and assign additional complexity. Calculate the final task complexity In this embodiment, the task complexity level is 1-10.

[0027] The process of collecting task scenario attributes is as follows: Establish a keyword database that includes both classified and unclassified keywords. Use the TF-IDF algorithm to calculate the similarity between the task text and the keyword database. When the similarity exceeds a preset threshold, it is marked as the corresponding scenario, and the task scenario attributes are output. ; In this embodiment, the established keyword database includes classified keywords such as "financial," "privacy," and "classified," and non-classified keywords such as "customer service," "public," and "general." The TF-IDF algorithm is used to calculate the similarity between the task text and the keyword database. At 0.6, it is marked as the corresponding scenario.

[0028] Finally, based on the final task complexity and task scenario attributes Generate task feature vectors It is then sent to the intelligent scheduling module via the HTTP / JSON protocol.

[0029] The resource monitoring module is used to monitor the resource load of each node in the cluster in real time and generate resource load data; specifically: Deploy distributed monitoring agents on each private cloud node and each authorized public cloud node in the cluster; Real-time CPU utilization is collected via an agent. GPU utilization Memory usage Broadband network The sampling frequency is 10 seconds per sampling. Based on the weighted calculation of the overall load rate of the nodes, the expression is:

[0030] Generate node load vector The data is synchronized in real time to the intelligent scheduling module, and the cluster load distribution is visualized in the form of a heatmap. In this embodiment, the node load heatmap is represented by red. Yellow character Green characterization .

[0031] The intelligent scheduling module, based on an improved Antlion optimization algorithm, takes task-aware data and resource load data as input, and outputs the optimal call path through encoding mapping, fitness calculation, and iterative optimization, and then generates call instructions from the optimal call path; the encoding mapping process specifically includes: Obtain task feature vector and node load vector Determine the attributes of the task scenario The value, if If the characterization involves classified information, then only private cloud nodes will be included in the coding scope. If a node is not classified, it is included in all nodes. In this embodiment, the private cloud cluster has 3 high-performance GPU nodes with node IDs 1-3. The GPUs of the private cloud nodes are NVIDIA A100, with a computing power of 400 TFLOPS and 256 GB of memory, and are used for classified tasks. The authorized public cloud nodes have 2 lightweight computing nodes with node IDs 4-5. The GPUs of the public cloud nodes are NVIDIA T4, with a computing power of 80 TFLOPS and 128 GB of memory, and are used for non-classified tasks.

[0032] Based on the included nodes, several coded individuals are randomly generated. The expression is:

[0033] in, Indicates compliance with scene requirements; Indicates the node ID; The degree of fit between tasks and nodes is expressed as:

[0034] in, This indicates the GPU utilization of the current node. This represents the maximum GPU utilization among all available compliant nodes. This represents the maximum preset task complexity of the system. This represents a quantified value indicating the complexity of the current task. This represents the scene weight coefficient.

[0035] Meanwhile, during the encoding mapping process, after generating the initial encoding, invalid encodings need to be excluded through "double verification" to avoid affecting subsequent algorithm iterations. Specifically: 1. Scenario compliance verification: If This indicates that it involves classified information, but For the corresponding public cloud node, then If forced to be 0, the encoding is marked as invalid and removed from the initial population; 2. Load threshold verification: If Corresponding node If the load threshold is greater than or equal to the load threshold, then If forced to be 0, the encoding is marked as low-fit, and will be naturally eliminated during subsequent fitness calculations; 3. Correction rule: If the proportion of invalid codes in the initial population exceeds... Then, update the node state and regenerate the encoding to ensure that the proportion of valid encodings in the initial population exceeds [a certain percentage]. .

[0036] The fitness calculation process in the intelligent scheduling module is as follows: A fitness function is constructed based on three constraints: scenario compliance, performance adaptation, and load balancing. The expression is:

[0037] in, Indicates the compliance weight of the scenario. Indicates performance adaptation weights. Indicates the load balancing weight. Indicates compliance with scene requirements. This indicates the compatibility between tasks and nodes. Indicates the overall load rate of the nodes. Indicates the average load rate of the cluster. Indicates the maximum load rate of the cluster; ; Substitute each node into the fitness function Calculate scores and sort them before filtering. The top 10 individuals are selected as elite antlions; in this embodiment, the top 10 individuals are selected as elite antlions.

[0038] The iterative optimization process is as follows: Constructing a dynamic step size factor The optimization accuracy is adjusted based on the number of iterations and load fluctuations, expressed as:

[0039] in, Indicates the initial step size. This represents the current iteration number. Indicates the maximum number of iterations. Indicates the standard deviation of the load factor. Indicates the overall load rate of the nodes; The ants adjust their search direction based on the location of the elite antlion, and update the ant positions; the update process is as follows: During the iteration process, the node ID is updated, and the compatibility between the task and the node is updated, while the scenario compliance identifier remains unchanged. The expression for updating the node ID is:

[0040] The expression for updating the fit between tasks and nodes is:

[0041] in, This represents the node ID of the (t+1)th generation of ordinary ants. This represents the node ID of the t-th generation of ordinary ants. This represents the node ID of the t-th generation elite ant. This is the dynamic step size factor. These are random numbers distributed according to a standard normal distribution. This represents the fitness of ordinary ants in generation t+1. This represents the fitness of the t-th generation of common ants; This represents the rounding function.

[0042] Finally, calculate the fitness of the updated ants. If it is higher than the current elite antlion, replace it and simultaneously... and This serves as a guiding value for subsequent iterations, ensuring the algorithm converges to the optimal path; after the maximum number of iterations, the individual with the highest fitness is selected for decoding. For the target node ID, generate the optimal call path from the business system to the target node to the corresponding large model; and generate the call instruction from the optimal call path.

[0043] The execution and feedback module is used to send call instructions to the target node based on the protocol to execute large model calls; simultaneously, it collects call feedback data, generates feedback vectors, and performs reverse iterative adjustments to the algorithm parameters in the intelligent scheduling module; in this embodiment, the execution and feedback module includes a call execution unit, a data acquisition unit, and a feedback optimization unit, wherein: The execution unit is used to schedule the server to send a large model invocation instruction to the target node according to the gRPC protocol. The instruction includes task data and model parameters. It also controls the target node to load the corresponding large model and perform inference calculations after receiving the invocation instruction. For example, a self-developed financial large model is loaded for classified tasks, and a general large model is loaded for non-classified tasks. The data acquisition unit is used to collect call response time, node resource utilization, and task completion rate; where call response time is the duration from command issuance to result return, and node resource utilization is the time during the call period. The average value is used to represent the task completion rate, which is represented by 1 and 0 respectively, where 1 represents completion and 0 represents failure.

[0044] The feedback optimization unit is used to adjust the initial step size of the dynamic step size factor in the improved Antlion optimization algorithm to the next iteration when the call response time is greater than a preset call response time threshold or the node resource utilization is less than a preset node resource utilization threshold; and to optimize the fitness function weights based on the fitness function value and historical data; for example, when the call response time is greater than 1.5 seconds or the node resource utilization is less than 30%. Then adjust the initial step size of the improved Antlion optimization algorithm. This improves the optimization accuracy in the next iteration; simultaneously, if the fitness function value is greater than 0.95 for 5 consecutive iterations, the current parameters are fixed to reduce computational overhead; and a scheduling report is generated daily to optimize the fitness function weights based on historical data. , , .

[0045] like Figure 2 As shown, in another embodiment of this example, a large model call path intelligent scheduling method based on task load and scenario is further included, applied to the above-mentioned large model call path intelligent scheduling system based on task load and scenario, including: S1: Real-time acquisition of task complexity and task scenario attributes to generate task-aware data; S2: Real-time monitoring of resource load on each node in the cluster and generation of resource load data; S3: Based on the improved Antlion optimization algorithm, it takes task-aware data and resource load data as input, outputs the optimal calling path through encoding mapping, fitness calculation and iterative optimization, and generates calling instructions from the optimal calling path; S4: Based on the protocol, the call command is sent to the target node to execute the large model call; at the same time, the call feedback data is collected, and a feedback vector is generated to perform reverse iterative adjustment of the algorithm parameters in the intelligent scheduling module.

[0046] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A large-scale model call path intelligent scheduling system based on task load and scenario, characterized in that: It includes a task awareness module, a resource monitoring module, an intelligent scheduling module, and an execution and feedback module, among which: The task awareness module is used to collect task complexity and task scenario attributes in real time and generate task awareness data. The resource monitoring module is used to monitor the resource load of each node in the cluster in real time and generate resource load data; The intelligent scheduling module is used to optimize the antlion algorithm based on the improved algorithm. It takes task perception data and resource load data as input, outputs the optimal calling path through encoding mapping, fitness calculation and iterative optimization, and generates calling instructions from the optimal calling path. The execution and feedback module is used to send the call command to the target node based on the protocol to execute the large model call; at the same time, it collects the call feedback data, generates feedback vectors, and performs reverse iterative adjustment of the algorithm parameters in the intelligent scheduling module.

2. The intelligent scheduling system for large model call paths based on task load and scenario as described in claim 1, characterized in that: The task complexity and task scenario attributes collected in real time by the task awareness module are specifically as follows: Deploy a text parsing service to receive large model call requests and task description texts submitted by business systems; Calculate the number of tokens in the task description text. Calculate the basic complexity The expression is: Identify the reasoning requirements of the task and assign additional complexity. Calculate the final task complexity ; Establish a keyword database that includes both classified and unclassified keywords. Use the TF-IDF algorithm to calculate the similarity between the task text and the keyword database. When the similarity exceeds a preset threshold, it is marked as the corresponding scenario, and the task scenario attributes are output. ; Based on the final task complexity and task scenario attributes Generate task feature vectors And it is sent to the intelligent scheduling module via the protocol.

3. The intelligent scheduling system for large model call paths based on task load and scenario as described in claim 2, characterized in that: The resource monitoring module monitors the resource load of each node in the cluster in real time and generates resource load data as follows: Deploy distributed monitoring agents on each private cloud node and each authorized public cloud node in the cluster; Real-time CPU utilization is collected via an agent. GPU utilization Memory usage Broadband network ; Based on the weighted calculation of the overall load rate of the nodes, the expression is: Generate node load vector The data is synchronized in real time to the intelligent scheduling module and the cluster load distribution is visualized in the form of a heatmap.

4. The intelligent scheduling system for large model call paths based on task load and scenario as described in claim 3, characterized in that: The intelligent scheduling module, based on the improved Antlion optimization algorithm, takes task-aware data and resource load data as input, and outputs the optimal call path through encoding mapping, fitness calculation, and iterative optimization. Specifically, it does so as follows: Obtain task feature vector and node load vector The initial population is generated using a mixture of three-dimensional binary and integer encoding. ;in Indicates compliance with scene requirements. Indicates the node ID. This indicates the compatibility between tasks and nodes; A fitness function is constructed based on three constraints: scenario compliance, performance adaptation, and load balancing. The expression is: in, Indicates the compliance weight of the scenario. Indicates performance adaptation weights. Indicates the load balancing weight. Indicates compliance with scene requirements. This indicates the compatibility between tasks and nodes. Indicates the overall load rate of the nodes. Indicates the average load rate of the cluster. Indicates the maximum load rate of the cluster; ; Substitute each node into the fitness function Calculate scores and sort them before filtering. Individuals are considered elite antlions; Constructing a dynamic step size factor The optimization accuracy is adjusted based on the number of iterations and load fluctuations, expressed as: in, Indicates the initial step size. This represents the current iteration number. Indicates the maximum number of iterations. Indicates the standard deviation of the load factor. Indicates the overall load rate of the nodes; Adjust the ants' search direction based on the location of the elite antlion, and update the ants' positions; Calculate the fitness of the updated ant; if it is higher than the current elite antlion, replace it; continue this process until the maximum number of iterations, then select the individual with the highest fitness and decode. For the target node ID, generate the optimal call path from the business system to the target node to the corresponding large model.

5. The intelligent scheduling system for large model call paths based on task load and scenario as described in claim 4, characterized in that: The acquisition of task feature vectors and node load vector The initial population is generated using a mixture of three-dimensional binary and integer encoding. Specifically: Obtain task feature vector and node load vector Determine the attributes of the task scenario The value, if If the characterization involves classified information, then only private cloud nodes will be included in the coding scope. If the identifier is not classified, then all nodes are included; Based on the included nodes, several coded individuals are randomly generated. The expression is: in, Indicates compliance with scene requirements; Indicates the node ID; The degree of fit between tasks and nodes is expressed as: in, This indicates the GPU utilization of the current node. This represents the maximum GPU utilization among all available compliant nodes. This represents the maximum preset task complexity of the system. This represents a quantified value indicating the complexity of the current task. This represents the scene weight coefficient.

6. The intelligent scheduling system for large model call paths based on task load and scenario as described in claim 5, characterized in that: The specific steps for adjusting the ants' search direction based on the location of the elite antlion and updating the ant positions are as follows: During the iteration process, the node ID is updated, and the compatibility between the task and the node is updated, while the scenario compliance identifier remains unchanged. The expression for updating the node ID is: The expression for updating the fit between tasks and nodes is: in, This represents the node ID of the (t+1)th generation of ordinary ants. This represents the node ID of the t-th generation of ordinary ants. This represents the node ID of the t-th generation elite ant. This is the dynamic step size factor. These are random numbers distributed according to a standard normal distribution. This represents the fitness of ordinary ants in generation t+1. This represents the fitness of the t-th generation of common ants; This represents the rounding function.

7. The intelligent scheduling system for large model call paths based on task load and scenario as described in claim 6, characterized in that: The execution and feedback module includes a call execution unit, a data acquisition unit, and a feedback optimization unit, wherein: The execution unit is used to schedule the server to send the large model invocation instruction to the target node according to the gRPC protocol, and to control the target node to load the corresponding large model and perform inference calculation after receiving the invocation instruction; The data acquisition unit is used to collect call response time, node resource utilization, and task completion rate; The feedback optimization unit is used to determine whether the call response time is greater than the preset call response time threshold or the node resource utilization rate is less than the preset node resource utilization threshold. When this condition is met, the initial step size of the dynamic step size factor in the improved Antlion optimization algorithm is adjusted to the next iteration. The fitness function weights are also optimized based on the fitness function value and historical data.

8. A method for intelligent scheduling of large model call paths based on task load and scenario, applied to the intelligent scheduling system for large model call paths based on task load and scenario as described in any one of claims 1-7, characterized in that: include: S1: Real-time acquisition of task complexity and task scenario attributes to generate task-aware data; S2: Real-time monitoring of resource load on each node in the cluster and generation of resource load data; S3: Based on the improved Antlion optimization algorithm, it takes task-aware data and resource load data as input, outputs the optimal calling path through encoding mapping, fitness calculation and iterative optimization, and generates calling instructions from the optimal calling path; S4: Based on the protocol, the call command is sent to the target node to execute the large model call; at the same time, the call feedback data is collected, and a feedback vector is generated to perform reverse iterative adjustment of the algorithm parameters in the intelligent scheduling module.

Images