Method for intelligent multi-model calling of net body based on multi-dimensional feature constraint and product

Abstract

The embodiment of the application provides a kind of based on multi-dimensional feature constraint's net body intelligent multi-model calling method and product, it is related to artificial intelligence technical field.The method comprises: obtaining directed graph structure;Wherein, directed graph structure includes multiple graph nodes and the node feature and the edge feature corresponding to each graph node;According to user task request, generate task request feature;According to task request feature and the multiple graph nodes in directed graph structure and the node feature and the edge feature corresponding to each graph node, one or more graph nodes in directed graph structure is determined as target graph node;According to target graph node, determine task processing path;Wherein, task processing path is used to call the model corresponding to target graph node according to the model calling order corresponding to task processing path when processing user task request;Solved the problem that the model called in prior art scheme exists poor matching, improves data processing efficiency and resource utilization.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and product for calling multiple intelligent models of a network body based on multi-dimensional feature constraints. Background Technology

[0002] With the rapid development of large model technology, models of different modalities (such as text, image, and voice), different parameter scales (such as small models, medium models, and ultra-large models), and different deployment locations (such as central cloud, edge nodes, and heterogeneous computing power pools) are widely deployed in operators, government and enterprises, and large cloud environments.

[0003] In existing technologies, when multiple models need to be called to process a task, an external "router" or "selector" is usually used to select whether to call a large or small model based on the difficulty of the query. For example, decisions are made based on semantic analysis or static cost rules to select to call multiple large or small models to process the task.

[0004] However, the existing solutions suffer from poor model matching. Summary of the Invention

[0005] The method and product for calling intelligent multi-models based on multi-dimensional feature constraints provided in this application are used to solve the problem of poor matching of models called by existing technical solutions.

[0006] In a first aspect, embodiments of this application provide a method for intelligent multi-model invocation based on multi-dimensional feature constraints, comprising: obtaining a directed graph structure; wherein the directed graph structure includes multiple graph nodes and node features and edge features corresponding to each graph node, the graph nodes are used to indicate models in a model cluster, the node features are model features of the model corresponding to the graph node, and the edge features are used to indicate the data transmission relationship between the models corresponding to the graph nodes; generating a task request feature according to a user task request; determining one or more graph nodes in the directed graph structure as target graph nodes according to the task request feature and the multiple graph nodes and node features and edge features corresponding to each graph node; determining a task processing path according to the target graph node; wherein the task processing path is used to invoke the model corresponding to the target graph node according to the model invocation order corresponding to the task processing path when processing the user task request.

[0007] In one possible implementation, the graph nodes include a start node, one or more intermediate nodes, and an end node. The step of determining one or more graph nodes in the directed graph structure as target graph nodes based on the task request features and the multiple graph nodes in the directed graph structure, as well as the node features and edge features corresponding to each graph node, includes: determining multiple first candidate nodes from one or more intermediate nodes and the end node based on the start node; wherein data transfer is allowed between the model corresponding to the first candidate node and the model corresponding to the start node; determining the start node as the current node; determining corresponding candidate edges based on the current node and each first candidate node; inputting the task request features, the node features of the current node, the node features of each first candidate node, and the edge features of the candidate edges corresponding to each first candidate node into a node scoring model, and outputting the estimated node score corresponding to each first candidate node; determining the highest node score from the estimated node scores corresponding to each first candidate node, and determining the first candidate node corresponding to the highest node score as the target node; and determining the target graph node based on the target node.

[0008] In one possible implementation, determining the target graph node based on the target node includes: determining whether the target node is the termination node; if the target node is not the termination node, then determining the target node as a path node, and determining multiple second candidate nodes from one or more intermediate nodes and termination nodes based on the target node; wherein data transfer is allowed between the model corresponding to the second candidate node and the model corresponding to the target node; determining the target node as the current node, re-determining each second candidate node as a first candidate node, and returning to execute the step of determining the corresponding candidate edge based on the current node and each first candidate node; if the target node is the termination node, then determining the starting node and the termination node as the target graph node, or determining the starting node, the termination node, and the path node as the target graph node.

[0009] In one possible implementation, after processing the user task request according to the task processing path, the method further includes: collecting processing path data corresponding to each target graph node in the task processing path; wherein the processing path data includes path latency, invocation cost, and result accuracy of the processing result corresponding to the user task request; for each target graph node: generating an actual node score corresponding to the target graph node based on the path latency, the invocation cost, and the result accuracy; determining whether the difference between the highest node score and the actual node score corresponding to the target graph node is within a preset difference range; if the difference between the highest node score and the actual node score is not within the preset difference range, then training the node scoring model based on the path latency, the invocation cost, and the result accuracy to update the model parameters of the node scoring model.

[0010] In one possible implementation, the node features include node capability features, node network features, and node cost features; the node scoring model includes a feature correction sub-model and a feature scoring sub-model; the step of inputting the task request features, the node features of the current node, the node features of each first candidate node, and the edge features of the candidate edges corresponding to each first candidate node into the node scoring model, and outputting the estimated node score corresponding to each first candidate node, includes: determining, based on the task request features, a first correction coefficient corresponding to the node capability features, a second correction coefficient corresponding to the node network features, and a third correction coefficient corresponding to the node cost features through the feature correction sub-model; and through... The feature correction sub-model corrects the node features of the current node based on the first correction coefficient, the second correction coefficient, and the third correction coefficient to obtain the corrected node features of the current node; the feature correction sub-model corrects the node features of each first candidate node based on the first correction coefficient, the second correction coefficient, and the third correction coefficient to obtain the corrected node features of each first candidate node; the feature scoring sub-model generates the estimated node score corresponding to each first candidate node based on the task request feature, the corrected node features of the current node, the corrected node features of each first candidate node, and the edge features of the candidate edges corresponding to each first candidate node.

[0011] In one possible implementation, the task request features include multiple sub-task request features; generating task request features based on user task requests includes: determining initial task request features based on the user task request; determining the number of splits for the user task request based on the initial task request features; splitting the user task request based on the number of splits to obtain the number of user sub-task requests; for each user sub-task request: performing feature parsing on the user sub-task request to generate the sub-task request features; correspondingly, the target graph nodes include multiple first target graph nodes; the step of generating task request features based on the task request features and the directed graph structure... The process of identifying one or more graph nodes in a directed graph structure as target graph nodes, based on multiple graph nodes and their corresponding node and edge features, includes: for each subtask request feature: determining one or more graph nodes in the directed graph structure as first target graph nodes based on the task request features and the multiple graph nodes in the directed graph structure and their corresponding node and edge features; correspondingly, determining a task processing path based on the target graph nodes includes: for each first target graph node: determining a corresponding first task processing path based on the first target graph node; summarizing each first task processing path to obtain the task processing path.

[0012] Secondly, embodiments of this application provide a network-based intelligent multi-model invocation device based on multi-dimensional feature constraints, comprising: an acquisition module, configured to acquire a directed graph structure; wherein the directed graph structure includes multiple graph nodes and node features and edge features corresponding to each graph node, the graph nodes are used to indicate models in a model cluster, the node features are model features of the model corresponding to the graph node, and the edge features are used to indicate the data transmission relationship between the models corresponding to the graph nodes; a processing module, configured to generate a task request feature according to a user task request; and determine one or more graph nodes in the directed graph structure as target graph nodes according to the task request feature and the multiple graph nodes and node features and edge features corresponding to each graph node; and a determination module, configured to determine a task processing path according to the target graph node; wherein the task processing path is used to invoke the model corresponding to the target graph node according to the model invocation order corresponding to the task processing path when processing the user task request.

[0013] In one possible implementation, the graph nodes include a start node, one or more intermediate nodes, and an end node. When the processing module determines one or more graph nodes in the directed graph structure as target graph nodes based on the task request features and the multiple graph nodes in the directed graph structure, as well as the node features and edge features corresponding to each graph node, the specific steps are as follows: Based on the start node, determine multiple first candidate nodes from one or more intermediate nodes and the end node; wherein data transfer is allowed between the model corresponding to the first candidate node and the model corresponding to the start node; determine the start node as the current node; determine corresponding candidate edges based on the current node and each first candidate node; input the task request features, the node features of the current node, the node features of each first candidate node, and the edge features of the candidate edges corresponding to each first candidate node into the node scoring model, and output the estimated node score corresponding to each first candidate node; determine the highest node score from the estimated node scores corresponding to each first candidate node, and determine the first candidate node corresponding to the highest node score as the target node; and determine the target graph node based on the target node.

[0014] In one possible implementation, when the processing module determines the target graph node based on the target node, it is specifically configured to: determine whether the target node is the termination node; if the target node is not the termination node, determine the target node as a path node, and determine multiple second candidate nodes from one or more intermediate nodes and termination nodes based on the target node; wherein, data transfer is allowed between the model corresponding to the second candidate node and the model corresponding to the target node; determine the target node as the current node, re-determine each second candidate node as each first candidate node, and return to execute the step of determining the corresponding candidate edge based on the current node and each first candidate node; if the target node is the termination node, determine the starting node and the termination node as the target graph node, or determine the starting node, the termination node, and the path node as the target graph node.

[0015] In one possible implementation, after processing the user task request according to the task processing path, the processing module is further configured to: collect processing path data corresponding to each target graph node in the task processing path; wherein, the processing path data includes path latency, call cost, and result accuracy of the processing result corresponding to the user task request; for each target graph node: generate an actual node score corresponding to the target graph node according to the path latency, the call cost, and the result accuracy; determine whether the difference between the highest node score and the actual node score corresponding to the target graph node is within a preset difference range; if the difference between the highest node score and the actual node score is not within the preset difference range, then train the node scoring model according to the path latency, the call cost, and the result accuracy to update the model parameters of the node scoring model.

[0016] In one possible implementation, the node features include node capability features, node network features, and node cost features; the node scoring model includes a feature correction sub-model and a feature scoring sub-model; when the processing module inputs the task request features, the node features of the current node, the node features of each first candidate node, and the edge features of the candidate edges corresponding to each first candidate node into the node scoring model and outputs the estimated node score corresponding to each first candidate node, it is specifically used to: determine, based on the task request features, a first correction coefficient corresponding to the node capability features, a second correction coefficient corresponding to the node network features, and a third correction coefficient corresponding to the node cost features using the feature correction sub-model. The node features of the current node are corrected using the feature correction sub-model based on the first correction coefficient, the second correction coefficient, and the third correction coefficient to obtain the corrected node features of the current node. The node features of each first candidate node are also corrected using the feature correction sub-model based on the first correction coefficient, the second correction coefficient, and the third correction coefficient to obtain the corrected node features of each first candidate node. Finally, the predicted node score for each first candidate node is generated using the feature scoring sub-model based on the task request features, the corrected node features of the current node, the corrected node features of each first candidate node, and the edge features of the candidate edges corresponding to each first candidate node.

[0017] In one possible implementation, the task request features include multiple sub-task request features; when generating task request features based on user task requests, the processing module is specifically configured to: determine initial task request features based on the user task request; determine the number of splits for the user task request based on the initial task request features; split the user task request based on the number of splits to obtain the number of user sub-task requests; for each user sub-task request: perform feature parsing on the user sub-task request to generate the sub-task request features; correspondingly, the target graph nodes include multiple first target graph nodes; the processing module, based on the task request features and the directed graph structure... When determining one or more graph nodes in the directed graph structure as target graph nodes based on multiple graph nodes and their corresponding node and edge features, the determination module is specifically used for: For each subtask request feature: based on the task request features and the multiple graph nodes in the directed graph structure and their corresponding node and edge features, determining one or more graph nodes in the directed graph structure as the first target graph node; correspondingly, when determining the task processing path based on the target graph node, the determination module is specifically used for: For each first target graph node: based on the first target graph node, determining the corresponding first task processing path; summarizing each first task processing path to obtain the task processing path.

[0018] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0019] The memory stores computer-executed instructions;

[0020] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0021] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0022] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0023] The method and product for intelligent multi-model invocation based on multi-dimensional feature constraints provided in this application solves the problem of poor model matching in existing technical solutions by introducing a directed graph structure with multi-dimensional feature constraints, specifically a directed graph structure constrained by node and edge features. Specifically, explicit modeling of node and edge features enables the model cluster corresponding to the directed graph structure to possess intrinsic perception capabilities. Furthermore, dynamic selection of target graph nodes is achieved by combining task features, node features, and edge features, thereby determining the task processing path based on the target graph nodes. The task processing path is used to invoke the models corresponding to the target graph nodes according to the model invocation order corresponding to the task processing path when processing user task requests. The method provided in this application constructs a unified model scheduling system with multi-dimensional feature constraints, enabling the model cluster corresponding to the directed graph structure to adapt to the wide area network environment, dynamically balance the constraints between different dimensional features, solve the problem of poor model matching in existing technical solutions, and improve data processing efficiency and resource utilization. Attached Figure Description

[0024] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0025] Figure 1 A schematic diagram illustrating a scenario for the intelligent multi-model invocation method for net body based on multi-dimensional feature constraints provided in this application;

[0026] Figure 2 A flowchart of a method for invoking intelligent multi-models of a net body based on multi-dimensional feature constraints, provided in one embodiment of this application;

[0027] Figure 3 This is a schematic diagram of a directed graph structure including node features provided in one embodiment of this application;

[0028] Figure 4 A flowchart of a method for invoking a network body intelligent multi-model based on multi-dimensional feature constraints, provided in another embodiment of this application;

[0029] Figure 5 A schematic diagram of the structure of a smart multi-model calling device for net body based on multi-dimensional feature constraints provided in one embodiment of this application;

[0030] Figure 6 A schematic diagram of the structure of the electronic device provided in this application.

[0031] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0032] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0033] The technical solution of this application involves the collection, storage, use, processing, transmission, provision and disclosure of user personal information and data, which comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0034] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0035] With the rapid development of large-scale model technology, models of different modalities (such as text, image, and speech), different parameter scales (such as small, medium, and ultra-large-scale models), and different deployment locations (such as central cloud, edge nodes, and heterogeneous computing pools) are widely deployed in carrier, government, and large-scale cloud environments. In existing technologies, when multiple models need to be invoked to process a task, an external "router" or "selector" is typically used to select which model to invoke based on query difficulty, such as semantic analysis or static cost rules. However, existing solutions suffer from poor model matching.

[0036] First, let me explain the terms used in this application:

[0037] Netbody Intelligence: Netbody Intelligence architecture is an AI architecture based on multiple AI models and a large-scale distributed computing environment. Based on a directed graph structure, it treats AI models / modules as "multi-brain subsystems" distributed in different regions and network nodes, forming an intelligent agent with node and edge characteristics through network connections. Furthermore, the intelligent agent based on the Netbody Intelligence architecture achieves sparse collaboration and implicit aggregation through edge decision-making (learnable), and continuously evolves under real business feedback. Moreover, the system-level security governance layer ensures compliance, consistency and auditability of the entire chain. Among them, node characteristics include node capability characteristics, node network characteristics, and node cost characteristics; the core of the network-based intelligent architecture is to construct multiple artificial intelligence models / modules in different regions into a directed graph structure, and introduce multi-dimensional constraint characteristics such as "capability boundary (node ​​capability characteristics), network state (node ​​network characteristics), and call cost (node ​​cost characteristics)" for each model / module, so that the model is intrinsically sensitive to network, computing power, and task requirements, thereby realizing intelligent scheduling, collaboration, and aggregation of multiple models across domains. It can proactively adapt to the high cost-effectiveness inference requirements in wide area network and cross-regional environments. At the same time, multiple artificial intelligence models are connected through a directed graph, and intelligent model scheduling and aggregation modes are constructed by activating connection parameters.

[0038] The application scenarios of the embodiments of this application are explained below:

[0039] Figure 1 A schematic diagram illustrating the scenario of the intelligent multi-model invocation method for network body based on multi-dimensional feature constraints provided in this application, such as... Figure 1 As shown, the execution subject of the method provided in this application embodiment can be any form of electronic device, with a computer device being used as the execution subject for illustration. The directed graph structure includes multiple graph nodes and node features and edge features corresponding to each graph node. Graph nodes are used to indicate models in a model cluster, node features are model features of the model corresponding to the graph node, and edge features are used to indicate the data transfer relationship between the models corresponding to the graph nodes. Based on obtaining the directed graph structure, the computer device generates task request features according to the user task request. Based on the task request features and the multiple graph nodes in the directed graph structure and the node features and edge features corresponding to each graph node, one or more graph nodes in the directed graph structure are determined as target graph nodes. Based on the target graph nodes, a task processing path is determined. The task processing path is used to call the model corresponding to the target graph node according to the model call order corresponding to the task processing path when processing the user task request.

[0040] Specifically, such as Figure 1As shown, the directed graph structure includes multiple graph nodes: the starting node Vs, intermediate nodes V1, V2, V3, V4, V5, and the ending node Ve. Each graph node has its own node characteristics. If there are connecting lines between graph nodes, then there is data transfer between the models corresponding to the graph nodes, i.e., there are edge features. If there are no connecting lines between graph nodes, then there is no data transfer between the models corresponding to the graph nodes, i.e., there are no edge features. Furthermore, based on the obtained directed graph structure, the computer device generates task request features according to the user's task request. Based on the task request features and the multiple graph nodes in the directed graph structure, as well as the node features and edge features corresponding to each graph node, one or more graph nodes in the directed graph structure are identified as target graph nodes. For example, if Vs, V1, V4, V5, and Ve are identified as target graph nodes, then the task processing path (e.g., ...) is determined based on Vs, V1, V4, V5, and Ve. Figure 1 As shown by the black dashed line, after the starting node Vs generates the task request features based on the user's task request, the target graph nodes V1, V4, V5, and Ve process them in sequence. After the terminal node Ve completes its processing, the processing result corresponding to the user's task request can be output.

[0041] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0042] Figure 2 A flowchart of a method for invoking intelligent multi-models based on multi-dimensional feature constraints provided in one embodiment of this application is shown below. Figure 2 As shown, the execution subject of the intelligent multi-model invocation method based on multi-dimensional feature constraints provided in this embodiment can be any form of electronic device. For example, this embodiment uses a computer device as the execution subject of the method. The intelligent multi-model invocation method based on multi-dimensional feature constraints provided in this embodiment includes the following steps:

[0043] Step S201: Obtain the directed graph structure; wherein the directed graph structure includes multiple graph nodes and the node features and edge features corresponding to each graph node.

[0044] For example, graph nodes are used to indicate models in a model cluster. Node features are the model features corresponding to the graph node, and edge features are used to indicate the data transfer relationships between the models corresponding to the graph nodes. Specifically, the directed graph structure G = (V', E), where V' represents multiple graph nodes, and E represents the data transfer relationships between the models corresponding to the graph nodes; V' includes a starting node Vs, intermediate nodes Vi (i=1,2,3,…,n), and a ending node Ve; E includes the data transfer relationships Esi between the starting node and intermediate nodes, Ese between the starting node and the ending node, and Eij between at least two intermediate nodes (i.e., the case where there is no data transfer relationship between intermediate nodes Vi and Vj (j=1,2,3,…,n; j≠i), for example...). Figure 1 The data transfer relationship between V1 and V2 (where no data transfer is established) and between intermediate nodes and termination nodes is shown in Eie.

[0045] In one possible implementation, the termination node Ve establishes data transfer relationships with the starting node Vs and each intermediate node Vi, respectively, to realize a dynamic "end and output processing results at any time" mechanism in the subsequent data processing steps. That is, at any intermediate node, it is possible to choose to "end and output processing results at any time", and then the intermediate results of the current intermediate node will directly converge to the termination node to generate the processing results and send them to the user.

[0046] Furthermore, the node features corresponding to each graph node include node model features, node capability features, node network features, and node cost features. Specifically, node model features are used to indicate the inference computation of the model corresponding to the graph node; node capability features are used to indicate the capability boundaries of the model corresponding to the graph node in different task dimensions, such as general language processing capability, inference capability, domain expertise capability, and multimodal processing capability, reflecting the matching degree between the model's processing capability and the task request features; node network features are used to indicate the network status of the model corresponding to the graph node in the current computing power network, such as relative latency interval (e.g., the latency reference object of each graph node is the starting node), available bandwidth range, and current load or congestion level, reflecting the network reachability of each graph node when the model corresponding to the graph node determines the target graph node; node cost features are used to indicate the resource and economic costs of the model corresponding to the graph node, such as the unit call price or unit computing power cost, and the estimated GPU time consumption, reflecting the cost loss of the model corresponding to the graph node when processing task request features.

[0047] Furthermore, the edge features corresponding to each graph node are used to indicate that when there is a data transmission relationship between graph nodes, the corresponding graph nodes are compatible in terms of data format and strategy. Furthermore, the edge features include whether the corresponding graph nodes are geographically separated, the typical transmission cost level, and whether protocol / format conversion is required. Further, the edge features also include data processing and transformation features. Taking the current graph node and one or more of its predecessor graph nodes as an example, the data processing and transformation features include using the complete output of the predecessor graph node as the input of the current graph node, using the embedded vector representation of the predecessor graph node as the input vector of the current graph node, extracting structured fields from the complete output of the predecessor graph node as the input of the current graph node, and fusing the outputs of multiple predecessor graph nodes as the input of the current graph node. The processing method of "fusing the outputs of multiple predecessor graph nodes as the input of the current graph node" avoids the complexity caused by setting up independent aggregation modules in existing technical solutions and reduces data processing resource consumption.

[0048] Step S202: Generate task request features based on the user's task request.

[0049] For example, a user task request includes text, a combination of text and graphics, and a multi-turn dialogue context; the user task request is parsed to generate task request features; wherein, the task request features include task type and modal requirements, preferences or constraints on response quality (accuracy of the output results of the model corresponding to the graph node), latency, cost, and task semantic feature representation.

[0050] Step S203: Based on the task request characteristics and multiple graph nodes in the directed graph structure, as well as the node characteristics and edge characteristics corresponding to each graph node, one or more graph nodes in the directed graph structure are determined as target graph nodes.

[0051] Specifically, the graph nodes include a start node, one or more intermediate nodes, and an end node; therefore, the specific implementation of step S203 includes:

[0052] Step S2031: Based on the starting node, determine multiple first candidate nodes from one or more intermediate nodes and ending nodes; wherein, data transfer is allowed between the model corresponding to the first candidate node and the model corresponding to the starting node.

[0053] Step S2032: Determine the starting node as the current node.

[0054] Step S2033: Determine the corresponding candidate edges based on the current node and each first candidate node.

[0055] For example, refer to Figure 1The directed graph structure shown has the following first candidate nodes: V1, V2, V3, V4, V5, and Ve, based on the starting node Vs; correspondingly, the candidate edges include Es1 (the edge corresponding to the starting node Vs and the first candidate node V1), Es2 (the edge corresponding to the starting node Vs and the first candidate node V2), Es3 (the edge corresponding to the starting node Vs and the first candidate node V3), Es4 (the edge corresponding to the starting node Vs and the first candidate node V4), Es5 (the edge corresponding to the starting node Vs and the first candidate node V5), and Ese (the edge corresponding to the starting node Vs and the first candidate node Ve).

[0056] Step S2034: Input the task request features, the node features of the current node, the node features of each first candidate node, and the edge features of the candidate edges corresponding to each first candidate node into the node scoring model, and output the estimated node score corresponding to each first candidate node.

[0057] For example, the node features corresponding to the graph nodes include node capability features, node network features, and node cost features. The node scoring model is a pre-trained scoring model. Then, the node scoring model calculates the input task request features, the node features of the current node, the node features of each first candidate node, and the edge features of the candidate edges corresponding to each first candidate node, thereby obtaining the estimated node score corresponding to each first candidate node.

[0058] For example, refer to Figure 1 The directed graph structure shown, Figure 3 This is a schematic diagram of a directed graph structure including node features provided in one embodiment of this application; as shown... Figure 3 As shown, As represents the node capability characteristics of the starting node Vs; Ns represents the node network characteristics of the starting node Vs; Cs represents the node cost characteristics of the starting node Vs; Ae represents the node capability characteristics of the ending node Ve; Ne represents the node network characteristics of the ending node Ve; Ce represents the node cost characteristics of the ending node Ve; Am represents the node capability characteristics of the intermediate node Vm; Mm represents the node network characteristics of the intermediate node Vm; Cm represents the node cost characteristics of the intermediate node Vm; where m = 1, 2, 3, 4, 5.

[0059] Taking the current node as the starting node as an example, the estimated node score corresponding to each first candidate node is S_s,n =Gscore(R, As, Ns, Cs, An, Nn, Cn, f_s,n); where n=1,2,3,4,5,e; R is the task request feature; As is the node capability feature of the starting node Vs; Ns is the node network feature of the starting node Vs; Cs is the node cost feature of the starting node Vs; An is the node capability feature of the first candidate node Vn; Nn is the node network feature of the first candidate node Vn; Cn is the node cost feature of the first candidate node Vn; and f_s,n is the edge feature of the candidate edge corresponding to the starting node Vs and the first candidate node Vn.

[0060] It is understood that this application does not specifically limit the basic model architecture of the node scoring model, which can be a long short-term memory network model, a multilayer perceptron model, or a graph neural network model.

[0061] In one possible implementation, the node scoring model calculates the input task request features, the node features of the current node, the node features of each first candidate node, and the edge features of the candidate edges corresponding to each first candidate node to obtain the estimated node score corresponding to each first candidate node, as shown in Equation (1).

[0062] (1)

[0063] in, First candidate node The estimated node score; For task request characteristics; First candidate node The node capability characteristics; First candidate node The characteristics of the node network; First candidate node The node cost characteristics; For the current node With the first candidate node The edge features of the corresponding candidate edges; Task request characteristics With node capability characteristics cosine similarity, This is used to determine whether a node can perform this task; Network health scoring function, for example This item measures how quickly data is transmitted. , This is a dynamic adjustment coefficient; : The normalized value of the node call cost, which is used to measure the first candidate node. Is it expensive or not? : Set the current node With the first candidate node The static features of the corresponding candidate edges are converted into specific scoring items, such as discrete mapping based on the "feature lookup table method" or linear weighted features; The coefficient is dynamically adjustable and can be adjusted according to the user's configured preferences (such as "high-speed mode" or "saving mode").

[0064] Step S2035: Determine the highest node score from the estimated node scores corresponding to each first candidate node, and determine the first candidate node corresponding to the highest node score as the target node.

[0065] Step S2036: Determine the target graph nodes based on the target nodes.

[0066] Specifically, the specific implementation steps of step S2036 include:

[0067] Step S20361: Determine whether the target node is a termination node.

[0068] Step S20362: If the target node is not the termination node, then the target node is determined as a path node, and based on the target node, multiple second candidate nodes are determined from one or more intermediate nodes and termination nodes; wherein, data transfer is allowed between the model corresponding to the second candidate node and the model corresponding to the target node.

[0069] For example, refer to Figure 1 The directed graph structure shown takes the target node as an intermediate node V1 as an example. Based on the target node V1, multiple second candidate nodes are determined from one or more intermediate nodes and the termination node. The second candidate nodes are intermediate nodes V3 and V4, and the termination node Ve.

[0070] Step S20363: Determine the target node as the current node, re-determine each second candidate node as the first candidate node, and return to execute step S2033.

[0071] For example, based on the embodiment shown in step S20362 and the processing procedure in step S20363, the candidate edges obtained by executing step S2033 are returned, including E13 (the edge corresponding to the current node V1 and the first candidate node V3), E14 (the edge corresponding to the current node V1 and the first candidate node V4), and E1e (the edge corresponding to the current node V1 and the first candidate node Ve); and then the subsequent steps are executed based on the return to executing step S2033.

[0072] Step S20364: If the target node is the termination node, then the start node and the termination node are determined as target graph nodes, or the start node, the termination node and the path node are determined as target graph nodes.

[0073] Specifically, for example, after determining multiple first candidate nodes from one or more intermediate nodes and end nodes based on the starting node, if the first candidate node corresponding to the highest node score is the end node (i.e., the target node), then step S20364, "determining the starting node and end node as target graph nodes," is executed. If, after determining multiple first candidate nodes from one or more intermediate nodes and end nodes based on the starting node, if the first candidate node corresponding to the highest node score is an intermediate node (i.e., the target node), then steps S20362 and S20363 are executed to return to step S2033, until the target node is the end node, and then step S20364, "determining the starting node, end node, and path node as target graph nodes," is executed.

[0074] In this embodiment, a node scoring model is used to achieve dynamic target graph node selection under constraints of multi-dimensional features (task request features, node capability features, node network features, node cost features, and edge features). The node scoring model selects target graph nodes in real-time and with high accuracy based on task request features, the current node's node features, the node features of each first candidate node, and the edge features of the candidate edges corresponding to each first candidate node. For example, in latency-sensitive tasks, the system prioritizes high-bandwidth target graph nodes to reduce task response time; in cost-sensitive tasks, it avoids excessive calls to high-cost models by calling low-cost target graph nodes. This improves the joint decision-making capability under multi-dimensional feature constraints, further enhancing scheduling efficiency and resource utilization in cross-domain multi-model scenarios.

[0075] Step S204: Determine the task processing path based on the target graph nodes; wherein, the task processing path is used to call the model corresponding to the target graph node according to the model calling order corresponding to the task processing path when processing user task requests.

[0076] For example, refer to Figure 1 The directed graph structure shown, based on the implementation steps, determines the starting node Vs, the ending node Ve, and the path nodes V1, V4, and V5 as target graph nodes. Then, based on the target graph nodes Vs, V1, V4, V5, and Ve, the task processing path is determined, namely Vs→V1→V4→V5→Ve. In other words, when processing user task requests, the models corresponding to the target graph nodes are called according to the model calling order corresponding to the task processing path Vs→V1→V4→V5→Ve.

[0077] Specifically, the starting node Vs serves as the entry point for task request features or user task requests, participating in decision-making but not requiring heavy inference; the target graph node V1 receives the task request features sent by the starting node Vs, processes them, and generates intermediate result R1; the target graph node V4 receives the intermediate result R1 sent by the target graph node V1, processes it, and generates intermediate result R4; the target graph node V5 receives the intermediate result R4 sent by the target graph node V4, processes it, and generates intermediate result R5; the terminating node Ve receives the intermediate result R5 sent by the target graph node V5, performs result convergence, and then outputs the processing result corresponding to the user task request.

[0078] In this embodiment, a directed graph structure with multi-dimensional feature constraints—specifically, a directed graph structure constrained by node and edge features—is introduced to address the problem of poor model matching in existing technical solutions. Specifically, explicit modeling of node and edge features enables the model cluster corresponding to the directed graph structure to possess intrinsic perception capabilities. Furthermore, dynamic selection of target graph nodes is achieved by combining task features, node features, and edge features, thereby determining the task processing path based on the target graph nodes. The task processing path is used to invoke the models corresponding to the target graph nodes according to the model invocation order corresponding to the task processing path when processing user task requests. The method provided in this application constructs a unified model scheduling system with multi-dimensional feature constraints, enabling the model cluster corresponding to the directed graph structure to adapt to the wide area network environment, dynamically balance constraints between different dimensional features, solve the problem of poor model matching in existing technical solutions, and improve data processing efficiency and resource utilization.

[0079] Furthermore, after processing the user task request according to the task processing path, the method provided in this application embodiment further includes:

[0080] Step S2037: Collect processing path data corresponding to each target graph node in the task processing path; wherein, the processing path data includes path delay, call cost, and result accuracy of the processing result corresponding to the user task request.

[0081] Among them, the accuracy of the processing results corresponding to the user task request corresponds to the node capability characteristics, the path delay corresponds to the node network characteristics, and the call cost corresponds to the node cost characteristics.

[0082] Step S2038: For each target graph node: Generate the actual node score corresponding to the target graph node based on path delay, call cost and result accuracy.

[0083] For example, for each target graph node: call the node scoring model so that the node scoring model calculates the actual node score corresponding to the target graph node based on the task request characteristics, the node characteristics of the current node, the path delay corresponding to the target graph node, the call cost and the result accuracy, and the edge characteristics of the candidate edges corresponding to the target graph node.

[0084] Step S2039: Determine whether the difference between the highest node score and the actual node score corresponding to the target graph node is within the preset difference range.

[0085] Step S20310: If the difference between the highest node score and the actual node score is not within the preset difference range, then the node scoring model is trained based on path delay, call cost and result accuracy to update the model parameters of the node scoring model.

[0086] For example, the node scoring model is trained based on path latency, invocation cost, and result accuracy to update the model parameters, ensuring that the difference between the highest node score and the actual node score is within a preset range. For instance, an online stochastic gradient descent algorithm can be used to update only the weight matrix of the node scoring model, eliminating the need to retrain the entire model.

[0087] In this embodiment, incremental training is used to adapt the node scoring model to dynamic changes. For example, when node congestion leads to increased path latency, the model parameters of the node scoring model are updated through incremental training. This allows for the selection of target graph nodes that constitute low-latency task processing paths, improving the path matching degree of task processing paths and the real-time performance and accuracy of scheduling decisions. This solves the problem of poor model matching in existing technical solutions and enhances data processing efficiency and resource utilization.

[0088] Furthermore, in one possible implementation, node features include node capability features, node network features, and node cost features; the node scoring model includes a feature correction sub-model and a feature scoring sub-model; then, the specific implementation of step S2034 includes:

[0089] Step S20341: Based on the task request features, the feature correction sub-model determines the first correction coefficient corresponding to the node capability features, the second correction coefficient corresponding to the node network features, and the third correction coefficient corresponding to the node cost features.

[0090] For example, by invoking the feature correction sub-model, the weight allocation ratio of node capability features, node network features, and node cost features for dynamically adjusting task request features can be achieved. That is, to obtain the first correction coefficient corresponding to the node capability characteristics. The second correction coefficient corresponding to the node network characteristics The third correction coefficient corresponding to the node cost characteristics .

[0091] For example, if a user's task request is determined to be a path delay-sensitive task based on the task request characteristics, then the constraint weight vector is [0.3, 0.6, 0.1], which strengthens the weight of node network characteristics and weakens the weight of node cost characteristics; if a user's task request is determined to be a cost-sensitive task based on the task request characteristics, then the constraint weight vector is [0.3, 0.1, 0.6], which weakens the weight of node network characteristics and strengthens the weight of node cost characteristics.

[0092] Step S20342: The node features of the current node are corrected by the feature correction sub-model based on the first correction coefficient, the second correction coefficient and the third correction coefficient to obtain the corrected node features of the current node.

[0093] For example, Ac represents the node capability feature of the current node Vc; Nc represents the node network feature of the current node Vc; and Cc represents the node cost feature of the current node Vc. By calling the feature correction sub-model, the node features of the current node are corrected according to the first correction coefficient, the second correction coefficient, and the third correction coefficient to obtain the corrected node features of the current node, that is, the corrected node capability feature of the current node Vc. Ac, Correction node network characteristics of the current node Vc Nc, the corrected node cost characteristics of the current node Vc Cc.

[0094] Step S20343: The node features of each first candidate node are corrected by the feature correction sub-model according to the first correction coefficient, the second correction coefficient and the third correction coefficient, so as to obtain the corrected node features of each first candidate node.

[0095] For example, for each first candidate node: An is the node capability feature of the first candidate node Vn; Nn is the node network feature of the first candidate node Vn; Cn is the node cost feature of the first candidate node Vn; by calling the feature correction sub-model, the node features of each first candidate node are corrected according to the first correction coefficient, the second correction coefficient, and the third correction coefficient, to obtain the corrected node features of each first candidate node, that is, the corrected node capability features of the first candidate node Vn. An, characteristics of the correction node network of the first candidate node Vn The correction node cost characteristics of Nn and the first candidate node Vn Cn.

[0096] Step S20345: Based on the task request features, the correction node features of the current node, the correction node features of each first candidate node, and the edge features of the candidate edges corresponding to each first candidate node, the feature scoring sub-model generates the estimated node score corresponding to each first candidate node.

[0097] For example, by inputting the task request features, the corrected node features of the current node, the corrected node features of each first candidate node, and the edge features of the candidate edges corresponding to each first candidate node into the feature scoring sub-model, the estimated node score corresponding to each first candidate node can be generated. The estimated node score corresponding to each first candidate node is S_c,n = Gscore(R, Ac, Nc, Cc, An, Nn, Cn, f_c,n); where n is used to indicate the node number of each first candidate node, referring to the aforementioned implementation step S2034; R is the task request feature; f_c,n is the edge feature of the candidate edge corresponding to the current node Vc and the first candidate node Vn.

[0098] In this embodiment, by dynamically adjusting the constraint weight coefficients corresponding to node capability characteristics, node network characteristics, and node cost characteristics based on task request features, the rationality of calling multiple models in scenarios involving complex user task requests is improved. This addresses the problem of poor model matching in existing technical solutions. For example, in latency-sensitive tasks, the system strengthens the weight of node network characteristics (such as bandwidth and latency) and weakens the weight of cost characteristics; in cost-sensitive tasks, it strengthens the weight of node cost characteristics and weakens the weight of capability characteristics. This step, through dynamic weight coefficient allocation, allows path selection to adapt to the needs of different user task requests, achieving a flexible balance between capability, network, and cost.

[0099] Figure 4 A flowchart of a method for invoking intelligent multi-models based on multi-dimensional feature constraints for net body, as provided in another embodiment of this application, is shown below. Figure 4 As shown, the intelligent multi-model invocation method for network body based on multi-dimensional feature constraints provided in this embodiment... Figure 2 The above embodiment provides a further refinement of the intelligent multi-model invocation method for network body based on multi-dimensional feature constraints. The task request features include multiple sub-task request features, and the target graph nodes include multiple first target graph nodes. Therefore, the intelligent multi-model invocation method for network body based on multi-dimensional feature constraints provided in this embodiment includes the following steps:

[0100] Step S301: Obtain the directed graph structure; wherein the directed graph structure includes multiple graph nodes and the node features and edge features corresponding to each graph node.

[0101] Graph nodes are used to indicate models in a model cluster, node features are the model features of the model corresponding to the graph node, and edge features are used to indicate the data transmission relationship between the models corresponding to the graph nodes.

[0102] Step S302: Determine the initial task request characteristics based on the user's task request.

[0103] For example, the initial task request features are used to indicate the task complexity of the user's task request; for example, if the user's task request is "analyze the content, shooting style and audience of video movie_1", then the initial task request features include video content information feature analysis, video content expression form feature analysis and video content audience matching feature analysis.

[0104] Step S303: Determine the number of user task requests to be split based on the initial task request characteristics.

[0105] For example, the number of splits for the user task request is determined based on the task complexity of the user task request indicated by the initial task request features; for example, if the initial task request features include three independent features, namely video content information feature analysis, video content expression form feature analysis, and video content audience matching degree feature analysis, then the number of splits for the user task request is 3.

[0106] Step S304: Based on the number of splits, split the user task request to obtain the number of user sub-task requests.

[0107] For example, based on the embodiment of step S303, the resulting three user subtask requests are split as follows: user subtask request task_1 is to analyze the content expressed by video movie_1, user subtask request task_2 is to analyze the shooting style of video movie_1, and user subtask request task_3 is to analyze the audience of video movie_1.

[0108] Step S305: For each user's subtask request: perform feature parsing on the user's subtask request to generate subtask request features.

[0109] For example, based on the embodiment of step S304, feature parsing is performed on user subtask request task_1, and the generated subtask request feature is "extracting the theme positioning, core narrative, expressive content, key information points, and content logic line of video movie_1"; feature parsing is performed on user subtask request task_2, and the generated subtask request feature is "analyzing the video's camera language, editing rhythm, visual presentation, audio matching (e.g., dubbing / background music / sound effects), and subtitle design"; feature parsing is performed on user subtask request task_3, and the generated subtask request feature is "combining the characteristics of the video's target audience group, analyzing the degree of matching between the theme, expressive style, information density of video movie_1 and the audience's needs, aesthetic preferences, and comprehension ability".

[0110] Among them, steps S302-S305 are the requirements of this application. Figure 2 A refined version of step S202 in the illustrated embodiment is that the task request features include multiple sub-task request features, and the number of sub-task request features is the number of user task request splits; the target graph nodes include multiple first target graph nodes, and the number of first target graph nodes is the number of user task request splits.

[0111] Step S306, for each subtask request feature: based on the request features of each task and multiple graph nodes in the directed graph structure, as well as the node features and edge features corresponding to each graph node, one or more graph nodes in the directed graph structure are determined as the first target graph node.

[0112] Specifically, regarding the request characteristics of each subtask: "Based on the request characteristics of each task and multiple graph nodes in the directed graph structure, as well as the node characteristics and edge characteristics corresponding to each graph node, the implementation method of determining one or more graph nodes in the directed graph structure as the first target graph node" is consistent with this application. Figure 2 The implementation of step S203 in the illustrated embodiment is the same, and will not be described in detail here.

[0113] Step S307: For each first target graph node: Determine the corresponding first task processing path based on the first target graph node.

[0114] For example, refer to Figure 1 The directed graph structure shown has the following paths: the first task processing path road_1 is Vs→V1→V4→V5→Ve, the first task processing path road_2 is Vs→V2→Ve, and the first task processing path road_3 is Vs→V5→V4→Ve. Specifically, the first task processing path road_1 corresponds to the user subtask request task_1, the first task processing path road_2 corresponds to the user subtask request task_2, and the first task processing path road_3 corresponds to the user subtask request task_3.

[0115] Specifically, the implementation method for "determining the corresponding first task processing path based on the first target graph node" for each first target graph node is the same as that in this application. Figure 2 The implementation of step S204 in the illustrated embodiment is the same, and will not be described in detail here.

[0116] Step S308: Summarize the processing paths of each first task to obtain the task processing path.

[0117] For example, based on the first task processing paths road_1, road_2, and road_3 determined in the embodiment shown in step S307, road_1, road_2, and road_3 are summarized. Each first task processing path is independent of each other and does not interfere with each other during the corresponding user subtask request processing. At the termination node Ve, the processing results can be summarized after each user subtask request is processed and the summarized processing results can be fed back to the user. Alternatively, the corresponding processing results can be fed back to the user immediately after any user subtask request is processed, so that the user can obtain the processing results corresponding to different user subtask requests more quickly.

[0118] In this embodiment, user task requests are split using a hierarchical routing strategy, which improves data processing efficiency. For example, in a video content analysis task, by splitting the video content analysis task and selecting sub-paths independently, the scheduling efficiency of multiple models is improved, solving the problem of poor model matching in existing technical solutions; it also solves the problems of low model scheduling efficiency and path conflicts under complex tasks, improving the ability to handle complex tasks.

[0119] based on Figures 2-4 The technical solutions of any of the method embodiments shown in this application can be applied to different types of model clusters. Each graph node included in the directed graph structure can correspond to various models or specific modules in models with different parameter levels, different modalities, different architectures, different manufacturers, different functions, etc.

[0120] For example, the first type of model cluster involves cross-network fusion / selection between language models with different parameter levels, such as: a small parameter model V_small (suitable for simple tasks and low-cost scenarios), a medium parameter model V_middle (suitable for medium-complexity tasks), and a large parameter model V_large (suitable for high-difficulty and high-value tasks). For the same user request, the task processing path is Vs→V_small→Ve, where a low-cost and fast answer is completed in the small parameter model; or Vs→V_small→V_large→Ve, where a brief analysis is performed in the small parameter model before calling the large parameter model for deep inference.

[0121] The second type of model cluster: cross-network fusion / selection between models of the same parameter scale but different languages, such as the Chinese model V_zh_LLM and the English model V_en_LLM; then for language mixing or cross-language scenarios, the task processing path may be: Vs→V_zh_LLM→Ve (Chinese task); or Vs→V_en_LLM→Ve (English task); or Vs→V_zh_LLM→V_en_LLM→Ve (cross-language task, understand first and then translate / polish).

[0122] The third type of model cluster: cross-network fusion / selection between different modal models, such as image understanding / generation model V_image, speech recognition / synthesis model V_asr, text model V_text, video model V_video, and native multimodal large model V_multimodal; for user task requests such as "image and text question answering", "speech question answering", and "video content summarization", the task processing path may be: Vs→V_image→V_text→Ve (image understanding + text generation); Vs→V_asr→V_text→Ve (speech transcription + text answer); Vs→V_video→V_text→Ve (video content analysis + text summary); Vs→V_multimodal→Ve (directly calling the native multimodal model).

[0123] The fourth type of model cluster: cross-network fusion between different modal sub-modules within a multimodal model. When resources are limited or cross-regional utilization is required, different modal sub-modules can be deployed in different regions or different computing centers. For example, the vision sub-module V_vision, the speech sub-module V_audio, and the text sub-module V_text; the task processing path may be: Vs→V_vision→V_text→Ve; or Vs→V_audio→V_text→Ve.

[0124] The fifth type of model cluster: different expert submodules V_{exp_ within a hybrid expert model }(in, Cross-network fusion between different expert submodules V_{exp_ }(in, When deployed in different regions or different computing centers, the task processing path may be as follows: Vs→V_{exp_2}→V_{exp_5}→V_{exp_7}→Ve.

[0125] Figure 5 This is a schematic diagram of the structure of a network body intelligent multi-model invocation device based on multi-dimensional feature constraints provided in one embodiment of this application, as shown below. Figure 5As shown, the intelligent multi-model invocation device 40 based on multi-dimensional feature constraints provided in this embodiment includes: an acquisition module 401, used to acquire a directed graph structure; wherein the directed graph structure includes multiple graph nodes and node features and edge features corresponding to each graph node, the graph nodes are used to indicate models in the model cluster, the node features are the model features of the model corresponding to the graph node, and the edge features are used to indicate the data transmission relationship between the models corresponding to the graph nodes; a processing module 402, used to generate task request features according to user task requests; and determine one or more graph nodes in the directed graph structure as target graph nodes according to the task request features and the multiple graph nodes and node features and edge features corresponding to each graph node; and a determination module 403, used to determine a task processing path according to the target graph node; wherein the task processing path is used to invoke the model corresponding to the target graph node according to the model invocation order corresponding to the task processing path when processing user task requests.

[0126] In one possible implementation, the graph nodes include a start node, one or more intermediate nodes, and an end node. When the processing module 402 determines one or more graph nodes in the directed graph structure as target graph nodes based on task request features and multiple graph nodes in the directed graph structure, as well as the node features and edge features corresponding to each graph node, it specifically performs the following steps: Based on the start node, determine multiple first candidate nodes from one or more intermediate nodes and the end node; wherein, data transfer is allowed between the model corresponding to the first candidate node and the model corresponding to the start node; determine the start node as the current node; determine the corresponding candidate edges based on the current node and each first candidate node; input the task request features, the node features of the current node, the node features of each first candidate node, and the edge features of the candidate edges corresponding to each first candidate node into the node scoring model, and output the estimated node scores corresponding to each first candidate node; determine the highest node score from the estimated node scores corresponding to each first candidate node, and determine the first candidate node corresponding to the highest node score as the target node; and determine the target graph node based on the target node.

[0127] In one possible implementation, when determining the target graph node based on the target node, the processing module 402 is specifically configured to: determine whether the target node is a termination node; if the target node is not a termination node, determine the target node as a path node, and determine multiple second candidate nodes from one or more intermediate nodes and termination nodes based on the target node; wherein, data transfer is allowed between the model corresponding to the second candidate node and the model corresponding to the target node; determine the target node as the current node, re-determine each second candidate node as a first candidate node, and return to execute the step of determining the corresponding candidate edge based on the current node and each first candidate node; if the target node is a termination node, determine the starting node and the termination node as target graph nodes, or determine the starting node, the termination node, and the path node as target graph nodes.

[0128] In one possible implementation, after processing the user task request according to the task processing path, the processing module 402 is further configured to: collect processing path data corresponding to each target graph node in the task processing path; wherein, the processing path data includes path latency, call cost, and result accuracy of the processing result corresponding to the user task request; for each target graph node: generate an actual node score corresponding to the target graph node based on path latency, call cost, and result accuracy; determine whether the difference between the highest node score and the actual node score corresponding to the target graph node is within a preset difference range; if the difference between the highest node score and the actual node score is not within the preset difference range, then train the node scoring model based on path latency, call cost, and result accuracy to update the model parameters of the node scoring model.

[0129] In one possible implementation, node features include node capability features, node network features, and node cost features; the node scoring model includes a feature correction sub-model and a feature scoring sub-model; when the processing module 402 inputs the task request features, the node features of the current node, the node features of each first candidate node, and the edge features of the candidate edges corresponding to each first candidate node into the node scoring model and outputs the estimated node score corresponding to each first candidate node, it is specifically used to: determine, based on the task request features, a first correction coefficient corresponding to the node capability features, a second correction coefficient corresponding to the node network features, and a third correction coefficient corresponding to the node cost features using the feature correction sub-model. The third correction coefficient; the feature correction sub-model corrects the node features of the current node based on the first, second, and third correction coefficients to obtain the corrected node features of the current node; the feature correction sub-model corrects the node features of each first candidate node based on the first, second, and third correction coefficients to obtain the corrected node features of each first candidate node; the feature scoring sub-model generates the estimated node score for each first candidate node based on the task request features, the corrected node features of the current node, the corrected node features of each first candidate node, and the edge features of the candidate edges corresponding to each first candidate node.

[0130] In one possible implementation, the task request features include multiple sub-task request features; when generating task request features based on user task requests, the processing module 402 is specifically used to: determine initial task request features based on user task requests; determine the number of splits for the user task request based on the initial task request features; split the user task request based on the number of splits to obtain the number of user sub-task requests; for each user sub-task request: perform feature parsing on the user sub-task request to generate sub-task request features; correspondingly, the target graph nodes include multiple first target graph nodes; the processing module 402, based on the task request features and multiple sub-task nodes in the directed graph structure... When determining one or more graph nodes in a directed graph structure as target graph nodes, based on the graph nodes and their corresponding node and edge features, the determination module 403, when determining the task processing path based on the target graph nodes, specifically performs the following: For each subtask request feature: Based on the task request feature and multiple graph nodes in the directed graph structure, as well as their corresponding node and edge features, determine one or more graph nodes in the directed graph structure as first target graph nodes; Correspondingly, when determining the task processing path based on the target graph nodes, the determination module 403, based on each first target graph node: Determine the corresponding first task processing path based on the first target graph node; Summarize each first task processing path to obtain the task processing path.

[0131] The intelligent multi-model calling device 40 based on multi-dimensional feature constraints provided in this embodiment can perform actions such as... Figures 2-4The technical solutions of any of the method embodiments shown are similar in implementation principle and technical effect, and will not be described again here.

[0132] Figure 6 A schematic diagram of the structure of the electronic device provided in this application. Figure 6 As shown, the electronic device 50 provided in this embodiment includes at least one processor 501 and a memory 502. Optionally, the device 50 further includes a communication component 503. The processor 501, memory 502, and communication component 503 are connected via a bus.

[0133] In a specific implementation, at least one processor 501 executes computer execution instructions stored in memory 502, causing at least one processor 501 to perform the above-described method.

[0134] The specific implementation process of processor 501 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0135] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0136] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0137] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0138] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0139] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0140] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0141] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0142] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0143] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0144] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0145] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0146] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0147] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for intelligent multi-model calling of a net body based on multi-dimensional feature constraints, characterized in that, include: Obtain a directed graph structure; wherein the directed graph structure includes multiple graph nodes and node features and edge features corresponding to each graph node, the graph nodes are used to indicate models in the model cluster, the node features are the model features of the model corresponding to the graph node, and the edge features are used to indicate the data transmission relationship between the models corresponding to the graph nodes; Generate task request characteristics based on the user's task request; Based on the task request features and multiple graph nodes in the directed graph structure, as well as the node features and edge features corresponding to each graph node, one or more graph nodes in the directed graph structure are determined as target graph nodes. Based on the target graph node, a task processing path is determined; wherein, the task processing path is used to call the model corresponding to the target graph node according to the model call order corresponding to the task processing path when processing the user task request.

2. The method of claim 1, wherein, The graph nodes include a start node, one or more intermediate nodes, and an end node; The step of determining one or more graph nodes in the directed graph structure as target graph nodes based on the task request features, multiple graph nodes in the directed graph structure, and the node features and edge features corresponding to each graph node includes: Based on the starting node, multiple first candidate nodes are determined from one or more intermediate nodes and ending nodes; wherein, data transfer is allowed between the model corresponding to the first candidate node and the model corresponding to the starting node; The starting node is determined as the current node; Based on the current node and each first candidate node, determine the corresponding candidate edges; The task request features, the node features of the current node, the node features of each first candidate node, and the edge features of the candidate edges corresponding to each first candidate node are respectively input into the node scoring model, and the estimated node score corresponding to each first candidate node is output. The highest node score is determined from the estimated node scores corresponding to each first candidate node, and the first candidate node corresponding to the highest node score is determined as the target node. Based on the target node, determine the target graph node.

3. The method of claim 2, wherein, Determining the target graph node based on the target node includes: Determine whether the target node is the termination node; If the target node is not the termination node, then the target node is determined as a path node, and based on the target node, multiple second candidate nodes are determined from one or more intermediate nodes and termination nodes; wherein, data transfer is allowed between the model corresponding to the second candidate node and the model corresponding to the target node; The target node is determined as the current node, each second candidate node is redefined as a first candidate node, and the process returns to the step of determining the corresponding candidate edge based on the current node and each first candidate node. If the target node is the termination node, then the starting node and the termination node are determined as the target graph node, or the starting node, the termination node, and the path node are determined as the target graph node.

4. The method of claim 2, wherein, After processing the user task request according to the task processing path, the method further includes: Collect processing path data corresponding to each target graph node in the task processing path; wherein, the processing path data includes path delay, call cost, and result accuracy of the processing result corresponding to the user task request; For each target graph node: Based on the path delay, the call cost, and the result accuracy, generate the actual node score corresponding to the target graph node; Determine whether the difference between the highest node score and the actual node score corresponding to the target graph node is within a preset difference range; If the difference between the highest node score and the actual node score is not within a preset difference range, the node scoring model is trained based on the path delay, the call cost, and the result accuracy to update the model parameters of the node scoring model.

5. The method of claim 2, wherein, The node characteristics include node capability characteristics, node network characteristics, and node cost characteristics; The node scoring model includes a feature correction sub-model and a feature scoring sub-model; The step of inputting the task request features, the node features of the current node, the node features of each first candidate node, and the edge features of the candidate edges corresponding to each first candidate node into the node scoring model, and outputting the estimated node score corresponding to each first candidate node, includes: The feature correction sub-model determines, based on the task request features, a first correction coefficient corresponding to the node capability feature, a second correction coefficient corresponding to the node network feature, and a third correction coefficient corresponding to the node cost feature. The feature correction sub-model corrects the node features of the current node based on the first correction coefficient, the second correction coefficient, and the third correction coefficient to obtain the corrected node features of the current node. The feature correction sub-model corrects the node features of each first candidate node based on the first correction coefficient, the second correction coefficient, and the third correction coefficient, thereby obtaining the corrected node features of each first candidate node. The feature scoring sub-model generates the estimated node score for each first candidate node based on the task request features, the corrected node features of the current node, the corrected node features of each first candidate node, and the edge features of the candidate edges corresponding to each first candidate node.

6. The method according to any one of claims 1 to 5, characterized in that, The task request features include multiple sub-task request features; The step of generating task request features based on user task requests includes: Based on the user's task request, determine the initial task request characteristics; Based on the initial task request characteristics, determine the number of splits for the user task request; The user task request is split according to the number of splits to obtain the number of user sub-task requests. For each user's subtask request: Perform feature parsing on the user's subtask request to generate the subtask request features; Accordingly, the target graph node includes multiple first target graph nodes; the step of determining one or more graph nodes in the directed graph structure as target graph nodes based on the task request features and the multiple graph nodes in the directed graph structure, as well as the node features and edge features corresponding to each graph node, includes: Based on the request characteristics of each subtask: Based on the task request features and the multiple graph nodes in the directed graph structure, as well as the node features and edge features corresponding to each graph node, one or more graph nodes in the directed graph structure are determined as the first target graph node. Accordingly, determining the task processing path based on the target graph nodes includes: For each node in the first target graph: Based on the first target graph node, determine the corresponding first task processing path; The task processing paths are obtained by summarizing the processing paths of each first task.

7. A device for intelligent multi-model calling of a net body based on multi-dimensional feature constraints, characterized in that, include: An acquisition module is used to acquire a directed graph structure; wherein, the directed graph structure includes multiple graph nodes and node features and edge features corresponding to each graph node, the graph nodes are used to indicate models in a model cluster, the node features are model features of the model corresponding to the graph node, and the edge features are used to indicate the data transmission relationship between the models corresponding to the graph nodes. The processing module is used to generate task request features based on user task requests; and to determine one or more graph nodes in the directed graph structure as target graph nodes based on the task request features, multiple graph nodes in the directed graph structure, and the node features and edge features corresponding to each graph node. The determination module is used to determine the task processing path based on the target graph node; wherein, the task processing path is used to call the model corresponding to the target graph node according to the model call order corresponding to the task processing path when processing the user task request.

8. An electronic device, comprising: include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 6.

Images