Computing power optimization method and device based on multi-agent cooperation
By building intention identification large models, business models and troubleshooting small models, and reasonably allocating GPU and CPU resources, the problem of low computing resource utilization rate when multiple agents work together is solved, and higher concurrent processing capabilities and resources are achieved Utilization.
Patent Information
- Application Number
- CN202411371987.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-29
- Publication Date
- 2025-01-10
AI Technical Summary
In the prior art, when multiple agents work together, the utilization rate of computing power resources is low, resulting in large models occupying too much resources when performing a single task, and other models cannot fully utilize limited computing power resources to support more concurrent users.
Build a big model of intention identification, business model and small model of troubleshooting, and use user intention identification and distribution to reasonably allocate GPU and CPU resources to avoid a single large model from continuously occupying GPU resources and realize concurrent processing tasks for multiple models.
It significantly improves the overall utilization rate of computing power resources when multiple agents work together, avoids inefficient modes, and ensures the concurrent processing capabilities of more users.
Smart Images
Figure CN119292779A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of network intelligent operation and maintenance, and specifically to a computing power optimization method and device based on multi-agent collaboration. Background Art
[0002] With the rapid development of artificial intelligence technology, large models have shown powerful capabilities in solving complex problems, especially in the fields of natural language processing and image recognition. However, the high performance of large models is often accompanied by high demand for computing resources, especially in the model reasoning stage, which requires huge computing power support. This demand is particularly evident in artificial intelligence assistants in the field of control, because artificial intelligence assistants in the field of control usually need to process a large number of concurrent requests in real time, and each request may involve complex problem analysis and decision making.
[0003] In order to overcome the above problems, the industry has begun to explore ways to improve the concurrent processing capabilities and resource utilization of AI assistants in the field of control by optimizing model usage strategies. Among them, an effective strategy is to adopt a multi-agent collaborative approach. In this way, the system will use multiple models to jointly process tasks, thereby improving the concurrent capabilities of AI assistants in the field of control.
[0004] However, existing methods lack an efficient resource allocation mechanism. When executing a single task, they will call large and powerful models multiple times, causing other models to lack computing resources and fall into a waiting state, making it impossible to fully utilize limited computing resources to support more concurrent users. Summary of the invention
[0005] The present application provides a computing power optimization method and device based on multi-agent collaboration, which can solve the problem of low computing power resource utilization when multiple agents work together in the prior art.
[0006] In a first aspect, an embodiment of the present application provides a computing power optimization method based on multi-agent collaboration, the method comprising:
[0007] Build a large model for intent recognition, a large model for business, and a small model for troubleshooting;
[0008] Generate user intent according to a natural language request sent by the user; the large intent recognition model occupies GPU resources, and releases GPU resources after generating a recognition result according to the user intent;
[0009] Distribute the user intention to the large business model, the small troubleshooting model, and the control interface in the existing control system according to the recognition result;
[0010] The large business model, small troubleshooting model and management and control interface perform tasks according to the user's intention.
[0011] In combination with the first aspect, in one implementation, constructing a large model for intention recognition includes:
[0012] According to the preset control documents, the preset model is pre-trained and fine-tuned to obtain a large intent recognition model.
[0013] In combination with the first aspect, in one implementation, a preset model is pre-trained and fine-tuned based on a preset control document to obtain a large intent recognition model, including:
[0014] Pre-training the preset model for a second time according to the preset control document to obtain a first model;
[0015] The first model is fine-tuned according to the preset QA question and answer data to obtain a large intent recognition model.
[0016] In conjunction with the first aspect, in one implementation, the building of the large business model and the small troubleshooting model includes:
[0017] According to the preset business configuration documents and protocol documents, a large business model is trained;
[0018] Based on the preset fault handling documents and expert experience documents, a small troubleshooting model is trained.
[0019] In combination with the first aspect, in one implementation, after constructing the intention recognition large model, the business large model and the troubleshooting small model, the method further includes:
[0020] The intention recognition big model, business big model and troubleshooting small model are deployed to the AI engine of the existing management and control system.
[0021] In combination with the first aspect, in one implementation, after the intention recognition large model, the business large model and the troubleshooting small model are deployed to the AI engine of the existing management and control system, the method further includes:
[0022] When the preset control document is changed, the intent recognition large model is retrained and deployed according to the changed control document;
[0023] When the preset business configuration document and protocol document are changed, the business big model is retrained and deployed according to the changed business configuration document and protocol document;
[0024] When the preset fault handling documents and expert experience documents are changed, the small troubleshooting model is retrained and deployed according to the changed fault handling documents and expert experience documents.
[0025] In combination with the first aspect, in one implementation, distributing the user intention to the large business model, the small troubleshooting model, and the management and control interface in the existing management and control system according to the recognition result includes:
[0026] According to the recognition result, the user intention can be divided into business intention, troubleshooting intention and query intention;
[0027] The business intention is sent to the large business model, the troubleshooting intention is sent to the small troubleshooting model, and the query intention is sent to the management and control interface in the existing management and control system.
[0028] In combination with the first aspect, in one implementation, when the business large model, the troubleshooting small model, and the management and control interface perform tasks according to the user intention, the following further includes:
[0029] When the business big model executes a task, it occupies GPU resources, and releases GPU resources after completing the task;
[0030] When the troubleshooting small model and the control interface in the existing control system perform tasks, no GPU resources are occupied.
[0031] In combination with the first aspect, in one implementation, after the business large model, the troubleshooting small model, and the management and control interface execute tasks according to the user intention, the following further includes:
[0032] The large business model, small troubleshooting model and management and control interface send the execution results of the task to the operation interface of the existing management and control system.
[0033] In a second aspect, an embodiment of the present application provides a computing power optimization device based on multi-agent collaboration based on any of the above methods, the device comprising:
[0034] Construction module, used to build large intent recognition models, large business models, and small troubleshooting models;
[0035] An analysis module, used to generate a user intent according to a natural language request sent by a user; the large intent recognition model occupies GPU resources, and the GPU resources are released after generating a recognition result according to the user intent;
[0036] A distribution module, used to distribute the user intention to the large business model, the small troubleshooting model, and the control interface in the existing control system according to the recognition result;
[0037] A processing module is used for the large business model, small troubleshooting model and management and control interface to perform tasks according to the user's intention.
[0038] The beneficial effects brought by the technical solution provided by the embodiments of the present application include:
[0039] This application identifies user intent through a large intent recognition model, and then distributes the user intent to the large business model, the small troubleshooting model, and the management and control interface in the existing management and control system, allowing multiple models to process tasks concurrently, avoiding the inefficient mode of a single large model processing tasks alone while continuously occupying GPU (Graphics Processing Unit) resources, thereby significantly improving the overall utilization of computing resources when multiple intelligent agents work together. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Figure 1 A flowchart of a computing power optimization method based on multi-agent collaboration according to an embodiment of the present application;
[0041] Figure 2 A schematic diagram of user usage of the computing power optimization method based on multi-agent collaboration in an embodiment of the present application. DETAILED DESCRIPTION
[0042] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0043] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below in conjunction with the accompanying drawings.
[0044] First, please refer to Figure 1 , Figure 1 The flowchart of the computing power optimization method based on multi-agent collaboration in the embodiment of the present application is shown in FIG. The computing power optimization method provided in this embodiment includes the following steps:
[0045] Step S1: Build a large intent recognition model, a large business model, and a small troubleshooting model.
[0046] Step S2: Generate user intent based on the natural language request sent by the user, and then the large intention recognition model occupies GPU resources. After generating the recognition result based on the user intent, the GPU resources are released.
[0047] Step S3: Distribute the user intent to the large business model, the small troubleshooting model, and the management and control interface in the existing management and control system based on the recognition results.
[0048] Step S4: The large business model, small troubleshooting model and management and control interface execute tasks according to user intent.
[0049] This application uses a large intent recognition model to identify user intent, and then distributes the recognition results to a large business model, a small troubleshooting model, and the management and control interface in the existing management and control system to achieve concurrent processing of multiple models. The large intent recognition model releases GPU resources after generating recognition results based on user intent, effectively avoiding the inefficient mode of a single large model continuously occupying GPU resources and processing tasks alone. The large business model, the small troubleshooting model, and the management and control interface in the existing management and control system can quickly solve problems and release resources for other users to handle other problems. By reasonably allocating roles and processes, the computing power utilization rate when multiple agents work together is significantly improved.
[0050] In some embodiments, in the above step S1, constructing the intention recognition model, the business model and the troubleshooting model includes the following steps:
[0051] According to the preset control documents, the preset model is pre-trained and fine-tuned to obtain a large intent recognition model.
[0052] It should be noted that the above-mentioned preset control documents include but are not limited to standards, manuals, books and documents in the control field and their respective sub-fields. The above-mentioned preset models include but are not limited to Qwen2-72B and 200B-level natural language processing models.
[0053] In some embodiments, in the above steps, the preset model is pre-trained and fine-tuned according to the preset control document to obtain the intent recognition model, including the following steps:
[0054] First, the preset model is pre-trained for a second time according to the preset management and control document to obtain the first model.
[0055] The first model is fine-tuned according to the preset QA (Question Answer) data to obtain the large intent recognition model.
[0056] It should be noted that the purpose of the secondary pre-training here is to enable the large intent recognition model to have a preliminary understanding of the knowledge in the field of communication industry control. The purpose of fine-tuning training based on QA question and answer data is to enable the large intent recognition model to have the ability to recognize user intent.
[0057] Compared to pre-trained models trained on general datasets, models that have been pre-trained twice can adapt faster and show higher performance when handling tasks in specific fields. During the fine-tuning stage, because the model already has certain control domain knowledge, it can adapt and converge to the best performance faster. This means that with the same number of training iterations, a model that has been pre-trained twice can reach or exceed its performance faster than a model that has not been pre-trained.
[0058] In this embodiment, specifically, the process of training to obtain the large model for intent recognition can refer to the following examples, for example:
[0059] Question about the intent recognition model: I need to build a SPN (Slicing Packet Network) service from Hankou to Wuchang. How do I configure it?
[0060] Intent recognition big model answer: This is a business question, please visit the business big model answer, and call the business big model interface to answer the question.
[0061] Question to the intention recognition model: How to handle the RLOS (Receive Loss of Signal) alarm that occurs in a certain network element?
[0062] Intent recognition large model answer: This is a troubleshooting question. Please visit the troubleshooting small model answer and call the troubleshooting small model interface to answer the question.
[0063] Question to the intent recognition model: Query the free ports of a certain network element.
[0064] The intent recognition large model answers: This is a query interface problem. Please query the free port interface and call the interface to execute.
[0065] In some embodiments, in the above step S1, building a large business model and a small troubleshooting model includes the following steps:
[0066] Based on the preset business configuration documents and protocol documents, a large business model is trained. Based on the preset fault handling documents and expert experience documents, a small troubleshooting model is trained.
[0067] It should be noted that the large business model obtained through training has the ability to handle business problems in the field of management and control. The small troubleshooting model obtained through training has the ability to handle business troubleshooting problems in the field of management and control.
[0068] In this embodiment, specifically, the following QA example may be referred to:
[0069] Question about the business model: I need to build an SPN service from Hankou to Wuchang. How do I configure it?
[0070] The business model answers: The first step is to find an existing device in Hankou and Wuchang that meets the needs; the second step is to find the path; the third step is to build a certain SPN service on the routing path; the fourth step is to issue certain configuration parameters, ... and so on; each step has the ability to call the existing management and control system to specifically execute the corresponding interface, and return the execution result to the next step, thereby dynamically adjusting the parameters of the next step.
[0071] Question to the troubleshooting model: How to handle the RLOS alarm of a certain network element?
[0072] The troubleshooting model answers: The first step is to check whether there is an RLOS alarm on the other end; the second step is to use OTDR (Optical Time Domain Reflectometer) to scan the wave; the third step is to determine whether the fiber is broken and give repair suggestions, ... and so on; each step can return detailed information for users to refer to for troubleshooting operations.
[0073] In some embodiments, in the above step S1, after constructing the intention recognition large model, the business large model and the troubleshooting small model, the following steps are also included:
[0074] Deploy the intention recognition big model, business big model and troubleshooting small model to the AI (Artificial Intelligence) engine of the existing management and control system.
[0075] In some embodiments, in the above steps, after the intention recognition big model, the business big model and the troubleshooting small model are deployed to the AI engine of the existing management and control system, it also includes:
[0076] When the preset control document changes, retrain and deploy the intent recognition model based on the changed control document.
[0077] When the preset business configuration documents and protocol documents are changed, the business model is retrained and deployed according to the changed business configuration documents and protocol documents.
[0078] When the preset fault handling documents and expert experience documents are changed, the small troubleshooting model is retrained and deployed based on the changed fault handling documents and expert experience documents.
[0079] It should be noted that with the changes in the business environment and the iteration of technology, the original document content may become outdated or no longer applicable. By timely retraining the model based on the latest documents, it can be ensured that the model is always based on the latest information processing tasks, thus maintaining its accuracy and timeliness.
[0080] In some embodiments, in the above step S2, if the preset model used for the secondary pre-training of the large intention recognition model is Qwen2-72B, the GPU resource video memory occupied by the large intention recognition model when recognizing the user's intention is 144GB.
[0081] In some embodiments, in the above step S3, according to the recognition result, distributing the user intention to the business big model, the troubleshooting small model, and the management and control interface in the existing management and control system includes the following steps:
[0082] First, based on the recognition results, the user intent is divided into business intent, troubleshooting intent and query intent. Then, the business intent is sent to the large business model, the troubleshooting intent is sent to the small troubleshooting model, and the query intent is sent to the management and control interface in the existing management and control system.
[0083] It should be noted that by subdividing the intent and sending it to the large business model, the small troubleshooting model, and the control interface in the existing control system, it helps to allocate resources more reasonably. For example, GPU resources can be allocated first to the large business model that requires high-performance computing, while CPU resources can be used to process troubleshooting intent and query intent. This resource allocation method ensures that this method can efficiently utilize limited computing resources to meet different business needs.
[0084] In some embodiments, in the above step S4, when the business large model, the troubleshooting small model and the management and control interface perform tasks according to the user's intention, it also includes:
[0085] When the large business model executes a task, it occupies GPU resources and releases GPU resources after the task is completed.
[0086] When the troubleshooting model and the control interface in the existing control system perform tasks, they do not occupy GPU resources.
[0087] Specifically, if the parameter size of the large business model is 7B, the GPU resource memory occupied is only 14GB, and the speed of executing business intent tasks is higher than that of the large intent recognition model. The small troubleshooting model and the management and control interface processing tasks do not require the high-performance computing capabilities of the GPU, so they do not occupy GPU resources.
[0088] When the task is completed, the business model releases GPU resources, which allows other tasks that require GPU acceleration (such as the business model's tasks) to be executed smoothly without being affected by GPU resource limitations. This dynamic resource allocation mechanism ensures the maximum utilization of computing resources.
[0089] In some embodiments, in the above step S4, after the business large model, the troubleshooting small model and the management and control interface perform tasks according to the user's intention, the following steps are also included:
[0090] The large business model, small troubleshooting model and management and control interface send the execution results of the task to the operation interface of the existing management and control system.
[0091] In some embodiments, please refer to Figure 2 , Figure 2 This is a user usage flow chart of an embodiment of a computing power optimization method based on multi-agent collaboration in this application. The specific process of users using this method to solve problems is as follows:
[0092] First, the user inputs a question in natural language in the operation interface, and the intention understanding module generates the user intention based on the natural language request issued by the user, and transmits the user intention to the intention recognition model. In this embodiment, the parameter scale of the intention recognition model is 200B.
[0093] Afterwards, the large intent recognition model occupies GPU resources, generates recognition results based on user intent, releases GPU resources, and feeds the recognition results back to the intent understanding module.
[0094] Then, the intent understanding module divides the user intent into business intent, troubleshooting intent, and query intent based on the recognition results, and distributes the business intent to the business big model, the troubleshooting intent to the troubleshooting small model, and the query intent to the management and control interface in the existing management and control system. In this embodiment, the parameter scale of the business big model is 7B.
[0095] Finally, the large business model, the small troubleshooting model and the management and control interface execute tasks according to the user's intention and return the execution results to the operation interface. In this embodiment, the large business model occupies GPU resources when executing tasks, and the small troubleshooting model occupies CPU resources when executing tasks.
[0096] In a second aspect, an embodiment of the present application provides a computing power optimization device based on multi-agent collaboration, which includes a construction module, an analysis module, a distribution module and a processing module.
[0097] The above-mentioned building modules are used to build a large model for intent recognition, a large business model and a small model for troubleshooting.
[0098] The above analysis module is used to generate user intent based on the natural language request sent by the user. The large model for intent recognition occupies GPU resources, and the GPU resources are released after the recognition result is generated according to the user intent.
[0099] The distribution module is used to distribute user intent to the large business model, the small troubleshooting model, and the management and control interface in the existing management and control system according to the recognition results.
[0100] The above processing modules are used for large business models, small troubleshooting models and management and control interfaces to perform tasks according to user intentions.
[0101] It should be noted that the serial numbers of the above-mentioned embodiments of the present application are only for description and do not represent the advantages or disadvantages of the embodiments.
[0102] The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices. The terms "first", "second" and "third" are used to distinguish different objects, etc., and do not represent a sequence, nor do they limit "first", "second" and "third" to different types.
[0103] In the description of the embodiments of the present application, "exemplary", "for example" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary", "for example" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary", "for example" or "for example" is intended to present related concepts in a specific way.
[0104] In the description of the embodiments of the present application, unless otherwise specified, “ / ” means or, for example, A / B can mean A or B; the “and / or” in the text is merely a description of the association relationship of associated objects, indicating that three relationships may exist, for example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, in the description of the embodiments of the present application, “multiple” refers to two or more than two.
[0105] In some processes described in the embodiments of the present application, multiple operations or steps that appear in a specific order are included, but it should be understood that these operations or steps may not be executed in the order in which they appear in the embodiments of the present application or in parallel, and the sequence number of the operation is only used to distinguish the different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed in sequence or in parallel, and these operations or steps may be combined.
[0106] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, disk, CD) as described above, and includes a number of instructions for a terminal device to execute the methods described in each embodiment of the present application.
[0107] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.
Claims
1. A computing power optimization method based on multi-agent collaboration, characterized in that: The method comprises: Build a large model for intent recognition; Generate user intent based on natural language requests made by users; The intention recognition big model generates a recognition result according to the user intention; Distribute the user intention to a corresponding execution unit according to the recognition result.
2. The computing power optimization method based on multi-agent collaboration as claimed in claim 1, characterized in that: The construction of the intention recognition large model includes: According to the preset control documents, the preset model is pre-trained and fine-tuned to obtain a large intent recognition model.
3. The computing power optimization method based on multi-agent collaboration as claimed in claim 2, characterized in that: According to the preset control documents, the preset model is pre-trained and fine-tuned to obtain the intent recognition model, including: Pre-training the preset model for a second time according to the preset control document to obtain a first model; The first model is fine-tuned according to the preset QA question and answer data to obtain a large intent recognition model.
4. The computing power optimization method based on multi-agent collaboration as claimed in claim 1, characterized in that: The execution unit includes a large business model and a small troubleshooting model, wherein: According to the preset business configuration document and protocol document, the business big model is obtained by training; The small troubleshooting model is trained based on preset fault handling documents and expert experience documents.
5. The computing power optimization method based on multi-agent collaboration as claimed in claim 4, characterized in that: After building a large model for intent recognition, it also includes: The intention recognition big model, business big model and troubleshooting small model are deployed to the AI engine of the existing management and control system.
6. The computing power optimization method based on multi-agent collaboration as claimed in claim 5, characterized in that: After the intention recognition big model, business big model and troubleshooting small model are deployed to the AI engine of the existing management and control system, it also includes: When the preset control document is changed, the intent recognition large model is retrained and deployed according to the changed control document; When the preset business configuration document and protocol document are changed, the business big model is retrained and deployed according to the changed business configuration document and protocol document; When the preset fault handling documents and expert experience documents are changed, the small troubleshooting model is retrained and deployed according to the changed fault handling documents and expert experience documents.
7. The computing power optimization method based on multi-agent collaboration as claimed in claim 4, characterized in that: Distributing the user intention to a corresponding execution unit according to the recognition result includes: According to the recognition result, the user intention can be divided into business intention, troubleshooting intention and query intention; The business intention is sent to the large business model, the troubleshooting intention is sent to the small troubleshooting model, and the query intention is sent to the management and control interface in the existing management and control system.
8. The computing power optimization method based on multi-agent collaboration as claimed in claim 7, characterized in that: After distributing the user intention to the corresponding execution unit according to the recognition result, the method further includes: The large business model, the small troubleshooting model, and the management and control interfaces in the existing management and control system perform tasks based on the distributed user intent.
9. The computing power optimization method based on multi-agent collaboration as claimed in claim 8, characterized in that: After the business model, troubleshooting model, and control interfaces in the existing control system perform tasks according to the distributed user intent, they also include: The large business model, the small troubleshooting model and the management and control interface in the existing management and control system send the execution results of the task to the operation interface of the existing management and control system.
10. A computing power optimization device based on multi-agent collaboration based on the method according to any one of claims 1 to 9, characterized in that: The device comprises: Building modules for building large models of intent recognition; An analysis module for generating user intent based on a natural language request issued by a user; A recognition module, used for the intention recognition model to generate a recognition result according to the user intention; A distribution module is used to distribute the user intention to a corresponding execution unit according to the recognition result.