AI large-scale model deployment methods and systems for enterprise scenarios

Abstract

This invention discloses a method and system for deploying large-scale AI models for enterprise scenarios. The method includes: S1, acquiring the business dataset of the target enterprise scenario and preprocessing the business dataset to generate a business scenario requirement set; S2, acquiring multiple sets of private datasets based on the business scenario requirement set and constructing a mapping between the business scenario requirement set and the multiple sets of private data; S3, writing the multiple sets of private datasets into various models in a pre-set large model library; S4, acquiring the hardware dataset of the target enterprise scenario, generating multiple sets of hardware containers based on the hardware data, with each hardware container configured with a corresponding hardware interface for the large model; S5, generating multiple simulated business requests to the pre-set large model library based on the business dataset and receiving simulated responses from various models; S6, determining the deployment scheme for the current business dataset based on the simulated business requests and responses. This method adapts to future business scenarios of enterprises, ensures model performance, and reduces enterprise deployment costs.

Description

Technical Field

[0001] This invention relates to the field of data intelligence technology, specifically to a method and system for deploying large-scale AI models for enterprise scenarios. Background Technology

[0002] With the development of artificial intelligence technology, large AI models have significant potential for application in enterprise business. However, enterprises face multiple problems when deploying them: general-purpose large AI models are poorly adapted to specific business scenarios, making it difficult to meet the accuracy and efficiency requirements of business operations; uploading private and sensitive enterprise data to third-party platforms can easily lead to security and privacy leaks; enterprise deployment environments are complex, general solutions have poor compatibility, resulting in high deployment costs and difficult operation and maintenance; large models consume a lot of resources, with response delays during peak business periods and idle and wasted resources during off-peak periods.

[0003] In existing technologies, some methods only optimize the model itself without customizing it to suit specific business scenarios, some local deployment solutions lack compatibility, and some resource scheduling methods fail to consider the business load characteristics of large models, resulting in low scheduling accuracy. Therefore, there is an urgent need for a method for deploying large AI models that achieves business scenario adaptation, data security assurance, multi-environment compatibility, and efficient resource utilization. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a method and system for deploying large-scale AI models for enterprise scenarios.

[0005] A first aspect of the present invention provides a method for deploying large-scale AI models for enterprise scenarios, comprising the following steps:

[0006] S1. Obtain the business dataset of the target enterprise scenario, and preprocess the business dataset to generate a business scenario requirement set;

[0007] S2. Obtain multiple sets of private datasets based on the business scenario requirement set, and construct a mapping between the business scenario requirement set and the multiple sets of private datasets;

[0008] S3. Write the multiple sets of the private datasets into the major models in the preset large model library;

[0009] S4. Obtain the hardware dataset of the target enterprise scenario, generate multiple sets of hardware containers based on the hardware dataset, and set the hardware interface corresponding to the large model in each hardware container.

[0010] S5. Based on the business dataset, generate multiple simulated business requests to the preset large model library, and receive the simulated responses from each of the large models;

[0011] S6. Determine the deployment scheme of the current business dataset based on the simulated service request and the simulated response.

[0012] As a preferred embodiment, the business dataset includes:

[0013] Based on the business modules of the target enterprise scenario, collect historical business data, real-time business data and corresponding business rule data for each business module;

[0014] The set of business scenario requirements was obtained through the following steps:

[0015] Determine the calling probability and calling frequency of business terms in the historical business data and real-time business data under the current business rule data in the business dataset;

[0016] Obtain the probability call sequence and frequency call sequence of the first business rule regarding time;

[0017] The probability call sequence is sorted by probability value with timestamp as the sequence axis;

[0018] The frequency call sequence is sorted by frequency value with timestamp as the sequence axis;

[0019] The probability call sequence and the frequency call sequence are combined to obtain the business scenario requirement set.

[0020] As a preferred method, the private dataset is obtained through the following steps:

[0021] Obtain the probability call sequence and the frequency call sequence;

[0022] Obtain the valid timestamps of each business rule in the business dataset;

[0023] Determine multiple future timestamps after the current valid timestamp, exhaustively enumerate multiple samples, determine the probability of each business rule occurring at different timestamps, and obtain the rule probability sequence;

[0024] Each rule in the rule probability sequence obtains a first sample and a second sample of business data from the probability call sequence and the frequency call sequence;

[0025] The first sample and the second sample are merged to obtain a private dataset.

[0026] As a preferred approach, step S3, writing multiple sets of private datasets into various models in a pre-defined large model library, includes the following steps:

[0027] Parse the rule probability sequence in the private dataset and extract the probability distribution features of each business rule at different timestamps;

[0028] Based on the aforementioned probability distribution characteristics, the input embedding layer of the large model is dynamically weighted, wherein the weight values ​​are positively correlated with the probability values ​​of the corresponding timestamps in the regular probability sequence.

[0029] The dynamic weight adjustment ensures that the large model prioritizes responding to requests from high-probability business rules during inference.

[0030] As a preferred approach, in step S4, generating multiple sets of hardware containers includes the following steps:

[0031] Analyze the probability and frequency call sequences in the business scenario demand set to determine the peak time window for business term calls and resource-intensive business modules;

[0032] Based on the peak time window and resource-intensive business modules, the computing resources of the hardware dataset are dynamically allocated to generate the initial configuration parameters of the hardware container.

[0033] The hardware interface of the hardware container is configured to allocate low-latency I / O channels to the business modules corresponding to the high-frequency call sequence, and to reserve elastic computing resources for high-probability timestamps in the probability call sequence.

[0034] As a preferred approach, in step S5, generating multiple simulated service requests includes:

[0035] Based on the frequency call sequence of business scenario requirements, a weight distribution model of business terms is constructed, where the weight value is determined by normalizing the frequency value of the timestamp unit of the call frequency sequence.

[0036] Based on the weight distribution model, real-time business data in the business dataset is sampled to generate simulated request clusters of high-weight business terms;

[0037] The simulated request clusters are injected into a pre-set large model library in the order of timestamps of the probability call sequence to simulate the temporal characteristics of real business scenarios.

[0038] As a preferred approach, step S6, determining the deployment scheme for the current business dataset, includes the following steps:

[0039] Calculate the conformity score between each simulated response and the business rule data. The conformity score is calculated based on the weighted probability sequence of the rules in the private dataset, and the weight is set to the probability of the response unit corresponding to the rule.

[0040] By combining the response time and compliance score of simulated business requests, deployment priorities are generated;

[0041] The deployment scheme selects the hardware container and its corresponding large model with a compliance score exceeding the threshold and the shortest response time as the target deployment entity.

[0042] A second aspect of the present invention provides an AI large-scale model deployment system for enterprise scenarios, comprising:

[0043] The data acquisition module is configured to acquire business datasets for the target enterprise scenario;

[0044] The data processing module is configured to determine the set of business scenario requirements based on the business dataset;

[0045] The server is configured to pre-store a large model library and a private dataset, and is also configured to generate simulation requests and receive simulation responses.

[0046] Compared with the prior art, the present invention has the following advantages:

[0047] By constructing probabilistic call sequences and frequency call sequences, historical and real-time business data are transformed into dynamic rule distributions on the timeline and rule probability sequences are generated. The business rule probabilities of future timestamps are enumerated. Furthermore, by using sequence data to construct a weight distribution model, time-aware simulated request clusters are generated. The time sensitivity of business rules is quantified into computable features, meeting dynamic business needs and significantly improving the fit between the model and enterprise scenarios. Attached Figure Description

[0048] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0049] Figure 1 This is a flowchart illustrating the method provided in an embodiment of the present invention.

[0050] Figure 2 This is a schematic diagram of the system provided in an embodiment of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] The first aspect of this embodiment provides a method for deploying large-scale AI models for enterprise scenarios, such as... Figure 1 As shown, it includes the following steps:

[0053] S1. Obtain the business dataset of the target enterprise scenario, and preprocess the business dataset to generate a business scenario requirement set;

[0054] S2. Obtain multiple sets of private datasets based on the business scenario requirement set, and construct a mapping between the business scenario requirement set and the multiple sets of private datasets;

[0055] S3. Write the multiple sets of the private datasets into the major models in the preset large model library;

[0056] S4. Obtain the hardware dataset of the target enterprise scenario, generate multiple sets of hardware containers based on the hardware data, and set up a hardware interface corresponding to the large model in each hardware container.

[0057] S5. Based on the business dataset, generate multiple simulated business requests to the preset large model library, and receive the simulated responses from each of the large models;

[0058] S6. Determine the deployment scheme of the current business dataset based on the simulated service request and the simulated response.

[0059] As a preferred embodiment of this disclosure, the business dataset includes: collecting historical business data, real-time business data, and corresponding business rule data for each business module according to the business modules of the target enterprise scenario.

[0060] Furthermore, in this embodiment of the disclosure, the set of business scenario requirements is obtained through the following steps:

[0061] In the business dataset, under the current business rule data, the calling probability and calling frequency of business terms in the historical and real-time business data are determined. Specifically, in this embodiment, the calling frequency is the frequency of occurrence of a business term when a business module initiates a call under the current business rule, representing the number of times a specific business term appears in the real-time and historical business data; the calling probability is the probability of occurrence of a business term when a business module initiates a call under the current business rule. Since at least one business term corresponds to one call, multiple business terms corresponding to the same call may appear simultaneously. The probability of each term appearing in all calls is calculated based on the frequency at this time. It should be noted that when any number of calls occur, the corresponding business term is always one. At this time, the calling probability of the business term under the business rule is equal to the calling frequency, that is, the calling probability is a value estimated based on the calling frequency. In addition, in this embodiment, the call is the occurrence of one event or response of the business module.

[0062] Obtain the probability call sequence and frequency call sequence of the first business rule regarding time;

[0063] The probability call sequence is sorted by probability value with timestamp as the sequence axis;

[0064] The frequency call sequence is sorted by frequency value with timestamp as the sequence axis;

[0065] The probability call sequence and the frequency call sequence are combined to obtain the business scenario requirement set.

[0066] As a preferred method, the private dataset is obtained through the following steps:

[0067] Obtain the probability call sequence and the frequency call sequence;

[0068] Obtain the valid timestamps of each business rule in the business dataset;

[0069] Determine multiple future timestamps after the current valid timestamp, exhaustively enumerate multiple samples, determine the probability of each business rule occurring at different timestamps, and obtain the rule probability sequence;

[0070] Each rule in the rule probability sequence obtains a first sample and a second sample of business data from the probability call sequence and the frequency call sequence;

[0071] The first sample and the second sample are merged to obtain a private dataset.

[0072] It should be noted that in this embodiment, the effective timestamp is the effective time of the business rule. The effective timestamp is determined by recording the effective time of each business rule in historical business data and real-time business data. In this implementation, the probability call sequence and frequency call sequence associated with the business rule both have time characteristics. Specifically, the value recorded in the probability call sequence is not overwritten by the effective timestamp of the current business rule. That is, the unit value recorded in the future timestamp is calculated from the value in the historical business data. Then, the probability of each business rule occurring at different timestamps is determined by exhaustive sampling to obtain the rule probability sequence. The rule probability sequence and the probability call sequence and frequency call sequence obtained according to the future timestamp are written into the mapping (i.e., constructing the mapping between the business scenario requirement set and multiple sets of the private data) to obtain the response that occurs when the current enterprise calls the business rule at a future time. In this way, the first sample and the second sample of business rules and their responses at any time are obtained as a private dataset and written into the preset large model library to complete the deployment.

[0073] It should also be noted that the exhaustive sample consists of historical data within a preset time period for similar or related enterprise scenarios.

[0074] As a preferred approach, step S3, writing multiple sets of private datasets into various models in a pre-defined large model library, includes the following steps:

[0075] Parse the rule probability sequence in the private dataset and extract the probability distribution features of each business rule at different timestamps;

[0076] Based on the aforementioned probability distribution characteristics, the input embedding layer of the large model is dynamically weighted, wherein the weight values ​​are positively correlated with the probability values ​​of the corresponding timestamps in the regular probability sequence.

[0077] The dynamic weight adjustment ensures that the large model prioritizes responding to requests from high-probability business rules during inference.

[0078] As a preferred approach, in step S4, generating multiple sets of hardware containers includes the following steps:

[0079] Analyze the probability and frequency call sequences in the business scenario demand set to determine the peak time window for business term calls and resource-intensive business modules;

[0080] Based on the peak time window and resource-intensive business modules, the computing resources of the hardware dataset are dynamically allocated to generate the initial configuration parameters of the hardware container.

[0081] The hardware interface of the hardware container is configured to allocate low-latency I / O channels to the business modules corresponding to the high-frequency call sequence, and to reserve elastic computing resources for high-probability timestamps in the probability call sequence.

[0082] Further explanation of the hardware container in this embodiment is needed. In this embodiment, the hardware dataset is the hardware ID corresponding to each business module, and the hardware container corresponds to the hardware ID of each business module. It serves as a data receiving, sending, and relay storage module so that after the large model interacts directly, the data can be exchanged through the hardware container to the hardware device with the corresponding hardware ID.

[0083] As a preferred approach, in step S5, generating multiple simulated service requests includes:

[0084] Based on the frequency call sequence of business scenario requirements, a weight distribution model of business terms is constructed, where the weight value is determined by normalizing the frequency value of the timestamp unit of the call frequency sequence.

[0085] Based on the weight distribution model, real-time business data in the business dataset is sampled to generate simulated request clusters of high-weight business terms;

[0086] The simulated request clusters are injected into a pre-set large model library in the order of timestamps of the probability call sequence to simulate the temporal characteristics of real business scenarios.

[0087] As a preferred approach, step S6, determining the deployment scheme for the current business dataset, includes the following steps:

[0088] Calculate the conformity score between each simulated response and the business rule data. The conformity score is calculated based on the weighted sum of the rule probability sequences in the private dataset. The weights are set to the probability of the response unit corresponding to each business rule.

[0089] By combining the response time and compliance score of simulated business requests, deployment priorities are generated;

[0090] The deployment scheme selects the hardware container and its corresponding large model with a compliance score exceeding the threshold and the shortest response time as the target deployment entity.

[0091] A second aspect of this embodiment provides an AI large-scale model deployment system for enterprise scenarios, such as... Figure 2 As shown, it includes:

[0092] The data acquisition module is configured to acquire business datasets for the target enterprise scenario;

[0093] The data processing module is configured to determine the set of business scenario requirements based on the business dataset;

[0094] The server is configured to pre-store a large model library and a private dataset, and is also configured to generate simulation requests and receive simulation responses.

[0095] The foregoing description and accompanying drawings fully illustrate embodiments of this disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operations may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Moreover, the terminology used in this application is for describing embodiments only and is not intended to limit the claims. As used in the description of embodiments and claims, the singular forms “a,” “an,” and “the” are intended to equally include the plural forms unless the context clearly indicates otherwise. Similarly, the term “and / or” as used in this application means including one or more of the associated listed items and all possible combinations thereof. Additionally, when used in this application, the term "comprise" and its variations "comprises" and / or "comprising" refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Without further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element. In this document, each embodiment may focus on the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, the relevant parts can be referred to the description of the method section.

[0096] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented using electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods for each specific application to achieve the described functions, but such implementation should not be considered beyond the scope of the embodiments of this disclosure. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the described devices, apparatuses, and units can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0097] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, function, and operation of possible implementations of the apparatus, method, and computer program product according to embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in a different order than those disclosed in the description; sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. Each block in a block diagram and / or flowchart, and combinations of blocks in a block diagram and / or flowchart, can be implemented using a dedicated hardware-based device that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

Claims

1. An AI large model deployment method for an enterprise scenario, characterized in that, Comprising the following steps: S1, obtaining a business data set of a target enterprise scene, preprocessing the business data set to generate a business scene requirement set; The business data set comprises: According to the business module of the target enterprise scene, collect the historical business data, real-time business data and their corresponding business rule data of each business module; The business scene requirement set is obtained by the following steps: Determine the calling probability and calling frequency of the business vocabulary in the historical business data and real-time business data under the current business rule data in the business data set; Get the probability calling sequence and frequency calling sequence of the first business rule about time; The probability calling sequence sorts each unit according to the probability value with timestamp as the sequence axis; The frequency calling sequence sorts each unit according to the frequency value with timestamp as the sequence axis; Merge the probability calling sequence and the frequency calling sequence to get the business scene requirement set; S2, obtaining a plurality of private data sets according to the business scene requirement set, and constructing a mapping between the business scene requirement set and the plurality of private data sets; The private data set is obtained by the following steps: Obtain the probability calling sequence and the frequency calling sequence; Obtain the valid timestamp of each business rule in the business data set; Determine a plurality of future time stamps after the current valid time stamp, exhaust a plurality of samples, determine the probability of each business rule occurring at different time stamps, and obtain a rule probability sequence; Determine the first sample and the second sample of each rule in the rule probability sequence to obtain business data from the probability calling sequence and the frequency calling sequence; Merge the first sample and the second sample to obtain the private data set S3, write the plurality of private data sets into each large model in the preset large model library; S4, obtaining a hardware data set of a target enterprise scene, generating a plurality of hardware containers according to the hardware data set, and each hardware container is provided with a hardware interface corresponding to the large model; S5, according to the business data set, generate a plurality of simulated business requests to the preset large model library, and receive the simulated response of each large model; S6, according to the simulated business request and the simulated response, determine the deployment scheme of the current business data set.

2. The enterprise-scene-oriented AI large model deployment method according to claim 1, characterized in that, In S3, the plurality of private data sets are written into each large model in the preset large model library, comprising the following steps: Analyze the rule probability sequence in the private data set, extract the probability distribution characteristics of each business rule at different time stamps; Based on the probability distribution characteristics, dynamically adjust the input embedding layer of the large model, wherein the weight value is positively correlated with the probability value of the corresponding time stamp in the rule probability sequence; The dynamic weight adjustment ensures that the large model responds to the request of high probability business rule preferentially during inference.

3. The enterprise-scene-oriented AI large model deployment method according to claim 2, characterized in that, In S4, generating a plurality of hardware containers comprises the following steps: Analyze the probability calling sequence and the frequency calling sequence in the business scene requirement set, determine the peak time window of business vocabulary calling and the resource intensive business module; According to the peak time window and the resource intensive business module, dynamically divide the computing resources of the hardware data set to generate the initial configuration parameters of the hardware container; The hardware interface of the hardware container is configured to assign a low-latency I / O channel to a service module corresponding to a high-frequency call sequence, and reserve elastic computing resources for a high-probability timestamp in a probabilistic call sequence.

4. The enterprise-scene-oriented AI large model deployment method according to claim 3, characterized in that, In the S5, the generating of the multiple simulation service requests comprises: Based on the frequency call sequence in the service scenario demand set, a weight distribution model of the service vocabulary is constructed, wherein the weight value is determined by the frequency value of the timestamp unit of the call frequency sequence. According to the weight distribution model, the real-time service data in the service data set is sampled to generate a simulation request cluster of high-weight service vocabulary; The simulation request cluster is injected into a preset large model library according to the timestamp order of the probabilistic call sequence to simulate the timing characteristics of the real service scenario.

5. The enterprise-scene-oriented AI large model deployment method according to claim 4, characterized in that, In the S6, the determining of the deployment scheme of the current service data set comprises the following steps: Calculate the compliance score of each simulation response and service rule data, wherein the compliance score is calculated based on the rule probability sequence in the private data set, and the weight size is set as the probability size of the response unit corresponding to the rule; Generate a deployment priority by combining the response time and the compliance score of the simulation service request; The deployment scheme selects the hardware container and the corresponding large model with the compliance score exceeding the threshold value and the minimum response time as the target deployment entity.

6. An AI large model deployment system for enterprise scenarios, adapted to perform the method of any one of claims 1-5, characterized in that, Comprise: A data acquisition module configured to acquire a service data set of a target enterprise scenario; A data processing module configured to determine a service scenario demand set according to the service data set; The server is configured to pre-store a large model library and a private data set, and is configured to execute the generation of simulation requests and receive simulation responses.

Images