Powerful model performance optimization method and system based on declarative self-prompting

Abstract

The application discloses a power large model performance optimization method and system based on a declarative self-prompting, comprising the following steps: mathematical modeling of a large model performance optimization task; constructing a task optimizer based on the large model performance optimization task; using a bootstrap random optimization method, a module-level prompt optimization method and a multi-prompt instruction optimization method in sequence on a training set to perform parameter optimization through the task optimizer, to obtain optimal parameters; and using the optimal parameters to complete power large model performance optimization. The method automatically generates an optimized LLM calling scheme and corresponding prompts by using a programming idea, enables the system to automatically learn and adjust the prompts without human intervention, so as to adapt to different tasks and context environments, widens the application range of the LLM, and also enhances the robustness and generalization ability of the LLM.

Description

TECHNICAL FIELD

[0001] The present application belongs to the cross field of artificial intelligence and power system, and particularly relates to a power large model performance optimization method and system based on declarative self-prompting. BACKGROUND

[0002] The rise of large language models (LLMs) marks a major advance in the field of natural language processing (NLP). These models, due to their massive parameter size and powerful pre-training capabilities, allow researchers to work at a higher level of abstraction when developing NLP applications, reducing the reliance on large amounts of labeled data. In practice, this means that developers can build complex multi-stage pipelines, where each stage can be completed by calling an LLM for a specific task, simplifying the overall design of the system.

[0003] To enable LLMs to efficiently perform these tasks, researchers need to carefully design "prompt words". An effective prompt is like providing LLM with a clear set of instructions on how to generate the desired output based on the input. For example, in a translation task, a good prompt can help the model understand the relationship between the source language and the target language; while in a code writing task, it can guide the model to generate program code that meets the syntax and logic requirements.

[0004] However, current power large model prompts often rely on static, manually set prompt templates, which have certain limitations. First, fixed prompts are difficult to adapt to changes in task requirements, as they are created based on best practices in specific scenarios. Second, if the prompt is not clear or misleading, it may lead to inaccurate or even completely wrong results for the power LLM. In addition, for new or different power LLMs, existing prompts may not achieve the expected effect, which limits the versatility and flexibility of the model. SUMMARY

[0005] To address the shortcomings of the prior art, one of the purposes of the present application is to provide a power large model performance optimization method based on declarative self-prompting, which automatically generates optimized LLM calling schemes and corresponding prompts using programming ideas, allowing the system to automatically learn and adjust prompts without human intervention, to adapt to different tasks and context environments, widening the application range of LLMs, and enhancing their robustness and generalization ability.

[0006] The second purpose of the present application is to provide a system for implementing the power large model performance optimization method based on declarative self-prompting.

[0007] The application provides a power large model performance optimization method based on declarative self-prompting, comprising the following steps:

[0008] S1. Mathematically model the large model performance optimization task;

[0009] S2. Based on the large model performance optimization task, construct a task optimizer;

[0010] S3. Based on the task optimizer obtained in step S2, sequentially use bootstrap random optimization method, module-level prompt optimization method and multi-prompt instruction optimization method on the training set to optimize the parameters, and obtain the optimal parameters;

[0011] S4. Complete the power large model performance optimization by using the optimal parameters obtained in step S3.

[0012] Step S1 comprises the following steps:

[0013] Define the composition and running process of the power large model;

[0014] Based on the data definition of the power large model, the large model performance optimization task is mathematically modeled.

[0015] Step S1 is specifically:

[0016] Define the power large model program as , which is composed of m sub-modules; each sub-module i is defined by a prompt word template protecting a group of variables v; define V as the variable set used in the prompt word template in the large model program ; use to represent the running of the program under the assignment condition of assigning the variable set V to the string S; each prompt word template p includes 1 instruction variable and K example variables, and other variables are set as constants;

[0017] Based on the definition of the power large model, the goal of the large model performance optimization task is to find a global assignment in the training set D composed of the input data set X and the metadata set to optimize the performance of the power large model program on the index ;

[0018] The goal of the large model performance optimization task is represented by the objective function :

[0019]

[0020] wherein, is the maximum function value under the assignment condition of assigning the variable set V to the string S; Let x be the modulus of the training set D; let x be an input data point in the input dataset X. Meta dataset One of the metadata in the document.

[0021] In step S2, the task optimizer includes an example bootstrapping module and a credit allocation module;

[0022] The example bootstrapping module uses a rejection sampling algorithm to generate an example sample set;

[0023] The credit allocation module determines the contribution of specific hyperparameters to the power large model program based on the example sample set, realizes credit allocation for different combinations of hyperparameters, and completes the performance optimization of the power large model.

[0024] The example bootstrap module operates as follows:

[0025] From power-related training sets The sampled input x is input into the power large model program and run to generate the input-output trajectory of each sub-module in the power large model program. All values ​​on the trajectory are stored as example samples.

[0026] Traversing the training set If all inputs or sampling times reach a preset threshold, a valid sample set is obtained.

[0027] The operation of the credit allocation module specifically includes the following steps:

[0028] Define the credit allocation target EI;

[0029] Kernel density estimation is used to model the probability distribution of the objective function value for a given hyperparameter;

[0030] Select the hyperparameter combination that maximizes the credit allocation objective as the next evaluation point, perform iterative calculations, and obtain the optimal hyperparameter combination.

[0031] The credit allocation module operates as follows:

[0032] The credit allocation target EI is defined using the following formula:

[0033]

[0034] in, The observation threshold; It is the objective function value, i.e., the index. The score; Given hyperparameters Lower objective function value The conditional probability distribution; To be in a given hyperparameter a lower credit allocation target value;

[0035] a training set is divided into two groups in a preset ratio ; wherein the training set is sorted from large to small according to the target function value, the first combination of hyperparameters, the remaining combination of hyperparameters; the kernel density estimation is used to model and probability, which is expressed by the following formula:

[0036]

[0037]

[0038] wherein, is the kernel density estimation of ; is the kernel density estimation of ; is the proportion of and in the training set D; is the probability distribution function of the target function value;

[0039] The EI value is calculated explicitly by the following formula:

[0040]

[0041] wherein, is the mean of ;

[0042] The combination of hyperparameters x that maximizes is selected as the next evaluation point, and the iteration is calculated to finally obtain the optimal combination of hyperparameters.

[0043] Step S3 is specifically:

[0044] The bootstrap random optimization method is adopted, and the example bootstrap module in the task optimizer is used to sample the instruction input x from the power-related training set , input into the power large model program, run in the power large model program, and generate the input-output trajectory of each sub-module in the power large model program. All values on the trajectory are stored as example samples; the all values include the parameters and final score results of each sub-module running process;

[0045] The training set all inputs or sampling times in the reach a preset threshold, obtaining an example sample set;

[0046] The module-level prompt optimization method is adopted, a Qwen-Plus large model is introduced as a judge, for the input of each sub-module stored in the example sample set and the corresponding score, the judge is used to generate a new instruction input, and the new instruction input is input into the corresponding sub-module to run and obtain a result score, the obtained new instruction and the corresponding score are also stored in the example sample set, and are classified and stored, to obtain an instruction set of each sub-module;

[0047] The multi-prompt instruction optimization method is adopted, according to the instruction set of each sub-module, for the parameter instruction and example of each sub-module, a credit distribution component is used for calculation, and finally the parameter with the highest probability is obtained as the optimal parameter.

[0048] The module-level prompt optimization method comprises the following steps:

[0049] An example sample set obtained by the bootstrap random optimization method is obtained; the sample set comprises a plurality of instruction inputs and scores of each sub-module after the large model runs the instruction input;

[0050] A Qwen-Plus large model is set as a judge, and the instruction input and the score of each sub-module in the example sample set are input into the judge, and the judge generates a new instruction input for each sub-module according to the input;

[0051] The instruction input generated by the judge is input into the corresponding sub-module of the large model program, and the instruction input and the corresponding score are stored according to different sub-modules, to obtain an instruction set of each sub-module;

[0052] The above steps are repeated based on the instruction set of each sub-module, new instruction inputs and corresponding scores are continuously generated and stored and updated in the instruction set of each sub-module, until the number of iterations reaches a preset maximum number of iterations.

[0053] The multi-prompt instruction optimization method comprises the following steps:

[0054] According to the instruction set of each sub-module obtained by the module-level prompt optimization method, each module comprises T instructions, and the example bootstrap module generates K task examples for each sub-module;

[0055] For the instruction and example of each sub-module, a credit distribution module is used for iteration, and after final convergence or a preset maximum number of iterations is reached, the parameter instruction with the highest probability is selected as the optimal parameter.

[0056] The application also provides a system for realizing the power large model performance optimization method based on the declarative self-prompting, comprising a mathematical modeling module, a task optimizer construction module, an optimal parameter acquisition module and a power large model optimization module.

[0057] The mathematical modeling module performs mathematical modeling on the large model performance optimization task and uploads data to the task optimizer construction module.

[0058] The task optimizer construction module constructs a task optimizer based on the received data and the large model performance optimization task, and uploads data to the optimal parameter acquisition module.

[0059] The optimal parameter acquisition module obtains optimal parameters by using the task optimizer to sequentially perform parameter optimization on the training set using the bootstrap random optimization method, the module-level prompt optimization method and the multi-prompt instruction optimization method, and uploads data to the power large model optimization module.

[0060] The power large model optimization module completes power large model performance optimization using the optimal parameters.

[0061] The application discloses a power large model performance optimization method and system based on declarative self-prompting, which automatically generates optimized LLM calling schemes and corresponding prompts using programming ideas, automatically learns and adjusts prompts without human intervention to adapt to different tasks and context environments, improves the application range of LLM, and enhances the robustness and generalization ability of LLM. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 It is a flowchart of the method of the application;

[0063] Figure 2 It is a structural schematic diagram of the system of the application;

[0064] Figure 3 It is a flowchart of an example bootstrap module in the task optimizer. DETAILED DESCRIPTION

[0065] The application provides a power large model performance optimization method based on declarative self-prompting, and a flowchart thereof is as shown in Figure 1 The application provides a power large model performance optimization method based on declarative self-prompting, and a flowchart thereof is as shown in

[0066] S1. Mathematical modeling is performed on a large model performance optimization task.

[0067] Step S1 comprises the following steps:

[0068] The composition and running process of the power large model are defined.

[0069] Based on the data definition of the power large model, a mathematical model is established for the large model performance optimization task.

[0070] Step S1 is specifically:

[0071] The power large model program is defined as , which is composed of m sub-modules; each sub-module i is defined by a prompt template protecting a group of variables v; V is defined as the power large model program , which is the union of all variables used in the prompt templates in the power large model program; the program is run under the assignment condition that the variables in V are assigned to the string S; each prompt template p includes 1 instruction variable and K example variables, and other variables are set to constants.

[0072] Based on the power large model definition, the goal of the large model performance optimization task is to find a global assignment that can optimize the performance of the power large model program on the index under the training set D composed of the input data set X and the metadata set .

[0073] The goal of the large model performance optimization task is represented by the objective function :

[0074]

[0075] where is the maximum function value under the assignment condition that the variables in V are assigned to the string S; is the length of the training set D; x is an input data in the input data set X; is a metadata in the metadata set .

[0076] S2. Based on the large model performance optimization task, a task optimizer is constructed;

[0077] In step S2, the task optimizer includes an example bootstrap module and a credit allocation module.

[0078] The example bootstrap module uses the rejection sampling algorithm to generate an example sample set, and its running flowchart is shown in Figure 3 .

[0079] The credit allocation module determines the contribution of a specific hyperparameter to the power large model program based on the example sample set, realizes credit allocation for different hyperparameter combinations, and completes the performance optimization of the power large model.

[0080] The running of the example bootstrap module is specifically:

[0081] ​Sample input x from the power-related training set Run the input power large model program, generate the input-output trajectory of each sub-module in the power large model program, and store all values on the trajectory as example samples;

[0082] Traverse all inputs in the training set or the number of sampling times reaches the preset threshold to obtain an effective example sample set.

[0083] The operation of the credit allocation module specifically includes the following steps:

[0084] Define the credit allocation target EI;

[0085] Model the probability distribution of the objective function value under a given hyperparameter using kernel density estimation;

[0086] Select the hyperparameter combination that currently maximizes the credit allocation target as the next evaluation point, perform iterative calculations, and obtain the optimal hyperparameter combination.

[0087] The operation of the credit allocation module is specifically:

[0088] Define the credit allocation target EI, which is represented by the following formula:

[0089]

[0090] Wherein, is the observation threshold; is the objective function value, i.e., the score of the indicator ; is the conditional probability distribution of the objective function value under a given hyperparameter ; is the credit allocation target value under a given hyperparameter ;

[0091] Divide the training set into two groups and at a preset ratio; wherein, the training set is sorted from large to small according to the objective function value, is the hyperparameter combination of the first , and is the remaining hyperparameter combination; model the probabilities of and using kernel density estimation, which is represented by the following formula:

[0092]

[0093] ​

[0094] wherein, is a kernel density estimation; is a kernel density estimation; is and the proportion of the training set D; is a probability distribution function of the objective function value;

[0095] The EI value is explicitly calculated using the following formula:

[0096]

[0097] wherein, is the mean value;

[0098] The hyperparameter combination x that maximizes the value of is selected as the next evaluation point, and iterative calculation is performed to finally obtain the optimal hyperparameter combination. S3. Based on the task optimizer obtained through step S2, sequentially use the bootstrap random optimization method, the module-level prompt optimization method and the multi-prompt instruction optimization method on the training set to perform parameter optimization, and obtain the optimal parameters;

[0099] Step S3 is specifically:

[0100] The bootstrap random optimization method is adopted, and the example bootstrap module in the task optimizer is used to sample instruction input x from the power-related training set

[0101] , input into the power large model program for running, and generate the input-output trajectory of each sub-module in the power large model program. All values on the trajectory are stored as example samples. The all values include the parameters and final score results of each sub-module running process; All inputs in the training set

[0102] or the sampling times reach the preset threshold, and the example sample set is obtained; The module-level prompt optimization method is adopted, and the Qwen-Plus large model is introduced as a judge. For the input of each sub-module stored in the example sample set and the corresponding score, the judge generates a new instruction input, inputs it into the corresponding sub-module for running and obtains a result score, and stores the obtained new instruction and the corresponding score in the example sample set and classifies and stores them to obtain the instruction set of each sub-module;

[0103]

[0104] ​The multi-prompt instruction optimization method is used, the credit distribution component is used to calculate the parameter instruction and the example of each sub-module according to the instruction set of each sub-module, and finally the parameter with the highest probability is obtained as the optimal parameter.

[0105] The module-level prompt optimization method comprises the following steps:

[0106] Obtain the example sample set obtained by the bootstrap random optimization method, the sample set comprising a plurality of instruction inputs and scores of each sub-module after the instruction input large model is run;

[0107] The Qwen-Plus large model is set as a judge, the instruction input and the score of each sub-module in the example sample set are input into the judge, and the judge generates a new instruction input for each sub-module according to the input;

[0108] The instruction input generated by the judge is input into the corresponding sub-module of the large model program, and the instruction input and the corresponding score are stored according to different sub-modules, so that the instruction set of each sub-module is obtained;

[0109] The above steps are repeated based on the instruction set of each sub-module, new instruction inputs and corresponding scores are generated and stored and updated in the instruction set of each sub-module until the number of iterations reaches the preset maximum number of iterations.

[0110] The multi-prompt instruction optimization method comprises the following steps:

[0111] According to the instruction set of each sub-module obtained by the module-level prompt optimization method, each module comprises T instructions, and K task examples are generated for each sub-module using the example bootstrap module;

[0112] For the instruction and the example of each sub-module, the credit distribution module is used to calculate iterations, and finally the parameter instruction with the highest probability is selected as the optimal parameter after convergence or the preset maximum number of iterations is reached.

[0113] S4. The optimal parameter obtained in step S3 is used to complete the performance optimization of the power large model.

[0114] The application also provides a system for realizing the power large model performance optimization method based on the declarative self-prompting, a structure diagram of which is shown in the figure and comprises a mathematical modeling module, a task optimizer construction module, an optimal parameter acquisition module and a power large model optimization module. Figure 2

[0115] The mathematical modeling module performs mathematical modeling on the large model performance optimization task and uploads data to the task optimizer construction module.

[0116] ​The task optimizer construction module constructs a task optimizer based on the received data and the performance of the large model, and uploads the data to the optimal parameter acquisition module;

[0117] The optimal parameter acquisition module obtains optimal parameters by using the task optimizer to sequentially use the bootstrap random optimization method, the module-level prompt optimization method, and the multi-prompt instruction optimization method on the training set based on the received data, and uploads the data to the power large model optimization module;

[0118] The power large model optimization module completes the performance optimization of the power large model using the optimal parameters based on the received data.

[0119] The following further describes the method of the application in combination with an embodiment:

[0120] Taking a fault diagnosis task for power relay protection as an example, a large model diagnosis program is constructed to solve the target task, which includes a retrieval module, a thinking chain module, and a generation module. The prompt words of each module will contain three variables, namely, instructions, examples, and target questions, wherein the target question is an optional item. The training set D uses the relay protection diagnosis knowledge base constructed by the application.

[0121] There are three evaluation indicators for this task. The indicator for the retrieval module is knowledge coverage, that is, the coverage of the returned text content for the query. The indicator for the thinking chain module is logic, that is, whether the thinking process is reasonable. The indicator for the generation module is accuracy, that is, whether the generated content accurately answers the question. The task optimizer is constructed based on these evaluation indicators.

[0122] First, the bootstrap random optimization method is used to sample an input instruction “CT wire breakage alarm signal appears” from the relay protection diagnosis knowledge base training set. The instruction enters the large model diagnosis program and records the parameters and final score results of the retrieval module, the thinking chain module, and the generation module during operation, which will be stored as a sample. Repeat the sampling from the training set N times to obtain N groups of samples.

[0123] Then, the Qwen-Plus large model is introduced as a judge. For the input instruction “CT wire breakage alarm signal appears” of the retrieval module and the corresponding retrieval result score, the judge generates new instructions: “The alarm signal indicates that there is a CT wire breakage”, “The system detects CT wire breakage and issues an alarm signal”, “What should be done about the alarm signal indicating CT wire breakage”, and similarly generates new instructions for the thinking chain module and the generation module and obtains the result score.

[0124] Finally, the multi-prompt instruction optimization method is used. According to the instruction set of each sub-module, a plurality of examples are generated by using the example bootstrap component in the optimizer. Taking the thought chain module as an example, two examples are shown in Table 1. For the parameter instructions and examples of each sub-module, the credit allocation component is used for calculation, and finally the parameter with the highest probability is obtained as the optimal parameter.

[0125] Table 1 Two examples generated by the example bootstrap component

[0126]

Claims

1. The present application provides a power large model performance optimization method based on declarative self-prompting, characterized in that, The method comprises the following steps: S1. Mathematical modeling of large model performance optimization tasks; S2. Based on the large model performance optimization task, a task optimizer is constructed; S3. Based on the training set, the task optimizer obtained in step S2 is used to sequentially use bootstrap random optimization method, module-level prompt optimization method and multi-prompt instruction optimization method for parameter optimization, and the optimal parameters are obtained; S4. Using the optimal parameters obtained in step S3, the performance of the power large model is optimized; In step S2, the task optimizer includes an example bootstrap module and a credit allocation module; The example bootstrap module uses the rejection sampling algorithm to produce an example sample set; The credit allocation module determines the contribution of a given hyperparameter to the power large model program based on the example sample set, allocates credit to different hyperparameter combinations, and optimizes the performance of the power large model; Step S1 is specifically: Definition of the power large model program is composed of m sub-modules; each sub-module i is defined by a prompt template protecting a set of variables v; define V as the large model program of all variables used in the prompt template, using to represent the program running under the assignment condition of assigning the variable set V to the string S; each prompt template p includes 1 instruction variable and K example variables, and other variables are set to constants; Based on the power large model definition, the target of the large model performance optimization task is to find a global assignment in the training set D composed of the input data set X and the metadata set , which can optimize the performance of the power large model program on the indicators . The target of the large model performance optimization task uses a target function Indicates: wherein, is the maximized function value under the assignment condition that the variable set V is assigned the string S; is the modulus of the training set D; x is one input data in the input data set X; is one metadata in the metadata set ; In step S3, the training set is a relay protection diagnosis knowledge base; Using the bootstrap random optimization method, the example bootstrap module in the task optimizer is used to sample instruction input text x from the power-related training set, input it into the power large model program for running, and generate the input-output trajectory of each sub-module in the power large model program. Store all values on the trajectory as example samples; The values include the parameters and final score results of each sub-module running process.

2. The declarative self-prompt based power macro model performance optimization method according to claim 1, wherein, The operation of the example bootstrap module is specifically: Sample input x from the power-related training set Run in the power large model program, and generate the input-output trajectory of each submodule in the power large model program. Store all values on the trajectory as example samples; traversing the training set until a preset threshold of all inputs or sampling times in the training set is reached, obtaining an effective example sample set.

3. The declarative self-prompt based power macro model performance optimization method according to claim 1, wherein, The operation of the credit allocation module is specifically: Define the credit allocation target EI, which is represented by the following formula: wherein, is an observation threshold; is a target function value; is a conditional probability distribution of a target function value under given hyperparameters ; and is a credit allocation target value under given hyperparameters ; and training set Divided into two groups according to a preset ratio and Among them, the training set Sort by objective function value from largest to smallest. For the front Hyperparameter combinations, For the remainder Hyperparameter combinations; kernel density estimation used for and Probability is modeled using the following formula: wherein, is a kernel density estimate; is a kernel density estimate; is a kernel density estimate; is a kernel density estimate; is a kernel density estimate; and the proportion of the training set D; is a probability distribution function of the objective function value; The EI value is calculated explicitly using the following formula: wherein is the mean of Choose to The largest hyperparameter combination x is used as the next evaluation point for iterative calculation, and finally the optimal hyperparameter combination is obtained.

4. The declarative self-prompt based power macro model performance optimization method of claim 1, wherein, Step S3 is specifically: Using a bootstrap random optimization method, using an example bootstrap module in the task optimizer, sampling instruction input x from the power-related training set In the power large model program, the input-output trajectory of each sub-module is generated by running the input power large model program, and all values on the trajectory are stored as example samples; the all values include the parameters and final score results of each sub-module running process; traversing the training set until a preset threshold of all inputs or sampling times in the training set is reached, obtaining an example sample set; Using the module-level prompt optimization method, introduce the Qwen-Plus large model as the judge, for each sub-module input and the corresponding score stored in the example sample set, use the judge to generate new instruction input, and input it into the corresponding sub-module for running and get the result score, the new instruction and the corresponding score are also stored in the example sample set, and are classified and stored, to get the instruction set of each sub-module; Using the multi-prompt instruction optimization method, according to the instruction set of each sub-module, for each sub-module parameter instruction and example, using the credit allocation component to calculate, finally the parameter with the highest probability is the optimal parameter.

5. The declarative self-prompt based power macro model performance optimization method according to claim 4, characterized in that, The module-level prompt optimization method comprises the following steps: Obtain the example sample set obtained by the bootstrap random optimization method; The sample set includes a plurality of instruction inputs and the scores of each sub-module after the large model runs the instruction input; Set the Qwen-Plus large model as the judge, input the instruction input and score of each sub-module in the example sample set into the judge, and the judge generates a new instruction input for each sub-module according to the input; Input the instruction input generated by the judge into the corresponding sub-module of the large model program, and store the instruction input and the corresponding score according to different sub-modules, to get the instruction set of each sub-module; Repeat the above steps based on the instruction set of each sub-module, constantly generate new instruction input and corresponding score and store and update to the instruction set of each sub-module, until the number of iterations reaches the preset maximum number of iterations.

6. The declarative self-prompt based power macro model performance optimization method of claim 4, wherein, The multi-prompt instruction optimization method comprises the following steps: According to the instruction set of each sub-module obtained by the module-level prompt optimization method, each module contains T instructions, and K task examples are generated for each sub-module using the example bootstrap module; For the instructions and examples of each sub-module, the credit allocation module is used to calculate iterations, and after final convergence or reaching a preset maximum number of iterations, the parameter instruction with the highest probability is selected as the optimal parameter.

7. A system for implementing the method for power large model performance optimization based on declarative self-prompting according to any one of claims 1-6, characterized in that, It includes a mathematical modeling module, a task optimizer construction module, an optimal parameter acquisition module, and a power large model optimization module. The mathematical modeling module performs mathematical modeling on the large model performance optimization task and uploads the data to the task optimizer construction module. The task optimizer construction module constructs a task optimizer based on the received data and the large model performance optimization task, and uploads the data to the optimal parameter acquisition module. The optimal parameter acquisition module uses the task optimizer to sequentially optimize the parameters using the bootstrap random optimization method, the module-level prompt optimization method, and the multi-prompt instruction optimization method on the training set, obtains the optimal parameters, and uploads the data to the power large model optimization module. The power large model optimization module uses the optimal parameters to complete the performance optimization of the power large model according to the received data.

Images