Large model processing method and device, equipment, storage medium and program product

Abstract

The invention provides a large model processing method and device, equipment, a storage medium and a program product, and relates to the field of artificial intelligence. The method comprises the steps of obtaining a basic large model of a to-be-processed large model; acquiring sample data for training the basic large model; according to the sample data, training the basic large model to obtain corresponding training data; according to the training data and the basic large model, generating a target differential small model corresponding to the to-be-processed large model; wherein the target difference small model is used for representing the model version difference of the to-be-processed large model in the training process. The problems that in the prior art, multiple model versions need to be stored in model training, and a large number of storage resources are consumed are solved; the problem that the processing efficiency of a large model is low due to the fact that model deployment of multiple application scenes consumes a large amount of computing power resources is solved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a method, apparatus, device, storage medium, and program product for processing large models. Background Technology

[0002] Currently, the use of large models is basically based on incremental training or fine-tuning of the basic large model to generate new models to provide model services for application scenarios.

[0003] Because large models have enormous parameters and feature sets, multiple versions of the model obtained during training are stored to obtain the best-performing model. Furthermore, deploying a model for each application scenario requires the computing power of one node. Each computing resource can only be used to deploy the model on an idle (unused) node.

[0004] However, existing large model processing techniques consume a lot of storage and computing resources, resulting in low efficiency in processing large models. Summary of the Invention

[0005] This application provides a method, apparatus, device, storage medium, and program product for processing large models, in order to solve the problem that the processing of existing large models requires a large amount of storage and computing resources, resulting in low efficiency in the processing of large models.

[0006] Firstly, this application provides a method for processing large models, including:

[0007] Obtain the base model of the large model to be processed;

[0008] Obtain sample data for training the base large model;

[0009] Based on the sample data, the basic large model is trained to obtain the corresponding training data;

[0010] Based on the training data and the base large model, a target difference small model corresponding to the large model to be processed is generated; the target difference small model is used to characterize the model version differences of the large model to be processed during the training process.

[0011] In one possible design, the sample data includes the overall sample data, which includes the overall training samples and the corresponding overall label results.

[0012] In one possible design, a basic large model is trained based on sample data to obtain corresponding training data, including:

[0013] The entire training sample is input into the basic large model to obtain the first inference result output by the basic large model.

[0014] In one possible design, based on the training data and the base large model, a target difference small model corresponding to the large model to be processed is generated, including:

[0015] Obtain the initial difference small model;

[0016] Based on the first inference result and the overall sample data, second sample data is generated for training the initial differential small model;

[0017] Based on the second sample data, the initial difference small model is trained to obtain the target difference small model.

[0018] In one possible design, after generating the target difference small model corresponding to the large model to be processed based on the training data and the base large model, the following steps are also included:

[0019] Obtain the sample to be predicted;

[0020] Input the sample to be predicted into the basic large model to obtain the basic prediction results;

[0021] Based on the basic prediction results and the sample to be predicted, determine the first differential prediction sample;

[0022] The first difference prediction sample is input into the target difference small model to obtain the target prediction result corresponding to the sample to be predicted.

[0023] In one possible design, the sample data includes application-specific sample data.

[0024] In one possible design, a basic large model is trained based on sample data to obtain corresponding training data, including:

[0025] Based on application scenario-specific sample data, the basic large model is trained to obtain the target parameters and target features of the corresponding target large model;

[0026] Obtain the basic parameters and features of the basic large model;

[0027] Based on the target parameters, target features, basic parameters, and basic features, a target difference small model is generated.

[0028] In one possible design, after generating the target difference small model corresponding to the large model to be processed based on the training data and the base large model, the following steps are also included:

[0029] Obtain the sample to be predicted;

[0030] Input the sample to be predicted into the basic large model to obtain the basic prediction results;

[0031] The sample to be predicted is input into the target difference small model to obtain the difference prediction result;

[0032] Based on the basic prediction results and the differential prediction results, the target prediction results corresponding to the samples to be predicted are generated.

[0033] Secondly, this application provides a large-scale model processing apparatus, comprising:

[0034] The first acquisition module is used to acquire the basic large model of the large model to be processed;

[0035] The second acquisition module is used to acquire sample data for training the base large model;

[0036] The first processing module is used to train the basic large model based on the sample data to obtain the corresponding training data.

[0037] The second processing module is used to generate a target difference small model corresponding to the large model to be processed based on the training data and the basic large model; wherein, the target difference small model is used to characterize the model version differences of the large model to be processed during the training process.

[0038] Thirdly, this application provides an electronic device, including: a processor, and a memory communicatively connected to the processor;

[0039] The memory stores instructions that the computer executes;

[0040] The processor executes computer-executable instructions stored in memory to implement a large-scale processing method as described in the first aspect of the invention.

[0041] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement a large-scale model processing method as described in the first aspect of the invention.

[0042] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements a large-scale model processing method according to the first aspect of the invention.

[0043] This application provides a method, apparatus, device, storage medium, and program product for processing large models, including: acquiring a base large model of the large model to be processed; acquiring sample data for training the base large model; training the base large model based on the sample data to obtain corresponding training data; and generating a target difference small model corresponding to the large model to be processed based on the training data and the base large model. The target difference small model is used to characterize the model version differences of the large model to be processed during training. Compared to existing technologies, model training requires storing multiple model versions, consuming significant storage resources; and model deployment in multiple application scenarios consumes substantial computing resources, resulting in low processing efficiency for large models. This application combines the base large model and sample data to train a smaller difference model, preserving the differences in parameters and features between the target model and the base large model, saving computing and storage resources, thereby improving the processing efficiency of large models. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 A schematic diagram of the system architecture for a large model processing method provided in an embodiment of this application;

[0046] Figure 2 A flowchart illustrating a large model processing method provided in this application embodiment. Figure 1 ;

[0047] Figure 3 A flowchart illustrating a large model processing method provided in this application embodiment. Figure 2 ;

[0048] Figure 4 A flowchart illustrating a large model processing method provided in this application embodiment. Figure 3 ;

[0049] Figure 5 A flowchart illustrating a large model processing method provided in this application embodiment. Figure 4 ;

[0050] Figure 6 A flowchart illustrating a large model processing method provided in this application embodiment. Figure 5 ;

[0051] Figure 7A schematic diagram of a large model processing device provided in an embodiment of this application;

[0052] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application;

[0053] Figure 9 A flowchart illustrating the model training process provided in the embodiments of this application. Figure 1 ;

[0054] Figure 10 A schematic diagram of the model inference process provided in the embodiments of this application. Figure 1 ;

[0055] Figure 11 A flowchart illustrating the model training process provided in the embodiments of this application. Figure 2 ;

[0056] Figure 12 A schematic diagram of the model inference process provided in the embodiments of this application. Figure 2 . Detailed Implementation

[0057] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0058] In the embodiments of this application, the terms "first" and "second" are used to distinguish identical or similar items with substantially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, nor do they necessarily imply difference. It should be noted that in the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner. In the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more.

[0059] It should be noted that the phrase "at...time" in the embodiments of this application can refer to the instant at which a certain situation occurs, or to a period of time after the occurrence of a certain situation; the embodiments of this application do not specifically limit this. Furthermore, the large model processing method provided in the embodiments of this application is merely an example; large model processing methods may also include more or fewer elements.

[0060] It should also be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation access points for users to choose to authorize or refuse.

[0061] To facilitate a clear description of the technical solutions in the embodiments of this application, some terms and technologies involved in the embodiments of this application will be briefly introduced below:

[0062] Difference model: A model that retains only the differences between the target model and the base model. The core idea is to generate a new difference sequence by calculating the differences between adjacent data points. This process removes linear trends and seasonal fluctuations from the original data, transforming a non-stationary sequence into a stationary one, facilitating subsequent modeling and analysis.

[0063] Due to the enormous scale of parameters and features in large models, several significant problems arise during development and deployment:

[0064] On the one hand, during model training, multiple versions of the model obtained during the training process are saved in order to obtain the best-performing model. Since large models themselves consume a lot of resources, multiple versions of the model will consume a lot of storage resources, thus requiring a significant investment in the construction of storage resources.

[0065] On the one hand, because large models have extremely large parameters and feature scales, deploying a model for a single application scenario often requires the computing power resources of a single node, and the consumption of computing power resources will increase linearly with the continuous increase of business scenarios. Therefore, when empowering multiple application scenarios, it is necessary to invest heavily in building computing power resources.

[0066] On the other hand, it is necessary to reserve the computing resources of one node for each large model application. Deployment can only be flexibly performed on any node when all nodes in a cluster are idle. If some nodes in the cluster are already partially occupied, the model can only be deployed on other idle nodes (where computing resources are not being used). Therefore, application deployment is limited by the number of idle nodes in the cluster.

[0067] Based on this, embodiments of this application provide a method, apparatus, device, storage medium, and program product for processing large models, which can be used in the field of artificial intelligence and are intended to solve the above-mentioned technical problems of the prior art.

[0068] The inventive concept of this application lies in addressing the aforementioned problems. During research on the efficiency of processing large models, the inventors discovered that training large models currently consumes significant storage resources, while deploying large models across multiple application scenarios requires substantial computing power. Based on this, the inventors combined a base large model with sample data to train a smaller difference model, preserving the differences in parameters and features between the target model and the base large model, thereby saving computing and storage resources. Based on this, this application proposes a method for processing large models to further improve their processing efficiency.

[0069] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0070] Figure 1 This is a schematic diagram of the system architecture for a large model processing method provided in an embodiment of this application. Figure 1 In the above architecture, at least one of data acquisition device 101, processing device 102 and display device 103 is included.

[0071] It is understood that the structures illustrated in the embodiments of this application do not constitute a specific limitation on the architecture of a large-scale model processing system. In other feasible embodiments of this application, the above architecture may include more or fewer components than illustrated, or combine some components, or divide some components, or arrange different components, which can be determined according to the actual application scenario and is not limited here. Figure 1 The components shown can be implemented in hardware, software, or a combination of both.

[0072] In the specific implementation process, the data acquisition device 101 may include an input / output interface or a communication interface. The data acquisition device 101 can be connected to the processing device through the input / output interface or the communication interface to obtain data corresponding to multiple large models to be processed.

[0073] The processing device 102 can train the basic large model based on the data corresponding to the large model to be processed, obtain the corresponding training data, and generate the target difference small model corresponding to the large model to be processed based on the training data and the basic large model.

[0074] The display device 103 can also be a touch screen or the screen of a terminal device, used to store data while displaying the above-mentioned content, so as to realize interaction with the user.

[0075] It should be understood that the aforementioned processing device can be implemented by a processor reading instructions from memory and executing those instructions, or it can be implemented by a chip circuit.

[0076] Furthermore, the network architecture and business scenarios described in the embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided in the embodiments of this application. As those skilled in the art will know, with the evolution of network architecture and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0077] The technical solution of this application will be described in detail below with reference to specific embodiments:

[0078] Figure 2 A flowchart illustrating a large model processing method provided in this application embodiment. Figure 1 ,like Figure 2 As shown, the method includes:

[0079] S201. Obtain the basic large model of the large model to be processed.

[0080] S202. Obtain sample data for training the base large model.

[0081] The sample data includes the overall sample data, which includes the overall training samples and the corresponding overall label results.

[0082] S203. Based on the sample data, train the basic large model to obtain the corresponding training data.

[0083] Optionally, the entire training sample can be input into the base model to obtain the first inference result output by the base model.

[0084] Specifically, intermediate results are obtained by inferring from the sample data of the training samples (including sample data and sample label results) through the basic large model.

[0085] S204. Based on the training data and the basic large model, generate the target difference small model corresponding to the large model to be processed.

[0086] Among them, the target difference small model is used to characterize the model version differences of the large model to be processed during the training process.

[0087] It should be noted that the difference small model is smaller in size, so compared to storing the same number of large models, storing the difference small model consumes much less storage resources, which can save a lot of storage resources and reduce the construction cost of the large model training platform.

[0088] It should be noted that because the differential small model is smaller in scale, it uses less computing resources during deployment. When different application scenarios are trained based on the same basic large model, compared to the traditional large model deployment where only one application can be deployed per node, the differential small model deployment method allows multiple applications to be deployed on one node, improving the efficiency of computing resource utilization and significantly reducing the construction cost of computing resources.

[0089] It should also be noted that, due to the smaller scale of the differential small model, as long as a node deploys the basic large model and has sufficient remaining computing resources to deploy the differential small model, the corresponding application model can be deployed. If all nodes in a cluster have already deployed the basic large model, the differential small model can be flexibly deployed to any available node, making application deployment of the differential small model very flexible. When using Scheme 1 for deployment, it can even support cross-node deployment of the basic large model and the differential small model, further enhancing the flexibility of application deployment.

[0090] This embodiment provides a method for processing large models, including: obtaining a base large model of the large model to be processed; obtaining sample data for training the base large model; training the base large model based on the sample data to obtain corresponding training data; and generating a target difference small model corresponding to the large model to be processed based on the training data and the base large model. The target difference small model is used to characterize the model version differences of the large model to be processed during training. Compared to existing technologies, model training requires storing multiple model versions, consuming significant storage resources; and model deployment in multiple application scenarios consumes substantial computing resources, resulting in low processing efficiency for large models. This application combines the base large model and sample data to train a smaller difference model, preserving the differences in parameters and features between the target model and the base large model, saving computing and storage resources, thereby improving the processing efficiency of large models.

[0091] Figure 3 A flowchart illustrating a large model processing method provided in this application embodiment. Figure 2 ,like Figure 3 As shown, the specific implementation steps of S204 above include:

[0092] S301. Obtain the initial difference small model.

[0093] S302. Based on the first inference result and the overall sample data, generate the second sample data for training the initial differential small model.

[0094] S303. Based on the second sample data, train the initial difference small model to obtain the target difference small model.

[0095] First, the basic large model is used to infer the sample data of the training samples (including sample data and sample label results) to obtain intermediate results.

[0096] Secondly, the model training samples, sample label results, and intermediate results are used as training samples for the difference model.

[0097] Finally, for the input sample data, the difference between the sample label results and intermediate results is analyzed through training to obtain the difference small model.

[0098] In one possible embodiment, Figure 9 A flowchart illustrating the model training process provided in the embodiments of this application. Figure 1 ,like Figure 9 As shown, the process starts with training sample data. First, it performs inference through the basic large model to generate inference results. Then, it uses these inference results to train a differential small model, which aims to capture the differences between the target model and the basic large model. Finally, the trained differential small model is generated, which can reduce resource consumption while ensuring model performance and is optimized for specific tasks.

[0099] It's important to note that the differential small model is optimized for specific tasks. By capturing the differences between the target model and the base large model, it achieves the goal of reducing resource consumption while ensuring model performance. The design of the differential small model considers the allocation and deployment flexibility of computing resources across cluster nodes. This allows for flexible adjustments to the deployment method of the differential small model according to specific needs in practical applications, adapting to different application scenarios.

[0100] In this embodiment, a differential model is obtained based on the inference results of the basic large model and the actual results of the samples. This reduces training time and computational resource consumption, enhances the richness and diversity of sample data, and enables the differential small model to more accurately capture the differences between the target model and the basic large model, thereby improving the processing efficiency of the large model.

[0101] Figure 4 A flowchart illustrating a large model processing method provided in this application embodiment. Figure 3 ,like Figure 4 As shown, the above S204 includes the following:

[0102] S401. Obtain the sample to be predicted.

[0103] S402. Input the sample to be predicted into the basic large model to obtain the basic prediction results.

[0104] S403. Based on the basic prediction results and the sample to be predicted, determine the first differential prediction sample.

[0105] S404. Input the first difference prediction sample into the target difference small model to obtain the target prediction result corresponding to the sample to be predicted.

[0106] In one possible embodiment, Figure 10 A schematic diagram of the model inference process provided in the embodiments of this application. Figure 1 ,like Figure 10 As shown, the data to be predicted for multiple scenarios first undergoes inference through a basic large model to generate inference results for each scenario. Then, the basic inference results for each scenario are input into the corresponding differential small models for further inference. Finally, each differential small model outputs the prediction results for its specific scenario. The entire process achieves detailed inference and prediction for each scenario through the combination of the basic large model and the differential small model.

[0107] It's important to note that the base model captures general patterns and underlying features in the data, providing a robust predictive foundation for the target difference model. By determining the first difference prediction sample based on the base prediction results and the sample to be predicted, the prediction task can be further refined, allowing the target difference model to focus more on the parts where its predictions differ from those of the base model. The target difference model then fine-tunes these differences, generating more accurate target predictions.

[0108] In this embodiment, a base model performs initial predictions, and a differenced small model is used for further predictions. Simultaneously, the differenced small model can be optimized for specific scenarios or tasks, capturing the differences between the target model and the base model. This allows the model to adapt more flexibly to different scenarios and tasks, quickly adapting to new prediction requirements, improving the accuracy and refinement of prediction results, and thus increasing the processing efficiency of the large model.

[0109] Figure 5 A flowchart illustrating a large model processing method provided in this application embodiment. Figure 4 The sample data includes application-specific sample data, such as... Figure 5 As shown, the specific implementation steps of S203 above include:

[0110] S501. Based on the application scenario-specific sample data, train the basic large model to obtain the target parameters and target features of the corresponding target large model.

[0111] S502. Obtain the basic parameters and basic features of the basic large model.

[0112] S503. Generate a target difference small model based on the target parameters, target features, basic parameters, and basic features.

[0113] In this embodiment, sample data specific to the application scenario is used to perform incremental training or fine-tuning training based on the basic large model, resulting in a target large model. Then, by analyzing the differences between the features of the target large model and the features of the basic large model, a differential small model is constructed.

[0114] In one possible embodiment, Figure 11 A flowchart illustrating the model training process provided in the embodiments of this application. Figure 2 ,like Figure 11 As shown, the basic large model and training sample data are used as initial inputs, with the training sample data used for model training. During the model training phase, the basic large model is trained using the training sample data to generate the target large model. Subsequently, the differential small model is generated from the target large model. In the entire process, the data flow starts from the training sample data, goes through the model training phase, and finally forms the target large model and the differential small model, reflecting the complete training steps from the basic model to the target model and then to the differential small model.

[0115] For example, when generating a small difference model, to avoid the small feature model becoming too large, features can be extracted according to the following principles:

[0116] Optionally, assuming that the feature value of feature A of the target large model and the base large model differs by a, when a > δ (δ can be set as a hyperparameter), feature A is considered to be a feature of the difference small model; when a ≤ δ, feature A is considered not to be a feature of the difference small model.

[0117] Optionally, if the target large model has a feature B that is not in the base large model, then when the weight b of feature B is greater than ε (ε can be set as a hyperparameter), feature B is considered to be a feature of the difference small model; when b ≤ ε, feature B is considered not to be a feature of the difference small model.

[0118] It's important to note that by setting thresholds δ and ε, the number of features included in the small-difference model can be effectively controlled, thus preventing the model from becoming too large. This helps reduce the model's computational complexity and storage requirements. Selecting only features that significantly impact the model's prediction results (i.e., features with large feature value differences or high weights for new features) allows the small-difference model to focus more on key information, reducing unnecessary computation and improving the model's inference speed and efficiency.

[0119] In this embodiment, a differenced small model is constructed based on the differences between the parameters and features of the base large model and the parameters and features of the target model. By retaining features that show significant differences between the target large model and the base large model, the differenced small model can capture the key features of the target task, thereby maintaining or even improving the model's predictive performance. Simultaneously, the thresholds δ and ε can be adjusted as hyperparameters, allowing the differenced small model to flexibly adapt to different application scenarios and task requirements. By adjusting these thresholds, the model's size and performance can be balanced to meet specific application needs, thus improving the processing efficiency of the large model.

[0120] Figure 6 A flowchart illustrating a large model processing method provided in this application embodiment. Figure 5 The sample data includes application-specific sample data, such as... Figure 6 As shown, the above S204 includes the following:

[0121] S601. Obtain the sample to be predicted.

[0122] S602. Input the sample to be predicted into the basic large model to obtain the basic prediction results.

[0123] S603. Input the sample to be predicted into the target difference small model to obtain the difference prediction result.

[0124] S604. Based on the basic prediction results and the differential prediction results, generate the target prediction results corresponding to the samples to be predicted.

[0125] In one possible embodiment, Figure 12 A schematic diagram of the model inference process provided in the embodiments of this application. Figure 2 ,like Figure 12 As shown, the basic large model serves as the starting point of the entire process, responsible for preliminary data processing or feature extraction. Subsequently, the data is split into prediction data for different scenarios. The prediction data for each scenario is input into its corresponding differential small model. These differential small models perform independent inference based on the specific needs of their respective scenarios, and finally output the prediction results for each scenario. This process effectively realizes customized processing and prediction of data for different scenarios.

[0126] It's important to note that the base model captures general patterns and underlying features in the data, providing a robust predictive foundation for the target-difference model. The target-difference model, on the other hand, focuses on subtle differences specific to a particular scenario or task, fine-tuning the predictions of the base model. By combining the predictions of both, the generalization ability of the base model and the specificity of the target-difference model can be fully utilized, thereby improving the overall prediction accuracy.

[0127] It's important to note that when the base model makes initial predictions, it can quickly filter out a large amount of irrelevant information, providing more targeted input for subsequent smaller difference models. The target smaller difference model is typically more lightweight and computationally less complex than the base model, thus generating prediction results faster. This staged prediction approach helps reduce unnecessary computation and improves overall prediction efficiency.

[0128] In this embodiment, predictions are made based on a base large model and a target difference small model. The base large model provides an initial prediction, while the difference small model makes more refined adjustments based on this. The base large model can capture general patterns and underlying features in the data, providing a robust prediction foundation for the target difference small model; the target difference small model focuses on the subtle differences of a specific scenario or task, making fine adjustments to the prediction results of the base large model. By fully utilizing the generalization ability of the large model and the specificity of the small model, the overall prediction accuracy is improved, computational complexity is reduced, and model flexibility is enhanced, thereby improving the processing efficiency of the large model.

[0129] Figure 7 A schematic diagram of a large-scale model processing device provided in an embodiment of this application is shown below. Figure 7 As shown, the device includes: a first acquisition module 71, a second acquisition module 72, a first processing module 73, and a second processing module 74.

[0130] The first acquisition module 71 is used to acquire the basic large model of the large model to be processed;

[0131] The second acquisition module 72 is used to acquire sample data for training the base large model;

[0132] The first processing module 73 is used to train the basic large model based on the sample data to obtain the corresponding training data.

[0133] The second processing module 74 is used to generate a target difference small model corresponding to the large model to be processed based on the training data and the basic large model; wherein, the target difference small model is used to characterize the model version differences of the large model to be processed during the training process.

[0134] In one possible design, the sample data includes the overall sample data, which includes the overall training samples and the corresponding overall label results.

[0135] In one possible design, a basic large model is trained based on sample data to obtain corresponding training data, including:

[0136] The first processing module 73 is also used to input the overall training samples into the basic large model to obtain the first inference result output by the basic large model.

[0137] In one possible design, based on the training data and the base large model, a target difference small model corresponding to the large model to be processed is generated, including:

[0138] The second processing module 74 is also used to obtain the initial difference small model;

[0139] Based on the first inference result and the overall sample data, second sample data is generated for training the initial differential small model;

[0140] Based on the second sample data, the initial difference small model is trained to obtain the target difference small model.

[0141] In one possible design, after generating the target difference small model corresponding to the large model to be processed based on the training data and the base large model, the following steps are also included:

[0142] Obtain the sample to be predicted;

[0143] Input the sample to be predicted into the basic large model to obtain the basic prediction results;

[0144] Based on the basic prediction results and the sample to be predicted, determine the first differential prediction sample;

[0145] The first difference prediction sample is input into the target difference small model to obtain the target prediction result corresponding to the sample to be predicted.

[0146] In one possible design, the sample data includes application-specific sample data.

[0147] In one possible design, a basic large model is trained based on sample data to obtain corresponding training data, including:

[0148] The first processing module 73 is also used to train the basic large model based on the application scenario-specific sample data to obtain the target parameters and target features of the corresponding target large model.

[0149] Obtain the basic parameters and features of the basic large model;

[0150] Based on the target parameters, target features, basic parameters, and basic features, a target difference small model is generated.

[0151] In one possible design, after generating the target difference small model corresponding to the large model to be processed based on the training data and the base large model, the following steps are also included:

[0152] Obtain the sample to be predicted;

[0153] Input the sample to be predicted into the basic large model to obtain the basic prediction results;

[0154] The sample to be predicted is input into the target difference small model to obtain the difference prediction result;

[0155] Based on the basic prediction results and the differential prediction results, the target prediction results corresponding to the samples to be predicted are generated.

[0156] This embodiment provides a large model processing device that can execute a large model processing method of the above embodiment. Its implementation principle and technical effect are similar, and will not be described again here.

[0157] In a specific implementation of the aforementioned large-scale model processing method, each module can be implemented as a processor. The processor can execute computer execution instructions stored in memory, thereby enabling the processor to execute the aforementioned large-scale model processing method.

[0158] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 8 As shown, the electronic device 80 includes at least one processor 81 and a memory 82. The electronic device 80 also includes a communication component 83. The processor 81, memory 82, and communication component 83 are connected via a second bus 84.

[0159] In the specific implementation process, at least one processor 81 executes computer execution instructions stored in memory 82, causing at least one processor 81 to execute a large-scale processing method as executed on the electronic device side as described above.

[0160] The specific implementation process of processor 81 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0161] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0162] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage.

[0163] The second bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0164] The above description of the functions implemented by electronic devices and main control devices has introduced the solutions provided by the embodiments of the present invention. It is understood that, in order to implement the above functions, the electronic device or main control device includes hardware structures and / or software modules corresponding to the execution of each function. By combining the units and algorithm steps of the various examples described in the embodiments of the present invention, the embodiments of the present invention can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed by hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the technical solutions of the embodiments of the present invention.

[0165] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the processing method of the large model described above.

[0166] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0167] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in application-specific integrated circuits (ASICs). Alternatively, the processor and the readable storage medium can exist as discrete components in an electronic device or a host device.

[0168] This application also provides a computer program product, comprising: a computer program stored in a readable storage medium, wherein at least one processor of an electronic device can read the computer program from the readable storage medium, and the at least one processor executes the computer program to cause the electronic device to perform the scheme provided in any of the above embodiments.

[0169] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disk, or optical disk.

[0170] The technical solutions of this application have been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it is readily understood by those skilled in the art that the scope of protection of this application is obviously not limited to these specific embodiments. The above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A processing method of a large model, comprising: The method comprises the following steps: obtaining a basic large model of a to-be-processed large model; obtaining sample data used for training the basic large model; training the basic large model according to the sample data to obtain corresponding training data; generating a target differential small model corresponding to the to-be-processed large model according to the training data and the basic large model; wherein the target differential small model is used to represent the model version difference of the to-be-processed large model in the training process.

2. The method of claim 1, wherein, The sample data comprises overall sample data, and the overall sample data comprises overall training samples and corresponding overall label results.

3. The method of claim 2, wherein, The training of the basic large model according to the sample data to obtain corresponding training data comprises the following steps: inputting the overall training samples into the basic large model to obtain a first inference result output by the basic large model.

4. The method of claim 3, wherein, The generation of the target differential small model corresponding to the to-be-processed large model according to the training data and the basic large model comprises the following steps: obtaining an initial differential small model; generating second sample data used for training the initial differential small model according to the first inference result and the overall sample data; training the initial differential small model according to the second sample data to obtain the target differential small model.

5. The method according to any one of claims 2 to 4, characterized in that, After the generation of the target differential small model corresponding to the to-be-processed large model according to the training data and the basic large model, the method further comprises the following steps: obtaining a to-be-predicted sample; inputting the to-be-predicted sample into the basic large model to obtain a basic prediction result; determining a first differential prediction sample according to the basic prediction result and the to-be-predicted sample; inputting the first differential prediction sample into the target differential small model to obtain a target prediction result corresponding to the to-be-predicted sample.

6. The method of claim 1, wherein, The sample data comprises application scenario special sample data.

7. The method of claim 6, wherein, The training of the basic large model according to the sample data to obtain corresponding training data comprises the following steps: training the basic large model according to the application scenario special sample data to obtain target parameters and target features of a target large model; obtaining basic parameters and basic features of the basic large model; generating the target differential small model according to the target parameters, the target features, the basic parameters and the basic features.

8. The method according to claim 6 or 7, characterized in that, After the generation of the target differential small model corresponding to the to-be-processed large model according to the training data and the basic large model, the method further comprises the following steps: obtaining a to-be-predicted sample; inputting the to-be-predicted sample into the basic large model to obtain a basic prediction result; inputting the to-be-predicted sample into the target differential small model to obtain a differential prediction result; generating a target prediction result corresponding to the to-be-predicted sample according to the basic prediction result and the differential prediction result.

9. A processing device of a large model, characterized by, The method comprises the following steps: a first obtaining module is configured to obtain a basic large model of a to-be-processed large model; a second obtaining module is configured to obtain sample data used for training the basic large model; a first processing module is configured to train the basic large model according to the sample data to obtain corresponding training data; The second processing module is configured to generate a target differential small model corresponding to the to-be-processed large model according to the training data and the base large model, wherein the target differential small model is used to represent the model version difference of the to-be-processed large model in the training process.

10. An electronic device, comprising: Comprise: A processor and a memory connected with the processor in communication; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory to implement the method of any one of claims 1 to 8.

11. A computer readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are executed by the processor to implement the method of any one of claims 1 to 8.

12. A computer program product, characterised in that, The computer program is executed by the processor to implement the method of any one of claims 1 to 8.

Images