A method and system for collaborative training of large and small models

Abstract

The present invention relates to the field of model training technology, and specifically discloses a method and system for collaborative training of large and small models. It includes acquiring multi-dimensional characteristic data related to the collection training task, and analyzing the modal composition information of the data according to the training process stage based on a preset modal analysis strategy, obtaining the corresponding large model advantage parameters and small model advantage parameters according to the multimodal fusion data modal category, and determining the training time ratio coefficient according to the estimated time consumption of the large model priority training and the estimated time consumption of the small model priority training, thereby allocating the small model and the large model. The present invention solves the problem of irrational resource allocation caused by blindly selecting models due to lack of systematic analysis in traditional methods, avoids resource redundancy when large models process small data volumes or simple tasks, and avoids performance bottlenecks when small models process complex multimodal big data.

Description

Technical Field

[0001] The present invention relates to the field of model training technology, and in particular to a method and system for collaborative training of large and small models. Background Art

[0002] With the rapid development of artificial intelligence (AI) technology, model training in deep learning has become increasingly important and complex. When handling various training tasks, it is necessary to select the appropriate model for training based on the data characteristics and training requirements to achieve optimal training results and efficiency. Traditional model training methods often use a single large or small model for training, without considering the diversity of the data and the specific circumstances of the training task.

[0003] In practical applications, training data often has different data distribution types, such as centralized and distributed data types, or data modality categories dominated by a single modality and those that fuse multiple modalities. Different data types and modalities have significant differences in the training adaptability of large and small models. For example, for data dominated by a single modality, large models can leverage their powerful representation capabilities to deeply explore the information within that modality, but this can lead to redundant computations due to the unique nature of the data. For multimodal fusion data, small models struggle to handle the complex relationships between modalities and the large amount of data, making it difficult to achieve good training results.

[0004] Therefore, there is an urgent need for a method and system for collaborative training of large and small models to solve the above problems. Summary of the Invention

[0005] The present invention aims to provide a method for collaborative training of large and small models, comprising:

[0006] Acquire multi-dimensional characteristic data related to the acquisition training task, where the characteristic data includes modal composition information, total data size data, and expected training cycle data;

[0007] Obtaining floating interval data about training duration based on the expected training cycle information;

[0008] Based on the preset modal analysis strategy, the modal composition information is divided into single-modal dominant type or multi-modal fusion type;

[0009] Obtain the corresponding large model advantage parameters and small model advantage parameters based on the single-modality-dominant data mode category and the multimodal fusion data mode category, respectively. Combined with the large model advantage parameters, small model advantage parameters, total data size, and floating range of training time, calculate the estimated time consumption of large model priority training;

[0010] Determine the training time ratio coefficient based on the estimated time of prioritizing large model training and the estimated time of prioritizing small model training;

[0011] Determine whether the training time ratio coefficient exceeds the preset time ratio limit;

[0012] If the training time ratio coefficient exceeds the preset time ratio limit, it is determined that the training task characteristics are suitable for large-scale model-led training, and the large-scale model is deployed to carry out the training operation;

[0013] If the training time ratio coefficient does not exceed the preset time ratio limit, it is determined that the training task characteristics are suitable for small model-led training, and the small model is deployed to carry out the training operation.

[0014] Furthermore, the step of collecting multi-dimensional characteristic data related to the training task includes:

[0015] Obtain the source data for the training task, determine the data modality type based on the source data, and conduct in-depth analysis of the data modality type to obtain the data's attribute identification information and structural organization details;

[0016] Identify structural organization details according to a preset multimodal feature format and screen out multiple groups of multimodal feature format data;

[0017] Identify the structural organization details according to the preset model capability matching format and select a number of model capability matching format data;

[0018] The feature information corresponding to the multimodal feature format data and the model capability matching format data is integrated into data modality composition information.

[0019] Furthermore, the step of obtaining floating interval data about the training duration according to the expected training cycle information includes:

[0020] Determine the standard training duration corresponding to the total amount of data;

[0021] The source data is divided into modal categories according to a preset training process stage to obtain multiple modal category division results;

[0022] Obtaining modal processing times corresponding to multiple modal category classification results, and accumulating these modal processing times to obtain cumulative modal processing time;

[0023] Calculate the difference between the regular training duration benchmark and the accumulated modal processing duration, and define the difference as the floating range of the training duration.

[0024] Furthermore, the modal composition information of the data is analyzed according to the training process stage based on the preset modal analysis strategy to obtain the modal distribution of several sub-stages, and the data modal category is defined accordingly. The category includes a single modality-dominant type and a multimodal fusion type, including the following steps:

[0025] Extracting intrinsic structural characteristics of each data modality category based on the modal distributions of the multiple sub-stages, wherein the intrinsic structural characteristics include at least one of modal purity information, modal interaction information, and modal dispersion index information of the data;

[0026] Compare the internal structural characteristics one by one with the preset standard characteristics of multimodal fusion;

[0027] If the intrinsic structural characteristics are consistent with the preset standard characteristics of the multimodal fusion type, the corresponding data modality category is determined to be the multimodal fusion type, and the quantitative indicators of the intrinsic structural characteristics of the multimodal fusion type are counted;

[0028] If the intrinsic structural characteristics do not conform to the preset standard characteristics of the multimodal fusion type, the corresponding data modality category is determined to be a single-modality-dominated type, and the quantitative indicators of the intrinsic structural characteristics of the single-modality-dominated type are counted;

[0029] Calculating the modal quantity ratio relationship based on the quantitative index of the multimodal fusion type intrinsic structural characteristics and the quantitative index of the single modality dominant type intrinsic structural characteristics, and determining whether the modal quantity ratio relationship is greater than a preset modal quantity ratio threshold;

[0030] If the modality ratio is greater than a preset modality ratio threshold, the data modality category corresponding to the data modality composition information is defined as a multimodal fusion type;

[0031] If the modal quantity ratio is not greater than a preset modal quantity ratio threshold, the data modal category corresponding to the data modal composition information is defined as a single modality-dominant type.

[0032] Furthermore, the step of calculating the estimated time consumption of prioritizing training of the large model includes:

[0033] Based on the single-mode-dominant data mode category, the advantage parameters of the large model, the advantage parameters of the small model, the total data size, and the floating range of training time are determined to calculate the estimated time consumption of the large model priority training. The calculation formula is:

[0034] ;

[0035] in, It indicates the estimated time consumption of training large models first, D indicates the total amount of data,

[0036] represents the large model advantage parameter, It represents the computing resource consumption coefficient of the large model, and F represents the floating range of training time.

[0037] Furthermore, the step of calculating the estimated time consumption of the small model priority training includes:

[0038] Based on the multimodal fusion data modality category, the advantage parameters of the large model, the advantage parameters of the small model, the total data size, and the floating range of training time are determined to calculate the estimated time consumption for prioritizing training of the small model.

[0039] The calculation formula is:

[0040] ;

[0041] in, It represents the estimated time consumption of training small models first, D represents the total amount of data,

[0042] represents the small model advantage parameter, It represents the computing resource consumption coefficient of the small model, and F represents the floating range of training time.

[0043] The present invention also discloses a system for collaborative training of large and small models, comprising:

[0044] The first acquisition module is used to acquire multi-dimensional characteristic data related to the acquisition training task, wherein the characteristic data includes modal composition information, total data size data, and expected training cycle data;

[0045] The second acquisition module is used to obtain floating interval data about the training duration according to the expected training cycle information;

[0046] The analysis module is used to classify the modal composition information into a single-modality-dominant type or a multi-modal fusion type based on a preset modal analysis strategy;

[0047] The calculation module is used to obtain the corresponding large model advantage parameters and small model advantage parameters according to the single-mode-dominant data mode category and the multimodal fusion data mode category, and calculate the estimated time consumption of large model priority training by combining the large model advantage parameters, small model advantage parameters, total data size and floating range of training time;

[0048] A determination module is used to determine a training time ratio coefficient based on the estimated time of prioritizing large model training and the estimated time of prioritizing small model training;

[0049] A judgment module is used to judge whether the training time ratio coefficient exceeds a preset time ratio limit;

[0050] The first allocation module is used to determine that the training task characteristics are suitable for large-scale model-led training if the training time consumption ratio coefficient exceeds a preset time consumption ratio limit, and allocate the large-scale model to carry out the training operation;

[0051] The second allocation module is used to determine that the training task characteristics are suitable for small model-led training if the training time ratio coefficient does not exceed the preset time ratio limit, and allocate the small model to carry out training operations.

[0052] Furthermore, the first acquisition module includes:

[0053] The first acquisition unit is used to obtain the source data of the training task, determine the data modality type based on the source data, and perform in-depth analysis on the data modality type to obtain the attribute identification information and structural organization details of the data;

[0054] A first screening unit is used to identify structural organization details according to a preset multimodal feature format and screen out multiple groups of multimodal feature format data;

[0055] The second screening unit is used to identify the structural organization details according to a preset model capability matching format and screen out a number of model capability matching format data;

[0056] The integration unit is used to integrate the feature information corresponding to the multimodal feature format data and the model capability matching format data into data modality composition information.

[0057] The present application also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0058] The present application also provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above method when executed by a processor.

[0059] The beneficial effects of this application are:

[0060] The present invention first obtains multi-dimensional characteristic data, including modal composition, total data size and expected training cycle, etc. This enables accurate response when facing different data modal categories. For single-modality-dominated data, the estimated time consumption of prioritizing training of the large model can be calculated based on the corresponding large model and small model advantage parameters and other key factors. Similarly, the estimated time consumption of prioritizing training of the small model is calculated for multimodal fusion data, and then the training time ratio coefficient is determined and compared with the preset limit. In this way, the problem of irrational resource allocation caused by blindly selecting models due to lack of systematic analysis in traditional methods is solved, and resource redundancy when large models process small data volumes or simple tasks, as well as performance bottlenecks when small models process complex multimodal big data, are avoided. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a flowchart of a method proposed in one embodiment of the present application.

[0062] The purpose, features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0063] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0064] like Figure 1 As shown, the present application provides a method for collaborative training of large and small models, including:

[0065] S1, acquiring multi-dimensional characteristic data related to the acquisition training task, wherein the characteristic data includes modal composition information, total data size data, and expected training cycle data;

[0066] S2, obtaining floating interval data about the training duration according to the expected training cycle information;

[0067] S3, based on the preset modal analysis strategy, the modal composition information is divided into a single modality dominant type or a multimodal fusion type;

[0068] S4: Obtain the corresponding large model advantage parameters and small model advantage parameters based on the single-modality-dominant data mode category and the multimodal fusion data mode category, respectively. Calculate the estimated time required for large model priority training by combining the large model advantage parameters, small model advantage parameters, total data size, and the floating range of training time.

[0069] S5, determining a training time ratio coefficient based on the estimated time of prioritizing large model training and the estimated time of prioritizing small model training;

[0070] S6, determining whether the training time ratio coefficient exceeds a preset time ratio limit;

[0071] S7, if the training time ratio coefficient exceeds the preset time ratio limit, it is determined that the training task characteristics are suitable for large-scale model-led training, and the large-scale model is deployed to carry out the training operation;

[0072] S8: If the training time ratio coefficient does not exceed the preset time ratio limit, it is determined that the training task characteristics are suitable for small model-led training, and the small model is deployed to carry out the training operation.

[0073] As described in the above steps S1-S8, compared with the traditional single model training method, the present invention first obtains multi-dimensional characteristic data, including modal composition, total data size and expected training cycle, etc., which enables accurate response when facing different data modal categories. For single-modality-dominated data, the estimated time consumption of the large model priority training can be calculated based on the corresponding large model and small model advantage parameters and other key factors. Similarly, the estimated time consumption of the small model priority training is calculated for multimodal fusion data, and then the training time ratio coefficient is determined and compared with the preset limit. In this way, the problem of irrational resource allocation caused by blindly selecting models due to lack of systematic analysis in traditional methods is solved, and resource redundancy when large models process small data volumes or simple tasks, as well as performance bottlenecks when small models process complex multimodal big data, are avoided.

[0074] The present invention also rationally allocates large models or small models to lead training through intelligent judgment of training task characteristics, thereby improving training efficiency. While meeting training time constraints, it can optimize model selection based on data characteristics, thereby improving training effects and effectively responding to complex and changeable training scenarios. It enables model training to achieve balance and optimization in resource utilization, training speed, and training quality, and makes up for the shortcomings of existing technologies in processing different data modalities and the complexity of training tasks.

[0075] Specifically, the step of collecting multi-dimensional characteristic data related to the training task includes:

[0076] S9, obtaining the source data of the training task, determining the data modality type based on the source data, and performing in-depth analysis of the data modality type to obtain the attribute identification information and structural organization details of the data;

[0077] S10, identifying structural organization details according to a preset multimodal feature format, and screening out multiple groups of multimodal feature format data;

[0078] S11, identifying the structural organization details according to a preset model capability matching format, and screening out a number of model capability matching format data;

[0079] S12, integrating the feature information corresponding to the multimodal feature format data and the model capability matching format data into data modality composition information.

[0080] As described in steps S9-S12 above, by obtaining the source data of the training task and determining the data modality type to which it belongs, and then deeply analyzing the data modality type, the attribute identification information and structural organization details of the data are obtained. This processing method can fully and deeply understand the characteristics of the data. In model training, different data modality types require different processing methods. For example, the processing methods for image data (single modality) and multimodal data containing images and text are different. Through this analysis, a detailed data foundation can be provided for subsequent model selection and training strategy formulation.

[0081] Structural details are identified based on a preset multimodal feature format, and multiple sets of multimodal feature format data are screened. While multimodal data often contains rich information, it also introduces complexity. By screening multimodal feature format data, valuable data features for model training can be extracted. This facilitates more targeted model selection in multimodal fusion data scenarios (for example, a small model may be more suitable for processing filtered multimodal data in some cases), avoiding the interference of data redundancy on model training.

[0082] Structural organization details are screened based on preset model capability matching formats, and data in matching formats is selected. Different models have different capabilities and strengths. By selecting data that matches the model's capabilities, we ensure that the selected model can fully utilize its strengths. For example, large models exhibit strong learning capabilities when processing data with specific formats. This screening method helps allocate appropriate data to the appropriate model, improving training efficiency and effectiveness.

[0083] The feature information corresponding to multimodal feature format data and model capability matching format data is integrated into data modal composition information. This integration can centralize scattered data feature information to form a complete data modal composition information. This helps to grasp the characteristics of the data as a whole. In subsequent steps, such as determining whether the data is single-modality-dominated or multimodal fusion, more accurate decisions can be made, thereby rationally selecting a large model or a small model for training.

[0084] Specifically, the step of obtaining floating interval data about the training duration according to the expected training cycle information includes:

[0085] S13, determining a conventional training duration benchmark corresponding to the total amount of data;

[0086] S14, performing modality classification on the source data according to a preset training process stage to obtain multiple modality classification results;

[0087] S15, obtaining modal processing times corresponding to the multiple modal category classification results, and accumulating these modal processing times to obtain a cumulative modal processing time;

[0088] S16, calculating the difference between the regular training duration benchmark and the accumulated modal processing duration, and defining the difference as a floating range of the training duration.

[0089] As described in steps S13-S16 above, a conventional training time benchmark corresponding to the total amount of data is determined, which provides an important reference for the subsequent evaluation of the floating range of training time. In model training, the amount of data is one of the key factors affecting the training time. By determining this benchmark in advance, the time required for training can be more accurately estimated, which helps to reasonably arrange the training plan and avoid the problem of waste of resources or poor training results caused by too long or too short training time. The source data is divided into modal categories according to the preset training process stages to obtain multiple modal category division results. This division helps to gain a deeper understanding of the modal characteristics of the data at different training stages. Data of different modalities (for example, image data and text data) require different processing methods and models. By dividing the modal categories, corresponding training strategies can be better formulated for data of different modalities, thereby improving the pertinence and efficiency of training.

[0090] The modal processing times corresponding to multiple modal classification results are obtained and accumulated to form the cumulative modal processing time. This step accurately calculates the time spent processing different modal data. In actual training, different modal data have varying processing difficulty and time requirements. This precise calculation provides a better understanding of the contribution of each modal data to the total training time, helping to optimize the training process, such as allocating computing resources and adjusting the training sequence. The difference between the baseline training time and the cumulative modal processing time is calculated and defined as the training time fluctuation range. This fluctuation range provides flexibility in the training process. In actual training, various factors (such as hardware performance fluctuations and the complexity of data preprocessing) can affect training time, causing it to vary. By defining a fluctuation range, the training time can be allowed to fluctuate within a certain range while ensuring training effectiveness, improving the adaptability and fault tolerance of the training process.

[0091] Specifically, the modal composition information of the data is analyzed according to the training process stage based on the preset modal analysis strategy to obtain the modal distribution of several sub-stages, and the data modal category is defined accordingly. The category includes a single modality-dominant type and a multimodal fusion type, including the following steps:

[0092] S17, extracting intrinsic structural characteristics of each data modality category based on the modal distributions of the plurality of sub-stages, wherein the intrinsic structural characteristics include at least one of modal purity information, modal interaction information, and modal dispersion index information of the data;

[0093] S18, comparing the internal structural characteristics with the preset standard characteristics of the multimodal fusion type one by one to see whether they are consistent;

[0094] S19, if the intrinsic structural characteristics are consistent with the preset standard characteristics of the multimodal fusion type, the corresponding data modality category is determined to be the multimodal fusion type, and the quantitative indicators of the intrinsic structural characteristics of the multimodal fusion type are counted;

[0095] S20, if the intrinsic structural characteristics do not conform to the preset standard characteristics of the multimodal fusion type, the corresponding data modality category is determined to be a single-modality-dominant type, and the quantitative indicators of the intrinsic structural characteristics of the single-modality-dominant type are counted;

[0096] S21, calculating a modality quantity ratio relationship based on a quantity index of the multimodal fusion type intrinsic structural characteristics and a quantity index of the single modality dominant type intrinsic structural characteristics, and determining whether the modality quantity ratio relationship is greater than a preset modality quantity ratio threshold;

[0097] S22: If the modality ratio is greater than a preset modality ratio threshold, the data modality category corresponding to the data modality composition information is defined as a multimodal fusion type;

[0098] S23: If the modality ratio is not greater than a preset modality ratio threshold, the data modality category corresponding to the data modality composition information is defined as a single modality-dominant type.

[0099] As described in the above steps S17-S23, the intrinsic structural characteristics of each data modal category are extracted based on the modal distribution of several sub-stages, including modal purity information, modal interaction information and modal dispersion index information. These characteristics can comprehensively and deeply describe the modal characteristics of the data, which is very helpful for accurately understanding the nature and structure of the data. In the process of model training, understanding these characteristics of the data helps to select appropriate models and training strategies. For example, for data with high modal purity, a relatively simple model can be used for processing. Compare the intrinsic structural characteristics one by one with the preset standard characteristics of the multimodal fusion type. This comparison method can scientifically determine whether the data belongs to the multimodal fusion type. By presetting standard characteristics, the subjectivity of human judgment can be avoided, and the accuracy and consistency of data modal category judgment can be ensured, thereby providing a reliable basis for subsequent model selection and training.

[0100] For data determined to be multimodal fusion type or single modality dominant type, quantitative indicators of their inherent structural characteristics are counted respectively. These quantitative indicators can further quantify the modal characteristics of the data. For example, the specific degree of modal interaction of multimodal fusion data can be known. This helps to analyze the data more accurately. In subsequent model training, the parameters of the model can be adjusted or a more suitable model architecture can be selected based on these quantitative indicators. The modal quantity ratio relationship is calculated based on the quantitative indicators of the inherent structural characteristics of the multimodal fusion type and the single modality dominant type, and compared with the preset modal quantity ratio threshold to define the data modal category. Determining the data modal category through a quantitative ratio relationship is more objective and accurate. In actual training scenarios, it is possible to reasonably select a large model or a small model for training based on the actual modal composition of the data, avoiding the problem of low training efficiency or poor results caused by blindly selecting a model.

[0101] Specifically, the step of calculating the estimated time consumption of prioritizing training of the large model includes:

[0102] S24, based on the single-mode-dominant data mode category, determines the large model advantage parameters, small model advantage parameters, total data size, and floating range of training time to calculate the estimated time required for large model priority training. The calculation formula is:

[0103] ;

[0104] in, It indicates the estimated time consumption of training large models first, D indicates the total amount of data,

[0105] represents the large model advantage parameter, It represents the computing resource consumption coefficient of the large model, and F represents the floating range of training time.

[0106] Specifically, the step of calculating the estimated time consumption of the small model priority training includes:

[0107] S25, based on the multimodal fusion data modality category, determines the large model advantage parameters, small model advantage parameters, the total data size, and the floating range of training time to calculate the estimated time consumption for prioritizing small model training. The calculation formula is:

[0108] ;

[0109] in, It represents the estimated time consumption of training small models first, D represents the total amount of data,

[0110] represents the small model advantage parameter, It represents the computing resource consumption coefficient of the small model, and F represents the floating range of training time.

[0111] As described in the above steps S24-S25, by introducing the large model advantage parameter and small model advantage parameter , can quantify the advantages of large and small models when processing data dominated by a single modality. Different models have different efficiencies and effects when processing specific types of data. This parameter can reflect these differences and help to more accurately estimate the training time based on the specific model characteristics. Calculate the computing resource consumption coefficient of the large model and the computational resource consumption coefficient of the small model The introduction of takes into account the computing resource consumption of the model during training. Different model architectures and configurations utilize different computing resources. This coefficient incorporates these differences in the estimated training time, making the estimate more realistic and facilitating the optimal allocation of computing resources. The consideration of the variable training time interval F provides flexibility in the training time estimate. In actual training, training time can vary due to various uncertainties (such as hardware performance fluctuations and the complexity of data preprocessing). By incorporating this variable range, training time estimates can be more reasonable, avoiding overly precise estimates that may lead to discrepancies between the planned and actual training time.

[0112] The present invention also discloses a system for collaborative training of large and small models, comprising:

[0113] The first acquisition module is used to acquire multi-dimensional characteristic data related to the acquisition training task, wherein the characteristic data includes modal composition information, total data size data, and expected training cycle data;

[0114] The second acquisition module is used to obtain floating interval data about the training duration according to the expected training cycle information;

[0115] The analysis module is used to classify the modal composition information into a single-modality-dominant type or a multi-modal fusion type based on a preset modal analysis strategy;

[0116] The calculation module is used to obtain the corresponding large model advantage parameters and small model advantage parameters according to the single-mode-dominant data mode category and the multimodal fusion data mode category, and calculate the estimated time consumption of large model priority training by combining the large model advantage parameters, small model advantage parameters, total data size and floating range of training time;

[0117] A determination module is used to determine a training time ratio coefficient based on the estimated time of prioritizing large model training and the estimated time of prioritizing small model training;

[0118] A judgment module is used to judge whether the training time ratio coefficient exceeds a preset time ratio limit;

[0119] The first allocation module is used to determine that the training task characteristics are suitable for large-scale model-led training if the training time consumption ratio coefficient exceeds a preset time consumption ratio limit, and allocate the large-scale model to carry out the training operation;

[0120] The second allocation module is used to determine that the training task characteristics are suitable for small model-led training if the training time ratio coefficient does not exceed the preset time ratio limit, and allocate the small model to carry out training operations.

[0121] Furthermore, the first acquisition module includes:

[0122] The first acquisition unit is used to obtain the source data of the training task, determine the data modality type based on the source data, and perform in-depth analysis on the data modality type to obtain the attribute identification information and structural organization details of the data;

[0123] A first screening unit is used to identify structural organization details according to a preset multimodal feature format and screen out multiple groups of multimodal feature format data;

[0124] The second screening unit is used to identify the structural organization details according to a preset model capability matching format and screen out a number of model capability matching format data;

[0125] The integration unit is used to integrate the feature information corresponding to the multimodal feature format data and the model capability matching format data into data modality composition information.

[0126] The present application also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0127] The present application also provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above method when executed by a processor.

[0128] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, value library or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct RAM bus dynamic RAM (DRDRAM), and RAM bus dynamic RAM (RDRAM).

[0129] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, apparatus, article, or method comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, apparatus, article, or method. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, apparatus, article, or method comprising the element.

[0130] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent results or equivalent process transformations made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for collaborative training of large and small models, characterized in that: include: Acquire multi-dimensional characteristic data related to the acquisition training task, where the characteristic data includes modal composition information, total data size data, and expected training cycle data; Obtaining floating interval data about training duration based on the expected training cycle information; Based on the preset modal analysis strategy, the modal composition information is divided into single-modal dominant type or multi-modal fusion type; Obtain the corresponding large model advantage parameters and small model advantage parameters according to the single-mode-dominant data mode category and the multi-modal fusion data mode category respectively; Calculate the estimated time required to prioritize large model training, including: Based on the single-mode-dominated data mode category, the dominant parameters of the large model, the total data size, and the floating range of training time are determined to calculate the estimated time consumption of large model priority training. The calculation formula is: ； in, It indicates the estimated time consumption of training large models first, D indicates the total amount of data, represents the large model advantage parameter, Indicates the computing resource consumption coefficient of the large model, and F indicates the floating range of training time; Calculate the estimated time required to prioritize small model training, including: Based on the multimodal fusion data modality category, the small model advantage parameters, the total data size, and the floating range of training time are determined to calculate the estimated time consumption of small model priority training. The calculation formula is: ； in, It represents the estimated time consumption of training small models first, D represents the total amount of data, represents the small model advantage parameter, Indicates the computing resource consumption coefficient of the small model, and F indicates the floating range of training time; Determine the training time ratio coefficient based on the estimated time of prioritizing large model training and the estimated time of prioritizing small model training; Determine whether the training time ratio coefficient exceeds the preset time ratio limit; If the training time ratio coefficient exceeds the preset time ratio limit, it is determined that the training task characteristics are suitable for large-scale model-led training, and the large-scale model is deployed to carry out the training operation; If the training time ratio coefficient does not exceed the preset time ratio limit, it is determined that the training task characteristics are suitable for small model-led training, and the small model is deployed to carry out the training operation.

2. The method for collaborative training of large and small models according to claim 1, characterized in that: The step of collecting multi-dimensional characteristic data related to the training task includes: Obtain the source data for the training task, determine the data modality type based on the source data, and conduct in-depth analysis of the data modality type to obtain the data's attribute identification information and structural organization details; Identify structural organization details according to a preset multimodal feature format and screen out multiple groups of multimodal feature format data; Identify the structural organization details according to the preset model capability matching format and select a number of model capability matching format data; The feature information corresponding to the multimodal feature format data and the model capability matching format data is integrated into data modality composition information.

3. The method for collaborative training of large and small models according to claim 1, characterized in that: The step of obtaining floating interval data about the training duration according to the expected training cycle information includes: Determine the standard training duration corresponding to the total amount of data; The source data is divided into modal categories according to a preset training process stage to obtain multiple modal category division results; Obtaining modal processing times corresponding to multiple modal category classification results, and accumulating these modal processing times to obtain cumulative modal processing time; Calculate the difference between the regular training duration benchmark and the accumulated modal processing duration, and define the difference as the floating range of the training duration.

4. The method for collaborative training of large and small models according to claim 1, characterized in that: The modal analysis strategy based on the preset modal composition information of the data is analyzed according to the training process stage to obtain the modal distribution of several sub-stages, and the data modal category is defined accordingly. The category includes single modality dominant type and multimodal fusion type, including the following steps: Extract data modal distribution based on stage features of the training process; By analyzing the preset characteristic parameters in the modal distribution, the quantitative indicators of the single-modality dominant type and the multi-modal fusion type are calculated; Calculating the modal quantity ratio relationship based on the quantitative index of the multimodal fusion type intrinsic structural characteristics and the quantitative index of the single modality dominant type intrinsic structural characteristics, and determining whether the modal quantity ratio relationship is greater than a preset modal quantity ratio threshold; If the modality ratio is greater than a preset modality ratio threshold, the data modality category corresponding to the data modality composition information is defined as a multimodal fusion type; If the modal quantity ratio is not greater than a preset modal quantity ratio threshold, the data modal category corresponding to the data modal composition information is defined as a single modality-dominant type.

5. A system for collaborative training of large and small models, characterized by: include: The first acquisition module is used to acquire multi-dimensional characteristic data related to the acquisition training task, wherein the characteristic data includes modal composition information, total data size data, and expected training cycle data; The second acquisition module is used to obtain floating interval data about the training duration according to the expected training cycle information; The analysis module is used to classify the modal composition information into a single-modality-dominant type or a multi-modal fusion type based on a preset modal analysis strategy; A calculation module is used to obtain the corresponding large model advantage parameters and small model advantage parameters according to the single-mode dominant data mode category and the multimodal fusion data mode category respectively; Calculate the estimated time required to prioritize large model training, including: Based on the single-mode-dominated data mode category, the dominant parameters of the large model, the total data size, and the floating range of training time are determined to calculate the estimated time consumption of large model priority training. The calculation formula is: ； in, It indicates the estimated time consumption of training large models first, D indicates the total amount of data, represents the large model advantage parameter, Indicates the computing resource consumption coefficient of the large model, and F indicates the floating range of training time; Calculate the estimated time required to prioritize small model training, including: Based on the multimodal fusion data modality category, the small model advantage parameters, the total data size, and the floating range of training time are determined to calculate the estimated time consumption of small model priority training. The calculation formula is: ； in, It indicates the estimated time consumption of training small models first, D indicates the total amount of data, represents the small model advantage parameter, Indicates the computing resource consumption coefficient of the small model, and F indicates the floating range of training time; A determination module is used to determine a training time ratio coefficient based on the estimated time of prioritizing large model training and the estimated time of prioritizing small model training; A judgment module is used to judge whether the training time ratio coefficient exceeds a preset time ratio limit; The first allocation module is used to determine that the training task characteristics are suitable for large-scale model-led training if the training time consumption ratio coefficient exceeds a preset time consumption ratio limit, and allocate the large-scale model to carry out the training operation; The second allocation module is used to determine that the training task characteristics are suitable for small model-led training if the training time ratio coefficient does not exceed the preset time ratio limit, and allocate the small model to carry out training operations.

6. A system for collaborative training of large and small models according to claim 5, characterized in that: The first acquisition module includes: The first acquisition unit is used to obtain the source data of the training task, determine the data modality type based on the source data, and perform in-depth analysis on the data modality type to obtain the attribute identification information and structural organization details of the data; A first screening unit is used to identify structural organization details according to a preset multimodal feature format and screen out multiple groups of multimodal feature format data; The second screening unit is used to identify the structural organization details according to a preset model capability matching format and screen out a number of model capability matching format data; The integration unit is used to integrate the feature information corresponding to the multimodal feature format data and the model capability matching format data into data modality composition information.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for collaborative training of large and small models as described in any one of claims 1 to 4 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for collaborative training of large and small models described in any one of claims 1 to 4 are implemented.

Images