A method and system for distributing target data using a large model

Abstract

The present invention relates to the field of computer technology applications, and in particular to a method and system for allocating target data using a large model. The method comprises: obtaining target data to be processed, extracting sensitive word information and the domain to which it belongs from the target data, as well as queue information of each candidate large model and network information between the data transmission platform and the candidate large model as reference information; calculating a comprehensive score for each candidate large model, comprehensively considering factors such as sensitive word processing capability, domain adaptability, queue load, and network transmission; and finally selecting a target large model from multiple candidate large models based on the comprehensive score to process the target data. The present invention achieves optimal allocation of large model resources and precise processing of target data through multi-dimensional information evaluation and quantitative decision-making.

Description

Technical Field

[0001] The present invention relates to the field of computer technology application technology, and in particular to a distribution method and system for processing target data using a large model. Background Art

[0002] In the digital age, data is experiencing explosive growth. The target data generated in various business scenarios is massive, complex, and rapidly growing. From user behavior data from internet companies to transaction data from financial institutions to patient diagnosis and treatment data in the healthcare sector, this target data holds immense value. Efficient processing and analysis can provide strong support for decision-making, business optimization, and product innovation. However, traditional data processing methods are increasingly exposed to numerous shortcomings when dealing with massive amounts of target data. For one thing, traditional data allocation strategies are often based on fixed rules or simple load balancing algorithms, making it difficult to dynamically adjust to the real-time characteristics of data and processing requirements. This leads to inefficient data processing and uneven resource utilization. For example, in a data processing cluster, some compute nodes may become overloaded due to excessive data volume or processing complexity, while other nodes remain idle or underloaded, severely impacting the timeliness and accuracy of overall data processing. Furthermore, traditional data processing technologies lack sufficient flexibility and adaptability to the complex and diverse target data types and structures, making it difficult to fully exploit the potential information hidden in the data. In recent years, large models, leveraging their powerful feature extraction, learning, and reasoning capabilities, have achieved remarkable results in fields such as natural language processing and computer vision, bringing new insights and approaches to data processing. Large models can automatically learn the inherent patterns and regularities of data, enabling in-depth understanding and analysis of complex data. However, when applying large models to target data processing, the proper allocation of data is a key issue. Large models require enormous computational resources, and different types and sizes of target data present varying processing requirements for large models. Failure to scientifically and rationally allocate target data not only fails to fully exploit the performance advantages of large models, but can also lead to wasted computational resources and increased data processing costs. Therefore, developing a method and system for allocating target data using large models to achieve efficient target data processing and optimal utilization of large model computational resources is of great practical significance and application value. Summary of the Invention

[0003] In view of the above technical problems, the technical solution adopted by the present invention is:

[0004] According to a first aspect of the present invention, a method for allocating target data using a large model is provided. The method is implemented based on multiple candidate large models and comprises the following steps:

[0005] S100, obtaining target data to be processed.

[0006] S200 , obtaining sensitive word information in the target data, the field to which the target data belongs, queue information of each candidate large model, and network information between the data sending platform and each candidate large model as reference information.

[0007] S300 , based on the reference information, selecting a target large model from a plurality of candidate large models for processing the target data.

[0008] S300 specifically includes:

[0009] S310, obtain the comprehensive score S of the i-th candidate large model i =W sen ×S i-sen +W dom ×S i-dom +W que ×S i-que +W net ×S i-net ; Among them, S i-sen is the sensitive word information dimension evaluation score of the i-th candidate large model, W sen is the weight of the sensitive word information dimension, S i-dom is the domain dimension evaluation score of the i-th candidate large model, W dom is the weight of the domain dimension, S i-que is the queue information dimension evaluation score of the i-th candidate large model, W que is the weight of the queue information dimension, S i-net is the network information dimension evaluation score of the i-th candidate large model, W net is the weight of the network information, i ranges from 1 to n, and n is the number of candidate large models.

[0010] S320 , based on the comprehensive scores of the n candidate large models, select a target large model from the n candidate large models.

[0011] According to a second aspect of the present invention, there is provided a distribution system for processing target data using a large model, the system being implemented based on a plurality of candidate large models and comprising:

[0012] The data acquisition module is used to acquire the target data to be processed.

[0013] The reference information acquisition module is used to obtain sensitive word information in the target data, the field to which the target data belongs, the queue information of each candidate large model, and the network information between the data sending platform and each candidate large model as reference information.

[0014] The target model acquisition module is used to select a target large model from multiple candidate large models based on the reference information to process the target data.

[0015] The target model acquisition module specifically includes:

[0016] Comprehensive score acquisition unit, used to obtain the comprehensive score S of the i-th candidate large model i =W sen ×S i-sen +W dom ×S i-dom +W que ×S i-que +W net ×S i-net ; Among them, S i-sen is the sensitive word information dimension evaluation score of the i-th candidate large model, W sen is the weight of the sensitive word information dimension, S i-dom is the domain dimension evaluation score of the i-th candidate large model, W dom is the weight of the domain dimension, S i-que is the queue information dimension evaluation score of the i-th candidate large model, W que is the weight of the queue information dimension, S i-net is the network information dimension evaluation score of the i-th candidate large model, W net is the weight of the network information, i ranges from 1 to n, and n is the number of candidate large models.

[0017] The target model selection unit is used to select a target large model from the n candidate large models based on the comprehensive scores of the n candidate large models.

[0018] The present invention has at least the following beneficial effects:

[0019] An embodiment of the present invention provides a method and system for allocating target data using a large model, including: obtaining target data to be processed, extracting sensitive word information and the domain to which it belongs from the target data, as well as queue information of each candidate large model and network information between the data transmission platform and the candidate large model as reference information; calculating a comprehensive score for each candidate large model, comprehensively considering factors such as sensitive word processing capability, domain adaptability, queue load, and network transmission; and finally selecting a target large model from multiple candidate large models to process the target data based on the comprehensive score. Through multi-dimensional information evaluation and quantitative decision-making, the present invention can achieve optimal allocation of large model resources and accurate processing of target data.

[0020] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present invention, nor is it intended to limit the scope of the present invention. Other features of the present invention will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0022] Figure 1 A flow chart of a method for distributing target data using a large model provided in one embodiment of the present invention;

[0023] Figure 2 A flowchart of a method for distributing target data using a large model provided in another embodiment of the present invention. DETAILED DESCRIPTION

[0024] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains. The terms used herein in the specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0026] It should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of the steps can be performed in parallel, concurrently, or simultaneously. In addition, the order of the steps can be rearranged. A process can be terminated when its operation is completed, but can also have additional steps not included in the accompanying drawings. A process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0027] (Example 1)

[0028] This embodiment provides a method for allocating target data using a large model. The method is implemented based on multiple candidate large models. The candidate large models are all existing mature large models, including pre-trained models such as the GPT series, BERT, and LLaMA in the field of natural language processing, ResNet and YOLO series models in the field of computer vision, and multimodal large models that integrate multimodal data processing capabilities. These candidate large models can be open sourced or commercially provided by mainstream technology companies and research institutions. Their algorithm architecture, training methods, and functional characteristics have been verified in a large number of practical application scenarios. The number of candidate large models can be flexibly set based on actual needs. When setting the specific number, the following factors can be comprehensively considered: the scale and complexity of the data processing task. For example, when processing massive amounts of text data or high-precision image recognition tasks, the number of candidate large models can be appropriately increased to cover different processing requirements; computing resource limitations, including the server's CPU, GPU computing power, and memory capacity, to ensure that the number of candidate large models matches the existing computing resources to avoid resource overload caused by an excessive number of models; and the diversity of business scenarios. For data processing in multiple fields such as finance, healthcare, and education, large models with expertise in different fields can be added as candidates. For example, when processing user behavior data, product image data, and customer service conversation text data from large e-commerce platforms, 8-12 candidate large models can be set up, including 3-4 natural language processing models that excel at sequence data analysis, 3-4 computer vision models with excellent image recognition performance, and 2-4 models with multimodal processing capabilities to meet the target data processing needs in complex business scenarios.

[0029] like Figure 1 As shown, the present embodiment provides a method for allocating target data using a large model, which may include the following steps:

[0030] S100, obtaining target data to be processed.

[0031] In this embodiment of the present invention, target data to be processed can be acquired from data sources (such as databases, file systems, and real-time data streams) through a data acquisition interface. Adaptive data acquisition protocols and parsing methods are employed for different data sources to ensure complete and accurate data acquisition. Simultaneously, preliminary data cleansing is performed on the acquired target data to remove duplicate, erroneous, or incomplete data records, laying the foundation for subsequent processing.

[0032] S200 , obtaining sensitive word information in the target data, the field to which the target data belongs, queue information of each candidate large model, and network information between the data sending platform and each candidate large model as reference information.

[0033] In an embodiment of the present invention, a data sending platform refers to an initiator or intermediate carrier responsible for transmitting target data to a candidate large model. Its core function is to initiate a data processing request and pass the target data to be processed to the large model. If the target data is a question entered by a user in a chat software, the data sending platform can be the back-end server of the chat software: after the server receives the text (target data) entered by the user, it needs to send it to the large model for semantic understanding and answer generation. In this case, the server is the data sending platform. If the target data is a document parsing task within an enterprise, the data sending platform can be the enterprise's document management system: after the system collects the document to be parsed (target data), it sends a processing request to multiple candidate large models (such as a large model that specializes in processing legal documents, a large model that processes technical documents). In this case, the document management system is the data sending platform.

[0034] In this embodiment of the present invention, sensitive word information in the target data can be obtained through a pre-trained natural language processing model. The pre-trained natural language processing model is mainly composed of an input layer, an embedding layer, a bidirectional long short-term memory network layer, an attention mechanism layer, and an output layer. Each layer works together to process sensitive word information in the target data. The specific structure is as follows:

[0035] Input layer: Serving as the model's entry point, it receives preprocessed target data. Since target data typically exists as a text sequence, the input layer segments the text into individual units, organized into sequence vectors, and then feeds them into the model, providing the data foundation for subsequent processing. For example, for text data like "User ID number is 123456789012345678," the input layer breaks it down into a sequence consisting of word units such as "user," "ID number," and "is."

[0036] Embedding layer: Its core function is to map the words or characters in the input layer into a low-dimensional, dense, and continuous vector representation, namely the word embedding vector. In this way, discrete text information is converted into a numerical form that the model can process while retaining the semantic information of the words. Common methods include Word2Vec and GloVe. In this model, the embedding layer maps each word into a vector of fixed dimension by learning the relationship between words in the training corpus. For example, "ID number" is mapped into a 128-dimensional vector. This makes the distance between words with similar semantics closer in the vector space, which facilitates the extraction and processing of semantic information by subsequent network layers.

[0037] Bidirectional Long Short-Term Memory (Bi-LSTM) layer: This layer consists of two long short-term memory networks (LSTMs) operating in opposite directions, processing the input sequence from front to back and back to front, respectively. As a special type of recurrent neural network, the LSTM network effectively addresses the vanishing and exploding gradient problems found in traditional RNNs by introducing gating mechanisms (input gate, forget gate, and output gate), thereby better capturing long-range dependencies. When processing text, the Bi-LSTM can simultaneously consider both preceding and following context. For example, when determining whether "medical record" is a sensitive word, it can more accurately understand the semantics of the word itself and incorporate contextual information such as "patient" and "diagnosis" to identify the sensitive word. The Bi-LSTM layer processes the input word embedding vectors sequentially and outputs a hidden state vector containing contextual semantic information.

[0038] Attention Mechanism Layer: The introduction of the attention mechanism aims to allow the model to focus more on important information related to sensitive word recognition and ignore less important content. It calculates the similarity between the hidden state vectors output by the Bi-LSTM layer and assigns an attention weight to each hidden state. The weight reflects the importance of the position information for sensitive word recognition. For example, in a text containing multiple words, the hidden state vectors corresponding to words directly related to sensitive words and their context will receive higher attention weights, allowing the model to focus on these areas in subsequent processing. Through a weighted summation, the hidden state vectors with attention weights are aggregated into a new feature vector, highlighting key information and enhancing the model's ability to recognize sensitive words.

[0039] Output layer: Based on the feature vector output by the attention mechanism layer, the output layer processes the data through a fully connected layer and activation function, outputting the sensitive word recognition results. The fully connected layer maps the feature vector to a specified output dimension. In this model, the output dimension corresponds to the number of sensitive word categories (such as high, medium, and low risk levels, as well as non-sensitive word categories). An activation function, such as the Softmax function, converts the output of the fully connected layer into a probability distribution, where each probability value represents the probability that a word in the input text belongs to a different sensitive word category. Ultimately, the model determines the number, category, and location of sensitive words in the target data based on the probability distribution, and then categorizes them into sensitivity levels.

[0040] Those skilled in the art should understand that the natural language processing model is trained using a dedicated corpus containing industry-sensitive words and privacy data keywords. The training process is based on conventional technical means in the field of deep learning, and specifically includes steps such as data preprocessing, model parameter initialization, forward propagation, backpropagation and parameter update. Its core algorithm logic, training framework construction and optimization strategy all fall within the scope of the existing technology. During the training phase, by inputting text data from the corpus into the model, the weight parameters in the model are iteratively updated using optimization algorithms such as stochastic gradient descent, so that the model can learn the semantic features, grammatical structure and expression of sensitive words in different contexts, thereby achieving accurate identification of sensitive word information in the target data and risk level classification. Throughout the training process, operations such as model hyperparameter adjustment and training round setting can be implemented with reference to the mature experience and technical solutions of existing natural language processing model training.

[0041] In an embodiment of the present invention, feature vectors such as the text content, data field names, and association relationships of the target data can be input into a domain classification model. The domain classification model is built based on a convolutional neural network (CNN) or a Transformer architecture, and is trained on historical data in the fields of finance, medical care, education, etc. to learn the characteristic patterns of each field. The model outputs the probability value that the target data belongs to different fields, and selects the field with the highest probability value as the main field. If there are multiple fields whose probability values ​​differ by less than a preset threshold (such as 0.1), the fields with the top two or three probability values ​​are determined as the main field and the secondary field.

[0042] Furthermore, in an exemplary embodiment, the domain classification model is constructed based on a convolutional neural network, specifically including:

[0043] Input layer: This layer receives feature vectors such as the target data's text content, data field names, and relationships, and converts them into tensor forms that the model can process. For example, for text content, words can be mapped into fixed-dimensional vectors using word embedding technology, forming a two-dimensional matrix. Data field names and relationships can be encoded into specific numerical vectors, ultimately integrated into a multidimensional tensor input model.

[0044] Convolutional layer: Consists of multiple convolution kernels of different sizes, performing convolution operations on the input tensor using a sliding window. Convolution kernels can automatically extract local features in the data, such as phrase structures in the text and local correlation patterns between data fields. Convolution kernels of different sizes can capture features at different scales. Small convolution kernels focus on extracting detailed features, while large convolution kernels can capture more macroscopic semantic information. For example, a 3×3 convolution kernel can identify pairwise collocations in text, while a 5×5 convolution kernel can capture longer phrase structures. After the convolution operation, multiple feature maps are generated, each representing a specific feature pattern.

[0045] Pooling layers: Common pooling methods include max pooling and average pooling. These downsample the feature maps output by the convolutional layer, reducing data dimensionality and computational complexity while preserving important features. For example, max pooling selects the maximum value within each pooling window, highlighting the most significant features; average pooling calculates the average value within the window, smoothing the data. Pooling makes the model more robust to local variations in the data.

[0046] Fully connected layer: After flattening the feature map output by the pooling layer into a one-dimensional vector, it is mapped to a dimension equal to the number of domain categories through a fully connected layer. Neurons in the fully connected layer are connected to all neurons in the previous layer, enabling comprehensive feature processing and nonlinear transformation, and learning the mapping relationship between different feature combinations and domain categories.

[0047] Output layer: The Softmax activation function is used to convert the output of the fully connected layer into a probability distribution belonging to different fields. Each probability value represents the possibility that the target data belongs to the corresponding field. The field with the highest probability value is finally output as the main field.

[0048] Furthermore, in an exemplary embodiment, the domain classification model is constructed based on the Transformer architecture, specifically including:

[0049] The input layer also receives feature vectors of the target data. After embedding the text content into words, it adds a positional encoding to each word vector to represent the word's position in the sequence. Because the Transformer architecture itself does not capture sequence order, positional encoding enables the model to distinguish between words at different positions. Data field names and their relationships are also encoded and fused with text features before being input into the model.

[0050] Multi-head attention layer: This is the core module of the Transformer. Through parallel computation across multiple heads (attention units), it captures dependencies and semantic associations between words or data fields in the input data from different perspectives. Each head independently calculates attention weights, focusing on different parts of the input sequence. The results from multiple heads are then concatenated and linearly transformed to produce a feature representation rich in semantic information. For example, when processing text data, the multi-head attention mechanism can simultaneously focus on the relationship between the subject and predicate, the relationship between modifiers and the central word, and other factors, effectively capturing long-range dependencies.

[0051] Feedforward neural network layer: This layer performs further nonlinear transformation and feature extraction on the features output by the multi-head attention mechanism, enhancing the model's expressiveness. The feedforward neural network consists of two fully connected layers: the first layer increases the network's nonlinearity, and the second layer maps the features to the appropriate dimension.

[0052] Layer normalization layer: After the multi-head attention mechanism layer and the feedforward neural network layer, layer normalization is used to normalize the features of each sample, making the data distribution more stable, accelerating model training convergence, and improving the model's generalization ability.

[0053] Output layer: Similar to the CNN-based architecture, the Softmax activation function is used to output the probability value of the target data belonging to different fields. The main field is determined based on the probability. When the difference in the probability values ​​of multiple fields is less than the preset threshold, the main field and secondary field are determined.

[0054] In an embodiment of the present invention, the queue information of each candidate large model can be obtained in the following manner: the relevant data of each candidate large model task queue can be collected in real time through a monitoring program deployed on the large model server; the monitoring program polls the queue status at a fixed time interval (such as 5 seconds) to obtain information such as the amount of data waiting to be processed in the queue, the estimated processing time of each task, the task priority label, etc.; at the same time, based on historical task processing records, the exponential smoothing method is used to predict the load change trend of the queue within a set time period in the future, such as 10 minutes.

[0055] In an embodiment of the present invention, the network information between the data sending platform and each candidate large model can be obtained in the following manner: a network monitoring node can be deployed in the communication link between the data sending platform and each candidate large model, and ICMP or TCP probe packets can be sent periodically (e.g., every 30 seconds) to obtain indicators such as network delay, packet loss rate, and available bandwidth; and based on the network performance data of the past 24 hours, an autoregressive integrated moving average model (ARIMA) is used to predict the fluctuation range of network performance indicators within a set time period in the future, such as 30 minutes.

[0056] S300 , based on the reference information, selecting a target large model from a plurality of candidate large models for processing the target data.

[0057] Furthermore, S300 specifically includes:

[0058] S310, obtain the comprehensive score S of the i-th candidate large model i =W sen ×S i-sen +W dom ×S i-dom +W que ×S i-que +W net ×S i-net ; Among them, S i-sen is the sensitive word information dimension evaluation score of the i-th candidate large model, W sen is the weight of the sensitive word information dimension, S i-dom is the domain dimension evaluation score of the i-th candidate large model, Wdom is the weight of the domain dimension, S i-que is the queue information dimension evaluation score of the i-th candidate large model, W que is the weight of the queue information dimension, S i-net is the network information dimension evaluation score of the i-th candidate large model, W net is the weight of the network information, i ranges from 1 to n, and n is the number of candidate large models.

[0059] In the embodiment of the present invention, the weight of the sensitive word information dimension is positively correlated with the risk level of the sensitive word. The specific relationship is as follows:

[0060] Risk level classification: The risk level of sensitive words in the target data is classified into three levels: high, medium, and low. The classification is based on the following:

[0061] High risk level: When the number of sensitive words in the target data accounts for ≥10% of the total vocabulary, or contains core sensitive information such as citizen ID numbers, bank card numbers, medical diagnosis records, etc., it is judged to be a high risk level.

[0062] Medium risk level: If the proportion of sensitive words in the total vocabulary is between 3% (inclusive) and 10%, and it involves general sensitive information such as common names, phone numbers, addresses, etc., it is classified as a medium risk level.

[0063] Low risk level: The number of sensitive words accounts for less than 3% of the total vocabulary, and only a small amount of non-critical sensitive information is included, or the sensitive information has been desensitized to a certain extent. It is defined as a low risk level.

[0064] Weight value:

[0065] When the risk level of sensitive words is high, the weight of the sensitive word information dimension can be set to 0.3, indicating that in the comprehensive evaluation model, this dimension has a high degree of influence on the selection of large models. Priority is given to screening candidate large models with high-level data security protection capabilities, ISO27001 information security management system certification, and mature data encryption and desensitization functions.

[0066] If the risk level of sensitive words is medium risk, the weight can be set to 0.2. During the evaluation process, the focus is on whether the large model has the basic ability to identify and process sensitive information.

[0067] When at a low risk level, the weight can be set to 0.1. In this case, the influence of the sensitive word information dimension in the selection of the large model is relatively low, but it is still included in the evaluation system as an important reference factor.

[0068] Through the above settings, we can achieve a precise match between the weights of sensitive word information dimensions and the risk levels of sensitive words, ensuring that when processing target data with different levels of sensitivity, we can reasonably adjust the importance of each evaluation dimension and scientifically select an appropriate large model.

[0069] In an embodiment of the present invention, the sum of the weights of each dimension is equal to 1. The weights of the sensitive word information dimension and the domain dimension may be greater than the weights of the queue information dimension and the network information dimension. In one exemplary embodiment, the weight ratio of the domain dimension, the queue information dimension, and the network information dimension may be 3:2:2.

[0070] Furthermore, S i-sen The following conditions must be met:

[0071] If the i-th candidate large model is suitable for processing sensitive word information in the target data, S i-sen =1, otherwise, S i-sen =0.

[0072] In this embodiment of the present invention, if the i-th candidate large model meets the following conditions, it is suitable for processing sensitive word information in the target data:

[0073] The candidate large model has passed the national information security level protection level 3 and above certification;

[0074] Built-in mature automatic sensitive word recognition algorithm, which can accurately identify sensitive words corresponding to the risk level of target data;

[0075] It has complete data encryption and desensitizing functions, which can provide full-process security protection for sensitive data and ensure the security and compliance of data during processing.

[0076] Furthermore, S i-dom The following conditions must be met:

[0077] Let N i is the amount of training data of the i-th candidate large model in the main domain to which the target data belongs, C i is the processing accuracy of the i-th candidate large model in the historical tasks of its main field;

[0078] If N i >N1 and C i ≥C1, indicating that the model has sufficient training data and high processing accuracy in the main domain of the target data, S i-dom =1.

[0079] If N2≤N i ≤N1 and C i ≥C2, or N i >N1 and C3≤C i≤C1, indicating that the model is close to the optimal standard in terms of training data volume or processing accuracy, S i-dom =0.5; wherein, N1 is the first preset value, N2 is the second preset value, C1 is the first preset accuracy, C2 is the second preset accuracy, C3 is the third preset accuracy, and C3<C2<C1.

[0080] The rest of the cases indicate that the model is less applicable to the domain of the target data. i-dom =0.

[0081] In the embodiment of the present invention, N1 may be 100,000, N2 may be 50,000, C1 may be 90%, C2 may be 85%, and C3 may be 80%.

[0082] Furthermore, S i-que The following conditions must be met:

[0083] Assume NW i is the current queue waiting data volume of the i-th candidate large model, NC i is the queue capacity of the i-th candidate large model, L i Predict the task load for the first set time period in the future for the i-th candidate large model;

[0084] If NW i <(k1×NC i ), L i ≤(k2×NC i ), and the task priority and task type are the same as those of the target data, indicating that the model has sufficient current and future processing capabilities and is suitable for the target data. i-que =1;

[0085] If (k1×NC i )≤NW i ≤(k3×NC i ), and the task priority is the same as the priority of the target data, or the task type is the same as the task type of the target data, indicating that the model has processing capabilities under some conditions, S i-que =0.5; wherein k1 is the first coefficient, k2 is the second coefficient, k3 is the third coefficient, and k1<k3<k2 is satisfied.

[0086] In other cases, S i-que =0.

[0087] Tasks in a large model's task queue are typically assigned different priorities based on factors such as urgency and importance, with common grading levels being high, medium, and low. When the target data processing requirements are also assigned a corresponding priority, if the priority of the tasks currently waiting in the candidate large model's queue matches the priority of the target data, the "same priority" condition is met. For example, if the target data is a financial transaction risk assessment task that requires immediate processing and is marked as high priority, a large model that has fewer high-priority tasks in its task queue and is able to quickly respond to this data processing requirement will meet the "same priority" requirement. Different target data corresponds to different processing task types, such as text classification, data prediction, and image recognition. If the task type currently in the candidate large model's queue matches the target data's processing task type, it indicates that the model is more suitable for the target data in terms of processing flow and resource allocation. For example, if the target data is a batch of text reviews requiring sentiment analysis, a text classification task in natural language processing, then a large model with extensive experience in similar sentiment analysis tasks in its task queue and optimized resources for this task will meet the "same task type" condition.

[0088] In an embodiment of the present invention, the first set time period can be 10 minutes. k1 can be set to 0.3, which is used to measure the reasonable threshold of the current queue waiting data volume. k2 can be set to 0.6, which is used to define the acceptable upper limit of the future task load. k3 can be set to 0.5.

[0089] Furthermore, S i-net The following conditions must be met:

[0090] Assume T i-delay is the network delay between the data sending platform and the i-th candidate large model, D i is the packet loss rate of the i-th candidate large model, AW i is the available bandwidth of the i-th candidate large model, and AW0 is the basic bandwidth required for target data transmission.

[0091] If T i-delay <T1 delay 、D i <d1、AW i > h1×AW0, and if the network performance prediction model determines that the network performance in the second set time period in the future still meets the above conditions, it indicates that the network condition is good and stable, S i-net =1.

[0092] If T1 delay ≤T i-delay ≤T2 delay , or, d1≤D i ≤d2, or h2×AW0≤AW i≤h1×AW0, and if the network performance prediction model determines that there is no deterioration trend in the network performance within the second set time period in the future, it means that the network performance is in an acceptable critical state, S i-net =0.5; where T1 delay The first preset network delay, T2 delay is the second preset network delay, d1 is the first preset packet loss rate, d2 is the second preset packet loss rate, h1 is the first bandwidth coefficient, and h2 is the second bandwidth coefficient.

[0093] In other cases, S i-net =0.

[0094] In an embodiment of the present invention, the second set time period may be 30 minutes. delay Can be 50ms, T1 delay It can be 80ms, d1 can be 1%, d2 can be 3%, h1 can be 1.5, and h2 can be 1.2.

[0095] In this embodiment of the present invention, "no trend in network performance deterioration" means, based on existing network performance data and prediction models, estimating the network status between the data transmission platform and the candidate large model over the next period of time (e.g., 30 minutes), and determining that the network performance indicators will not deteriorate. In the network information dimension evaluation score, when network performance reaches a critical state such as "network latency between 50-80ms, or packet loss rate between 1%-3%, or bandwidth between 1.2-1.5 times the required value," in addition to meeting the current indicators, it is necessary to use time series analysis and machine learning prediction algorithms (such as the Autoregressive Integrated Moving Average (ARIMA) model and LSTM network) combined with historical network performance data and real-time monitoring data to predict that within the next 30 minutes, network latency will not continue to rise above 80ms, packet loss rate will not further increase by more than 3%, and available bandwidth will not drop below 1.2 times the target data transmission demand. For example, if the current network latency is 70ms, which is within the 50-80ms range, an ARIMA model is used to analyze the network latency data for the past 24 hours and predict the latency value for the next 30 minutes. If the prediction shows that the latency will not exceed 80ms, or even shows a decreasing trend, it indicates that network performance is not deteriorating. Conversely, if the predicted latency exceeds 80ms, even if the 50-80ms condition is currently met, the large model cannot be considered to receive a score of 0.5 in the network information dimension, but only a score of 0. This requirement ensures that the network environment can continuously and stably support data transmission while the selected large model processes the target data, avoiding the impact of sudden network performance drops on data processing efficiency and quality.

[0096] In an embodiment of the present invention, the network performance prediction model may be a time series prediction model, a machine learning model, a deep learning model, or the like.

[0097] S320 , based on the comprehensive scores of the n candidate large models, select a target large model from the n candidate large models.

[0098] Furthermore, S320 specifically includes:

[0099] S321, obtain the maximum comprehensive score among the comprehensive scores of n candidate large models. If the number of maximum comprehensive scores P=1, the candidate large model corresponding to the maximum comprehensive score is used as the target large model; if P>1, execute S322.

[0100] S322, obtain the maximum sensitive word information dimension evaluation score in P comprehensive scores. If the number of maximum sensitive word information dimension evaluation scores P1=1, then the candidate large model corresponding to the maximum sensitive word information dimension evaluation score is used as the target large model. If P1>1, execute S323.

[0101] S323, obtain the maximum domain dimension evaluation score among P comprehensive scores. If the number of maximum domain dimension evaluation scores P2=1, then the candidate large model corresponding to the maximum domain dimension evaluation score is used as the target large model. If P2>1, execute S324.

[0102] S324, obtain the maximum queue information dimension evaluation score among P comprehensive scores. If the number of maximum queue information dimension evaluation scores P3=1, then the candidate large model corresponding to the maximum queue information dimension evaluation score is used as the target large model. If P3>1, execute S325.

[0103] S325, obtain the maximum network information dimension evaluation score among P comprehensive scores. If the number of maximum network information dimension evaluation scores P4=1, then the candidate large model corresponding to the maximum network information dimension evaluation score is used as the target large model. If P4>1, randomly select a candidate large model from the P candidate large models corresponding to the P maximum comprehensive scores as the target large model.

[0104] (Example 2)

[0105] Based on the same inventive concept, an embodiment of the present invention provides a distribution system for processing target data using a large model. The system is implemented based on multiple candidate large models and includes:

[0106] The data acquisition module is used to acquire the target data to be processed.

[0107] The reference information acquisition module is used to obtain sensitive word information in the target data, the field to which the target data belongs, the queue information of each candidate large model, and the network information between the data sending platform and each candidate large model as reference information.

[0108] The target model acquisition module is used to select a target large model from multiple candidate large models based on the reference information to process the target data.

[0109] The target model acquisition module specifically includes:

[0110] Comprehensive score acquisition unit, used to obtain the comprehensive score S of the i-th candidate large model i =W sen ×S i-sen +W dom ×S i-dom +W que ×S i-que +W net ×S i-net ; Among them, S i-sen is the sensitive word information dimension evaluation score of the i-th candidate large model, W sen is the weight of the sensitive word information dimension, S i-dom is the domain dimension evaluation score of the i-th candidate large model, W dom is the weight of the domain dimension, S i-que is the queue information dimension evaluation score of the i-th candidate large model, W que is the weight of the queue information dimension, S i-net is the network information dimension evaluation score of the i-th candidate large model, W net is the weight of the network information, i ranges from 1 to n, and n is the number of candidate large models.

[0111] The target model selection unit is used to select a target large model from the n candidate large models based on the comprehensive scores of the n candidate large models.

[0112] The system can be used to perform Figure 1 Therefore, for the functions that can be realized by each functional module of the system, please refer to Figure 1 The description of the illustrated embodiment is omitted for brevity.

[0113] (Example 3)

[0114] This embodiment provides a method for allocating target data using a large model. The method is implemented based on multiple candidate large models, such as Figure 2 As shown, the method may include the following steps:

[0115] S10, obtaining target data to be processed.

[0116] S20, if the number of characters in the target data is less than a set number, randomly selecting a large model from a plurality of candidate large models as a target large model for processing the target data; if the number of characters in the target data is greater than or equal to the set number, executing S30;

[0117] S30, obtaining sensitive word information in the target data, the field to which the target data belongs, queue information of each candidate large model, and network information between the data sending platform and each candidate large model as reference information;

[0118] S40: Select a target large model from a plurality of candidate large models based on the reference information.

[0119] S40 specifically includes:

[0120] S401, obtain the comprehensive score S of the i-th candidate large model i =W sen ×S i-sen +W dom ×S i-dom +W que ×S i-que +W net ×S i-net ; Among them, S i-sen is the sensitive word information dimension evaluation score of the i-th candidate large model, W sen is the weight of the sensitive word information dimension, S i-dom is the domain dimension evaluation score of the i-th candidate large model, W dom is the weight of the domain dimension, S i-que is the queue information dimension evaluation score of the i-th candidate large model, W que is the weight of the queue information dimension, S i-net is the network information dimension evaluation score of the i-th candidate large model, W net is the weight of the network information, i ranges from 1 to n, and n is the number of candidate large models.

[0121] S402 : Based on the comprehensive scores of the n candidate large models, a target large model is selected from the n candidate large models.

[0122] This embodiment provides an allocation method for processing target data using a large model. The main framework of the method provided in Example 1 is basically the same. The core difference is that this embodiment optimizes the large model selection process based on the number of characters in the target data. When the number of characters in the target data is small, the process is simplified to directly randomly select a large model to balance processing efficiency and resource consumption; when the number of characters is large, the same multi-dimensional reference information evaluation method as in Example 1 is used to select a large model to ensure the accuracy and efficiency of complex data processing. In this way, the large model allocation method can better adapt to data processing needs of different scales and enhance the flexibility and practicality of the overall data processing solution.

[0123] Furthermore, in an embodiment of the present invention, the set quantity can be obtained by the following steps:

[0124] S201: Input m sample input sentences into n candidate large models respectively, and obtain the processing time record set T of the i-th candidate large model. i ={T i1 , T i2 ,……,T ij ,……,T im}, where T ij It is the processing time of the i-th candidate large model for the j-th sample input sentence, and the value of j ranges from 1 to m.

[0125] In the embodiment of the present invention, the m sample input sentences need to cover different data scales, data characteristics, and business scenarios to ensure that the acquired processing time data is comprehensive and reliable.

[0126] In an embodiment of the present invention, when each sample input statement enters the large model for processing, a timing program is started. When the large model completes the processing task of the sample statement, the timing program stops and the time from the start of input to the end of processing is recorded as the processing time of the large model.

[0127] In the embodiment of the present invention, the time of entering the large model can be recorded in the following manner:

[0128] Code insertion marker: In the data transmission code logic, insert a time-recording code when the target data or sample input statement begins to be transmitted to the big model. For example, in a data transmission script written in Python, use the time module to record the current timestamp, import time, and add start_time=time.time() before the statement that starts data transmission to mark the starting time of data entering the big model.

[0129] Message queues and logging: If you use message queues to manage data transmission, you can log the current time when data is placed in the message queue and awaits retrieval by the large model. For example, if you use RabbitMQ as a message queue, after the producer sends data to the queue, use a logging tool (such as Python's logging module) to record "Data sent to message queue, time: " + str(datetime.now())." When the large model subsequently retrieves data from the message queue, it can use the logged time to determine when the data entered the large model.

[0130] In the embodiment of the present invention, the time for the large model to complete the processing task can be determined by the following method:

[0131] Output signal monitoring: After completing data processing, large models typically output processing results. A monitoring mechanism can be set up in the code that receives the processing results. When the complete processing result data is received, the current time is recorded as the processing completion time. For example, if the large model returns the processing result through an API interface, add end_time = time.time() to the response receiving function of the API call to obtain the time when the large model completed the processing task.

[0132] Task status callback: Some large model services provide a task status query or callback mechanism. For example, when submitting a data processing task, a task ID is assigned. The task status query interface is periodically called to check whether the task is in the "Completed" state. Alternatively, a callback address can be set. When the large model completes processing a task, a completion notification is sent to that address, and the time of receipt is recorded.

[0133] Resource Usage Monitoring: Large models consume a certain amount of computing resources (such as CPU and GPU usage) when processing data. System resource monitoring tools can be used to monitor the resource usage of large model processes in real time. When resource usage drops significantly and stabilizes at a low level for a period of time (e.g., 1-2 seconds), the large model can be considered to have completed its processing task and the time at which this occurred can be recorded. For example, in Linux, use the top or nvidia-smi commands to obtain resource usage information, and use scripts for real-time analysis and assessment.

[0134] In an embodiment of the present invention, the m sample input sentences can be layered according to the number of characters, and the samples in each layer are input into n candidate large models in a multi-threaded parallel manner, with the number of characters in the m sample input sentences increasing in sequence.

[0135] S202: Obtain the processing time record set T corresponding to the j-th sample input statement. j ={T 1j , T2j ,……,T ij ,……,T nj}, that is, to sort out the processing time sets of n candidate large models processing the same sample input sentence.

[0136] S203, obtain T j The corresponding time difference set △T j ={△T j 12 , △T j 23 ,……,△T j （i-1）i ,……,△T j （n -1）n}, △T j （i-1）i is △T j The i-1th time difference in j （i-1）i =T ij -T （i-1）j ;

[0137] By calculating the time difference between adjacent large models in processing the same statement, we can preliminarily reflect the efficiency differences when different large models process the same data.

[0138] S204, obtain ΔT j The corresponding degree of dispersion , △T jr is △T j The rth time difference in the time interval, r ranges from 1 to n-1, Avg (△T j ) is △T j The average of the n-1 time differences in .

[0139] In the embodiment of the present invention, the degree of dispersion is used to quantify the degree of fluctuation in the processing time difference when different large models process the same sample input sentence.

[0140] S205, traverse the discrete degree set DS={DS1, DS2, ..., DS j ,……,DS m}, for the discrete degree DS traversed j , if DS j >DS0, it means that the efficiency of different large models in processing the sample input sentence is quite different. jThe number of characters in the corresponding sample input sentence is used as the set number. This is because when the data volume reaches this number of characters, the processing efficiency of each candidate large model begins to diverge significantly. Using this as the threshold, in actual applications, for data with less than this number of characters, randomly selecting a large model for processing has a smaller impact on overall efficiency. However, for data with a number of characters or greater, selecting a large model through subsequent multi-dimensional evaluation can more effectively match data processing requirements and achieve a balance between efficiency and accuracy.

[0141] In this embodiment of the present invention, DS0 is a threshold for setting the degree of discreteness. This threshold can be set based on the actual business requirements for large-scale model processing efficiency consistency. For example, in scenarios where high efficiency and stability are required, the DS0 black value can be set to a smaller value. Conversely, in scenarios where efficiency and stability are less required, DS0 can be appropriately increased.

[0142] Those skilled in the art should understand that since the number of characters in the m sample input sentences increases in sequence, during the traversal process, as long as the DS j >DS0, the traversal of the discrete degree set will no longer continue and the traversal program will end.

[0143] (Example 4)

[0144] Based on the same inventive concept, an embodiment of the present invention provides a distribution system for processing target data using a large model. The system is implemented based on multiple candidate large models and includes:

[0145] The data acquisition module is used to acquire the target data to be processed.

[0146] A target model acquisition module is used to randomly select a large model from multiple candidate large models as a target large model for processing the target data when the number of characters in the target data is less than a set number, and to obtain reference information in the target data when the number of characters in the target data is greater than or equal to a set number, and select a target large model from multiple candidate large models based on the reference information.

[0147] The parameter information includes sensitive word information in the target data, the field to which the target data belongs, queue information of each candidate large model, and network information between the data sending platform and each candidate large model.

[0148] Furthermore, the target model acquisition module specifically includes:

[0149] a judging unit, configured to obtain the number of characters in the target data and to judge whether the number of characters in the target data is less than a set number;

[0150] Comprehensive score acquisition unit, used to obtain the comprehensive score S of the i-th candidate large modeli =W sen ×S i-sen +W dom ×S i-dom +W que ×S i-que +W net ×S i-net ; Among them, S i-sen is the sensitive word information dimension evaluation score of the i-th candidate large model, W sen is the weight of the sensitive word information dimension, S i-dom is the domain dimension evaluation score of the i-th candidate large model, W dom is the weight of the domain dimension, S i-que is the queue information dimension evaluation score of the i-th candidate large model, W que is the weight of the queue information dimension, S i-net is the network information dimension evaluation score of the i-th candidate large model, W net is the weight of the network information, i ranges from 1 to n, and n is the number of candidate large models.

[0151] A target model selection unit is used to randomly select a large model from multiple candidate large models as a target large model when the number of characters in the target data is less than a set number, and to select a target large model from n candidate large models based on the comprehensive scores of the n candidate large models.

[0152] The system can be used to perform Figure 2 Therefore, for the functions that can be realized by each functional module of the system, please refer to Figure 2 The description of the illustrated embodiment is omitted for brevity.

[0153] An embodiment of the present invention also provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are configured to execute the method described in the embodiment of the present invention.

[0154] An embodiment of the present invention further provides a computer-readable storage medium storing computer-executable instructions, wherein the computer instructions are used to execute the method described in the embodiment of the present invention.

[0155] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in the present invention can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in the present invention can be achieved. This is not limited herein.

[0156] The above specific embodiments do not limit the scope of protection of the present invention. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A method for distributing target data using a large model, characterized in that: The method is implemented based on multiple candidate large models and includes the following steps: S100, obtaining target data to be processed; S200, obtaining sensitive word information in the target data, the field to which the target data belongs, queue information of each candidate large model, and network information between the data sending platform and each candidate large model as reference information; S300, based on the reference information, selecting a target large model from a plurality of candidate large models for processing the target data; S300 specifically includes: S310, obtain the comprehensive score S of the i-th candidate large model i =W sen ×S i-sen +W dom ×S i-dom +W que ×S i-que +W net ×S i-net ; Among them, S i-sen is the sensitive word information dimension evaluation score of the i-th candidate large model, W sen is the weight of the sensitive word information dimension, S i-dom is the domain dimension evaluation score of the i-th candidate large model, W dom is the weight of the domain dimension, S i-que is the queue information dimension evaluation score of the i-th candidate large model, W que is the weight of the queue information dimension, S i-net is the network information dimension evaluation score of the i-th candidate large model, W net is the weight of the network information, i ranges from 1 to n, and n is the number of candidate large models; S320, selecting a target large model from the n candidate large models based on the comprehensive scores of the n candidate large models; S i-dom The following conditions must be met: Let N i is the amount of training data of the i-th candidate large model in the main domain to which the target data belongs, C i is the processing accuracy of the i-th candidate large model in the historical tasks of its main field; If N i >N1 and C i ≥C1, S i-dom =1; If N2≤N i ≤N1 and C i ≥C2, or N i >N1 and C3≤C i ≤C1,S i-dom =0.5; wherein N1 is the first preset value, N2 is the second preset value, C1 is the first preset accuracy rate, C2 is the second preset accuracy rate, C3 is the third preset accuracy rate, and C3<C2<C1; In other cases, S i-dom =0; S i-que The following conditions must be met: Assume NW i is the current queue waiting data volume of the i-th candidate large model, NC i is the queue capacity of the i-th candidate large model, L i Predict the task load for the first set time period in the future for the i-th candidate large model; If NW i <(k1×NC i ), L i ≤(k2×NC i ), and the task priority and task type are the same as those of the target data, S i-que =1; If (k1×NC i )≤NW i ≤(k3×NC i ), and the task priority is the same as the priority of the target data, or the task type is the same as the task type of the target data, S i-que =0.5; where k1 is the first coefficient, k2 is the second coefficient, and k3 is the third coefficient, and k1<k3<k2; In other cases, S i-que =0.

2. The method according to claim 1, characterized in that S320 specifically includes: S321, obtaining the maximum comprehensive score among the comprehensive scores of n candidate large models. If the number of maximum comprehensive scores P = 1, the candidate large model corresponding to the maximum comprehensive score is used as the target large model; if P>1, executing S322; S322: Obtain the maximum sensitive word information dimension evaluation score among the P comprehensive scores. If the number of maximum sensitive word information dimension evaluation scores P1 = 1, then use the candidate macro model corresponding to the maximum sensitive word information dimension evaluation score as the target macro model. If P1 > 1, execute S323. S323, obtaining the maximum domain dimension evaluation score among the P comprehensive scores. If the number of maximum domain dimension evaluation scores P2 = 1, then the candidate large model corresponding to the maximum domain dimension evaluation score is used as the target large model. If P2>1, execute S324; S324, obtaining the maximum queue information dimension evaluation score among the P comprehensive scores. If the number of maximum queue information dimension evaluation scores P3 = 1, then the candidate large model corresponding to the maximum queue information dimension evaluation score is used as the target large model. If P3>1, execute S325; S325, obtain the maximum network information dimension evaluation score among P comprehensive scores. If the number of maximum network information dimension evaluation scores P4=1, then the candidate large model corresponding to the maximum network information dimension evaluation score is used as the target large model. If P4>1, randomly select a candidate large model from the P candidate large models corresponding to the P maximum comprehensive scores as the target large model.

3. The method according to claim 1, characterized in that The weight of the sensitive word information dimension is positively correlated with the risk level of the sensitive word.

4. The method according to claim 1, wherein S i-sen The following conditions must be met: If the i-th candidate large model is suitable for processing sensitive word information in the target data, S i-sen =1, otherwise, S i-sen =0.

5. The method according to claim 1, wherein S i-net The following conditions must be met: Assume T i-delay is the network delay between the data sending platform and the i-th candidate large model, D i is the packet loss rate of the i-th candidate large model, AW i is the available bandwidth of the i-th candidate large model, and AW0 is the basic bandwidth required for target data transmission; If T i-delay <T1 delay 、D i <d1、AW i > h1×AW0, and if the network performance prediction model determines that the network performance in the second set time period in the future still meets the above conditions, S i-net =1; If T1 delay ≤T i-delay ≤T2 delay , or, d1≤D i ≤d2, or, h2×AW0≤AW i ≤h1×AW0, and based on the network performance prediction model, it is determined that there is no deterioration trend in the network performance within the second set time period in the future, S i-net =0.5; where T1 delay The first preset network delay, T2 delay is the second preset network delay, d1 is the first preset packet loss rate, d2 is the second preset packet loss rate, h1 is the first bandwidth coefficient, and h2 is the second bandwidth coefficient; In other cases, S i-net =0.

6. A distribution system for processing target data using a large model, characterized in that: The system is implemented based on multiple candidate large models, and includes: A data acquisition module, used to acquire target data to be processed; A reference information acquisition module is used to obtain sensitive word information in the target data, the field to which the target data belongs, queue information of each candidate large model, and network information between the data sending platform and each candidate large model as reference information; a target model acquisition module, configured to select a target large model from a plurality of candidate large models based on the reference information, for use in processing the target data; The target model acquisition module specifically includes: Comprehensive score acquisition unit, used to obtain the comprehensive score S of the i-th candidate large model i =W sen ×S i-sen +W dom ×S i-dom +W que ×S i-que +W net ×S i-net ; Among them, S i-sen is the sensitive word information dimension evaluation score of the i-th candidate large model, W sen is the weight of the sensitive word information dimension, S i-dom is the domain dimension evaluation score of the i-th candidate large model, W dom is the weight of the domain dimension, S i-que is the queue information dimension evaluation score of the i-th candidate large model, W que is the weight of the queue information dimension, S i-net is the network information dimension evaluation score of the i-th candidate large model, W net is the weight of the network information, i ranges from 1 to n, and n is the number of candidate large models; A target model selection unit, configured to select a target large model from the n candidate large models based on a comprehensive score of the n candidate large models; S i-dom The following conditions must be met: Let N i is the amount of training data of the i-th candidate large model in the main domain to which the target data belongs, C i is the processing accuracy of the i-th candidate large model in the historical tasks of its main field; If N i >N1 and C i ≥C1, S i-dom =1; If N2≤N i ≤N1 and C i ≥C2, or N i >N1 and C3≤C i ≤C1,S i-dom =0.5; wherein N1 is the first preset value, N2 is the second preset value, C1 is the first preset accuracy rate, C2 is the second preset accuracy rate, C3 is the third preset accuracy rate, and C3<C2<C1; In other cases, S i-dom =0; S i-que The following conditions must be met: Assume NW i is the current queue waiting data volume of the i-th candidate large model, NC i is the queue capacity of the i-th candidate large model, L i Predict the task load for the first set time period in the future for the i-th candidate large model; If NW i <(k1×NC i ), L i ≤(k2×NC i ), and the task priority and task type are the same as those of the target data, S i-que =1; If (k1×NC i )≤NW i ≤(k3×NC i ), and the task priority is the same as the priority of the target data, or the task type is the same as the task type of the target data, S i-que =0.5; where k1 is the first coefficient, k2 is the second coefficient, and k3 is the third coefficient, and k1<k3<k2; In other cases, S i-que =0.

Images