Task processing method, information processing method based on target processing model, task processing system, and model training platform

Abstract

Provided in the embodiments of the present disclosure are a task processing method, an information processing method based on a target processing model, a task processing system, and a model training platform. The task processing method comprises: acquiring task data of a target task; and inputting the task data into a target processing model, so as to obtain a task processing result for the target task, wherein the target processing model is obtained by means of training an initial processing model on the basis of sample data, a target model weight of the target processing model comprises an initial weight and an incremental compression weight, the initial weight refers to a weight in the target processing model that is the same as an initial model weight of the initial processing model, the incremental compression weight is obtained by means of performing sparse compression on an incremental weight, and the incremental weight refers to a weight in the target processing model that is different from the initial model weight. Structured grouped sparse compression is performed on an incremental weight, such that the task processing delay is reduced, the resource utilization rate is increased, and the model precision is also ensured.

Description

[0001] Task processing method, information processing method based on target processing model, task processing system and model training platform The present disclosure claims priority to Chinese Patent Application No. 202410452231.X, filed on April 15, 2024 with the Chinese Patent Office, entitled "Task processing method, information processing method based on target processing model, task processing system and model training platform", the entire contents of which are incorporated herein by reference. TECHNICAL FIELD Embodiments of the present disclosure relate to the technical field of computer technology, and in particular to a task processing method, an information processing method based on a target processing model, a task processing system and a model training platform. BACKGROUND With the development of computer technology, large models have shown excellent performance. Supervised fine-tuning (SFT) is also widely used to improve the performance of large language models in various downstream tasks. However, providing services for large models consumes a lot of storage resources and computing resources of graphics processing units (GPUs), which is very expensive. Currently, a dedicated GPU memory and computing resources are usually provided for each model, and in the task processing process, multiple models are swapped in and out of the GPU memory according to the traffic of each model. However, due to the large number of model parameters, the loading time will be long, which leads to a sharp increase in total request delay, so there is an urgent need for a low-delay task processing solution. CONTENT In view of this, embodiments of the present disclosure provide a task processing method. One or more embodiments of the present disclosure also relate to an information processing method based on a target processing model, a task processing system, a model training platform, a task processing device, an information processing device based on a target processing model, a computing device, a computer-readable storage medium, and a computer program product, to solve the technical defects in the prior art that "due to the large number of model parameters, the loading time in the task processing process will be long, which leads to a sharp increase in total request delay". According to a first aspect of embodiments of the present disclosure, a task processing method is provided, comprising: obtaining task data of a target task; inputting the task data into a target processing model to obtain a task processing result of the target task, wherein the target processing model is obtained by training an initial processing model based on sample data, the target model weight of the target processing model includes an initial weight and an incremental compression weight, the initial weight refers to the weight in the target processing model that is the same as the initial model weight of the initial processing model, and the incremental compression weight is obtained based on grouping and sparse compression of an incremental weight, the incremental weight refers to the weight in the target processing model that is different from the initial model weight.According to a second aspect of the embodiments of the present disclosure, a method for processing information based on a target processing model is provided, including: receiving a task generation request, wherein the task generation request includes request information; obtaining the target processing model based on the request information, wherein the target processing model is obtained by training an initial processing model based on sample data, the target model weight of the target processing model includes an initial weight and an incremental compression weight, the initial weight refers to a weight in the target processing model that is the same as an initial model weight of the initial processing model, the incremental compression weight is obtained by performing group sparse compression on an incremental weight, and the incremental weight refers to a weight in the target processing model that is different from the initial model weight; and generating task information based on the target processing model, wherein the task information is used to execute a target task. According to a third aspect of the embodiments of the present disclosure, a task processing apparatus is provided, including: a first obtaining module configured to obtain task data of a target task; and an input module configured to input the task data into a target processing model to obtain a task processing result of the target task, wherein the target processing model is obtained by training an initial processing model based on sample data, the target model weight of the target processing model includes an initial weight and an incremental compression weight, the initial weight refers to a weight in the target processing model that is the same as an initial model weight of the initial processing model, the incremental compression weight is obtained by performing group sparse compression on an incremental weight, and the incremental weight refers to a weight in the target processing model that is different from the initial model weight. According to a fourth aspect of the embodiments of the present disclosure, a method for processing information based on a target processing model is provided, including: a receiving module configured to receive a task generation request, wherein the task generation request includes request information; a second obtaining module configured to obtain the target processing model based on the request information, wherein the target processing model is obtained by training an initial processing model based on sample data, the target model weight of the target processing model includes an initial weight and an incremental compression weight, the initial weight refers to a weight in the target processing model that is the same as an initial model weight of the initial processing model, the incremental compression weight is obtained by performing group sparse compression on an incremental weight, and the incremental weight refers to a weight in the target processing model that is different from the initial model weight; and a generating module configured to generate task information based on the target processing model, wherein the task information is used to execute a target task.According to a fifth aspect of the embodiments of the present disclosure, a task processing system is provided, including a processing unit and a storage unit, the storage unit is configured to store initial weights and increment compression weights of a target processing model, the target processing model is obtained by training an initial processing model based on sample data, the initial weights refer to weights in the target processing model that are the same as initial model weights of the initial processing model, the increment compression weights are obtained based on grouping and sparse compression of increment weights, the increment weights refer to weights in the target processing model that are different from the initial model weights; the processing unit is configured to obtain task data of a target task; obtain the initial weights and the increment compression weights from the storage unit; construct the target processing model according to the initial weights and the increment compression weights; input the task data into the target processing model to obtain a task processing result of the target task. According to a sixth aspect of the embodiments of the present disclosure, a model training platform is provided, including a request interface unit and a model unit; the request interface unit is configured to receive a model training request, wherein the model training request includes request information; the model unit is configured to obtain an initial processing model; train the initial processing model based on the request information to obtain a target processing model after training, wherein target model weights of the target processing model include initial weights and increment compression weights, the initial weights refer to weights in the target processing model that are the same as initial model weights of the initial processing model, the increment compression weights are obtained based on grouping and sparse compression of increment weights, the increment weights refer to weights in the target processing model that are different from the initial model weights. According to a seventh aspect of the embodiments of the present disclosure, a computing device is provided, including a storage and a processor; the storage is configured to store computer programs / instructions, and the processor is configured to execute the computer programs / instructions, which realize the steps of the method provided in the first aspect or the second aspect when executed by the processor. According to an eighth aspect of the embodiments of the present disclosure, a computer readable storage medium is provided, which stores computer programs / instructions, which realize the steps of the method provided in the first aspect or the second aspect when executed by the processor. According to a ninth aspect of the embodiments of the present disclosure, a computer program product is provided, including computer programs / instructions, which realize the steps of the method provided in the first aspect or the second aspect when executed by the processor.The task processing method provided by one embodiment of the present disclosure includes: obtaining task data of a target task; inputting the task data into a target processing model to obtain a task processing result of the target task, wherein the target processing model is obtained by training an initial processing model based on sample data, target model weights of the target processing model include initial weights and incremental compression weights, the initial weights refer to weights in the target processing model that are the same as initial model weights of the initial processing model, and the incremental compression weights are obtained based on group sparse compression of incremental weights, the incremental weights refer to weights in the target processing model that are different from the initial model weights. Through the structural group sparse compression of the incremental weights, the incremental compression weights consume less memory resources due to the small amount of parameters of the incremental weights compared to the target model weights of the target processing model, thereby realizing the quick switching of the incremental weights in and out of the GPU memory, reducing the task processing delay, improving the resource utilization, and ensuring the model accuracy. BRIEF DESCRIPTION OF DRAWINGS FIG. 1 is an architecture diagram of a task processing system according to one embodiment of the present disclosure; FIG. 2 is a flowchart of a task processing method according to one embodiment of the present disclosure; FIG. 3 is a flowchart of an information processing method based on a target processing model according to one embodiment of the present disclosure; FIG. 4 is a structural schematic diagram of a model training platform according to one embodiment of the present disclosure; FIG. 5 is a structural schematic diagram of a task processing device according to one embodiment of the present disclosure; FIG. 6 is a structural schematic diagram of an information processing device based on a target processing model according to one embodiment of the present disclosure; and FIG. 7 is a structural block diagram of a computing device according to one embodiment of the present disclosure. DETAILED DESCRIPTION In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present disclosure. However, the present disclosure can be practiced without the specific details, other than in the examples described herein. Those skilled in the art will recognize that the present disclosure can be practiced with modifications other than those described herein, and would recognize the applicability of the present disclosure to other embodiments. Accordingly, the present disclosure is not limited to the embodiments described herein. The terminology used in one or more embodiments of the present disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of one or more embodiments of the present disclosure. As used in one or more embodiments of the present disclosure and the accompanying claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "and / or," as used in one or more embodiments of the present disclosure, are used to meanFor example, without departing from the scope of one or more embodiments of the present disclosure, the first can also be referred to as the second, and similarly, the second can also be referred to as the first. Depending on the context, the word "if" as used herein can be interpreted as "when" or "upon" or "in response to determining". In addition, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in one or more embodiments of the present disclosure are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation interfaces for users to choose authorization or refusal. In one or more embodiments of the present disclosure, a large model refers to a deep learning model with a large number of model parameters, usually containing hundreds of millions, tens of billions, hundreds of billions, thousands of billions or even tens of billions of model parameters. The large model can also be called a foundation model, which is pre-trained by a large amount of unlabeled corpus to output a pre-training model with hundreds of millions of parameters. Such a model can adapt to a wide range of downstream tasks, and the model has good generalization ability, such as large language models (LLM, Large Language Model), multi-modal pre-training models (multi-modal pre-training model), etc. In practical applications, the large model only needs a small amount of samples to fine-tune the pre-training model and can be applied to different tasks. The large model can be widely used in natural language processing (NLP, Natural Language Processing), computer vision and other fields, and can be applied to visual question answering (VQA, Visual Question Answering), image description.

[0002] (IC, Image Caption), image generation and other computer vision field tasks, and text-based sentiment classification, text summarization generation, machine translation and other natural language processing field tasks. The main application scenarios of the large model include digital assistants, intelligent robots, search, online education, office software, e-commerce, intelligent design, etc. First, the noun terms involved in one or more embodiments of the present disclosure are explained.

[0003] GPU: GPU is a microprocessor specially designed for efficient parallel processing of large amounts of graphics and image data. GPU was originally developed to accelerate the graphics rendering work of computer systems, especially for three-dimensional graphics and complex image processing algorithms. With the development of technology, GPU is not limited to processing graphics information, but also shows strong ability in general parallel computing field, widely used in scientific computing, machine learning, artificial intelligence and other fields. In addition to processing graphics data, GPU also includes GPGPU (General Purpose GPU), which tends to general computing, mainly used in high-performance computing, parallel computing scenarios, and is the main hardware platform for running large models.

[0004] SFT: refers to using labeled data to train a pre-trained pre-trained model to make it more suitable for a specific task.

[0005] SparseGPT (Generative Pretrained Transformer): A sparse technology for large models such as GPT models, designed to reduce the number of model parameters through pruning without significantly sacrificing model performance, thereby reducing the storage requirements and computational cost of the model. This technology can effectively sparse the parameters of GPT series models, i.e. remove a certain proportion of redundant or non-critical weights, so that the model can greatly reduce the required computing resources while maintaining high accuracy. Convolutional Neural Network (CNN, Convolutional Neural Networks) model: A multi-layer deep learning model with forward propagation and backpropagation, with convolution kernels (filter) for processing feature data. Recurrent Neural Network (RNN, Recurrent Neural Network) model: A recursive deep learning model that processes vectors in the direction of the chain connection of the intermediate layers. Long Short Term Memory (LSTM, Long Short Term Memory) model: A deep learning model with the ability to remember long and short term information, with convolution kernels (filter) for processing feature data. Deep Self-Attention Model (Transformer model): A deep learning architecture based on attention mechanism (Attention) for processing sequence data such as natural language. Bidirectional Encoder Representations from Transformers (BERT, Bidirectional Encoder Representations from Transformers): A special Transformer model that uses a bidirectional Transformer encoder and large-scale unlabeled text data for training. With the development of computer technology, large models have shown excellent performance and are widely used in various scenarios, such as chat robots, coding assistants, writing assistants, etc. In order to improve the performance of large models in various downstream scenarios (such as mathematics, programming, law, medicine), better results are usually obtained by fine-tuning a pre-trained model based on the downstream task, resulting in a full-parameter fine-tuned model that performs better on each downstream task. The structure of these models is the same, only the weight data is slightly different. However, efficient deployment of multiple full-parameter fine-tuned models in the inference stage has become a new challenge.On the one hand, a large number of full-parameter fine-tuned models are needed due to a large number of downstream tasks in real scenes. On the other hand, the request difference of different models is large, and if all models are deployed at the same time, it will lead to large resource demand and low utilization rate. If the model is loaded according to the request flow, the model with large flow is put into the GPU, and the model with small flow is put out of the GPU. Due to the large number of model parameters, the loading time will be long, which will lead to a sharp increase in total request delay. Since multiple full-parameter fine-tuned models are all obtained by supervised fine-tuning based on the same pre-trained model, if the incremental weight between the large model of each downstream task and the pre-trained model is compressed efficiently, the fine-tuned model can be quickly put into and put out of the GPU memory, the model switching is accelerated, and the utilization efficiency of resources is improved. Moreover, when deploying a service system of multiple full-parameter fine-tuned models, the model switching efficiency is mainly related to the compression rate of the incremental weight, but the compression rate is also related to the accuracy effect. The higher the compression rate, the higher the switching efficiency and the higher the utilization rate of computing resources, and vice versa. However, the higher the compression rate, the more information loss, the lower the accuracy of the model, and the worse the effect on the downstream task. Therefore, designing an algorithm that can maintain accuracy at a high compression rate is a very valuable challenge. Currently, the incremental weight of the model can be compressed in the following two ways: First, the SparseGPT sparse compression method can be used to compress the incremental weight of the model, which can compress the incremental weight of the model by 6 to 8 times, increase the service throughput by 1.5 to 3 times, and maintain high model quality at the same time, thereby effectively serving multiple full-parameter fine-tuned models at the same time. However, when the compression rate exceeds 75%, the accuracy loss of the model is obviously larger, and the effect cannot be close to that of the uncompressed model, and the sparse compression is unstructured sparse, which requires additional location index information when saving data, and in fact the theoretical compression rate cannot be achieved. Generally speaking, the delay of unstructured sparse on the system is also large. Second, random sparsification and rescaling of the incremental weight can be used to maintain low accuracy loss. However, this scheme mainly focuses on merging multiple models into one model, and uses unstructured sparse. If this method is directly applied to a multi-model system, the calculation speed of unstructured sparse is relatively slow. In addition, there is still a significant loss in accuracy at a compression rate of 0.9 and above.To solve the above problems, the embodiments of the present disclosure find that the intermediate result of the matrix multiplication calculation of the incremental weight of each full-parameter fine-tuning model has a very small variance and a minimum maximum range, based on which a scheme for incrementally compressing the weight of a fine-tuned large model based on distribution is proposed: the incremental weight of the large model relative to the base model is grouped and sparsely compressed (randomly selected weight parameters are rescaled), which can achieve a compression rate of up to 96.875% with low loss of accuracy. Specifically, task data of a target task is obtained; the task data is input into a target processing model to obtain a task processing result of the target task, wherein the target processing model is trained based on sample data from an initial processing model, the target model weight of the target processing model includes an initial weight and an incremental compressed weight, the initial weight refers to the weight in the target processing model that is the same as the initial model weight of the initial processing model, and the incremental compressed weight is obtained based on grouping and sparse compression of the incremental weight, the incremental weight refers to the weight in the target processing model that is different from the initial model weight. By grouping and sparsely compressing the incremental weight, the incremental weight can be quickly swapped in and out of the GPU memory, which reduces the task processing delay and improves the resource utilization rate while ensuring the model accuracy. In the present disclosure, a task processing method is provided, and the present disclosure also relates to a target processing model-based information processing method, a task processing system, a model training platform, a task processing device, a target processing model-based information processing device, a computing device, a computer-readable storage medium, and a computer program product, which are described in detail one by one in the following embodiments. Referring to FIG. 1, FIG. 1 shows an architecture diagram of a task processing system according to an embodiment of the present disclosure. The task processing system can include a processing unit 100 and a storage unit 200. The storage unit 200 is configured to store the initial weight and the incremental compressed weight of a target processing model. The target processing model is trained based on sample data from an initial processing model. The initial weight refers to the weight in the target processing model that is the same as the initial model weight of the initial processing model. The incremental compressed weight is obtained based on grouping and sparse compression of the incremental weight. The incremental weight refers to the weight in the target processing model that is different from the initial model weight. The processing unit 100 is configured to obtain task data of a target task; obtain the initial weight and the incremental compressed weight from the storage unit 200; construct the target processing model based on the initial weight and the incremental compressed weight; and input the task data into the target processing model to obtain a task processing result of the target task. By grouping and sparsely compressing the incremental weight, the scheme of the embodiments of the present disclosure can reduce the task processing delay, improve the resource utilization rate, and ensure the model accuracy.In an optional embodiment of the present disclosure, the number of storage units can be multiple. For example, the storage units include a first storage unit and a second storage unit, the first storage unit is configured to store the initial weights of the target processing model, and the second storage unit is configured to store the incrementally compressed weights of the target processing model. In another optional embodiment of the present disclosure, the processing unit can also be configured to store the initial weights of the target processing model, and the storage unit is configured to store the incrementally compressed weights of the target processing model. Further, there can be multiple target processing models, and since the multiple target processing models are all obtained by full-parameter fine-tuning based on the initial processing model, the initial weights of the multiple target processing models are the same, and when storing the target model parameters of each target processing model, the initial weights can be stored only once. For example, the processing unit is configured to store the initial weights of the target processing model, the first storage unit is configured to store the incrementally compressed weights A of the target processing model A, and the second storage unit is configured to store the incrementally compressed weights B of the target processing model B. For another example, the processing unit is configured to store the initial weights of the target processing model, and the storage unit is configured to store the incrementally compressed weights A of the target processing model A and the incrementally compressed weights B0 of the target processing model B. Referring to FIG. 2, FIG. 2 shows a flowchart of a task processing method according to an embodiment of the present disclosure, which specifically includes the following steps: Step 202: obtaining task data of a target task. In one or more embodiments of the present disclosure, when processing a task, the task data of the target task can be obtained first, so that the target processing model is used to process the task data of the target task, and a task processing result of the target task is obtained. Specifically, the target task can be different tasks in different scenarios, such as an intelligent question answering task in an intelligent question answering scenario, a code completion task in an intelligent coding scenario, an article generation task in an intelligent writing scenario, and the like. The task data can be data in different modalities, such as text data in a text modality, image data in an image modality, speech data in an audio modality, video data in a video modality, and the like. In actual applications, there are multiple ways to obtain the task data of the target task, which are specifically selected according to actual conditions, and the present disclosure does not make any limitation in this regard. In a possible implementation manner of the present disclosure, the task data of the target task sent by a front-end user through an end-side device can be received. In another possible implementation manner of the present disclosure, the task data of the target task can be read from other data acquisition devices or databases.Step 204: input the task data into the target processing model to obtain a task processing result of the target task, wherein the target processing model is obtained by training the initial processing model based on the sample data, the target model weight of the target processing model includes an initial weight and an incremental compression weight, the initial weight refers to a weight in the target processing model that is the same as the initial model weight of the initial processing model, the incremental compression weight is obtained based on grouping and sparse compression of the incremental weight, and the incremental weight refers to a weight in the target processing model that is different from the initial model weight. In one or more embodiments of the present disclosure, after obtaining the task data of the target task, the task data can be further input into the target processing model to obtain the task processing result of the target task. Specifically, the weight refers to a numerical variable used to represent the internal knowledge of the model, the relationship between the neurons of each layer, and the response mode of the model to the input data. In the case where the target processing model is a deep learning model obtained by training the initial processing model based on the sample data in a supervised fine-tuning manner, the target processing model can be referred to as a full-parameter fine-tuning model. The initial processing model can be a pre-trained large model or a deep learning model. The initial processing model includes but is not limited to a CNN model, an RNN model, an LSTM model, a Transformer model, and a BERT model. The task processing result is related to the target task. If the target task is an intelligent question answering task, the task processing result is an answer result. If the target task is a code completion task, the task processing result is a code completion result. The incremental weight is used to describe the part of the target processing model that is locally adjusted or expanded based on the existing weight of the initial processing model. Since the target processing model is trained based on the initial processing model, the incremental weight can be a weight in the target processing model that is different from the initial model weight. The incremental weight helps the model better adapt to new situations or optimize performance. Structured compression of the incremental weight obtains the incremental compression weight, which can reduce the storage or transmission overhead of the model. It should be noted that when compressing the incremental weight, a structured compression method can be used, that is, grouping and sparse compression of the incremental weight. Specifically, there are many ways to group and sparse compress the incremental weight, which can be selected according to actual conditions, and the present disclosure does not make any limitation on this. In one possible implementation of the present disclosure, the incremental weight can be grouped to obtain at least one weight sequence; the weight sequence is randomly weighted and rescaled to obtain the incremental compression weight.In another possible implementation of the present disclosure, the incremental weights can be grouped to obtain at least one weight sequence; the weight sequence is randomly weighted and rescaled to obtain incremental compressed weights; the incremental compressed weights are uniformly quantized to obtain quantized incremental compressed weights, and the quantized incremental compressed weights are determined as the final incremental compressed weights of the target processing model, wherein the uniform quantization refers to quantizing the incremental compressed weights so that the quantized weights are uniformly spaced in the entire value range. Compared with the target model weights of the target processing model, the intermediate output result of the matrix multiplication corresponding to the incremental weights of the target processing model has extremely small variance and numerical range. Therefore, the numerical value of the intermediate output can be restored by randomly selecting weights and rescaling in the incremental weights, thereby further improving the model accuracy. Meanwhile, the grouping and sparsification are used in the incremental weight compression process, which is more friendly to the hardware platform and is more conducive to obtaining faster calculation speed and improving the hardware resource utilization rate. In addition, in an optional embodiment, the uniform quantization method is used, so that the model can achieve higher accuracy under the condition of high compression rate of the model weights. In an optional embodiment of the present disclosure, before the task data is input into the target processing model to obtain the task processing result of the target task, the target processing model can be directly obtained, or the target processing model is constructed based on the initial weights and the incremental compressed weights of the target processing model, that is, before the task data is input into the target processing model to obtain the task processing result of the target task, the following steps can also be included: obtaining the initial weights and the incremental compressed weights of the target processing model; decompressing the incremental compressed weights to obtain the incremental weights; and constructing the target processing model according to the initial weights and the incremental weights. Specifically, the incremental compressed weights refer to the weights obtained by compressing the incremental weights of the target processing model relative to the initial processing model in a grouping and sparsification manner. There are various ways to obtain the initial weights and the incremental compressed weights of the target processing model, which are selected according to actual conditions, and the present disclosure does not make any limitation on this. In a possible implementation of the present disclosure, the initial weights of the target processing model and the incremental compressed weights of the target processing model can be read from a storage unit. In another possible implementation of the present disclosure, the initial weights of the processing model sharing one base model (initial processing model) with the target processing model and the incremental compressed weights of the target processing model can be read from the storage unit, and the initial weights of the processing model sharing one base model (initial processing model) with the target processing model are determined as the initial weights of the target processing model.In actual applications, there are multiple ways to decompress the increment compression weight to obtain the increment weight, and the specific selection is based on actual conditions, and the embodiments of the present disclosure do not make any limitation. In a possible implementation manner of the present disclosure, the increment compression weight can be sent to an end-side device, and the increment weight obtained by manual decompression of a front-end user is received. In another possible implementation manner of the present disclosure, the value of the position of the non-zero weight element can be reconstructed based on the storage format of the increment compression weight, and then the increment weight is recovered, wherein the storage format includes but is not limited to a row-priority-based sparse matrix storage format (CSR, Compressed Sparse Row), a column-priority-based sparse matrix storage format (CSC, Compressed Sparse Column). It should be noted that, according to the initial weight and the increment weight, the target model weight of the target processing model can be generated based on the initial weight and the increment weight when the target processing model is constructed, and the target processing model is constructed according to the target model weight. The way of generating the target model weight based on the initial weight and the increment weight includes but is not limited to directly adding the increment weight to the initial weight to obtain the target model weight, or combining the initial weight and the increment weight in a moving average or exponential smoothing manner. By applying the scheme of the embodiments of the present disclosure, the initial weight and the increment compression weight of the target processing model are obtained; the increment compression weight is decompressed to obtain the increment weight; and the target processing model is constructed according to the initial weight and the increment weight. By obtaining the increment compression weight of the target processing model, the weight transmission overhead is reduced, the task processing delay is reduced, and the resource utilization rate of the hardware is improved. In an optional embodiment of the present disclosure, the initial weight and the increment compression weight of the target processing model can be obtained from a storage unit, that is, the above-mentioned obtaining the initial weight and the increment compression weight of the target processing model can include the following steps: obtaining task description information of a target task; determining a target processing model corresponding to the target task according to the task description information; extracting the initial weight of the target processing model from a first storage unit, and extracting the increment compression weight of the target processing model from a second storage unit. Specifically, the initial weight and the increment compression weight can be stored in the same storage unit (GPU memory) or different storage units, and a dedicated GPU memory is provided for each model. The task description information is used to describe the target task, and includes but is not limited to the task scene identifier of the target task, the task model identifier, or the sample data of the target generation task. It should be noted that the implementation manner of “determining the target processing model corresponding to the target task according to the task description information” is the same as the implementation manner of “obtaining the target processing model based on the request information”, and the embodiments of the present disclosure will not be repeated.In actual application, taking that the initial weights and the incremental compressed weights are stored in different storage units as an example, the initial weights of the target processing model can be extracted from the first storage unit based on the initial weight identifiers of the initial weights, and the incremental compressed weights of the target processing model can be extracted from the second storage unit based on the incremental compressed weight identifiers of the incremental compressed weights. According to the scheme of the embodiment of the present disclosure, the task description information of the target task is obtained; the target processing model corresponding to the target task is determined according to the task description information; the initial weights of the target processing model are extracted from the first storage unit, and the incremental compressed weights of the target processing model are extracted from the second storage unit. Since the incremental compressed weights are stored in the storage units, the data storage amount is reduced, and the resource utilization rate is improved. In an optional embodiment of the present disclosure, the generation manner of the incremental compressed weights of the target processing model is described, that is, the above-mentioned obtaining the initial weights and the incremental compressed weights of the target processing model can include the following steps: obtaining the target model weights of the target processing model and the initial model weights of the initial processing model; comparing the target model weights and the initial model weights, and dividing the target model weights into the initial weights and the incremental weights according to the comparison result; and grouping and sparsely compressing the incremental weights to obtain the incremental compressed weights. It should be noted that there are various ways to obtain the target model weights of the target processing model and the initial model weights of the initial processing model, which are selected according to actual conditions, and the embodiment of the present disclosure does not make any limitation on this. In a possible implementation manner of the present disclosure, the target model weights of the target processing model and the initial model weights of the initial processing model sent by the front-end user through the terminal device can be received. In another possible implementation manner of the present disclosure, the target model weights of the target processing model and the initial model weights of the initial processing model can be read from other data acquisition devices or databases. Further, after obtaining the target model weights of the target processing model and the initial model weights of the initial processing model, the target model weights can be compared with the initial model weights, the same part of the target model weights as the initial model weights is determined as the initial weights, and the different part of the target model weights from the initial model weights is determined as the incremental weights. According to the scheme of the embodiment of the present disclosure, the target model weights of the target processing model and the initial model weights of the initial processing model are obtained; the target model weights are compared with the initial model weights, and the target model weights are divided into the initial weights and the incremental weights according to the comparison result; and the incremental weights are grouped and sparsely compressed to obtain the incremental compressed weights. In the incremental weight compression process, the grouping and sparse manner is used, which is more friendly to the deployment on the hardware platform, is more conducive to obtaining a faster calculation speed, and improves the hardware resource utilization rate.In an optional embodiment of the present disclosure, after the increment weight is grouped and sparsely compressed to obtain the increment compressed weight, the method further includes the following steps: storing the initial weight in a first storage unit and storing the increment compressed weight in a second storage unit. It should be noted that, after the increment weight is grouped and compressed to obtain the increment compressed weight, the initial weight and the increment compressed weight can be stored, so that the initial weight and the increment compressed weight can be directly obtained subsequently, and the target processing model is constructed based on the initial weight and the increment compressed weight to perform task processing. In actual application, there are various ways to store the initial weight and the increment compressed weight, which are selected according to actual conditions, and the present disclosure does not make any limitation on this. In a possible implementation manner of the present disclosure, the initial weight and the increment compressed weight can be stored in the same storage unit. In another possible implementation manner of the present disclosure, the initial weight and the increment compressed weight can be stored in different storage units, for example, the initial weight is stored in a first storage unit and the increment compressed weight is stored in a second storage unit. It should be noted that, the storage unit storing the increment compressed weight of the target processing model can store only the increment compressed weight of the target processing model, or can store the increment compressed weight of the processing model of the same base initial processing model as the target processing model. By storing the initial weight in the first storage unit and storing the increment compressed weight in the second storage unit, the initial weight and the increment compressed weight can be directly obtained subsequently, the target processing model is constructed based on the initial weight and the increment compressed weight to perform task processing, and the task processing efficiency is improved. In the present embodiment, based on the distribution of the weight data, it is assumed that there are multiple target processing models obtained based on the same initial processing model through supervised fine-tuning, the target model weight Wj of the i-th target processing model can be divided into two parts, which are the initial weight Wo and the increment weight AWj, wherein the target model weight Wj is equal to the sum of the initial weight Wo and the increment weight AWj. The increment compression refers to compressing the increment weight AWj. In model compression, the optimization problem of the entire model compression is usually converted into a sub-problem of optimization of compression of each model processing layer, wherein the model processing layer includes but is not limited to an embedding layer, a normalization layer and the like. It is assumed that the weight of the i-th layer of the target processing model Mj is Wj W; G R. dlXk (R represents a real number, d, X k represent that the weight is a two-dimensional matrix, one dimension is d, and the other dimension is k), and the compressed weight is Wj. The optimization target of this layer compression is to minimize the 1_2 loss of this layer, and the 1_2 loss is specifically shown in the following formula (1): Where, Layer represents the |_2 loss of the |th layer of the target processing model; | | • | |; represents the Euclidean modulus, the 2 in the lower right corner represents the k-order norm, and the 2 in the upper right corner represents the square of the norm; X; represents the input data of the i-th layer, X; e R d2Xk (R represents a real number, d2X k represents that the data is two-dimensional, with one dimension being d2 and the other being k); W represents the transpose of the weight matrix of layer I; MinD represents the transpose of the weight matrix after compression of layer I. Assume that the output of matrix multiplication X; W is A; , A; c R d2Xdl , matrix multiplication The above formula (1) can be calculated by the following formula (2):

[0006] L Iayer =| | X; W " - X Que Ding Hua = | | A; - Sebi = Z; *] (ap,q I' -a p , q W) ⑵ Among them, a P ,q I' represents the output matrix A; the value at the (p, q) position, a p , q I' represents the output matrix A; the value at position (p, q), p represents the pth row of the matrix, q represents the qth column of the matrix, a p q |; can be calculated by the following formula (3), where e al l Indicates that the uncompressed value is the sum of multiple elements: When sparse compression is performed on the model weights, it is equivalent to multiplying the weights to be retained by 1, that is, m Si ,- --, m Si+ki e {l} kl , multiply the weights to be sparse by 0, that is, Where m represents the mask matrix, s and k represent that the mask matrix is ​​a two-dimensional matrix, one dimension is s and the other dimension is k, then a p q |; can be calculated by the following formula (4): For ap,q |; intermediate results Wp,oXo,q, …, Wp,kXk,q and Aa p q |; (a p qThe intermediate results △Wp,oXo,q, …, △Wp,kXk,q of the incremental values) are analyzed, and it is found that the intermediate output results △Wp,oXo,q, …, △Wp’kXk,q of the matrix multiplication corresponding to the incremental weights have the characteristics of small variance and small numerical range, and the numerical values are similar, which explains that the smooth distribution of the incremental weights allows better compressibility compared with the original weights. In an optional embodiment of the present disclosure, since the intermediate output results of the matrix multiplication corresponding to the incremental weights have very small variance and numerical range, they have better compressibility and can obtain higher compression rate, therefore, the incremental compressed weights can be obtained by using the grouping random weight sparsification and rescaling method, that is, the above grouping sparsification of the incremental weights to obtain the incremental compressed weights can include the following steps: dividing the first matrix row of the incremental weights according to the first preset value to obtain at least one weight sequence, wherein the first matrix row is any matrix row of the incremental weights, and the weight sequence includes continuous weight parameters; for the first weight sequence, selecting second preset value target weight parameters from the first weight sequence, wherein the first weight sequence is any one of the at least one weight sequence; and compressing the target weight parameters according to the first preset value and the second preset value to obtain the incremental compressed weights. It should be noted that for any row of the incremental weights, first, the weight elements are grouped based on the first preset value (m), that is, m continuous weight parameters are divided into a weight sequence; second, the random weight sparsification is performed based on the second preset value (n), that is, n weight parameters are randomly selected from the weight sequence to be retained, and the remaining m-n weight parameters are deleted; then, the n:m grouping balanced sparse weight parameters are rescaled, that is, the parameter values of the retained n weight parameters are multiplied by m / n to obtain the incremental compressed weights. By using the scheme of the embodiments of the present disclosure, the numerical values of the intermediate output are restored by randomly selecting the weights and rescaling, thereby further improving the model accuracy. Moreover, by using the grouping sparsification method, the accuracy is ensured while being more friendly to the deployment on the hardware platform, which is more conducive to obtaining faster calculation speed. In an optional embodiment of the present disclosure, in order to avoid the decrease of the model inference speed, the incremental compressed weights can also be quantized, that is, after the above compression of the target weight parameters according to the first preset value and the second preset value to obtain the incremental compressed weights, the following steps can also be included: quantizing the incremental compressed weights to obtain quantized incremental compressed weights, and determining the quantized incremental compressed weights as the incremental compressed weights of the target processing model.It should be noted that there are multiple ways to quantize the incremental compression weights, including but not limited to uniform quantization using the minimum-maximum (min-max) method and uniform quantization by converting the incremental compression weights to a standard normal distribution based on the mean and standard deviation. The specific method to be selected depends on the actual situation and is not limited in the present embodiment. In practical applications, taking the use of the min-max method to uniformly quantize the incremental compression weights retained after sparseness as an example, when quantizing the incremental compression weights, uniform quantization processing can be performed using the following formulas (5), (6), and (7): clamp(0, 2kT,P) (5) s = max£| XJ) ⑹

[0007] twenty one

[0008] Lx< l clamp(tu, x) = < x, 1 < x < u (7) u, x > u wherein, S represents a scaling parameter; max represents a maximum value; XR represents real data of an input of a model; k represents a bit number of an integer; X represents a quantized k-bit integer; Q represents quantization; I and u respectively represent an upper limit and a lower limit of a quantization range, in formula (5), 1 =0, U=2I, X=[XR / S]; U represents rounding to the nearest integer; the clamp() function is used to limit a value between an upper limit and a lower limit. By applying the scheme of the embodiment of the present disclosure, the incremental compression weight is quantized to obtain the quantized incremental compression weight, and the memory usage and the wideband requirement are greatly reduced, and the calculation speed is accelerated. In actual application, three models (including a model A for mathematical reasoning, a model B for mathematical reasoning and a model C for programming) trained by the same base large model are tested, in the testing process, a dataset (GSM8, Grade School Math 8K) containing 8500 high-quality, language-diverse primary school mathematics application problems is used to test the accuracy of the model A and the model B, and a dataset (HumanEval) including 164 programming problems is used to test the accuracy of the model C. It can be tested that the task processing scheme proposed in the embodiment of the present disclosure can provide higher processing accuracy in most scenarios, and most models can achieve a compression rate of 96.875% with a small accuracy loss. In an optional embodiment of the present disclosure, after the task data is input into the target processing model to obtain the task processing result of the target task, the following steps can be further included: sending the task processing result to an end-side device; receiving result feedback information sent by the end-side device, wherein the result feedback information is information for feeding back the task processing result based on the task information of the target task; constructing model optimization data according to the result feedback information; and adjusting parameters of the target processing model by using the model optimization data. Specifically, the result feedback information can be information for feeding back the content, quality and completion degree of the task processing result, reflecting the real feelings and expectations of the front-end user on the task processing result, and the result feedback information includes but is not limited to result quality evaluation information, corrected accurate task processing result and optimization field of the model. The model optimization data refers to accurate optimization sample data for optimizing the target processing model. In actual application, there are various ways to construct the model optimization data according to the result feedback information, which is selected according to actual conditions, and the embodiment of the present disclosure does not make any limitation on this.In a possible implementation of the present disclosure, model optimization data can be constructed directly according to the result feedback information. In another possible implementation of the present disclosure, optimization prompt information can be generated based on the result feedback information, and the receiving end device sends model optimization data based on the optimization prompt information. It should be noted that, taking the example of constructing model optimization data directly according to the result feedback information, if the result feedback information is an accurate corrected task processing result, the model optimization data can be constructed according to the task data of the target task and the accurate corrected task processing result. If the result feedback information is an optimization field of the model, such as an XXX field, sample data in the XXX field can be obtained, and the sample data in the XXX field is determined as the model optimization data. The process of adjusting the parameters of the target processing model by using the model optimization data is the same as the training process of the target processing model, and the training can be performed in a supervised fine-tuning manner, which will not be described herein again. According to the scheme of the present disclosure, the task processing result is sent to the end device, the result feedback information sent by the receiving end device is received, the result feedback information is information for feeding back the task processing result based on the task information of the target task, the model optimization data is constructed according to the result feedback information, and the parameters of the target processing model are adjusted by using the model optimization data. By collecting and using the result feedback information, the performance of the target processing model is continuously optimized, the actual needs of the front-end user are more accurately met, and the quality and accuracy of the final task processing result are improved. In an optional embodiment of the present disclosure, the above-mentioned construction of the model optimization data according to the result feedback information can include the following steps: generating optimization prompt information according to the result feedback information, wherein the optimization prompt information is used to guide the end device to send model optimization data for optimizing the target processing model; sending the optimization prompt information to the end device, and receiving the model optimization data sent by the end device based on the optimization prompt information. It should be noted that there are various ways to generate the optimization prompt information according to the result feedback information, which is selected according to actual conditions, and the present disclosure does not make any limitation on this. In a possible implementation of the present disclosure, preset prompt information can be directly obtained, and the result feedback information is added to the preset prompt information to obtain the optimization prompt information. For example, the preset prompt information is “I am very sorry to bring you inaccurate information. Please point out where it is not accurate or provide the correct answer to the related question, and I will correct and optimize my answer as soon as possible to better serve you”. The result feedback information is “the result is not accurate”, and the optimization prompt information is “I am very sorry to bring you inaccurate information according to the problem of the result not being accurate that you feedback”.Please point out where it is not accurate or provide the correct answer to the relevant question, and I will correct and optimize my answer as soon as possible to better serve you. In another possible implementation of the present disclosure, the result feedback information can be identified by type, the information type of the result feedback information is determined, the information type is further matched with the prompt type of each prompt information in the prompt information library, and the prompt information with the same prompt type as the information type is determined as the optimized prompt information. According to the scheme of the embodiment of the present disclosure, the optimized prompt information is generated according to the result feedback information, wherein the optimized prompt information is used to guide the terminal side device to send model optimization data for optimizing the target processing model; the optimized prompt information is sent to the terminal side device, and the model optimization data sent by the terminal side device based on the optimized prompt information is received. The model optimization data is obtained through the interactive guidance mode, which improves the interaction between the user and improves the user satisfaction. In an optional embodiment of the present disclosure, after the above-mentioned task data is input into the target processing model to obtain the task processing result of the target task, the following steps can be further included: labeling the key information in the task processing result to obtain an updated task processing result; and sending the updated task processing result to the terminal side device. Specifically, the key information refers to the information in the task processing result that is helpful for the user to understand, such as keywords, key data, etc. In actual application, before the key information in the task processing result is labeled to obtain the updated task processing result, the key information in the task processing result can be identified to determine the key information in the task processing result. There are various ways to identify the key information in the task processing result, which are selected according to actual conditions, and the embodiments of the present disclosure do not make any limitation on this. In a possible implementation of the present disclosure, the task processing result can be matched with the preset key information, and the information in the task processing result that is the same as the preset key information is determined as the key information. In another possible implementation of the present disclosure, the task processing result can be input into a key information identification model to obtain the key information, wherein the key information identification model is trained based on multiple training data and key information labels of the training data. It should be noted that when labeling the key information in the task processing result, the key information can be subjected to operations such as bolding, highlighting, tilting, adding a frame, etc. The labeling method of the key information is selected according to actual conditions, and the embodiments of the present disclosure do not make any limitation on this. According to the scheme of the embodiment of the present disclosure, the key information in the task processing result is labeled to obtain the updated task processing result; and the updated task processing result is sent to the terminal side device, so that the front-end user can view the result, and the user experience is improved.In an optional embodiment of the present disclosure, after the task data is input into the target processing model to obtain the task processing result of the target task, the following steps can be further included: sending the task processing result to the terminal device; receiving the editing information sent by the terminal device, wherein the editing information is used to edit the task processing result; and editing the task processing result according to the editing information to obtain the edited task processing result. Specifically, the editing information is used to describe the editing requirement of the front-end user on the task processing result. The editing requirement includes but is not limited to translation requirement, synonym replacement requirement, and paraphrasing requirement. Further, the editing information can also be the editing requirement for the key information in the task processing result. For example, assuming that the key information is "the sun rises from the east" and the editing information is "please translate the key information into English", the key information in the task processing result can be translated into "The sun rises in the east" according to the editing information. It should be noted that there are various ways to edit the task processing result according to the editing information, which can be selected according to actual conditions, and the present disclosure does not make any limitation on this. In a possible implementation manner of the present disclosure, an editing template (such as a content replacement template or a content deletion template) corresponding to the editing information can be obtained, and the task processing result is edited based on the editing template. In another possible implementation manner of the present disclosure, a deep learning model (such as a translation model or a classification model) corresponding to the editing information can be determined, and the task processing result is edited by using the deep learning model corresponding to the editing information. By applying the scheme of the present disclosure, the task processing result is sent to the terminal device, the editing information sent by the terminal device is received, wherein the editing information is used to edit the task processing result, and the task processing result is edited according to the editing information to obtain the edited task processing result. By editing the task processing result according to the editing information sent by the terminal device, the human-computer interaction is increased, and the adaptability and flexibility of the task processing are also improved. Referring to FIG. 3, FIG. 3 shows a flowchart of an information processing method based on a target processing model according to an embodiment of the present disclosure, which specifically includes the following steps: step 302: receiving a task generation request, wherein the task generation request includes request information. Specifically, the information processing method based on the target processing model can be applied to a terminal device or a cloud device. The task generation request is used to request to generate the task information of the target task. The task generation request usually includes the task type, the expected output format, and the request information.For example, when a user selects the "intelligent question answering" function on the front-end interface and uploads task data, a task generation request can be constructed, which includes task data, task type (i.e., intelligent question answering), language type of the task data, and the like. The request information refers to the parameters or description information related to the target task carried in the task generation request. The request information includes but is not limited to the task scene identifier of the target task, the task model identifier, or multiple sample data of the target task. Step 304: Based on the request information, the target processing model is obtained, wherein the target processing model is obtained by training the initial processing model based on the sample data, the target model weight of the target processing model includes the initial weight and the incremental compression weight, the initial weight refers to the weight in the target processing model that is the same as the initial model weight of the initial processing model, and the incremental compression weight is obtained based on the grouping and sparse compression of the incremental weight, and the incremental weight refers to the weight in the target processing model that is different from the initial model weight. In an optional embodiment of the present disclosure, the above step of obtaining the target processing model based on the request information can include the following steps: determining the target scene template from the multiple preset scene templates based on the task scene identifier, and searching the target processing model from the model library based on the target scene template, wherein the model library stores multiple processing models, and the request information includes the task scene identifier of the target task; or searching the target processing model from the model library based on the task model identifier, wherein the request information includes the task model identifier of the target task. Specifically, the task scene identifier refers to a unique or specific label used to distinguish different task application scenarios. In the embodiments of the present disclosure, the task scene identifier is part of the request information. Through the task scene identifier, the target scene template matching the request information can be selected from a series of preset scene templates to generate task information. For example, the task scene identifier is "intelligent question answering", which means that the front-end user hopes to perform question answering processing on the uploaded task data, and then the intelligent question answering scene template can be selected from the multiple preset scene templates according to the task scene identifier. The preset scene template is a standard configuration scene template defined in advance for different task application scenarios, and each template includes model information and task processing flow information that match the task application scenario. Through a series of preset scene templates, different scene task generation requests can be quickly responded. Different preset scene templates correspond to different task types, model information, and processing flows. For example, there can be a scene template in the preset scene template that is specifically for intelligent question answering, which includes model information and processing flow of the trained intelligent question answering model. The target scene template refers to the scene template matching the task scene identifier.In the process of analyzing the task generation request, the corresponding target scene template can be located based on the task scene identifier, and the corresponding target processing model and other related configuration information can be selected from the model library according to the model information included in the target scene template. For example, when the task scene identifier is "intelligent question and answer", the target scene template is a template including the model information of the intelligent question and answer model and the related configuration parameters. The model library is a resource library that centrally stores deep learning models, and these models are trained and optimized to solve different tasks. The processing models stored in the model library include but are not limited to intelligent question and answer models, article generation models, and code completion models. The task model identifier refers to a unique or specific label used to distinguish models applicable to different tasks. For example, the task model identifier can be "intelligent question and answer". Based on the task model identifier, the intelligent question and answer model applicable to processing the intelligent question and answer task can be found in the model library. According to the scheme of the present disclosure, the target scene template is determined from a plurality of preset scene templates based on the task scene identifier, and the target processing model is found in the model library based on the target scene template, wherein the model library stores a plurality of processing models, and the request information includes the task scene identifier of the target task; or, the target processing model is found in the model library based on the task model identifier, wherein the request information includes the task model identifier of the target task. Through the pre-defined task scene template, task model identifier, and model library resources, the acquisition process of the target processing model is more flexible, efficient, and standard. In another optional embodiment of the present disclosure, in addition to selecting a pre-trained target processing model from the model library, the target processing model can also be trained specifically according to the sample data in the request information, that is, the request information includes a plurality of sample data of the target task; the above acquisition of the target processing model based on the request information can include the following steps: training the initial processing model corresponding to the target task based on the plurality of sample data to obtain the trained target processing model. It should be noted that the way of training the initial processing model corresponding to the target task based on the plurality of sample data is supervised fine-tuning. Specifically, the plurality of sample data can be input into the initial processing model to obtain the prediction results corresponding to the plurality of sample data respectively; the loss value is calculated according to the sample result labels and the prediction results carried by the plurality of sample data, the model parameters of the initial processing model are adjusted according to the loss value, until the preset stopping condition is reached, and the trained target processing model is obtained. According to the scheme of the present disclosure, the initial processing model corresponding to the target task is trained based on the plurality of sample data to obtain the trained target processing model, so that the target processing model is more in line with user demand, and the accuracy of the target processing model is ensured.Step 306: generating task information based on the target processing model, where the task information is used to execute the target task. Specifically, the task information contains model configuration and processing flow required for executing the target task. The terminal device or other server components can correctly use the target processing model to process the target task based on the task information. It should be noted that when generating the task information based on the target processing model, the model parameters of the target processing model can be directly packaged to obtain the task information. The model information of the target processing model can also be obtained, and the task information is constructed based on the model information, where the model information is, for example, model parameter configuration, processing mode of input data, specification of expected output result, and other auxiliary information that may be involved. Exemplarily, in an intelligent question answering task, the task information can include address information of the selected intelligent question answering model, storage location of the input task data, target path of the processing result of the output, and other environment configurations required for running the intelligent question answering model, etc. These information enables the intelligent question answering model to be correctly loaded on a local or remote server and execute the intelligent question answering task. By generating the task information of the target task, the scheme of the embodiment of the present disclosure can reduce the system deployment and operation and maintenance cost while ensuring the processing quality and efficiency of the target task, and provide convenient and efficient task processing service for users. Referring to FIG. 4, FIG. 4 shows a structural schematic diagram of a model training platform provided by an embodiment of the present disclosure. The model training platform includes a request interface unit 402 and a model unit 404. The request interface unit 402 is configured to receive a model training request, where the model training request includes request information. The model unit 404 is configured to obtain an initial processing model, train the initial processing model based on the request information, and obtain a target processing model after training is completed, where the target model weight of the target processing model includes an initial weight and an incremental compression weight. The initial weight refers to the weight of the target processing model that is the same as the initial model weight of the initial processing model. The incremental compression weight is obtained based on grouping and sparse compression of the incremental weight. The incremental weight refers to the weight of the target processing model that is different from the initial model weight. It should be noted that the request information includes, but is not limited to, task scene identification of the target task, task model identification, or multiple sample data of the target task.The model unit can determine a target scene template from a plurality of preset scene templates based on the task scene identifier, and find the initial processing model from the model library based on the target scene template, wherein the model library stores a plurality of processing models, and the request information includes the task scene identifier; or find the initial processing model from the model library based on the task model identifier, wherein the request information includes the task model identifier. The implementation manner of "training the initial processing model based on the request information to obtain the target processing model trained" is the same as the implementation manner of "training the initial processing model corresponding to the target task based on a plurality of sample data to obtain the target processing model trained", and the embodiments of the present disclosure will not be described again. In an optional embodiment of the present disclosure, the model training platform further includes a model library, and the model unit is further configured to store the target processing model trained into the model library. It should be noted that after the model unit trains the initial processing model based on the request information to obtain the target processing model trained, the target processing model trained can also be stored into the model library, so that the target processing model trained can be directly obtained from the model library subsequently, without the need to repeat the model training. The manner of obtaining the target processing model trained from the model library can refer to the implementation manner of "determining a target scene template from a plurality of preset scene templates based on a task scene identifier, and finding a target processing model from a model library based on the target scene template, wherein the model library stores a plurality of processing models, and the request information includes a task scene identifier of a target task; or finding a target processing model from a model library based on a task model identifier, wherein the request information includes a task model identifier of a target task", and the embodiments of the present disclosure will not be described again. Corresponding to the above task processing method embodiments, the present disclosure further provides a task processing device embodiment, and FIG. 5 shows a structural schematic diagram of a task processing device according to an embodiment of the present disclosure. As shown in FIG. 5, the device includes: a first obtaining module 502 configured to obtain task data of a target task; and an input module 504 configured to input the task data into a target processing model to obtain a task processing result of the target task, wherein the target processing model is obtained by training an initial processing model based on sample data, a target model weight of the target processing model includes an initial weight and an incremental compression weight, the initial weight refers to a weight of the target processing model that is the same as an initial model weight of the initial processing model, and the incremental compression weight is obtained based on grouping and sparse compression of an incremental weight, and the incremental weight refers to a weight of the target processing model that is different from the initial model weight.Optionally, the apparatus further comprises: a construction module configured to obtain initial weights and incremental compressed weights of a target processing model; decompress the incremental compressed weights to obtain incremental weights; and construct the target processing model according to the initial weights and the incremental weights. Optionally, the construction module is further configured to obtain task description information of a target task; determine a target processing model corresponding to the target task according to the task description information; extract the initial weights of the target processing model from a first storage unit, and extract the incremental compressed weights of the target processing model from a second storage unit. Optionally, the construction module is further configured to obtain target model weights of the target processing model and initial model weights of an initial processing model; compare the target model weights and the initial model weights, and divide the target model weights into the initial weights and the incremental weights according to a comparison result; and group and sparsely compress the incremental weights to obtain the incremental compressed weights. Optionally, the construction module is further configured to, for a first matrix row of the incremental weights, divide the first matrix row according to a first preset number to obtain at least one weight sequence, wherein the first matrix row is any matrix row of the incremental weights, and the weight sequence comprises continuous weight parameters; for a first weight sequence, filter out second preset number of target weight parameters from the first weight sequence, wherein the first weight sequence is any one of the at least one weight sequence; and compress the target weight parameters according to the first preset number and the second preset number to obtain the incremental compressed weights. Optionally, the apparatus further comprises: a quantization module configured to quantize the incremental compressed weights to obtain quantized incremental compressed weights, and determine the quantized incremental compressed weights as the incremental compressed weights of the target processing model. Optionally, the apparatus further comprises: a storage module configured to store the initial weights to the first storage unit, and store the incremental compressed weights to the second storage unit. Optionally, the apparatus further comprises: an adjustment module configured to send a task processing result to an end-side device; receive result feedback information sent by the end-side device, wherein the result feedback information is information for feeding back the task processing result based on task information of the target task; construct model optimization data according to the result feedback information; and adjust parameters of the target processing model by using the model optimization data. Optionally, the adjustment module is further configured to generate optimization prompt information according to the result feedback information, wherein the optimization prompt information is used to guide the end-side device to send model optimization data for optimizing the target processing model; send the optimization prompt information to the end-side device, and receive model optimization data sent by the end-side device based on the optimization prompt information.Optionally, the apparatus further comprises a labeling module configured to label key information in the task processing result, to obtain an updated task processing result; and send the updated task processing result to the terminal-side device. Optionally, the apparatus further comprises an editing module configured to send the task processing result to the terminal-side device; receive editing information sent by the terminal-side device, wherein the editing information is used for editing the task processing result; and edit the task processing result according to the editing information to obtain an edited task processing result. By grouping and sparsely compressing the incremental weights, the scheme of the embodiment of the present disclosure reduces the task processing delay, improves the resource utilization rate, and ensures the model accuracy. The above is a schematic scheme of a task processing apparatus of the embodiment. It should be noted that the technical scheme of the task processing apparatus belongs to the same concept as the technical scheme of the task processing method described above, and the details of the technical scheme of the task processing apparatus that are not described in detail can be referred to the description of the technical scheme of the task processing method. Corresponding to the above embodiment of the information processing method based on the target processing model, the present disclosure further provides an embodiment of an information processing apparatus based on the target processing model, and FIG. 6 shows a structural schematic diagram of an information processing apparatus based on a target processing model according to an embodiment of the present disclosure. As shown in FIG. 6, the apparatus comprises: a receiving module 602 configured to receive a task generation request, wherein the task generation request comprises request information; a second obtaining module 604 configured to obtain a target processing model based on the request information, wherein the target processing model is obtained by training an initial processing model based on sample data, the target model weight of the target processing model comprises an initial weight and an incremental compressed weight, the initial weight refers to the weight in the target processing model that is the same as the initial model weight of the initial processing model, the incremental compressed weight is obtained by grouping and sparsely compressing the incremental weight, and the incremental weight refers to the weight in the target processing model that is different from the initial model weight; and a generating module 606 configured to generate task information based on the target processing model, wherein the task information is used to execute a target task. Optionally, the second obtaining module 604 is further configured to determine a target scene template from a plurality of preset scene templates based on a task scene identifier, and find the target processing model from a model library based on the target scene template, wherein the model library stores a plurality of processing models, and the request information comprises a task scene identifier of the target task; or find the target processing model from the model library based on a task model identifier, wherein the request information comprises a task model identifier of the target task.Optionally, the request information includes a plurality of sample data of the target task; the second obtaining module 604 is further configured to train the initial processing model corresponding to the target task based on the plurality of sample data, and obtain the target processing model after training. By generating the task information of the target task, the scheme of the embodiment of the present disclosure can reduce the system deployment and operation and maintenance cost while ensuring the processing quality and efficiency of the target task, and provide convenient and efficient task processing service for users. The above is a schematic scheme of the information processing apparatus based on the target processing model. It should be noted that the technical scheme of the information processing apparatus based on the target processing model and the technical scheme of the information processing method based on the target processing model belong to the same concept, and the details of the technical scheme of the information processing apparatus based on the target processing model that are not described in detail can be referred to the description of the technical scheme of the information processing method based on the target processing model. FIG. 7 shows a structural block diagram of a computing device according to an embodiment of the present disclosure. The components of the computing device 700 include, but are not limited to, a memory 710 and a processor 720. The processor 720 is connected with the memory 710 through a bus 730, and a database 750 is used to save data. The computing device 700 further includes an access device 740, which enables the computing device 700 to communicate via one or more networks 760. Examples of these networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet.The access device 740 can include one or more of any type of network interface (e.g., network interface card (NIC)) wired or wireless, such as an IEEE 802.11 wireless local area networks (WLAN) wireless interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a near field communication (NFC) interface, and the like. In one embodiment of the present disclosure, the above-mentioned components of the computing device 700 and other components not shown in FIG. 7 can also be connected to each other, for example, through a bus. It should be understood that the computing device structure block diagram shown in FIG. 7 is only for the purpose of example, and is not a limitation on the scope of the present disclosure. Those skilled in the art can add or replace other components as needed. The computing device 700 can be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smartphone), a wearable computing device (e.g., a smartwatch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or a personal computer (PC). The computing device 700 can also be a mobile or stationary server. The processor 720 is configured to execute computer programs / instructions that implement the steps of the task processing method or the information processing method based on the target processing model when the computer programs / instructions are executed by the processor. The above is a schematic scheme of a computing device according to the present embodiment. It should be noted that the technical scheme of the computing device belongs to the same concept as the technical scheme of the task processing method and the information processing method based on the target processing model, and the details of the technical scheme of the computing device that are not described in detail can be referred to the description of the technical scheme of the task processing method or the information processing method based on the target processing model.An embodiment of the present disclosure further provides a computer readable storage medium storing computer programs / instructions, which, when executed by a processor, implement the steps of the task processing method or the information processing method based on the target processing model. The above is a schematic scheme of the computer readable storage medium of the embodiment. It should be noted that the technical scheme of the storage medium and the technical scheme of the task processing method and the information processing method based on the target processing model belong to the same concept, and the details of the technical scheme of the storage medium which are not described in detail can be referred to the description of the technical scheme of the task processing method or the information processing method based on the target processing model. An embodiment of the present disclosure further provides a computer program product including computer programs / instructions, which, when executed by a processor, implement the steps of the task processing method or the information processing method based on the target processing model. The above is a schematic scheme of the computer program product of the embodiment. It should be noted that the technical scheme of the computer program product and the technical scheme of the task processing method and the information processing method based on the target processing model belong to the same concept, and the details of the technical scheme of the computer program product which are not described in detail can be referred to the description of the technical scheme of the task processing method or the information processing method based on the target processing model. The above describes a specific embodiment of the present disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions recited in the claims can be performed in a different order and still achieve desirable results. Additionally, the processes depicted in the figures do not necessarily require the particular order shown, or sequential order to achieve desirable results. In certain implementations, multitasking and parallel processing can be advantageous. The computer instructions include computer program code, which can be in source code, object code, executable code or some intermediate form. The computer readable medium can include any entity or apparatus capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer storage, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content included in the computer readable medium can be appropriately increased or decreased according to the requirements of patent practice, for example, in some regions, according to the patent practice, the computer readable medium does not include electrical carrier signals and telecommunication signals.It should be noted that, for the foregoing method embodiments, for the convenience of description, they are all described as a combination of a series of actions, but those skilled in the art should know that the disclosed embodiments are not limited to the action sequence described, because according to the disclosed embodiments, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the disclosure are all preferred embodiments, and the actions and modules involved are not necessarily essential to the disclosed embodiments. In the above embodiments, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments. The preferred embodiments of the disclosure disclosed above are only used to help explain the disclosure. The optional embodiments do not describe all the details and limit the disclosure to the specific embodiments described. Obviously, according to the content of the disclosed embodiments, many modifications and changes can be made. The disclosure selects and describes these embodiments in order to better explain the principles and practical applications of the disclosed embodiments, so that those skilled in the art can well understand and use the disclosure. The disclosure is limited only by the claims and their full scope and equivalents.

Claims

CLAIM 1. A method of task processing, comprising: obtain task data of a target task; input the task data into a target processing model to obtain a task processing result of the target task, wherein the target processing model is obtained by training an initial processing model based on sample data, target model weights of the target processing model include initial weights and incremental compression weights, the initial weights refer to weights of the target processing model that are the same as initial model weights of the initial processing model, and the incremental compression weights are obtained by performing group sparse compression on incremental weights that refer to weights of the target processing model that are different from the initial model weights.

2. The method of claim 1, before the inputting the task data into a target processing model, and obtaining a task processing result of the target task, further comprising: obtain initial weights and incremental compression weights of the target processing model; decompress the incremental compression weights to obtain incremental weights; construct the target processing model according to the initial weights and the incremental weights.

3. The method of claim 2, wherein the obtaining the initial weights and the delta compression weights of the target processing model comprises: obtain task description information of the target task; determine a target processing model corresponding to the target task according to the task description information; extract initial weights of the target processing model from a first storage unit and extract incremental compression weights of the target processing model from a second storage unit.

4. The method of claim 2, wherein the obtaining the initial weights and the delta compression weights of the target processing model comprises: obtain target model weights of the target processing model and initial model weights of the initial processing model; compare the target model weights and the initial model weights, and divide the target model weights into initial weights and incremental weights according to a comparison result; perform group sparse compression on the incremental weights to obtain incremental compression weights.

5. The method of claim 4, wherein the grouping and sparsely compressing the delta weights, obtaining delta compressed weights, comprises: for a first matrix row of the incremental weights, divide the first matrix row according to a first preset number of values to obtain at least one weight sequence, wherein the first matrix row is any matrix row of the incremental weights, and the weight sequence includes continuous weight parameters; for a first weight sequence, filter out second preset number of target weight parameters from the first weight sequence, wherein the first weight sequence is any one of the at least one weight sequence; and compress the target weight parameters according to the first preset number of values and the second preset number of values to obtain the incremental compression weights.

6. The method of claim 5, wherein after the target weight parameter is compressed according to the first preset value and the second preset value to obtain the incrementally compressed weight, the method further comprises: quantize the incremental compression weights to obtain quantized incremental compression weights, and determine the quantized incremental compression weights as the incremental compression weights of the target processing model.

7. The method of any one of claims 4-6, wherein the incrementally weighted are group sparsely compressed, and after obtaining the incrementally compressed weights, further comprising: store the initial weights to a first storage unit and store the incremental compression weights to a second storage unit.

8. The method of any one of claims 1-7, wherein the task data is input into the target processing model, obtaining a task processing result of the target task, and send the task processing result to an end-side device; receive result feedback information sent by the end-side device, wherein the result feedback information is information that feeds back the task processing result based on task information of the target task; construct model optimization data according to the result feedback information; and adjust parameters of the target processing model by using the model optimization data.

9. The method of claim 8, wherein constructing model optimization data based on the result feedback information comprises: generating optimization prompt information based on the result feedback information, wherein The optimization prompt information is used to guide the terminal device to send model optimization data for optimizing the target processing model. The optimization prompt information is sent to the terminal device, and the model optimization data sent by the terminal device based on the optimization prompt information is received.

10. The method of any one of claims 1-9, after the inputting the task data into a target processing model to obtain a task processing result of the target task, further comprising: Key information in the task processing result is labeled to obtain an updated task processing result. The updated task processing result is sent to a terminal device.

11. The method of any one of claims 1-10, after the inputting the task data into the target processing model, and obtaining the task processing result of the target task, further comprising: The task processing result is sent to a terminal device. Editing information sent by the terminal device is received, wherein the editing information is used to edit the task processing result; and the task processing result is edited according to the editing information to obtain an edited task processing result.

12. An information processing method based on a target processing model, comprising: A task generation request is received, wherein the task generation request includes request information; a target processing model is obtained based on the request information, wherein the target processing model is obtained by training an initial processing model based on sample data, target model weights of the target processing model include initial weights and incremental compression weights, the initial weights refer to weights of the target processing model that are the same as initial model weights of the initial processing model, the incremental compression weights are obtained by performing group sparse compression on incremental weights, and the incremental weights refer to weights of the target processing model that are different from the initial model weights; and task information is generated based on the target processing model, wherein the task information is used to execute a target task.

13. The method of claim 12, wherein the obtaining the target processing model based on the request information comprises: A target scene template is determined from a plurality of preset scene templates based on a task scene identifier, and a target processing model is searched from a model library based on the target scene template, wherein the model library stores a plurality of processing models, and the request information includes the task scene identifier of the target task; or, a target processing model is searched from the model library based on a task model identifier, wherein the request information includes the task model identifier of the target task.

14. The method of claim 12, wherein the request information comprises a plurality of sample data of a target task; and the obtaining, based on the request information, of the target processing model comprises: An initial processing model corresponding to the target task is trained based on the plurality of sample data to obtain a trained target processing model. ​ 15. A task processing system, comprising a processing unit and a storage unit, the storage unit being configured to store initial weights and incremental compression weights of a target processing model, the target processing model being trained based on sample data on an initial processing model The initial weights refer to weights in the target processing model that are the same as initial model weights of the initial processing model, the delta compression weights are obtained based on group-wise sparse compression of delta weights, and the delta weights refer to weights in the target processing model that are different from the initial model weights. The processing unit is configured to obtain task data of a target task; obtain the initial weights and the incremental compression weights from the storage unit; construct the target processing model according to the initial weights and the incremental compression weights; and input the task data into the target processing model to obtain a task processing result of the target task.

16. A model training platform, comprising a request interface unit and a model unit; the request interface unit is configured to receive a model training request, wherein, The model training request comprises request information; the model unit is configured to obtain an initial processing model; based on the request information, the initial processing model is trained to obtain a target processing model after training, wherein a target model weight of the target processing model comprises an initial weight and an incremental compression weight, the initial weight refers to a weight of the target processing model that is the same as an initial model weight of the initial processing model, the incremental compression weight is obtained based on grouping and sparse compression of an incremental weight, and the incremental weight refers to a weight of the target processing model that is different from the initial model weight.

17. The model training platform of claim 16, further comprising a model library; and the model unit is further configured to store the target processing model after training into the model library.

18. A computing device comprising: a memory and a processor; The memory is configured to store computer programs / instructions, and the processor is configured to execute the computer programs / instructions, which realize the steps of the method of any one of claims 1 to 11 or any one of claims 12 to 14 when executed by the processor.

19. A computer readable storage medium, which stores computer programs / instructions, which realize the steps of the method of any one of claims 1 to 11 or any one of claims 12 to 14 when executed by the processor.

20. A computer program product, which comprises computer programs / instructions, which realize the steps of the method of any one of claims 1 to 11 or any one of claims 12 to 14 when executed by the processor.

Images