Information processing method and device, equipment, storage medium and computer program product

Abstract

The invention discloses an information processing method and device, equipment, a storage medium and a computer program product. The method comprises the steps that first network equipment acquires first information, and the first information comprises description related information of one or more tasks; second information is determined by utilizing the first information, the second information represents a task model and a corresponding resource allocated for each task in the one or more tasks, and the task model is used for executing the task; third information is obtained through the second information, and the third information comprises execution results of the one or more tasks.

Description

Technical Field

[0001] This application relates to wireless communication technology, and more particularly to an information processing method, apparatus, device, storage medium, and computer program product. Background Technology

[0002] Currently, the integration of communication and artificial intelligence (AI) has been identified as one of the scenarios in the sixth-generation mobile communication standard (6G) network. This requires the 6G network to provide AI services for users and the network itself. Therefore, the network architecture with endogenous AI has been proposed in related technologies.

[0003] In the network architecture of endogenous AI, the AI ​​model corresponding to the AI ​​task is usually monitored and managed by an external overlay method; that is, by setting up a processing unit on the terminal device, the life cycle of the AI ​​model can be managed. However, the above solution cannot meet the diverse needs of AI task scenarios in 6G networks. Summary of the Invention

[0004] To address the related technical problems, embodiments of this application provide an information processing method, apparatus, device, storage medium, and computer program product.

[0005] The technical solution of this application embodiment is implemented as follows:

[0006] This application provides an information processing method applied to a first network device, comprising:

[0007] Obtain first information, which includes descriptions of one or more tasks;

[0008] Using the first information, determine the second information, which represents the task model and corresponding resources allocated to each of the one or more tasks, and the task model is used to execute the task;

[0009] Using the second information, third information is obtained, which includes the execution results of the one or more tasks.

[0010] The method in the above scheme further includes:

[0011] Using the third information, fourth information is determined, wherein the fourth information characterizes the performance level of one or more task models corresponding to the one or more tasks;

[0012] If the performance level of one or more first task models in the one or more task models represented by the fourth information does not meet a first condition, it is determined that the one or more first task models should be retrained, wherein the first condition includes conditions associated with the one or more tasks.

[0013] The method in the above scheme further includes:

[0014] The fifth piece of information is determined, wherein the fifth piece of information represents the resources of the first network device;

[0015] Using the fifth piece of information, network devices for retraining the one or more first task models are identified.

[0016] In the above scheme, the step of using the fifth information to determine the network device for retraining the one or more first task models includes:

[0017] Obtain sixth information, which represents the size of one or more datasets associated with the one or more first task models;

[0018] Using the fifth and sixth information, network devices for retraining the one or more first task models are identified.

[0019] In the above scheme, the step of using the fifth information to determine the network device for retraining the one or more first task models includes:

[0020] Determine whether the resources represented by the fifth information satisfy the second condition, the second condition including conditions associated with the one or more first task models, and the resources represented by the fifth information including computing resources and / or storage resources;

[0021] If the resources represented by the fifth information satisfy the second condition, the network device used to retrain the one or more first task models is determined to be the first network device.

[0022] In the above scheme, determining whether the resource represented by the fifth information satisfies the second condition includes:

[0023] If the resources represented by the fifth information include computing resources, then the seventh information is determined, which includes the duration information for retraining the one or more first task models.

[0024] Using the seventh piece of information, determine whether the computing resources represented by the fifth piece of information satisfy the second condition.

[0025] The method in the above scheme further includes:

[0026] If the resources represented by the fifth information do not meet the second condition, an eighth message is sent to the second network device. The eighth message is used to request retraining of the one or more first task models. The second network device includes network devices adjacent to the first network device.

[0027] Receive the ninth information sent by the second network device, the ninth information being used to indicate whether the resources of the second network device meet the second condition;

[0028] If the resources of the second network device meet the second condition, the network device used to retrain the one or more first task models is determined to be the second network device.

[0029] The method in the above scheme further includes:

[0030] If the resources of the second network device do not meet the second condition, the tenth information is sent to the first network element. The first network element is used to retrain the one or more first task models. The tenth information includes at least the one or more first task models.

[0031] In the above scheme, obtaining the third information using the second information includes:

[0032] Obtain eleventh information, which includes inference data associated with the one or more tasks;

[0033] The third information is obtained by using the eleventh information and the second information.

[0034] This application also provides an information processing apparatus, including:

[0035] A receiving unit is configured to acquire first information, the first information including description-related information of one or more tasks;

[0036] The first processing unit is configured to use the first information to determine the second information, wherein the second information represents the task model and corresponding resources allocated to each of the one or more tasks, and the task model is used to execute the task.

[0037] The second processing unit is used to obtain third information using the second information, the third information including the execution results of the one or more tasks.

[0038] This application embodiment also provides a first network device, including: a processor and a communication interface; wherein,

[0039] The communication interface is used to obtain first information, which includes description-related information of one or more tasks.

[0040] The processor is configured to use the first information to determine second information, the second information representing a task model and corresponding resources allocated to each of the one or more tasks, the task model being used to execute the task; and to use the second information to obtain third information, the third information including the execution results of the one or more tasks.

[0041] This application also provides a first network device, including: a processor and a memory for storing computer programs capable of running on the processor.

[0042] When the processor runs the computer program, it executes the steps of any of the information processing methods described above.

[0043] This application also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of any of the information processing methods described above.

[0044] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the information processing methods described above.

[0045] The information processing method, apparatus, device, storage medium, and computer program product provided in this application embodiment include: a first network device acquiring first information, the first information including description-related information of one or more tasks; using the first information to determine second information, the second information representing a task model and corresponding resources allocated to each of the one or more tasks, the task model being used to execute the task; and using the second information to obtain third information, the third information including the execution results of the one or more tasks. The technical solution provided in this application embodiment, for one or more tasks (such as AI tasks), allows the network device to uniformly schedule resources and allocate models for different tasks based on the task description-related information, in order to achieve management of the model lifecycle, thus meeting the diverse needs of task scenarios in 6G networks. Attached Figure Description

[0046] Figure 1 This is a schematic flowchart of an information processing method according to an embodiment of this application;

[0047] Figure 2 This is a schematic diagram illustrating the structure of a lifecycle management system for an AI model, serving as an application example of this application.

[0048] Figure 3 This is a schematic diagram illustrating a method for lifecycle management of an AI model, serving as an application example of this application.

[0049] Figure 4 This is a schematic diagram of the information processing device structure according to an embodiment of this application;

[0050] Figure 5 This is a schematic diagram of the structure of the first network device according to an embodiment of this application. Detailed Implementation

[0051] The present application will now be described in further detail with reference to the accompanying drawings and embodiments.

[0052] In the network architecture of endogenous AI, the data may change, making the data used for inference (which can also be understood as performing tasks through the AI ​​model) different from the data used during model training. This may cause a decrease in the inference performance of the AI ​​model. Therefore, it is necessary to establish an AI lifecycle management system that can be optimized to improve the performance of AI models and user experience.

[0053] In related technologies, one AI lifecycle management solution is designed for AI models deployed in the cloud. Since the cloud has sufficient computing power and storage resources, there is no need to consider the problem of insufficient resources during the management process (specifically, the performance optimization process). Therefore, the lifecycle management solution for cloud-based AI models is not suitable for endogenous AI.

[0054] Another AI lifecycle management solution involves setting up processing units on the terminal to monitor and manage the performance of AI models. However, this solution is only for specific AI tasks, meaning different processing units are needed for different AI tasks, which is costly and cannot meet the diverse AI task scenarios required in 6G networks. In addition, the management of AI models does not consider task offloading based on the cloud, which may not be able to meet the needs of model optimization when local resources are insufficient.

[0055] Based on this, in various embodiments of this application, for different tasks, network devices (such as base stations) uniformly schedule network resources and allocate corresponding models for different tasks in order to manage the life cycle of the models. In this way, under the network architecture of endogenous AI, flexible scheduling of tasks can be achieved, thereby meeting the diverse needs of different task scenarios.

[0056] This application provides an information processing method, such as... Figure 1 As shown, applied to a first network device, the method includes:

[0057] Step 101: Obtain first information, which includes descriptions of one or more tasks;

[0058] Step 102: Using the first information, determine the second information, which represents the task model and corresponding resources allocated to each of the one or more tasks, and the task model is used to execute the task;

[0059] Step 103: Using the second information, obtain the third information, which includes the execution results of the one or more tasks.

[0060] In practical applications, the first network device may include a base station, such as a base station in a 6G network. This application embodiment does not limit this, as long as its function is implemented.

[0061] In practical applications, in step 101, the first network device can obtain the first information from the terminal. That is, when the terminal initiates one or more tasks, the first network device can receive the first information sent by the terminal. The terminal can be called a user equipment (UE) or a user, etc. This application embodiment does not limit the name of the terminal, as long as its function is implemented. Of course, the first network device can also obtain the first information locally. That is, when the first network device initiates one or more tasks, the first network device can generate the first information. This application embodiment does not limit the method of obtaining the first information.

[0062] Here, regarding the first information, the description information corresponding to each task may include one or more of the following (or at least one): task objective, data format, model size, and lifecycle; wherein, the task objective is used for model training and / or model performance monitoring, and may specifically include a loss function; the data format is used to indicate the data format to be input and the data format to be output; the model size is used to indicate the model size corresponding to the task, and can be set according to experience; the lifecycle represents the expected usage time of the model corresponding to the task, such as 5 minutes.

[0063] In practical applications, after obtaining the first information, the first network device can use the first information to allocate corresponding task models and corresponding resources for different tasks, thereby obtaining the second information; wherein, the allocated task model can be understood as a trained task model associated with the task, that is, one task corresponds to one task model, and the allocated resources can be understood as the resources required in the process of executing the task, such as storage resources and / or computing resources.

[0064] For example, assuming that the core network stores trained task models for different tasks, the first network device can obtain the task models and corresponding model size information for different tasks from the core network based on the first information; based on the model size information, the first network device can allocate appropriate physical resources for the task models.

[0065] In practical applications, once the second information is determined, the first network device can initialize (or load the initial model) the one or more task models based on the allocated resources, so as to execute the one or more tasks based on the initialized task models.

[0066] Specifically, in one embodiment, the implementation of step 103 may include:

[0067] Obtain eleventh information, which includes inference data associated with the one or more tasks;

[0068] The third information is obtained by using the eleventh information and the one or more task models.

[0069] Here, when the first information is obtained from the terminal, the first network device can obtain the eleventh information from the terminal; specifically, after completing the initialization of the task model, the first network device can send the twelfth information to the terminal, the twelfth information being used to instruct the execution of the one or more tasks; after receiving the twelfth information, the terminal can collect inference data associated with the one or more tasks to obtain the eleventh information, and send the eleventh information to the first network device.

[0070] In addition, if the first information is obtained from the first network device, the first network device can obtain the eleventh information from the first network device; specifically, after completing the initialization of the task model, the first network device can collect inference data associated with the one or more tasks to obtain the eleventh information.

[0071] In practical applications, after obtaining the eleventh information, the first network device can input the inference data associated with the one or more tasks into the corresponding task model (which can also be understood as forward propagation) to obtain the third information. Then, the first network device can send the third information to complete the execution (inference) of the one or more tasks. In the case of obtaining the first information from the terminal, the third information can be sent to the terminal. In the case of obtaining the first information from the first network device, the first network device can locally save the third information.

[0072] In addition, the first network device can also monitor the usage of the one or more task models; if a task model is not used to execute a task within a first time period (which can be set as needed) (which can also be understood as not being called for a long time), the first network device can release the resources corresponding to the task model to avoid wasting resources.

[0073] In practical applications, during the execution of the one or more tasks, the first network device can also monitor the performance of the one or more task models so as to optimize the task models when their performance degrades. In this way, effective management of the lifecycle of the task models can be achieved.

[0074] Based on this, in one embodiment, the method may further include:

[0075] Using the third information, fourth information is determined, wherein the fourth information characterizes the performance level of one or more task models corresponding to the one or more tasks;

[0076] If the performance level of one or more first task models in the one or more task models represented by the fourth information does not meet a first condition, it is determined that the one or more first task models should be retrained, wherein the first condition includes conditions associated with the one or more tasks.

[0077] The conditions associated with the one or more tasks may include a first threshold (also known as a performance threshold) for each of the one or more tasks. The value of the first threshold can be set as needed, such as 70% to 80% of the ideal performance level of the task model. This application embodiment does not limit this.

[0078] In practical applications, the first network device can pre-set a first algorithm associated with each task, and use the first algorithm and the third information to determine whether to retrain the one or more task models; wherein, for the first algorithm, the third information can be used as input information, and the judgment result of whether to retrain the one or more task models can be used as output information; the first algorithm can be called a performance monitoring algorithm, and can include a calculation formula for a performance function and the first condition. The name of the first algorithm is not limited in this application embodiment, as long as its function is implemented.

[0079] Specifically, the first network device can use the third information and the calculation formula included in the first algorithm to obtain the performance level of the one or more task models; if the performance level of each task model in the one or more task models is higher than or equal to the corresponding first threshold, the first network device can determine that the performance level of the one or more task models has not decreased, that is, it is not necessary to retrain the one or more task models; if the performance level of the one or more first task models is lower than the corresponding first threshold, the first network device can determine that the performance level of the one or more first task models has decreased, in which case the one or more first task models can be retrained to optimize the performance of the task models.

[0080] In practical applications, when it is necessary to retrain one or more of the first task models, the first network device can determine the network device to be retrained based on local resource conditions.

[0081] Based on this, in one embodiment, the method may further include:

[0082] The fifth piece of information is determined, wherein the fifth piece of information represents the resources of the first network device;

[0083] Using the fifth piece of information, network devices for retraining the one or more first task models are identified.

[0084] The resources of the first network device can be understood as the resources currently available to the first network device (also known as local computing power), specifically including resource-related information of the first network device, such as storage resource information and / or computing resource information.

[0085] In practical applications, the first network device can compare the size of its own resources with the size of the resources required to retrain the one or more first task models to determine whether it is capable of retraining the one or more first task models. Therefore, the first network device needs to determine the resources required to retrain the one or more first task models.

[0086] Specifically, in one embodiment, determining the network device for retraining the one or more first task models using the fifth information includes:

[0087] Obtain sixth information, which represents the size of one or more datasets associated with the one or more first task models;

[0088] Using the fifth and sixth information, network devices for retraining the one or more first task models are identified.

[0089] Wherein, the one or more datasets are used to retrain the one or more first task models, and the sixth information can be understood as the size of the one or more datasets.

[0090] In practical applications, the first network device can determine the resources required for retraining the one or more first task models based on the sixth information. Of course, the first network device can further combine the model size information corresponding to the task models to determine the resources required for retraining the one or more first task models. In other words, the first network device can use the sixth information and the model size information corresponding to the one or more first task models to determine the resources required for retraining the one or more first task models.

[0091] Here, the first network device can pre-set a second algorithm and use the second algorithm to determine the resources required for retraining the one or more first task models, thereby determining whether the resources of the first network device are sufficient. In this way, it can be determined whether the network device used to retrain the one or more first task models can be the first network device. For the second algorithm, the fifth and sixth information can be used as input information, and the judgment result of whether the resources of the first network device are sufficient can be used as output information. The second algorithm can be called a computing power check algorithm. In this application embodiment, the name of the second algorithm is not limited, as long as its function is implemented.

[0092] Specifically, in one embodiment, determining the network device for retraining the one or more first task models using the fifth information includes:

[0093] Determine whether the resources represented by the fifth information satisfy the second condition, the second condition including conditions associated with the one or more first task models, and the resources represented by the fifth information including computing resources and / or storage resources;

[0094] If the resources represented by the fifth information satisfy the second condition, the network device used to retrain the one or more first task models is determined to be the first network device.

[0095] The second condition may include a second threshold (i.e. the size of the required resources) corresponding to the resources required for retraining the one or more first task models. The value of the second threshold may be set based on the sixth information and / or the model size information corresponding to the one or more first task models.

[0096] In practical applications, when the resources represented by the fifth information include at least storage resources, the first network device can use the second algorithm to determine the second threshold corresponding to the storage resources required for retraining the one or more first task models by utilizing the sixth information and / or the model size information corresponding to the one or more first task models; by comparing the second threshold corresponding to the storage resources with the size of the storage resources of the first network device, it can be determined whether the storage resources of the first network device are sufficient.

[0097] Here, if the size of the storage resource represented by the fifth information is lower than the second threshold corresponding to the storage resource, the first network device can determine that the storage resource of the first network device is insufficient. In this case, it can be determined that the network device used to retrain the one or more first task models cannot be the first network device.

[0098] Furthermore, if the size of the storage resource represented by the fifth information is higher than or equal to the second threshold corresponding to the storage resource, the first network device can determine that the storage resource of the first network device is sufficient. In this case, the first network device can further determine whether the computing resource of the first network device is sufficient; that is, the resource represented by the fifth information may also include computing resources.

[0099] Specifically, in one embodiment, determining whether the resource represented by the fifth information satisfies the second condition includes:

[0100] If the resources represented by the fifth information include computing resources, then the seventh information is determined, which includes the duration information for retraining the one or more first task models.

[0101] Using the seventh piece of information, determine whether the computing resources represented by the fifth piece of information satisfy the second condition.

[0102] The seventh piece of information can be understood as the time required for the one or more first task models to be retrained once.

[0103] In practical applications, the first network device can use the second algorithm to obtain the seventh information by utilizing the model size information corresponding to the one or more first task models and the size of computing resources represented by the fifth information; by comparing the seventh information with the second threshold (which can be understood as the retraining time that can be accepted or tolerated, and the value can be set as needed) corresponding to the computing resources required for retraining the one or more first task models, it can determine whether the computing resources of the first network device are sufficient.

[0104] For example, through the second algorithm and the model size information corresponding to the one or more first task models, the first network device can determine the number of layers of the first task model and the number of nodes corresponding to each layer, and thus determine the number of computations required for backpropagation once when the one or more first task models are retrained; based on the computing resources represented by the fifth information, the current available computing power of the first network device (which can be understood as the amount of computation per second) can be determined; based on the current available computing power and the determined number of computations, the time required for retraining the one or more first task models can be calculated; if the time required for retraining the one or more first task models is greater than or equal to the second threshold corresponding to the computing resources (e.g., 100ms), it can be determined that the computing resources of the first network device are insufficient; if the time required for retraining the one or more first task models is less than the second threshold corresponding to the computing resources, it can be determined that the computing resources of the first network device are sufficient.

[0105] In practical applications, when the storage resources and / or computing resources of the first network device meet the second condition, the first network device can determine that its local resources meet the retraining requirements of the one or more first task models, and thus determine the network device used to retrain the one or more first task models as the first network device.

[0106] In practical applications, when local resources are insufficient to meet the retraining requirements of one or more first task models, the first network device can query the resource status of adjacent network devices to retrain the one or more first task models through these adjacent network devices. In other words, the first network device can flexibly offload retraining tasks through computing power checks to avoid problems such as excessively long retraining times or no performance improvement in the retrained task models due to insufficient resources.

[0107] Based on this, in one embodiment, the method may further include:

[0108] If the resources represented by the fifth information do not meet the second condition, an eighth message is sent to the second network device. The eighth message is used to request retraining of the one or more first task models. The second network device includes network devices adjacent to the first network device.

[0109] Receive the ninth information sent by the second network device, the ninth information being used to indicate whether the resources of the second network device meet the second condition;

[0110] If the resources of the second network device meet the second condition, the network device used to retrain the one or more first task models is determined to be the second network device.

[0111] The eighth information may be referred to as a computing power support request. The eighth information includes at least the computing resources and / or storage resources required for the retraining of the one or more first task models, such as the second threshold corresponding to the storage resources required for the retraining of the one or more first task models or the seventh information. The second network device may be referred to as a neighboring network device or a neighboring base station, etc. This application embodiment does not limit this.

[0112] In practical applications, the first network device can send the eighth information to one or more second network devices respectively, so that each second network device can determine whether the resources of the second network device can meet the second condition based on the eighth information, and feed back the ninth information to the first network device; wherein, the second network device can make the judgment in a similar manner to the first network device.

[0113] It should be noted that if the resources of multiple second network devices meet the second condition, the first network device may randomly select one second network device from the multiple second network devices and determine the selected second network device as the network device for retraining the one or more first task models.

[0114] Then, the first network device can transmit the one or more first task models and the one or more datasets to the second network device, whereby the second network device retrains the one or more first task models. After retraining the one or more first task models, the second network device can transmit the relevant parameters of the retrained one or more first task models to the first network device, so that the first network device can update its local first task models.

[0115] In practical applications, if the resources of the second network device do not meet the second condition, the first network device can also retrain the one or more first task models through the core network.

[0116] Based on this, in one embodiment, the method may further include:

[0117] If the resources of the second network device do not meet the second condition, the tenth information is sent to the first network element. The first network element is used to retrain the one or more first task models. The tenth information includes at least the one or more first task models.

[0118] The first network element can be understood as a network element deployed in the core network, which can at least retrain the one or more first task models. In this embodiment of the application, the specific type of the first network element is not limited, as long as its function is implemented.

[0119] In practical applications, the first network device can transmit the one or more first task models and the one or more datasets to the first network element, and the first network element can retrain the one or more first task models; that is, the tenth information may also include the one or more datasets.

[0120] Here, after retraining the one or more first task models is completed, the first network element can transmit the relevant parameters of the retrained one or more first task models to the first network device so that the first network device can update the local first task model.

[0121] The information processing method provided in this application embodiment involves a first network device acquiring first information, which includes description-related information of one or more tasks; using the first information, determining second information, which represents a task model and corresponding resources allocated to each of the one or more tasks, the task model being used to execute the task; and using the second information, obtaining third information, which includes the execution results of the one or more tasks. The technical solution provided in this application embodiment, for one or more tasks (such as AI tasks), allows the network device to uniformly schedule resources and allocate models for different tasks based on the task description-related information, thereby enabling management of the model's lifecycle and meeting the diverse needs of task scenarios in 6G networks.

[0122] The following section provides a more detailed description of this application with reference to application examples.

[0123] This application example proposes a lifecycle management scheme for network-native AI models. By deploying a performance monitoring module and a computing resource management module on the base station, it can monitor the performance of the AI ​​model and the available computing resources on the network side in real time. By deploying an AI task orchestration module, it can uniformly schedule network resources for different AI tasks. The system architecture corresponding to the above lifecycle management scheme is as follows: Figure 2 As shown, it includes an AI task orchestration module, a performance monitoring module, a computing resource management module, a training control module, an inference module, an information acquisition module, and a training module; among them,

[0124] The AI ​​task orchestration module is used to orchestrate different AI tasks in a unified manner and allocate different physical resources to the corresponding AI models to avoid resource usage conflicts, thereby achieving flexible scheduling of AI tasks.

[0125] The information acquisition module is used to collect the data required for AI model inference and training. In the case where the AI ​​task is initiated by the base station, one working mode is to communicate with the base station to obtain the data required by the AI ​​model; in the case where the AI ​​task is initiated by the terminal, another working mode is to communicate with the terminal to obtain the data required by the AI ​​model.

[0126] The computing resource management module is used to uniformly schedule all computing resources of the base station. Specifically, firstly, it can initialize the AI ​​model; secondly, it can assess whether the base station can perform retraining when the AI ​​model performance degrades and retraining is required; and thirdly, it can release the corresponding computing space when the AI ​​model has not been used for a long time to avoid wasting resources.

[0127] The inference module is used to store the AI ​​model and perform forward propagation during the inference phase to obtain inference results, and perform forward and backward propagation during the training phase to update the parameters of the AI ​​model.

[0128] The performance monitoring module is used to obtain the inference results of the inference module in real time and monitor the performance of the AI ​​model so as to initiate a retraining request when the performance deteriorates.

[0129] The training control module is used to control the inference module to perform training functions when the performance of the AI ​​model deteriorates and the base station resources are sufficient, as well as to save the loss function required for training and pass the gradient of the loss function to the inference module.

[0130] The training module, deployed in the core network, has ample computing and storage resources for training and retraining AI models.

[0131] In practical applications, Figure 2 Under the system architecture shown, the process of AI model lifecycle management in multi-AI task scenarios is as follows: Figure 3 As shown, it includes the following steps:

[0132] Step 301: The terminal sends an AI task request to the AI ​​task orchestration module on the base station (i.e., the first network device mentioned above);

[0133] The AI ​​task request includes AI task description information (i.e., the first information mentioned above), which includes task objectives, input and output data formats, model size, lifecycle, etc.

[0134] Step 302: The AI ​​task orchestration module allocates resources to different tasks according to the AI ​​task requests (i.e., the second information mentioned above) and informs the computing resource management module to execute.

[0135] Step 303: The computing resource management module loads the initial model into the inference module;

[0136] Step 304: After the initial model loading is completed, the computing power resource management module informs the AI ​​task orchestration module that the model loading is complete;

[0137] Step 305: After the AI ​​task orchestration module learns that the model has been loaded, it sends an inference task execution signal to the terminal;

[0138] Step 306: After receiving the inference task execution signaling, the terminal collects the data used for inference (i.e., the eleventh information mentioned above) and transmits the data used for inference to the information collection module, which then sends the data used for inference to the inference module.

[0139] Step 307a: The inference module performs forward propagation based on the data used for inference to obtain the inference result (i.e., the third information mentioned above), and feeds the inference result back to the terminal;

[0140] Step 307b: The inference module sends the inference results to the performance monitoring module;

[0141] Step 308: After receiving the inference results, the performance monitoring module calculates the model performance based on the inference results and compares it with the required performance threshold. When the model performance meets the requirements, the performance monitoring module notifies the inference module to continue executing the inference function. Otherwise, when the model performance degrades, it sends a retraining request to the computing resource management module.

[0142] Here, the calculation formula of the performance function and the required performance threshold (i.e., the first condition mentioned above) can be set in advance on the performance monitoring module so that the performance monitoring module can obtain the inference performance of the current model (i.e., the fourth information mentioned above) based on the inference results and calculation formula. If the current inference performance is higher than the performance threshold, no retraining request is sent; if the current inference performance is lower than the performance threshold, a retraining request is sent.

[0143] Step 309: After receiving the retraining request, the computing power resource management module determines the computing and storage resources available in real time at the base station (i.e., the fifth information mentioned above), and compares the determined resources with the resources required for retraining (i.e., the second condition mentioned above); when the computing power and storage resources of the base station are insufficient, it informs the training control module to initiate retraining of the model (i.e., the first task model mentioned above);

[0144] Here, a computing power check algorithm can be set on the computing power resource management module so that the computing power resource management module can compare the model size, the size of the training dataset required for retraining (i.e., the sixth information mentioned above) with the storage resources available to the base station in real time; if the storage resources are insufficient, the computing power resource management module directly determines that the current base station has insufficient available resources; if the storage resources are sufficient, the computing power resource management module further determines whether the computing resources are sufficient, thereby determining whether the local base station's computing power is sufficient.

[0145] Specifically, the computing power resource management module calculates the number of computations required for one backpropagation during training based on the number of layers and nodes in each layer of the model. At the same time, it obtains the computing power currently available at the base station and then calculates the time required for one retraining (i.e., the seventh information mentioned above). If the time for one retraining is greater than 100ms, it is determined that the current base station has insufficient computing resources; otherwise, it is determined that the current base station has sufficient computing resources, that is, the local base station has sufficient computing power.

[0146] Step 310: If the local base station's computing power is insufficient, the computing power resource management module obtains computing power support from the adjacent base station (i.e., the second network device mentioned above);

[0147] Here, the computing power resource management module sends a computing power support request to the neighboring base station (i.e., the eighth information mentioned above). At the same time, it informs the computing power resource management module of the neighboring base station of the computing and storage resources required for retraining, so that the neighboring base station can assess whether it has sufficient resources to meet the retraining requirements (which can also be understood as whether the computing power is sufficient), and sends the feedback assessment result to the local base station (i.e., the ninth information mentioned above). The assessment result is used to indicate whether there are sufficient resources to meet the retraining requirements.

[0148] Step 311: If both local resources and the computing power of adjacent base stations are insufficient, the computing power resource management module obtains computing power support from the core network;

[0149] Here, the computing power resource management module sends a computing power support request to the core network; after receiving the computing power support response from the core network, the computing power resource management module can send the model and data for retraining (i.e., the tenth information mentioned above) to the core network for retraining.

[0150] Step 312: After completing the retraining, update the model-related parameters stored in the inference module.

[0151] Here, when the base station has completed retraining, the model-related parameters stored in the inference module are updated; when the neighboring base station has completed retraining, the model-related parameters after retraining sent by the neighboring base station are received, and the model-related parameters stored in the inference module are updated; when the core network has completed retraining, the model-related parameters after retraining sent by the core network are received, and the model-related parameters stored in the inference module are updated.

[0152] Application Example 1

[0153] Currently, machine question-answering technology based on large language models has been widely applied in people's production and daily lives. Lightweight large language models can now be deployed on the edge, meaning that 6G networks can utilize edge computing power to provide machine question-answering services to terminals. For machine question-answering tasks in the above scenarios, the corresponding lifecycle management process includes the following steps:

[0154] Step 1: The end user initiates a machine question-and-answer task request to the base station through the mobile application (APP). The machine question-and-answer task request carries a task-related description.

[0155] Step 2: After receiving the machine question answering task request, the AI ​​task orchestration module allocates physical space (i.e., physical resources) for the machine question answering task and informs the resource management module to load the initial model corresponding to the machine question answering task into the inference module.

[0156] Step 3: The information collection module obtains problem information from the end user in real time and sends the problem information to the inference module.

[0157] Step 4: After receiving the question information, the inference module performs forward propagation based on the question-answering model, generates the answer to the question required by the user and feeds it back to the end user. At the same time, the answer to the question is reported to the performance monitoring module.

[0158] Step 5: The performance monitoring module monitors performance based on the question-answering model; when the performance of the question-answering model does not meet the requirements, retraining is performed.

[0159] Application Example 2

[0160] With the further development of AI technology, the intelligence level of autonomous driving in vehicles has been greatly improved. Based on this, a large number of autonomous driving navigation assistance models can be deployed in 6G networks, and combined with a wealth of network perception information, more efficient path planning services can be provided for autonomous driving technology. For real-time navigation tasks in the above scenarios, the corresponding lifecycle management process includes the following steps:

[0161] Step 1: The vehicle system sends a real-time navigation task request to the base station and provides a task-related description through the real-time navigation task request.

[0162] Step 2: After receiving the real-time navigation task request, the AI ​​task orchestration module allocates physical space for the real-time navigation task and informs the resource management module to load the initial model corresponding to the real-time navigation task into the inference module.

[0163] Step 3: The information acquisition module obtains the vehicle's location and destination information from the end user in real time and sends the acquired information to the inference module. Simultaneously, the information acquisition module can also collect sensing information (such as relevant road images or video information collected by cameras deployed in the network) from local base stations and other base stations in the network, and send the collected sensing information to the inference module.

[0164] Step 4: After receiving the vehicle's location information, destination information, and perception information, the inference module performs forward propagation based on the real-time navigation model, generates the vehicle's planned path, and feeds it back to the vehicle system. At the same time, it reports the vehicle system's information to the performance monitoring module.

[0165] Step 5: The performance monitoring module monitors the performance based on the vehicle's planned path; when the performance of the real-time navigation model does not meet the requirements, retraining is performed.

[0166] In the application examples of this application, a unified lifecycle management framework is designed for AI tasks in different scenarios, which can uniformly schedule computing and storage resources for multiple AI tasks. In addition, when the performance of the AI ​​model degrades and retraining is required, the retraining task can be flexibly offloaded through computing power checks. That is, the network or cloud computing power can be flexibly utilized to retrain the model, avoiding problems such as excessively long retraining time and no improvement in retraining performance due to insufficient computing or storage resources at the base station. In this way, the diverse needs of AI task scenarios in 6G networks can be met.

[0167] To implement the method of the embodiments of this application, the embodiments of this application also provide an information processing apparatus, which is disposed on a first network device, such as... Figure 4 As shown, the device includes:

[0168] The receiving unit 401 is used to acquire first information, the first information including description-related information of one or more tasks;

[0169] The first processing unit 402 is used to determine second information using the first information, wherein the second information represents the task model and corresponding resources allocated to each of the one or more tasks, and the task model is used to execute the task.

[0170] The second processing unit 403 is used to obtain third information using the second information, the third information including the execution results of the one or more tasks.

[0171] In one embodiment, the second processing unit 403 is further configured to:

[0172] Using the third information, fourth information is determined, wherein the fourth information characterizes the performance level of one or more task models corresponding to the one or more tasks;

[0173] If the performance level of one or more first task models in the one or more task models represented by the fourth information does not meet a first condition, it is determined that the one or more first task models should be retrained, wherein the first condition includes conditions associated with the one or more tasks.

[0174] In one embodiment, the second processing unit 403 is further configured to:

[0175] The fifth piece of information is determined, wherein the fifth piece of information represents the resources of the first network device;

[0176] Using the fifth piece of information, network devices for retraining the one or more first task models are identified.

[0177] In one embodiment, the second processing unit 403 is configured to:

[0178] Obtain sixth information, which represents the size of one or more datasets associated with the one or more first task models;

[0179] Using the fifth and sixth information, network devices for retraining the one or more first task models are identified.

[0180] In one embodiment, the second processing unit 403 is configured to:

[0181] Determine whether the resources represented by the fifth information satisfy the second condition, the second condition including conditions associated with the one or more first task models, and the resources represented by the fifth information including computing resources and / or storage resources;

[0182] If the resources represented by the fifth information satisfy the second condition, the network device used to retrain the one or more first task models is determined to be the first network device.

[0183] In one embodiment, the second processing unit 403 is configured to:

[0184] If the resources represented by the fifth information include computing resources, then the seventh information is determined, which includes the duration information for retraining the one or more first task models.

[0185] Using the seventh piece of information, determine whether the computing resources represented by the fifth piece of information satisfy the second condition.

[0186] In one embodiment, the second processing unit 403 is further configured to:

[0187] If the resources represented by the fifth information do not meet the second condition, an eighth message is sent to the second network device. The eighth message is used to request retraining of the one or more first task models. The second network device includes network devices adjacent to the first network device.

[0188] Receive the ninth information sent by the second network device, the ninth information being used to indicate whether the resources of the second network device meet the second condition;

[0189] If the resources of the second network device meet the second condition, the network device used to retrain the one or more first task models is determined to be the second network device.

[0190] In one embodiment, the second processing unit 403 is further configured to send tenth information to the first network element when the resources of the second network device do not meet the second condition. The first network element is configured to retrain the one or more first task models, and the tenth information includes at least the one or more first task models.

[0191] In one embodiment, the second processing unit 403 is configured to:

[0192] Obtain eleventh information, which includes inference data associated with the one or more tasks;

[0193] The third information is obtained by using the eleventh information and the second information.

[0194] In practical applications, the receiving unit 401 can be implemented by the communication interface in the information processing device; the first processing unit 402 can be implemented by the processor in the information processing device; and the second processing unit 403 can be implemented by the processor in the information processing device in combination with the communication interface.

[0195] It should be noted that the information processing device provided in the above embodiments is only illustrated by the division of the above program modules. In practical applications, the above processing can be assigned to different program modules as needed, that is, the internal structure of the device can be divided into different program modules to complete all or part of the processing described above. In addition, the information processing device and the information processing method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the method embodiments, which will not be repeated here.

[0196] Based on the hardware implementation of the above program modules, and in order to implement the method of the embodiments of this application, the embodiments of this application also provide a first network device, such as... Figure 5 As shown, the first network device 500 includes:

[0197] The communication interface 501 enables information exchange with other devices;

[0198] The processor 502 is connected to the communication interface 501 to enable information interaction with other devices and to execute the methods provided by one or more of the above-mentioned technical solutions when running a computer program;

[0199] The computer program is stored in memory 503.

[0200] Specifically, the communication interface 501 is used to obtain first information, which includes description-related information of one or more tasks;

[0201] The processor 502 is configured to use the first information to determine second information, the second information representing a task model and corresponding resources allocated to each of the one or more tasks, the task model being used to execute the task; and to use the second information to obtain third information, the third information including the execution results of the one or more tasks.

[0202] In one embodiment, the processor 502 is further configured to:

[0203] Using the third information, fourth information is determined, wherein the fourth information characterizes the performance level of one or more task models corresponding to the one or more tasks;

[0204] If the performance level of one or more first task models in the one or more task models represented by the fourth information does not meet a first condition, it is determined that the one or more first task models should be retrained, wherein the first condition includes conditions associated with the one or more tasks.

[0205] In one embodiment, the processor 502 is further configured to:

[0206] The fifth piece of information is determined, wherein the fifth piece of information represents the resources of the first network device;

[0207] Using the fifth piece of information, network devices for retraining the one or more first task models are identified.

[0208] In one embodiment, the processor 502 is configured to:

[0209] The sixth information is obtained through the communication interface 501, and the sixth information represents the size of one or more datasets associated with the one or more first task models;

[0210] Using the fifth and sixth information, network devices for retraining the one or more first task models are identified.

[0211] In one embodiment, the processor 502 is configured to:

[0212] Determine whether the resources represented by the fifth information satisfy the second condition, the second condition including conditions associated with the one or more first task models, and the resources represented by the fifth information including computing resources and / or storage resources;

[0213] If the resources represented by the fifth information satisfy the second condition, the network device used to retrain the one or more first task models is determined to be the first network device.

[0214] In one embodiment, the processor 502 is configured to:

[0215] If the resources represented by the fifth information include computing resources, then the seventh information is determined, which includes the duration information for retraining the one or more first task models.

[0216] Using the seventh piece of information, determine whether the computing resources represented by the fifth piece of information satisfy the second condition.

[0217] In one embodiment, the processor 502 is further configured to:

[0218] If the resources represented by the fifth information do not meet the second condition, an eighth message is sent to the second network device through the communication interface 501. The eighth message is used to request retraining of the one or more first task models. The second network device includes network devices adjacent to the first network device.

[0219] The communication interface 501 receives the ninth information sent by the second network device, the ninth information being used to indicate whether the resources of the second network device meet the second condition;

[0220] If the resources of the second network device meet the second condition, the network device used to retrain the one or more first task models is determined to be the second network device.

[0221] In one embodiment, the processor 502 is further configured to send tenth information to a first network element through the communication interface 501 when the resources of the second network device do not meet the second condition. The first network element is configured to retrain the one or more first task models, and the tenth information includes at least the one or more first task models.

[0222] In one embodiment, the processor 502 is configured to:

[0223] Eleventh information is obtained through the communication interface 501, the eleventh information containing inference data associated with the one or more tasks;

[0224] The third information is obtained by using the eleventh information and the second information.

[0225] It should be noted that the specific processing procedures of the processor 502 and the communication interface 501 can be understood by referring to the above method.

[0226] Of course, in practical applications, the various components in the first network device 500 are coupled together through a bus system 504. It can be understood that the bus system 504 is used to implement communication between these components. In addition to a data bus, the bus system 504 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in... Figure 5 The general designated all buses as Bus System 504.

[0227] The memory 503 in this embodiment is used to store various types of data to support the operation of the first network device 500. Examples of such data include any computer program used to operate on the first network device 500.

[0228] The methods disclosed in the embodiments of this application can be applied to the processor 502, or implemented by the processor 502. The processor 502 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in the processor 502 or by instructions in the form of software. The processor 502 may be a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 502 can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor, etc. The steps of the methods disclosed in the embodiments of this application can be directly reflected as being executed by a hardware decoding processor, or being executed by a combination of hardware and software modules in the decoding processor. The software modules may be located in a storage medium, which is located in the memory 503. The processor 502 reads the information in the memory 503 and combines its hardware to complete the steps of the aforementioned method.

[0229] In an exemplary embodiment, the first network device 500 may be implemented by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers (MCUs), microprocessors, or other electronic components to perform the aforementioned method.

[0230] It is understood that the memory (memory 503) in this embodiment of the application can be volatile memory or non-volatile memory, or both. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), ferromagnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM); the magnetic surface memory can be disk storage or magnetic tape storage. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Synchronous Static Random Access Memory (SSRAM), Dynamic Random Access Memory (DRAM), Synchronous Dynamic Random Access Memory (SDRAM), Double Data Rate Synchronous Dynamic Random Access Memory (DDRSDRAM), Enhanced Synchronous Dynamic Random Access Memory (ESDRAM), SyncLink Dynamic Random Access Memory (SLDRAM), and Direct Rambus Random Access Memory (DRRAM).The memories described in the embodiments of this application are intended to include, but are not limited to, these and any other suitable types of memories.

[0231] In an exemplary embodiment, this application also provides a storage medium, namely a computer storage medium, specifically a computer-readable storage medium, such as a memory 503 storing a computer program, which can be executed by the processor 502 of the first network device 500 to complete the steps described in the aforementioned first network device-side method. The computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, Flash Memory, magnetic surface memory, optical disc, or CD-ROM.

[0232] In an exemplary embodiment, this application also provides a computer program product, including a computer program that can be executed by the processor 502 of a first network device 500 to complete the steps described in the aforementioned first network device-side method.

[0233] It should be noted that terms such as "first" and "second" are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0234] Furthermore, the technical solutions described in the embodiments of this application can be combined arbitrarily without conflict.

[0235] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application.

Claims

1. An information processing method, characterized in that, Applied to the first network device, including: Obtain first information, which includes descriptions of one or more tasks; Using the first information, determine the second information, which represents the task model and corresponding resources allocated to each of the one or more tasks, and the task model is used to execute the task; Using the second information, third information is obtained, which includes the execution results of the one or more tasks.

2. The method according to claim 1, characterized in that, The method further includes: Using the third information, fourth information is determined, wherein the fourth information characterizes the performance level of one or more task models corresponding to the one or more tasks; If the performance level of one or more first task models in the one or more task models represented by the fourth information does not meet a first condition, it is determined that the one or more first task models should be retrained, wherein the first condition includes conditions associated with the one or more tasks.

3. The method according to claim 2, characterized in that, The method further includes: The fifth piece of information is determined, wherein the fifth piece of information represents the resources of the first network device; Using the fifth piece of information, network devices for retraining the one or more first task models are identified.

4. The method according to claim 3, characterized in that, The step of using the fifth information to determine the network device for retraining the one or more first task models includes: Obtain sixth information, which represents the size of one or more datasets associated with the one or more first task models; Using the fifth and sixth information, network devices for retraining the one or more first task models are identified.

5. The method according to claim 3, characterized in that, The step of using the fifth information to determine the network device for retraining the one or more first task models includes: Determine whether the resources represented by the fifth information satisfy the second condition, the second condition including conditions associated with the one or more first task models, and the resources represented by the fifth information including computing resources and / or storage resources; If the resources represented by the fifth information satisfy the second condition, the network device used to retrain the one or more first task models is determined to be the first network device.

6. The method according to claim 5, characterized in that, Determining whether the resource represented by the fifth information satisfies the second condition includes: If the resources represented by the fifth information include computing resources, then the seventh information is determined, which includes the duration information for retraining the one or more first task models. Using the seventh piece of information, determine whether the computing resources represented by the fifth piece of information satisfy the second condition.

7. The method according to claim 5, characterized in that, The method further includes: If the resources represented by the fifth information do not meet the second condition, an eighth message is sent to the second network device. The eighth message is used to request retraining of the one or more first task models. The second network device includes network devices adjacent to the first network device. Receive the ninth information sent by the second network device, the ninth information being used to indicate whether the resources of the second network device meet the second condition; If the resources of the second network device meet the second condition, the network device used to retrain the one or more first task models is determined to be the second network device.

8. The method according to claim 7, characterized in that, The method further includes: If the resources of the second network device do not meet the second condition, the tenth information is sent to the first network element. The first network element is used to retrain the one or more first task models. The tenth information includes at least the one or more first task models.

9. The method according to any one of claims 1 to 8, characterized in that, The process of obtaining the third information using the second information includes: Obtain eleventh information, which includes inference data associated with the one or more tasks; The third information is obtained by using the eleventh information and the second information.

10. An information processing device, characterized in that, include: A receiving unit is configured to acquire first information, the first information including description-related information of one or more tasks; The first processing unit is configured to use the first information to determine the second information, wherein the second information represents the task model and corresponding resources allocated to each of the one or more tasks, and the task model is used to execute the task. The second processing unit is used to obtain third information using the second information, the third information including the execution results of the one or more tasks.

11. A first network device, characterized in that, include: Processor and communication interface; among which, The communication interface is used to obtain first information, which includes description-related information of one or more tasks. The processor is configured to use the first information to determine second information, the second information representing a task model and corresponding resources allocated to each of the one or more tasks, the task model being used to execute the task; and to use the second information to obtain third information, the third information including the execution results of the one or more tasks.

12. A first network device, characterized in that, include: The processor and the memory used to store computer programs that can run on the processor. When the processor is used to run the computer program, it performs the steps of the method according to any one of claims 1 to 9.

13. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

14. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1 to 9.

Images